By the same author:
Intensive Science and Virtual Philosophy A New Philosophy of Socie Also available from Continuum
Being and Event, Alain Badiou Conditions, Alan Badiou Innite Thought, Alain Badiou Logics of Worlds, Alain Badiou Theoretical Writings Alan Badiou Theo of the Subject Alan Badiou Cinema I lles Deleuze
q
PHILOSOPHY AND SIMUATION
The The Eme Emerg rgen encce of Sy Synt nthe heti ticc Reason Manuel DeLanda
Deleuze
Deleuze Dialogues I Dierence and Repetition Gilles Deleuze The Fold Gilles Deleuze Foucault Gilles Deleuze Francis Bacon, Glles Deleuze Kant's Critical Philosophy, Gilles Deleuze Logic of Sense Gilles Deleuze Nietzsche and Philosophy Gilles Deleuze Proust and Signs Gilles Deleuze Ant-Oedipus, Gies Deleuze and Fx Guattai A Thousand Plateaues Gilles Deleuze and Flix Guattari Seeing the Invisible Michel Henry Future Christ Fran
.\ continuum
m l Pblsg G
Te Twer Biig 11 Yrk- SEI 7X
80 ie e ite 704 ew Yrk, Y 10038
Coe
www.ti www.tibk bk © e e, 2011
rigt reerve prt f ti pbiti y be repre r trite i y fr r by y e eeri r ei iig ppyig, rerig, r y irti rge r retriev yte, witt prir perii i writig fr te pbier \
s Lb gg--Pb D tge rer fr ti bk i vibe fr te Brii ibrry IB: B: 978-144117028-6
b f gss lgg--Pbl D
Introduction Eme�gence Histry Chaptr On he Strm in the Cmputer
7
Chaptr o Cellulr utmt nd Pttern Flw
Chaptr Thr rticil Chemistries nd the Preitic Sup
e , e ipy iti: e eergee eergee f teti teti re / e e p IB: 978-1-44117028-6 1. ergee (ipy) 2 ieeipy ieeipy e
Chaptr Four Genetic lgrithms nd the Prebitic Sup
Chaptr Fv Genetic lgrithms nd ncient Orgisms
5
QI732E44D4 QI732E44D4 2010 003dc22
Chaptr Six Neurl Nets nd Insect ntelligence
Chaptr Svn Neurl Nets nd min emry
9
201002032
Chaptr Eight utigents nd Primte Strtegies
1
Chaptr Nin ultigents nd Ste ge Ecnmics
pee by ewge Igig yte v , ei, i rie b i Ii by epik re v t v
m l Pblsg G
Te Twer Biig 11 Yrk- SEI 7X
80 ie e ite 704 ew Yrk, Y 10038
Coe
www.ti www.tibk bk © e e, 2011
rigt reerve prt f ti pbiti y be repre r trite i y fr r by y e eeri r ei iig ppyig, rerig, r y irti rge r retriev yte, witt prir perii i writig fr te pbier \
s Lb gg--Pb D tge rer fr ti bk i vibe fr te Brii ibrry IB: B: 978-144117028-6
b f gss lgg--Pbl D
Introduction Eme�gence Histry Chaptr On he Strm in the Cmputer
7
Chaptr o Cellulr utmt nd Pttern Flw
Chaptr Thr rticil Chemistries nd the Preitic Sup
e , e ipy iti: e eergee eergee f teti teti re / e e p IB: 978-1-44117028-6 1. ergee (ipy) 2 ieeipy ieeipy e
Chaptr Four Genetic lgrithms nd the Prebitic Sup
Chaptr Fv Genetic lgrithms nd ncient Orgisms
5
QI732E44D4 QI732E44D4 2010 003dc22
Chaptr Six Neurl Nets nd Insect ntelligence
Chaptr Svn Neurl Nets nd min emry
9
201002032
Chaptr Eight utigents nd Primte Strtegies
1
Chaptr Nin ultigents nd Ste ge Ecnmics
pee by ewge Igig yte v , ei, i rie b i Ii by epik re v t v
CONTENT
146
Chapter Eleven ultiagents and Archaic States
166
Append Links to Assemblage heory
1 84
Notes ndex
204 223
/
vi
INTRODUCION
Chapter Ten ultiagents and Primitive Language
Emergence in Hstory
he origin o the modern concet o emergence can be traced to the middle o the nineteenth century hen realist hilosohers rst began ondering the dee dissimiarities dissimiarities beteen causality in the elds o hysics and chemistry he classical examle o causality in hysics is a collision beteen to molecules or other rigid objects Even in the case o several colliding molecules the overall eect is a simle addition , or examle, one molecule is hit by a second one in one drection and by a third one in a dierent direction the com osite eect be the same as the sum o the to searate eecs the rst molecule il end u in the same nal osition i the other to hit it simultaneously or i one colsion haens beoe the other I short, in these causal interactions there are no surises nothing is roduced over and above hat is already there But hen to molecules interact chemically an entirely ne entity may emerge, as hen hydrogen and oxygen interact to orm ater Water has roer ties that are not ossessed by its comonent arts oxygen and hydro gen are gases at room temerature hile ater is liquid And ater has caacities distinct rom those o its arts adding oxygen or hydro gen to a re ues it hile adding ater extinguishes it he act that novel roerties and caacities emerge rom a causal interaction as beieved to have imortant hilosohical imications or the nature o scientic exlanation I articular, the absence o novelty in hysical interactions meant that exaining their eects could be reduced to deduction rom general rinciles or las Becaus deductive logic simly transers truth rom general sentences
PHLOOPHY AND MULATON
to particular ones without adding anything new it seemed like an ideal way of odeling the explanation of situations like those involv ing rigid collisions. But the synthesis of water does produce soe thing new, not new in the absolute sense of soething that has never existed before but only in the relative sense that soething eerges that was not in the interactg entities actg as causes This led soe philosophers to the erroneous conclusion that emergent eects could not be explained, or what amounts to the sae thing, that an effect is eergent only for as long as a law fro which it can be deduced has not yet been found2 This line of thought went on to become a full ledged philosophy in the early twentieth century a philosophy based on the idea that eergence ws intrinsically unexplainable This rst wave of "eergetist philosophers were not ystical thinkers but quite the opposite: they wanted to use te cncept of eergence / eliinate fro biology mystifying entities like a "life force or the "lan vital But their position toward explanation gave their views an inevitable mystical tone: eergent poperties they said ust be accepted with an attitude of intellectual resignation, that is they must be treated as brute facts toward which the only honest stance is one of natural piety3 Expressions like these were bound to ake the concept of eer gence suspect to future generations of philosophers was only the passage of me and the fact that atheatical laws like those of classical physics were not found in chemistry or biology-or for that matter, in the more historical elds of physics like geology or clima tologythat would rescue the concept fro intellectual oblivion Without simple laws acting as selfevident truths (axioms) from which all causal effects could be deduced as theores the axiomatic drea eventually withered away. Today a scientic explanation is identied not with soe logic operation but with the ore creative endeavor of eludating the echanss that produce a gien effect The early eergentists disissed this idea because they could not image any thing more complex than a linear clockwork echanism But there are many other hysical mechanss that are nonlinear Even in the real of human technology we have a plurality of exeplars to/guide our imagation steam engines, therostats, transistors And outside technology the diversity is even greater as illustrated by al the differ ent echanisms that have been discovered in chemistry and biology 2
EMERGENCE N HTORY
Armed with a richer concept of mechaism the eergent properties of a whole can now be explained as an effect of the causal interactions between its component parts A large portion o this book will be dedicated to describe the wide variety of mechanisms of emergence that have been lucidated in the decades since the origal eergentists rst wrote. Thus, what is different today fro the early twentieth century views is the epistemologial status of eergenc: it does not have to be accepted as a brute fact but can be explained without fearing that it will be explaed away What has reained the sae is the ontological status of emergence: it stil refers to soething that is objectively irreducible But what kids of entities displa this ontological irreduc ibility? The origal examples of irreducible wholes were entities lie "Life, "Mind, or even "Deity But these entities canot be consid ered legitmate inhabitants of objective reality because the are noth ing but reied generalities And even if one des not have a proble with an ontological commitent to entities like these it is hard to see how we could specify echansms of emergence for life or mind n general, as opposed to accountg for the eergent properties and capacities of concrete wholes like a metabolic circuit or an assembly of neurons. The only problem with focusing on concrete wholes is that this would seem to make philosophers redundant sce they do not play any role in the elucidation of the series of events that produce eergent effects This fear of redundancy ay expla the attach ment of philosophers to vague entities as a way of carving out a nche for theselves in this enterprise But realist philosophers need not fear irrelevance because they have plenty of work creatg an onto logy free of reied generalities within which the concept of emer gence can be correctly deployed What kids of concrete emergent wholes can we legitimately believe in? Wholes the identity of which is determined historically by the processes that intiated and sustain the interactions between their parts The historially contingent identity of these wholes is dened by their emergent properties, capacities, and tendencies Let's illus trate te distinction between properties and capacities with a smple exaple A kitchen kfe ay be either sharp or not, sharpness being an actual property of the kfe We can identify this property with the shape of the cross section of the knife's blade: if this cross section has 3
PHLOOPHY AND MULATON
a triangular shae then the knie is sha ese it is blunt his shae is emergent because the metalic atoms makg u the knie must be arranged in a very aricuar way or it to be trianguar here is on the other hand the caacity o the knie to ct things his is a very dierent thing because unike the roerty o sharness which is aways actual the caacity to cut may never be actua i the knie is never used I other words, a caacity may reman only otential i it is never actualy ercised his already oints to a very dierent ontoogical status etween roeries and caacities I addition, when the caacity oes become actual it is not as a state, ike the state o being shar, but as an event an event that is aways double to cut-to be cut. he reason or this is that the knie's caacity to aect is contingent on the existence o other things cuttabe things, that have the caacity to be aeced by it hus, while roerties can be seci ed without reerence to anything else caacities to aect must aways be thought in relation to caacities to be aected Finaly, the onto ogical relation between ro erties and caaci ties disays a comex symmetry On one hand, caacities deend on roerties a knie must be shar to be abe to cut On the other, the roerties o a whole emerge rom interactions between its comonent arts inter actions in which the arts must exercise their own caacities wthout metalic atoms exercising their caacity to bond with one another the knie's sharness would not exist similar distction can be made between emergent roeries and tendencies o stick to the same exame a knie has the roerty o soidity, a roerty that is stable within a wide range o temeratures Nevertheess there are always environents that exceed that range, environments in which the temerature becomes so intense that the knie is orced to manest the tendency to liquiy t even greater intensities the molten metal may gasiy hese tendencies are as emer gent as the shae o a knies blade a single metalc atom cannot be sad to be soid, iquid, or gas; we need a large enough oulation o interacting atoms the tendency to be in any o these states to emerge endencs are similar to caacities in their ontological status, that is they need not be actual to be real, and when they do become actual is as events to melt or to solidiy he man dierence between tendencies and caacities is that while the ormer are tyically nite the latter need not be We can enumerate, or examle the ossible
l
4
EMERGENCE N HTORY
states in which a material entity tend to be (sod iquid gas, lasma or the ossibe ways in which it may tend to low (unormly eriodically tubulenty But caacities to aect need not be nite because they deend on the caacities to be aected o nnumerabe other entities a kne has the caacity to cut when it interacts with something that has the caacit to be cut; but it aso has the caacity to ki i it interacts with large organisms with dierentiated organs that is, with entities that have the caacity to be kied Since neither tendenies nor caacities must be acua to be real it would be temting to give them the status o ossibilities But the concet o a ossibe event is ilosohicaly susec because it is amost indistinguishabe rom that o a rea event the only dierence being the ormers ack o reaity Rather what is needed is a way o seciying the structure of the space of possibilities that is dened by an entitys tendencies and caacities hilosohers ontological com tment shoud be to the obective existence o this structure and not to the ossibiities themseves since the latter exist ony when enter tained by a mind Some ossibiity saces are continuous having a wedened satia structure that can be nvestigated mathematicaly, whie others are discrete ossessing no inherent satia order but eing nevertheless caable o being studied through the imosition o a certain arrangement he sace o ossible regimes o low (un orm, eriodic, turbuent is an examle o a continuus ossibiity sace in which the ony discontinuities are the critical oints searat ing the dierent tendencies he sace o ossible genes, on the other hand, is an examle o a discrete sace that must be studied by imos ing an order on it, such as an arrangement wich every gene has as neighbors other genes diering rom it by a single mutation s we see in the dierent chaters o tis book the strucure o ossibil ity saces lays as great a role in the exlanation o emergence as do mechanisms he chaters are deiberately arranged in a way that dearts rom the ideas o the orignal emergentists hese hilosohers beieved that entities ike "Saceme, "Lie, "Mnd, and "Deity (not "god but the sense o the sacred that emerges in some ds ormed a yramid o rogressively ascending grades though the levels o this yramid were not suosed to imly any teleology it is hard ot to view each level as leading to the next ollowg a necessary sequece 5
PHOOPHY AND MUATON
To eliminate this possible interpretation an entirely different image is used here that of a contingent accumulation of layers or strata that may differ in complexity but that coexist and interact with each other in no particular order a biological entity may interact with a sub atomic one as when neurons manipulate concentrations of metallic ions or a psychological entity interact with a chemical one as when subective experience is moed by a drug The book begins wth purely physical entities thunderstorms that are already complex enough to avoid the idea thtir behavior can be deduced from a general law It then moves on to explore the pre biotic soup bacterial ecosystems insect intelligence mammalian memory primate social strategies and the emergence of trade language and institutionl organizations in human communities Each of these layers will be discussed in terms of the mechanisms of emergence involved drawing ideas and insights from the relevant elds of science as well as in tes of the structure of their possibty spaces using the results of both mathematical analysis and the outcomes of computer siulations Simulations are partly responsible for the restoration of te legiti macy of the concept of emergence becaus� they can stage interactions between virtual entities from which properties tendencies and capacities actually emerge Since this emergence is reproducible in many computers it can be probed and stued by diferent scientists as if it were a laboratory phenomeon In other words simulations can play the role of laboratory experiments in the study of emergence complementing the role of mathematics in deciphering the structure of possibility spaces nd philosophy can be the mechanism through which these insights can be synthesied into an emergent materialist worldview that nally does ustice to the creative powers of matter and energy
6
CHAPER ONE The Stom in the Compute
ets begin with the simplest emergent properties properties ike temperature or pressure characteriing wholes made out of a large number of identical parts such as a body of water in a container Being composed of billions of molecules that are qualitatively the same makes a body of water much simpler than say an ecosystem in which hundreds of fferent species constantly interact But this simplicity is what makes the mechanism of emergence behind tem perature r pressure a promising starting point for philosophical thought To begin with in what sense are these properties emergent? Temperature is dened as the average energy that a molecular popu 1ation has by virtue of the motion of its parts the more violent the motion the more intense the temperature Pressure is dened as the average degree to which the population pushes against the walls of the container by virtue of the momentum of its parts the faster and more massive the molecules the more intense the pressure exerted These denitions have tempted philosopher in the past to think that temperature and pressure can be reduced to kinetic energy and momentum that is that they are not emergent But temperature and pressure re in fact irreducible because they are the result of an obective averaging process that takes place spontaneously n molecular populations To understand how this works let's imagine two bodies of water at erent temperatures The moment we place these bodies in contact with each other energy will low from the body with higher tempera ture to the one wth lower temperature the low continuing until the 7
THE STORM IN THE COMPUT ER
PHILOSOPHY AND SIMULATION
temperature dfference disappears In other words the dfference in temperature wl dsplay a tendency to average itself out Thus saying that a body of water possess a certa temperature and that posses sion of that property denes an enduring state impies that depar tures from that state are constantly beunteractd by an objective tendency For the same reason dening "sameness of temperature can be done by placing two bodes of water into contact and verifying that no ow of energy is taking place between them Thus in ths imple case the irreducibe status of a property like temperature is estabished by elucidating a mechanism through which the property emerges a mechanism invoving the manifestation of tendency Accounting for this tendency in turn demands switching scales and focusing on the interactions between the parts of the whole, interac tions in whch the parts exercse their capacities to affect and be affected. particular, for a temperature dfference t o cancel itself the component molecules must exercise their capacity to collide and redistribute energy in those cosions We can visualize the seres of events leading tohe emergence of an average temperature by comparing the states of two bodes of water before and after the dissipating tendency has been manifested At the start of the process the exstence of a temperature dfference means that the water moecues are dstributed with a hgh degree of order, that is that they are neatly sorted out into two parts one hot and the other cold At the end of the process the entire population is uni formy warm and this order has disappeared. A disordered state is characterized by the fact that we can make a large number of changes in the molecular dstribution and eave the bulk state basically the same other words, a much arger number of combinations of ind vidual kinetic energies wil resut in the same warm body of water than the number that w yield the state in which one conainer is hot and the other one is cold This affects the probability that one or the other state w occur spontaneousy Because in this case the interactions between molecues are entirely random these odds make a the dfference in the world: the state characterizing a warm body of 'ater wl have a much higher probabiity of occurring as a resut of random collisions than the one in which they are sorted into hot and cold subpopuations. It s this dfference in the odds with which 8
the ordered and dsordered states can occur that explains the ten dency for one to become the other The mechanism of emergence just described for temperature is basicaly the same for pressure density and other intensive properties of molecular populations Despite their relative simplicity these prop erties are important because the spontaneous low of energy that takes place as intensive dfferences cancel themselves can be tapped into to fuel other processes The whole composed by two containers of water at different temperatures for example has the capacity to drive another process partly because the hgh temperature container stores a lot of energy much more than the ow temperature one and parly because we can extract that energy by placing the former in contact with the latter2 This capacty is reatively short ived how ever, because once the intensive difference disappears the energy left behind becomes much harder to extract But if the dfference is con tinuously refreshed, by placing the rst container on top of a re for instance then the whole formed by the hot and cold molecular popu lations can ecome a component part of a larger whoe playing the role that a gasolne tank or an eectric battery pay in an automobie or an electronic appliance The capacity of intensive dfferences to act as energy storage devices will play such a promnent roe in the explanation of emergence in many other exampes that it wl be useful to have a more compact term for them We will refer to them as gradients. addtion to serve as energy sources gradents can serve to gener ate the moving parts of larger wholes For example if a gradient is intense enough and if it is prevened from dsspating it can cause a molecular population to self-organize into a circuar motion pattern that w persist as long as the gradent persists. The coordinated movement of bilions of molecules needed to yield such a pattern is a highly unlikely event and yet it occurs spontaneously in the ocean and the atmosphere every single day Ths coherent circulaory low, referred to as a convection cell, is produced by the gradent as the means to cancel itself even as the imposed constraints prevent it from doing SO.3 The mechanism of emergence behind a convection cell can be explained using the same example of a water container: when the container is heated from below it becomes dvided into a warm botom 1
9
PHILOSOPHY AND SMULATION
and a cool top; as the bottom water warms up i expands and�mes less dense tending to rise, while the high density cold water on top tends to sink these up and down movements are counteracted by the internal friction generated by the viscosity of the water, but when the temperature difference becomes intense enough this resistance is overcome and the upward and downward lows join together to form a circular patte4 Because this pattern is very stable it can literally be used as a bulding block to construct larger emergent entities What kind of entities can be built using gradients as fuel tanks and convection cells as moving parts? The most dramatic example is a thunderstorm, a typical storm containng ve to eight convection cells each a few kilometers in ameter5 Viewed from the outside a thun derstorm appears as a lrge complex cloud with a welldened form. At the center of the storm is a massive columnlike structure called the "central pillar. This vertical structure adopts an asymmetric horizontal shape at its top called an "anvil for its resemblance to the metl block used by blackmths The central pilar often overshoots the anvil creating a do re at its top Finally, at the bottom of the pilar there ar lankng horizontal clouds lined up in the opposite direction to the anvi This complex form is one of the emergent properties of a thunderstorm, its directly observable property But behnd its observ able form there is the inteal machnery of the storm In addition to graents and convection this machinery uses phase transitions, the transition from gas to liquid or from liquid to solid, as energy mpl ers One of the fferences between a material such as water in its gas, iquid, or solid state is the degree to which its composing mole cules move around and therefore the amount of kinetic energy the material contains In a solid the molecules are relatively conned to xed positions so their activity is relatively calm. In liquids ths con nement is relaxed molecules still exert some restraining inluence over one another but they allow a more energetic movement In gases the molecules are even more excited since their movement is not constraned at all A gas therefore contains more energy than a liquid or a solid Whe rising vapor becomes rain some of this extra energy becomes available as a surplus that can be exported to the surround ing medium, increasing the amout of energy available to the thun derstorm This exportable energy is referred to as "latent heat.6 10
THE STORM IN THE COMPUTER
To understnd the mechanism of emergence behnd a thunderstorm we need to explain how these different components are coupled together Fst of all a difference in temperature between the surface of the ocean and that of the atmosphere must exist to get the process started This vertical gradent causes an upward ow of air and vapor forming one leg of a convection cell As the wam moist air moves up it becomes cooler eventualy reachng the crtical point at whch vapor becomes liquid water At rst ths phase transition produces very smll liquid droplets that become suspended in the surroundng air The concentration of these tiny droplets makes the upward air cur rent visible as a small caulowershaped cloud that becomes the base of the future thunderstorm Although at ths point the air should start turning sideways and head for the downward leg of the convection cell, the latent heat released by the phase transition increases the temperature of the air current adding buoyancy to it and propelling further up Ths selstmulating interaction is repeated several times allowing the updraft to reach great heights eventually becomig the gint cloud described above, with its central pilar, anvi, nd dome. The death of a thunderstorm, in turn, is linked to processes that coun teract its sustaining gradients: the hgher the air reaches the colder it gets the more saturated it becomes, and the larger the quid drops and ice crystals that condense from it. When the weight of these drops and crystals reach a ipping point-the point at which the downward force exerted by graity becomes stronger than that of the updraftit begins to fall in the form of rain and hail dragging air with it, stealing energy from the updraft and eventualy destroying the internal machinery of the storm Other features of ths emergent eteoroogical entity are lso explained by graents A severe thunderstorm is usually accompa ned by the roduction of ghtning either intensely brght ashes created withn the cloud or owerful bolts between the cloud and the ground. Lightning is the result of an electrical gradient, a fference n charge between the upper nd lower regions of he cloud, or between the cloud and the ground, the briliant scharge being the form cre ated by the graent to cancel itsef Thunderstorms are also the birth place of toadoes, whrling masses of a made visible by the matter (vaor, dust, debs) that they suck nto the ntensely raid circuation 11
PHLOOPHY AND MULATON
Tornadoes are born fro the sae vertical teperatre gradient that cases the pdraft to which a steep horizontal pressre graient is added The latter is cased by the fact that the pdraft scks the air fro the center of the tornado greatly redcing the pressre inside of it copared to that of the otside7 As an eergent whole a thnder stor is characterized by its properties, sch as the heights it reaches or the speed with which it oves y its tendencies, like its tendency to ove i a cerain irection or its tendency to conse all its energy and die and by the capacities it exercises when it interacts with other entities Fro a han point of view these interactions are ostly destrctive: its lightning kills people and starts brsh and forest res the heavy rain along its downdraft can reslt i loods; and the torna does it spawns can violently latten entire towns These capacities can srely inspire awe and respect at the destrctive potential of a thnderstor bt they shold not lead to an attitde of intellectal resignation or natral piety toward it: we can explain how it is born, how it lives, and how it ies Lets pase to consider the argent so far The objective reality of eergent properties can be established by elcidating the echa niss that prodce the at a one scale and by showing that eergent entities at that scale can becoe the coponent parts of a whole at a larger scale Mechaniss of eergence ay, of corse, ndergo revision or elaboration, and soe are better nderstood than others, bt the possibility of iproveent or change in the proposed echa niss shold not force s to take eergence at any scale as a brte fact There are, on the other hand, aspects of the concept of eer gence hat this argent does not address I particlar, siilar eergent effects can be prodced by entirely different echaniss sggesting that there is ore to the eergent properties of a whole than the interactions between its parts Lets retrn to the cas of convection cells The selforganized rhythmic ow characterizig convection eerges in any kids of aterials The lows of olten rock that lie ndernath the srface of the earth, for exaple, exhibit the sae coherent circlar otion as the air and water above it More iportantly, the sae selforganization is displayed by other rhythic pattes tht have nothing to do with the oveent of atter i space A good exaple coes fro the world of cheistry The gradients in this case are differences i the concentration of 12
THE TORM N THE COMPUTER
certain sbstances, that is, they are gradients of atter not of energy The rhyths are the rates at which new copond olecles are synthesizedthe cheical reacion switches spontaneosly fro the prodction of one type of olecle to the prodction of another fol lowing a prfec beatnot collective circlar otions Yet despite these ifferences a convecion cell and a chemical clock as these reac tions are called, are qalitatively the sae This iplies that a fll explanation of these eergent entities st possess a coponent that is idependent of any pariclar echanis cold be arged that the siilarity i rhythic behavir is spercial and that it does not deand coplicating he concet of explanation, bt there are other shared characteristics that cnot be so easily isissed I particlar, the periodic behavior in both cases is stable against pertrbations, that is, if a convection cell or a chei cal clock are distrbed by an otside shock they wi tend to retrn to their origial period and aplitde after a relatively short tie This tendency is referred to as asymptotic stabili Not all oscillating entities possess this kid of stability A pendl in which friction has been careflly eliinated, for exaple, will not react the sae way: a sall psh peranently change both the dration and intensity of its swing, the pendl acting as if it "reebered the pertrbation A convecion cell or a cheical clock, on the other hand, qickly "forget the pertrbation acting as if nothing had happened8 When we explaied convection by the casal echanis otlined abovea teperatre gradient that creates a density gradient that, in trn, aplies ctations ito a circlar lowwe were givig only part of the explanation becase the casal chain behid the eergence of a convection cell does not accont for the fact that its properties are stable against peJtrbations And siilarly for a chemical clock Ading to the explanation of eergence a mechanismindependent coponent will involve introdcing entirely new ideas so it will be sefl at this poit Qjstify the need for the extra coplexity So far the concept of eergnce has played an ontological role, showg why it is legitiate to believe in the existence of objective properties, ten dencies, and capacities Bt once we add the echanisidependent coponent the concept of eergence leads to two iportant episte ological conseqences: it explains why we can se partial odels to learn abot reality and it provides an accont for the capacity of those 13
PHLOOPHY AND MULATON
models to ic the behavior of the processes they mode. The rst consequence derives drecty from the notion of asymptotic stabilty. When the emergent properties of a whoe are stabe they cn survive changes in the detas of the interactions between its parts. A given degree of temperature in a body of water, for exmpe; may result from a number of dfferent interactions between the kinetic enegy of its component molecues. This imples that we can take the existence of temperature for granted when expaining the cculatory pattern n a convection ce that is that we can legitmatel leave out of the expnation a detaled census of the inetic enegy of each of the moecules in th· popuation To put this dfferently a stabe emergent propert s "ndifferent" to local changes in the interactions that give rise to it, nd this objective infference tansates nto n objective epanato irreevace of the detas of the inteactions: includg the atter in an expnation would be redundant because mny fferent interactions woud yield the same outcome9 Being abe to take for granted the existence of emergent properties at one scae n order to explain properties at another scae is arguably a basic requirement for scientic research. scientists had to bud modes that captured a scaes simultaneousl no scientic eld would ever have succeeded in expainng anthing We would be trapped in a lock universe in hich ever aspect is inextricabl reated to every other aspect and our incapacit to sepaate levels of emergence would leave us cogni tivel poweress. The second epistmological consequence derives from the ver noion of mechanism-independence: f processes as dfferent in deta as a convection ce and a chemical cock cn exhibit the sme behav ior perhaps mathematca equations cn also spay that behavor To set the stage for the argument let's rs give a simplied account of the reation btween mathematica models and aborator experi ments. Lets assume that we want to understand the behavor of the air currents forg the updraft and downdraft of a thunderstorm We can use a mathematica model of the dnamics of nonviscous luids that has xisted since the eighteenth centur: a setof fferen tia equations that relate the properties of densi pressure nterna energ, nd veocit to the ow of air Usng these equations we can generate a series of numbers that ncate the course of the modeed
14
THE TORM N THE COMPUTER
uid at dscrete intervals of space nd time nd then give this series of numerical soutions a vsua form such as a pot in a piece of paper The expression the ehavor of equations" refers to the pattern gen erated by its umerical soutions as presented graphicly b the pot Next we move to the aboratory and create an experimental situation in which n actua ow of air is affected ony y those same proper ties using an apparatus that cn excude any othe causal factor from affecting the low. We run he experiment and take measurements of the airow at dfferent points in space and instnts of time and pot the measured values on a piece of pape. To be abe to compare the two pots we must make sure tha the vaues of the variables in the equations and the values of the properties in the apparatus are approximately the same at the begnning of the run, that is that both the equations and the apparatus are set to the same inital contions If the mathematica model captured the eal dynamics then the two pots shoud be geometricay simiar 0 This is of course a high ideaized picure of the reation between theor and experiment but it points to the crucia question: the sarit etween the two graphic pots suggests that the behavor of the numerica soutions to the equations is isomorphic to the behav or of the phsica properties insde the apparatus, a high improba be behavioa isomorphism that ces out for expanation Moreover, the advent of computer simuatons has aowed scientsts to tace not just smpe air currents but entire thunderstorms and the fac that the geometric simiarit has persisted has made the underlying behavora isomorphism even more probematic. I recent decades the equations for nonviscus ow used in the previous exampe were couped t o another set modeng the phase transitions n water and were numericll soved for ver point of a threedmensona grid each point representing a ox one kiometer wide nd haf a kiomeer high. Enough of these boxes were incuded in the simua tion to t a reguar size thunderstorm. To add a tempora dimenson the equations were soved recursivey that is the soutions otaned as outputs at n one instnt were used as inpus fo the equations to get the soutions for the next me interva. This makes the expression the behavor of equations" ess metaphorica because recrsion transforms a static mathematica object into a dnamic comptatonal process.
5
THE STORM IN THE COMP UTER
PHILOSOPHY AND SIMULATION
A set of values to serve as initial conditions was obtained from actual measureents of wnd, temperature, and humidity of an area of the ocea where an actual thunderstorm had developed. After feeng the computer the initia vaues the recursive proce dure took over repeatedy generating populations of soutions for every time interval, the nal product rendered using standard com puter graphics software To everyone's surprise a centra pilar an anvil, a dome and a lanking ine of clouds spontaneousy emerged despite the fact that none of those features had been explicitly modeed The updraft and the downdraft forming the internal machin ery of the storm also emerged made visible by adding purey graphic entities (weightess spheres) that followed the simuated air currents Part of the expanation for the success of the simuation is the decom posabity of reality made possibe by emergent properti es The mic ro scopic interactions between moecules at the surface of the ocean and those in the air above it for exape, id not have to be modeled in detail The effect of friction between water and air moecules that starts the process of storm formation was introduced from the outside as a perturbatin of the interface between the two luids Similary macroscopic details ike the inluence of the earths rotation were simpy ignored. That the correct geometrical form emerged despite these simplications shows that natural phenomena exhibit a recur rent part-to-whole reation in which whoes at one scae become parts at the next larger scae and that interactions between scaes can be either eft out of a mode or added exogenousy12 On the other hand, the behaviora isomorphism between the soutions to the equa tions and the physica lows in a real thunderstor is not expained by the decomposability of reaity This isomorphism has mystied physicists for as ong as there has been evidence of its existence some of them resigning themselves to accept the unreasonable effectiveness of mathematics as miracuous.13 But as argued above an explanation of this "miracle can be given using the notion of echanisindependence Let's clarify this notion rst for the case of materia processes As mentioned in the Introduc tion the distinction beween properties on one hand and tendencies and capacities on the other is that the former are aways actua actua characteristics of the state of a whole at any given point in timewhile the latter need not be: the tendency of liquid water to II
16
solidify at a critical point o f temperature may not manifest itself if th e temperature aways remains above that point; and the capacity of liquid water to acts as a solvent may not be exercised if the water ever comes into contact with solube substances. The ontologica status of both tendencies and capacities is therefore ifferent from that of properties As the simpe case of temperature or pressure shows the part of the expanation of emergence that depends on mechaniss involves the actua manifestation of tendencies (the ten dency behind the objective averaging process) and the actual exercse of capacities (the capacity of olecues to code and reistribute energy) The mechanismindependent component of an expanation on the other hand demands carifying the status of tendencies and capacities when they are not actuay manifested or exercised We coud, of course characterize that status as that of a possibility but that woud be too vague an umanifested tendency and an unexer cised capacity are not just possible but dene a concrete space of possibilities with a denite structure.
Lets imagine this abstract space as a set of points each epresenting a ifferent possibility The structure of this space can be conceived as the subset of those points that have a much gher probabiity to become actual When we described the mechanism of emergence behnd the tendency of a graient to cance itsef we said that it was based on a probabilistic argument: the state in which the graient is alive is much more ordered than that in wch it is dissipated and in a moecuar popuation in wch al interacions are basicaly random the isordered state is a vastly more probabe outcome of those interactions A aternative way of sayng this is that in the space of possibilities for the moecular population there exists a specia point the point of maximum disorder, and that the popuation is attracted to that state because it is much more probabe than the others A simiar idea can be applied to convection cells and chemica clocks. We can imagine that in their space of possible states there is a set of points forming a closed loop that has the ghest probability of actu aly occurring forcing a physica or chemica process to repeat the same series of states over and over. If the process is subjected to an extea shock it will move away fro that loop existing momen tariy in ess probabe states but then it i tend to return to it. Ts informal argument points to the solution to our problem the stability 17
PHLOOPHY AND MULATON
of emgent properties is expained by the structure of a possibiity space and the fact that this stabity can be dispayed by entirey differ ent mechanisms is expained by the fact that their possibity spaces shae the same structure. The concept of a possibity space can be made rigorous using the resuts of severa centuries of mathematica investigation on the nature of abstract spaces I mathematics a basic distinction is made between metric spaces, the best known exampe of whch is Eucdean geometry, and nonmetric spaces exemped by a variety of other geometries projective, differentia, topoogica A reativey simpe way of distguishing metric from non metric spaces is by the way in which the component parts of a space, indivdua poi nts, are identi ed The meric souion is to give each point an address" by ocatg the space reative to a set of xed coordinates and deerminig he disance that eah point has from hose axes. But this famiiar proce dure is no the ony way of individuaig points One can, for exam pe, determie the rate a which the curvaure of a space changes a a given pot and use ths stananeous rate of change o identify i14 When we do this a space ceases to be a set of coordinate addresses and becomes a eld of rapidities and slownesses the rapidity or sowness wih which curvaure varies a each point. The strucure of a abstract space, in urn, can be characerized by hose properies tha remain unchanged when the space is transformed, when it is moved, roated, foded, streched Metric properties ke ength or area rema invari ant under he smaest set of transformations, whie hose of he east meric spaces say unchanged under the arges set For the purpose of undersading what sense two different mechanisms can share the same srucure we need hghy variant structura properies sce he meric deais of heir possibity spaces are bound o be different. A very impora exampe of these varian properies is the exisence and distribution of specia or remarkabe points (or sets of such pos) caed singlarities This is he concept ha we need to make he remarks in he prevous paragraph ess meaphorica the possibiiies with the highes probabiiy of occurrg are topoogica singuaities acing as atractors. Le's now appy ths ine of hough o mahematica modes To creae a mahemaica mode he rs step is to enumerate a he reevant ways whch the process o be modeed is free o change 18
THE TORM N THE COMPUTER
Let's imagine that we are modeng a simpe physica process charac terized by two changing properties, such as temperature and pressure These are caed its degrees of freedom" The process may, of course, also change in an inite number of irreevant ways, the art of math matica modeing being ased part on the abiity to udge what changes do, and what changes do not, make a difference. Once the reevant degrees of freedom of a physica process have been identied the mode can be give a spatia form by assignig each of them to a dimension of a topoogical space Each point in ths space be a combination of vaues of temperature and pressure representing an instantaneous sae of he process being modeed, we the set of pots as a whoe represents the space of al possibe states for the process. For thsreason he abstract space is referred to as state space (or phase space") Finay, since a given process changing time foows a seuence of sates its behavor appears in stae space as a series of points, hat is, as a curve or trajecory. was by observing the tendency of many of these trajecories to converge on specic areas of stae space, to converge on sguarities, that the exstence of asymptoic stabiity was rst estabished 15 Using hese ideas the expanation for he ureasonabe effective ness of mathematics can be phrased
PHLOOPHY AND MULATON
and energy o the universe ceased to exist, woud singuarities aso disappear (immanent) or woud they continue to exist (transcendent)? though these questions are not mathematica but phiosophica the practice o mathematicians can sti provide insights into the answers I singuarities are immanent they must be both irreducibe t any particar materia process whie at the same time requiring that some process or another actuay exists These two conditions are reected in the way singuarities are studied Topoogists, or exampe, do not study the singuarities strcturing the possibiity space·o � mode ree to change in its temperature and its pressure but o all models with two derees ofeedom whatever these may be can be proved or exam pe, that in a twodmensiona space ony certain kinds o singuarities exist our dierent types o point singuarities distinguished rom each other by the orm o the ow o nearby trajectories (nodes, sadde points, oci, and centers) as we as one type o periodic singu arity16 I three-densiona spaces, the our point singuarities are sti part o the repertoire but now periodic singuarities come in three dierent orms: stabe unstabe and sadde-shaped oops I addition, a new type o singuarity becomes avaiabe one that can be pictured as a oop that has been repeatedy stretched and oded (a socaed chaotic attractor) Y This impies that topoogica acts about possibiity spaces can be discovered without reerence to the nature o the degrees o reedm, ony to their number and without reerence to the nature o the gradient (therma, gravitationa, mechanica heca) ony to its existenceY But the act that the existence o a gradient, any gradient is necessary conrms the immanent status o singuarities Singuarities are, thereore, perecty acceptabe entities in a materi aist phiosophy. The main probem conronting us now is the extent to which we can generaize rom these ideas State space is ony one kind o possibiity space, a space useu to study endencies but not capacities Capacities invove a much arger set o possibiities than tendencies because entities can exercise their capacities in interaction with a potentiay innumerabe variety o other entities. The more compe possibiity spaces associated with capacities, and the nature o the singuarities that structure them, are not neary as we under stood as those o tendencies On the other hand� computers can sup py the means to expore these other possibiity spaces in a rigorous 20
THE TORM N THE COMPUTER
way because the interactions in which capacities are exercised can be staged in a simuation and varied in mutipe ways nti the singuar eatures o the possibiity space are made visibe Each o the oow ing chapters wi expore how staging a dierent type o simuated interaction (chemica bioogica, socia) can tease ot the singuar structure o their possibiity spaes The rst step in this exporation however, not address the question o the reation between modes and reaity We rst need to cariy the concept o emergence in the case o simuations Tis is what the oowing chapter wi attempt to do
21
CELL ULAR AUOMAA AND PAERN OF FLOW
CHAPTER TWO Cellular Automata and Patterns of Flow
Much ike he simpciy of emperaure and pressure make hem an ideal sarn pon for phosopcal reecion on emerence, he quesion of wha is an emeren propery n compuer simulaions can be bes approached sarn wih he leas complex cases The simles of all compuin machines, nite st ate autom ata can perform a compuaion by chann from one sae o anoher in a well dened sequence wihou havin o sore inermediae resus Le's imaine a popuaion of hese simpe maces emboded physically One way of ein hem inerac is o creae a lare sn ne made of elecrically conducin maeria and o place a ne sae auoma on wherever wo wires cross The neracions would ake place hrouh an elecrica curren lowin hrouh he wires alown he sae of a iven auomaon a any one ime o be deerned by he saes of neihborn auomaa accordin o some rue The same neracn populaion can be creaed wih sofware insead of hard ware if an absrac woensiona space is subdvided no "cells and a simulaed ne sae auomaon is paced n each cell The cells may be rianular, recanular, hexaonal, or any shape ha exacly iles he plane so ha he cells are conneced o one anoher by shar n edes and corners Boh he physica realizaion of he populaion of auomaa and is simulaion n sofware are referred o as cellular automata The quesion is wheher he same sraey used wih em peraure and pressure will sill work when we replace a populaion of molecues wh one of compuin machnes Tha is, wheher we can idenify emeren properies, endencies, and capaciies, and wheher 22
we can show ha emeren eniies a one scale can be used o compose emeren eniies a a larer scale The bes known example of a celular auomaon is he socalled Game of Le The cells housn ne sae auomaa n Life are recan ular in shape and he auomaa are capable of bein in only wo saes "on and "off This is usually expressed as if he wo saes beoned o he cells hemselves which are said o be "alive or "dead, respecivey Life is no a reular compuer ame snce here is no connuous neracion beween he user and he compuer: he user merely ses he sae of a few cells (ha is, makes some of hem "alive and hen passivey waches as he cells affec and are affeced by heir neihbors The ineracion rues of Life are decepvely simpe: if a a iven momen a cell has wo neihbors ha are alive i say in he same sae i was n he previous insan, ha is, if he cel was ave i will say aive and if i was dead i remain dead; if he cell is dead ad hree of is neihbors are ave, i come aive; nally, if he cell is ave and ony one neihbor is ave, or if more han hree are, i wll die The rules are applied o he enire popula ion simulaneously an repeaedly for as on as he simulaion is alowed o las The fac ha he ineracions are riidly specied by rules impes ha hey are not emergent. On he oher hand, as he popuaion ineracs paerns of saes in neihborn cels appear and hese are indeed emeren snce he paerns have properies, enden cies, and capaciies ha are no presen in he ndividual auomaa The mos basic emeren endencies of paerns of saes are he same as hose found in real physica processes: he endency o reman n a seady sae and he endency o oscillae beween saes When play wh Life one quickly discovers ha mos niia paerns ds appear afer a few eneraion So he endency o mainan he same shape, even houh hardy a form of behavior due o he absence of chane, is no ordinary Many paes n Life dsplay s endency, he simpes one bein composed of four neihborin cels smulane ously aive, a paern called a "block These saic bu endurn pa erns are caled "sill fes A second class of paerns, collecivey known as "oscillaors, dispays a endency o cycle hrouh a series of conuraions over an over aain The simpes one, called a "bnker, consiss of hree neihborin cells n a row ha becomes a coumn n he nex eneraion and hen becomes a row aain 23
PHLOOPHY AND MULATON
The bnker spy switches back and forth between two phases but other osciators undergo a number of changes before returning to the nitia phase and repeating the series again. A third cass of patterns caed spaceships," adds to this rhythmc behavior a tendency to move in space The spest one is made of ve ive ces arranged in a "V shape that cyce through four phases the ast of which is ike the nitia one but dispaced one ce iagonay. On a computer the pattern seems to gide across the screen so it is referred to as a gider." Gders and other spaceships provide the cearest exampe of emer gence in ceuar automata: whie the automata themseves remain xed in their ces a coherent patte of states moving across them is ceary a new entity that is easiy istinguishabe from them. Sti ifes, osciators, spaceships, and other emergent patterns can interact and in these interactions they revea their capacities As with moecuar popuations the interactions are spe coisions but the resuts of a coiding event are so varied that they must be stuied empiricay. o giders, for exampe, can coide with each other in 3 iffeent ways the outcome depening on their positions and the phase of their osciation at the te of the event: 28 of these coisions resut in the mutua annihiation of both giders six produce a bock as their outcome three resut in a bnker two eave a singe gider behind and the rest produce a variety of sti ifes and osciators2 The possibe outcomes become more numerous as we increase the number of giders and whe efforts to cataogue them exhaustivey continue a more promsing route to revea the capacities of muti gder coisions is to ook for specic combinations that synthesize interesting objects There is a coision of 1 3 giders, for nstance, that synthesizes a pattern more compex than the gider itsef: a glider gun, an osciating pattern that produces a continuous stream of giders. There is aso an interesting capacity ispayed by some sti ifes: they are capabe of coiding with another pattern destroying it whie at the same time reconstituting themseves This capacity to eat" other pat terns aso depends on the capacity of the pattes to be eaten: a sti fe shaped ike shhook, for exampe can eat most patterns but is not capabe of igesting a bock. 3 Knowedge of these emergent tendencies and capacities can be used in the construction of engineered" Life pattes. Bocks, binkers, and giders are referred to as natura" patterns because when one 24
CELL ULAR AUTOMATA AND PAERN OF FLOW
starts a Life session with a random popuation of ive and dead ces these three patterns (and other spe ones) emerge spontaneousy This spontaneity is expained both by the fact that the pattes are smapatterns made out of a few ive ces have a higher probabiity of occurring than arge onesand by the fact that they can arise fo owing severa sequences of predecessor patterns: the more numer ous the sequences converging on a pattern the higher the keihood that one of them wi occur by chance. Most arge and compex pat terns, on the other hand, must be handcrfted A gider gun is a good exampe of an engineered pattern The rst gun to be created was based on a shutte," a pattern that moves ke a spaceship but peri oicay changes irection The shutte is a simpe trianguar pattern that may occur naturay but it is typicay short-ived because as it changes irection it coides destructivey against its own debris But if bocks acting as eaters a re positioned at an e xact pace they con sume the debris stabizing the shutte. Once stabiized two shuttes can be made to coide with one another and if their position and timig is just right they wi prouce a gider every 30 generations.4 Egineered patterns are portant because they hod the key to the other test of emergence: the abiity of emergent entities at one scae to become component parts of arger entities with their own emer get properties Gider guns are aready an exampe of such arger scae emergence but for the test to be signicant we woud want to determine if using the patterns that emerge in a popuation of the simpest automata we can buid another automaton with much higher
omputational apai The space of a possibe automata be described in detai in Chapter 10 but at this point it can be characterized as structured by singuarities in computationa capacity: a nite state automaton represents a capacity minum whie a so-caed Turing mahine represents a capacity maximum The main factor contributing to this greater capacity is access to memory, a degree of access that varies from zero the case of nite state automata to absote in the case of a Tring machine and its inite memory tape" The gap between the minimum and maximum of computing capacity can be bridged because the ifferent automata stand in a reation of parttowhoe to one another A Tring machine, for exampe, needs a nite state automaton as a component part to contro the head" it uses to read 25
PHLOOPHY AND MULATON
the contents of its nte tape nd to write on it. On the other hand, what we need for our second test of emergence is not a way to buld a Tring machne out of nite state automata ut a means to buid it out of emergent patterns of automaton states ike gliders nd glder guns There are two different ways of approachng this chalenge One is to actualy build a working Tring machine, with its read/write head and its memory tape, nd make it carry out a caculation that coud not be performed by a te state automaton The other is to show that in principle such a construction could be perormed withou actualy carryng it out Both approaches rely on the fact that computng machnes can be bul out of basic digtal crcuits caed lgical gates A Ad gate, for example, has two nputs and one output: if the value of both nputs is the numer " the output will also be a "; any ot her combnation yields "a as an output A Not gate takes a sngle nput and outputs its opposite: a a i the input is " and vice versa Using these elemen tary circuits as components larger ones can be created, such as a Fip-op circuit that acts as the spest type o memory. Ay of these elementary crcuits cn be created out of trnsistors and in that form they exist in the chips that power most persona computers, but they can aso be created out of emergent Life pattes A And gate, or example, may e bult out of a glider gun and an eater The glider gun produces a contnuous stre of gliders that disappears once it reaches he eater The nputs to the gate are two additiona glider streams, produced lsewhere and aready coded with nformation: the presence of a glider n the stre represents a " whle its absence represents a "a These two nput streams code destructively with the main stre at 90 degrees beore it reaches the eater. If both nput streams have a glider in the same positon (if both nputs have a vaue of " the rst nput collide and eliminate a glider from the main stre creatng a ho e through which the corresponding glider n the second nput can ass Al other combnations do not let any glider through Thus, the assembly acs like an Ad gate, the nput gliders that manage o make t through the man stream constitutng its output5 The "n prnciple approach stops here because with this And gate and a spler to bud Not gate the circuits needed for a ful computer coud be ult if we had enough patience
26
CELLULA R AUTOMATA AND PATERN OF FLOW
The second approach, on the other hand, does not a at spy provng that the construction is feasible but at working out the dif cut detals A current plementation can sere as n ustration: a sple Trng machine capae of beng n thee dfferent states and usng thee symbols to write n its memory tape Ths implementaton s easer to describe startng at the top with the full automaton and worlng our way downwards The lrgest scale components that must be buit are a te state machine controling the readwrite head, the memory tape, and a signal detector connectng the two together Each of the three main components can be assemled out of spler ones: a te state machne cn be plemented with ne memory cells a tape with two stacks; and the signa detector can e made ou of coding glider streams and of units that duplicae each stream, send ng the copies to the sacks nd he nite state machne Each of these subcomponens can, in tu, be bul out of even spler ones A memory cell, for example, can e made ou of a glider gun nd a special ype of shuttle (caled a "queen ee) tha acts as a reector for gliders, and can e accessed by two sreams of spaceships each pro duced by its own gun6 When these and oher components are assem bed togeher the resu lt is a worlg Trng machne tha can actually carry ou computatons The computations are too sple nd the process s too slow for he emergent automaton o be of ny practcal use, ut watchng the Trng machne n operation induces n the viewer the same ld of awe created y the emergent thundersorm discussed n the previous chapter ths case, however, we are not tempted o view the result as "miraculous because the mechasms just described are aready part of he explanation, an explanation that may e considered to be ausal despie the fact that glider cosons nd other events are no physical But this st leaes par of he mystery unexplained the mechasmndependent par. What possiblity space is nvoled n hs oher part of the explana ton? Gven tha the neractions between the nite sae automata inhaitng he cels of Life are dened by rules wha must e deer mned is the sngularities structurng the space f all pssible rules Ad what these snguarities must account for is the exisence of natural paterns like locks and gliders since all the engneered ones used to bud the emergent Trng machne depend on them Sricly speakn
27
PHLOOPHY AND MULATON
gemetry the cels It is wel , r example that certan uantities (such as the uantity energy) are cnserved in physical prcesses energy cannt spntaneusy be ceated r destryed The rues latticegas autmata map energy nt the number states representing individual mlecues and this number is always cn served Symmetry reers t the indierence that the reguarities in the behavir mecuar ppulatins display t beng transrmed n certan ways Rtating a ppulatin, r exape, leaves its behav iral regularities unchanged, therwise experiments cnducted in ur sping panet wuld give resuts entirely dierent in abraties riented in dieren directins Latticegas autmata using ces wth a suare shape d nt have enugh rtatinal symmetry (they remain nvariant nly r rtatins 90 degrees) ntrducng ariacts int a simulatin Hexagna cells n the ther hand have enugh rta tina invariance t avid such artiacts S Because s me ismrphism is already built int latticegas autmata their abty t reprduce patterns w is ess prbematic than that the slutins t the euatins uid dynamics Nevertheless, the expanain still pr vides sme philsphical nsight because it identies sme key prp erties (nvertibty, cnservatin, and symmetry) that are shared by the simuatins and the euatins and that help explain why the slutins t the latter can exhbit the regularities that they d additin latticegas autmata cmplement mathematica mdels n a nvel way the mdels used n statistica mechanics, r example crtain assumptins must be made abt the lkehd dierent spatia arrangements a mdel gradient dissipatin r example, we mus t assume that the disdered state at the end the prcess is much mre prbabe than the rdered ne at the start an assumptin that must be justied n sparate grunds But with atticegas autmata we can actually lw the mlecular ppuatin unt it reaches a given spatia arrangement This is an imprtant capacity simuatins nt shared by mathematical euatins the abty t stage a prcess and track it as it unlds Mrever several runs the sam� simuatin can be caried ut changng the values assigned t certan paraeters and tracking the unlding prcess t check whether it arrives at the same nal state S nce the vaues the parameters represent assumptins made abut the envrnment n which the simuated prcess takes place varyng them in dierent 34
CELLULA R AUTOMATA AND PAERN OF FLOW
uns is euivaent t checking whether the utcme is rbust t changes n the assumptins ther wrds, each run a smulatin is lile an experiment cnducted n a labratry except that it uses numbers and rma peratrs as its raw materials Fr these and ther reasns cmputer simulatins may be thught as ccupying an intermediate psitin between that rmal thery and labratry experiment 19 Latticegas autmata prvide an imprtant lnk between the subect matter this chapter and that the pevius ne It is time, hw ever t mve beynd ppulatins identical eements whse iden tity is nt altered by their interactins ther wrds, we must leave the relative simplicity physics behnd and mve t the mre cm plex wrld chemistry, specicay the chemistry the prebitic sup The mlecular ppulatins cmpsing the primrdial medium in which rganisms and their genetic cde emerged depart in tw ways rm the nes examned s ar mst the members a ppu latin change identity in the curse their nteractins resultng n the synthesis entirely new cmpunds whie thse that retain their identity (cataysts) are capable aectng that synthesis by acceerating it r deceerating it Cntrlling the reative rapidity r slwness wth which new cmpuns are prduced is a very sngular capacity ne that as we will see is as imprtant r the emergence ivng ceatures as is the capacity r selreplicatin itsel.
35
ARTIFICIAL CH EMISTRIES AND THE PRE BIOTIC SOUP
CHAPTER THREE
Artfcal Chemistres and the Prebotc Soup
So far we have examined questions of emergence in the reativey smple case of a body of water in which gradients of properties le temperature, density, or speed cause the spontaneous organization of the moecues into collective patterns of ow We can continue to use a body of water as our basic environment but in order to add a new ayer of complexity we must make its composition less homogenous This means that new gradents must be introduced: gradients of con centration of substances other than water These new gradents are aso characterized by a tendency to dissipate but unke the od ones countering that tendency involves the injection of a ow of matter not just a ow of energy The effect of that injection w in tu, depend on the specic chemica substances being njected Some sub stances, for instance, have a tendency to donate or export some of their protons, basically hydrogen atoms deprived of their eectrons. These substances are caed acids Other substances have the ten dency to act as acceptors or mporters of protons and are caed "bases When concentrations of substances with these opposite ten denciess come nto contact, forming an acidbase or Ph gradent, a spontaneous ow of protons from one chemica species to another is generated as the me'ns used by the graent to cancel itsef Another exampe of a cemica graent is an oxidationreduction or redox graent created when a substance that has a tendency to oxidize, that is, to donate eectrons, comes into contact with one that has a tendency to reduce or accept eectrons A redox gradent can drive an electron ow across chemica speces when it occurs, for example, ,
36
in an aqueous environment in the presence of concentrations of metallic substances Chemica gradients lke these are thought to have played an impor tant roe in the dynamics of the prebiotic Up But in addition to new gradents exporing the prmordia medium in which living crea tures rst emerged involves considering interactions in which ffer ent molecues can bond and form a new compound or, on the contrary, n which compounds can be broken down into their component parts This leads not oy to much arger combinatoria spaces with atoms and moecules of fferent species entering nto innumerable possibe combinations, but more mportantly, forces us to invent new means to explore possbility spaces that are intrinsically open-ended: if chemical interactions lead to the formation of a compound not originally pres ent in a mixture, and if ts compound is produced in large enough quantities to yield a concentration gradent, then the very nature of the miture as we as that of its possibiity space is changed I other words, the nherent open-endedness of chemica interactions allows possiblties not originay present in a space to be subsequenty added to it The phiosophica interest of the prebiotic soup is that by deni tion it does not contain any iving creatures in it and yet it has the capacity to lead to a novel situation in which such creatures do exist Thus, both mechanisms and mechanism-independent structure change radically once we reach the evel of chemistry and its heterogeneous moecular popuations What types of moecues n adtion to water can we assume were part of these prebiotic popuations? The most promising candidates are moecues caled polymers that are arranged in ong near chans I contemporary organisms both genes and the proteins these genes code for are poymers dffering oy n their composition: proteins use amno acids as component parts while genes use nueotides Today most proteins are synthesized from nstructions contained n genes so we face a chicken-andegg emma when trying to decide what came rst genes are needed to create proteins but protens are needed to construct and maintain the bodes that carry those genes Proteins, on the other hand, can aso be synthesized without the help of genes and their component parts have a higher probabiity to form spntane ously than those of genes2 So we w assume that poymers made out of amino acids were the tpe of moecue originaly popuating the 37
PHLOOPHY AND MULATON
roducing one o he exising members and i heir own roduction is caalzed b anoher member his wa an autcatalytic lp ma sontaneous orm and then grow he closure o such a cooerative nework o catasts would endow it wth stabilit aganst sruion b erurbations and aow it to concentrae the resources needed or growth amino acids, shor omers nto itsel o the exten hat hese resources are used or the benet o all the cooeraing mem bers the loo as a whoe behaves as a rotomeabolism, traing and gesting "ood is these emergent autocataltic oos ha are he real solution o the omerization robem and hence consi ute he man characers n he re biotic dr ama Let's summarize what has been said so ar he robem o he emer gence o living creatures an norganic world has a welldened causal srucure he mechanisms volved are a chemical n naure the creaion o bonds between amino acids to create rotes; the acceeration o bond creaion to comensate or the sonaneous desrucion romoted b an aqueous environmen, an acceeration exlained b he maniulation o graiens; he use o a lockandke mechanism to erorm molecular recognition and ocus the eects o graient maniuation on secic targes; and nall, the orma tion o cooeraive networks o cataltic olmers acing as roto metabolisms Ahough these chemica mechanisms are relaivel well known there are still gas in our undersanding he recognition mechanism, or examle, has on been arl exlaed For enzmes to have is caacit heir onemensiona srucure mus be oded to a hreemensional shae with the righ geomeric eaures he olng roblem has not et been solved, its sud consumng more suercomuter tme than an other roblem wha oows he comexities o he oing rocess wil be ignored to concentrate on the increase in length o the chans oher words, we wl assume that longer chas have more comlex emergen cataltic and recog nition caaciies than shorer ones adtion as i has been argued n revious chaers a causal exlanaion in terms o mechanisms must be sueened wth a mechanismdeendent comonen an elucidaion o he sructure o he sace o ossibilities associaed wh he rebiotic sou to deterine wheher is singularities make he exisence o auocaalic loos not just ossible but also higl robable 40
ARTFCAL CHEMTRE AND THE PREBOTC OUP
ac, we need to examne two ossibt saces the sace o ossible olmers o ierent engths and comosiions and he sace o ossible chemical reactions tha can roduce hose olmers he robabt ha an auocatatic oo can sontaneousl orm can then be assessed b a theoreica exaination o these wo saces and the acual emergence o such a loo tracked usg comuer simulaions he rst ting that needs to be evauated is the size o he two ossibili saces he number o ossibe olmers o a given length is the number o available comonents raised to the maximum length he roteins used b conemorar organisms the reer oire o comonents is limited to onl 20 amino acids Neverheless even a ver short olmer ve amno acids ong can exis n over three mion ierent combnaions the number 20 raised o the th ower he number o ossibe omers 300 ao acids ong, the engh o a smal contemorar enzme, is litera nnie hus, we are considerng ossibilit saces hat grow exosivel as he length o he olmers ncreases Next we need o esmate the number o ossible chemical reacions that can roduce this nit o combations, that is the number o reacions tha can glue or condense two shorter chans o a onger one, and vice versa, cut or ceave a long chain into two sorer ones For an given omer there are man ierent reacions hat could roduce it a snge ao acid can be atached at one end o an existing chan; or wo shorer chas can be condensed ogeher; or three even shorter ones can be used, each erhas the roduct o the cleavage o longer ones he existence o multile was o roduce each olmer mlies that the number o ossibe cheica reacions grows even aster than he number o rote secies as the length o he chans increases4 Fall we need estimate he disribution o cataic caacities, a task that is imossible o erorm wthout solving he olding roblem A ossible wa out o is icult would be to assume tha given he ni o boh ossible roten secies and cemica reacions he secic sribuion o caatic caacities does not maer much We can assume random sribuions acknowledging our ignorance o the detais, and sil reach some valuable conclusions Al hat we need to low is he likelihood that a series o caalzed reactions inerconneced o orm a closed oo can emerge sonane ous is can be estimated b creatng a grah n wich each roein 41
PHLOOPHY AND M ULATON
speces becomes a point and each chemical reacton between two proteins becomes a edge joinng two points The fact that as length increases the number of possble reactons grows faster than that of protein species means that the number of ways of interconnecting the points grows faster than the number of ponts. At a cetan citical length the number of possbe connections relative to that of points is so large that the pobabty that some catalytic proteins are connected nto a closed loop becomes very high5 This agument s sar to the one used to explain the tendency of graents to ssipate But as n that case, our condence n the conclusion woud be geaty strength ened if we could actuay follow the process n a smulation, tracking the population o polymers as they get longe and checkng f closed loops have indeed omed. These smuations ae colectively own as "articia chemisties� Some artical chemstries are screte smuations that explicitly represent ea nvidual polyme and in wch the nteractions between chains depend on the composition Others use contnuous fferentia equations ignoing the detas of invdua nteactions and focusing instead on rates of change the ate at wch cleavage and condensation reactions take place, the rates at wch oncentrations of polymers of a given length fom, and so on The contnuos approach, called metadynamics modls the pi moal body of wate n wch polyme evolution took place as a chemcal reaction vessel or reactor ) The reacto s driven by a ga dent of "food" that is, by an external ow of monomes and short polymes At any one pont in time the reactor contains a large bt te number o chemcal reactions, a number that can be reduced if we mpose a theshod of concentration below which the presence of a paticar speces can be considered nsignicant. hus, at any given instant we can model the dynamcs of the reactor by a te number of equations one for each chemical reaction, the entire set of equa tions possessng a single state space Assuming that the chemca reactions ae revesbe alows us to further smplify thngs because n tha case the ossibity space for the reacto is structured by a single singuarity a st eady-state attracto dening a stable stbution of concentrations of polymers of different species On the other hand, unke the state spaces of physics n wch the number of dimensions is given n advance and does not change, n metadynamcs the numbe 42
ARTFCAL CHEMTRE AND THE PREBOTC OUP
of reactons does change and so does the number of dimensons of the space Keepng track of the ncrease n the number of dimensons can be done by couplng a state space to a graph e the one just discussed, a graph capturing the exsting polymer species and their chemical inteactions 6 Thus, metadynamic smulations expore o copled possibili spaces the graph gives us the stcture of the space of possible chemcal reactions while the singulaities of the state space provde the strcture of the space of possibe outcomes for a given set of reactions A typical run of a metadynamc simulation proceeds like this inject some food into the reactor to ceate a gadent and get the dynamics going; follow the process until it reaches the steady-state attractor nd check what poymer species are present above and below the theshod of concentration add to the gaph any new species that is above the threshod, ease any od speces that is below it, and update the connecting edges to nclude the new possible chemical reactions change the set of equations and its assocated state space) to eect the change in composition of the reactor follow the new process unt it reaches a new steady state, adjusting the graph and the equations to include new poymer species and reactions Ts series of steps is then epeat ed many tmes unt the composition of the reactor stops chang ing. At that point another steady-state attractor is reached caled a "metadynamica xed point") but ts tme it is a sngularity repre senting the longtem tendencies of the entire series of reactions, and it is therefore much moe mportant that all the ntervening ones7 Once at the metadynamical xed pont we look at the graph to se e if it contans nodes connected by edges into a cosed network. If such a closed network exists it implies that, had the smulated reactor been populated by invdual polymers they woud have formed an auto catalytic loop. Finally, the smulation is performed several tmes vary ing the values of certan control parametesthe degree to wch water affects the reactions the rate at wch food is fed into the soup the stribution of catalytic capacities in the polymer popuatonto check that the emergence of the oop occurs over a wde range of vales and it is ot just an artifact8 The results of the metadynamic simatons that have actuay been peformed show that the sponta neous emergence of a protometabolism s indeed a likely outcome, one that could have occued n prebiotc contions. 43
PHLOOPHY AND MULATON
o further reduc e the lkelhood that the resut s an artfact we can try to reproduce t usng other smulatons based on entrely dfferent prncples o be truly dfferent ths other smulaton should not rely on macropropertes lke overall rates of concentraton but rather enact mcro-nteractons between explct representatons of ndvd ua poymers and et those nteractons generate the concentratons. Lke metadynamcs the new smulaton can explot the smarty between lnear chans of moecues and lnear strngs of symbols but t must not empoy the strngs as mere labels for an entre speces of poymers. Instead t must use them to capture the behavor of ndvdual polymers. In partcuar, a symbo strng must capture the double fe of an enzyme: on one hand, an enzyme s smpy a chan of amno acds that can be acted upon to be ceaved nto parts or be condensed nto larer wholes; on the other hand, after ths chan has foded nto a three dmensonal form an enzyme s an entty wth the capaty to act on other chans breakng them up or glung them together he perfect canddate for a symbol strng wth such a double lfe would be a smal computer program. When a computer program s coped from one hard dsk to another, for exampe, t s treated as a mere chan of ones and zeroes but when t s executed t becomes a procedure that can perform a varety of operatons on other chas of ones and zeroes hus computer programs lead the knd of double lfe that we need for the new artcal chemstry Not all programs however lend themselves to ths task because we need to consder not only the capacty of a program to affect the data but also the capacty of the data to be afected by the program: pro grams that perform arthmetcal operatons, for example demand that the data conssts of numbers not words n Englsh whle pro grams that perform lngustc operatons need words not numbers. Smlarly n an artcal chemstry when a program s treated as a mere strng of symbols t must have the rght capactes to be affected: the symbol strng must be able to undergo random cleavages and con densatons and stll retan enough coherence to be executed as a pro gram. Programs wrtten n conventonal languages contan mechaca recpes speced step by stepusng speca constructons to repeat seres of steps and to transfer contro from one recpe to anotherand these recpes are completely destroyed f we add to them or delete from
44
ARTFCAL CHEMTRE AND THE PREBOTC OUP
them random peces of code. But a dfferent fy of programmng languages, called "recursve functon anguages, does not specfy computer programs as stepby-step recpes but smpy as transforma tons of nputs nto outputs that s as functons and moreover as functons that can take other functons as nputs and produce yet other functons as outputs hs means that a complex program can be bult usng a part-to-whoe relaton: complex functons can be created out of smpe functons whch are composed of even smpler ones a the way down to operatons that are drectly executable. he resutng compex functons can be expressed as symbo strngs, wth parentheses separatng the dfferent nested components, strngs that have a much hgher chance of remanng coherent after beng acted upon by other functons9 Usng a recursve functon language a foded poymer wth cataytc capactes can be represented by a strng of symbols wth the ablty to affect other strngs whle the unfoded poymer that s the target of the catalyst can be represented by a symbo strng wth the capacty to be affected by other strngs When these two strngs nteract a thrd strn results representng the product of the chemcal reacon cata lyzed by the enzyme. o smplfy the smulaton spata relatons are not modeed explctly, that s the smulated poymers are not related to one another by relatons of proxmty. hs s equvalent to assum ng that the chemcal reactor s subjecte to a constant strrng moton that destroys all neghborhood relatons between polymers. IQ a wel strred reactor all collsons between moecules are bascally random so nteractons between polymers can be modeled by randomly pck ng two symbo strngs from the exstng populaton, treatng one as a program (enzyme) and the other as data (target substrate) then releasng ther product back nto the populaton. hese chance encounters are nluenced by the changng concentratons of symbol strngs of the same "speces, just as n a welstrred reactor the fre quency of collson between two polymers of dfferent speces would depend on how many members of each speces exst n the chemcal soup Because recursve functon languages have the computatonal capacty of the most sophstcated automata and because of the ran dom character of the collsos ths artca chemstry s referred to as a Turing gas 0
4S
PHLOOPHY AND MULATON
A typial Tring gas simulation starts with a popuation of randomly onstruted funtions The population, onsisting of about 1000 members, is then alowed to interat produing new funtions. Some of these produts are identia to others aready present others are nove in the sense of not having ounterpars in the existing popua tion' whie yet others are absolutely novel in that they have never existed in the "history represented by a partiular run of the sula tion The population is deliberately kept onstant by radomly dran ing exess symbols strings from the gas and this ats as a seletion pressure: the ony way for a simulated polymer to remain n the popuation for extended periods of time is by beoming the produt of a stable series of nterations between funtions. And n tu the ony way for a series of interations to be stable is to dispay atalyti losureY The simplest autoatalyti loop is omposed of a funtion that ats on a symbol string to produe another symbol strng that when treated as a funtion produes the symbol string orresponding to the rst funtion. Suh losed sets of interating funtions have n fat been produed n Trng gas simulations and an be shown that they orrespond to singularities (nvariant xed points) of the dynamis of th system The results from Trng gases and metadynamis show that auto atalyti loops ating as protometabolisms are a reurrent feature in artiial hemistries using very different formal resoures Tis isomorphism in turn may be evidene of yet other isomorpism one between simulated poymer populations and the real polymer popu lations inhabiing present day hemial reators or anient prebioti soups. That is the spaes of possibilities of the sulations and the real environments may exibit sufient overlap so that what we lea from the former an be used as the mehanismindependent part of the explanation of the latter On the other hand, the spontane ous emergene of autoatalyti loops annot fuly aount for the orig of life beaus these loops do not have the means to store infor mation and to pass whatever advantages they may ave aquired in atalyti efiey and speiity to future generations A possible solution to tis problem relies on the fa tat the results of these simuations also apply to polymers made out of nuleotides, polymers that do have the ability to a as tempates in order to produe opies of themselves. Moreover while some has of nuleotides suh as
ARTFCAL CHEMTRE AND THE PREBOTC OUP
those haraterizing ontemporary DNA, do not have atayti apai ties the hains omposing some forms of A do have them and tis meas that aient A may have ombined the ability to trasmit nformation aross generations with the propery of atayti osure. The foowng hapter wil expore tis senario ad the new apaity that emerges when a popuation of sefrepliatng moeules is ou pled to ay proess that biases that repliation one way or another the apaity to perform a searh n a spae of possibilities.
7
GENETC ALGORTHM AND THE PREBOTC OUP
CHAPER FOUR Genetc Algothms and the Pebiotc Soup
A choice of a simple but signicant starting point for an examinatio of emergece has proved mportant in the case of both physics and chemistry. A similar strategy wil be sefl when we move to te realm of biology We mst remain in an aqueos environment pop lated by molecules but add a new capacity: the ability to selfrepicate The simplest self-replicating entities are RNA molecules existing in a free state, that is, not encapsulated by a membrane or sbordinated to the needs of an organism To explan their mechanism of self-repication we do not have to consider the complex process used by iving creatures, a process in which the genetic code and al its translation machinery are crucially involved All that we need to account for is the relatively simple process through which an RNA molecule c serve as a template for the creation of a copy of itself A population of these repicators in t introduces a new kind of gradient, a gradi ent of tness Fitness refers to the ifferential reproductive sccess of embodied organisms bt it can also be appied to molecular replica tors and their ifferent capacities to produce copies of themselves can beregarded as a gradient because tness differences act just ike temperatre differences: as long as the differences exist they fel a process of selection aVOring the replication of one kind of polymer over another; te moment the ifferences disappear the selection process stopS Focusing on " aked RNA molecules has therefore the advantage of addng oly a minimal layer of complexity to the world we are explor ing but a layer that already has many of the properties that separate 8
biology from chemistry. A RNA molecule is simply a sequence of for possible ncleotides, each member of the set displaying clear tendencies to bond with just another member we refer to the four nucleotides sing the conventional letters (A C G U) we now that s tend to bond with U's, and Gs with C's These tendencies, of corse, need to be explained by yet other mechanisms but for the prpose of acconting for self-repication they can simply be assumed to exist Given these tendencies a seqence of ncleotides can create a copy of itself sing a mechanism similar to that of conventional photography: rst a "negative is created as complementary mono mers attach themselves one by one to the template, and then a "posi tive print is prodced from that negative. We also now that the amont of energy involved in forming the complementary chemical bonds is not that large, rougly equal to that of thermal luctations in an aqeos environment and this impies that errors in the copy ing process tend to occr spontaneously other words the copying errors or mutatons that provide the necessary variability to prevent tness graients from isappearing do not need a separate explanation 2 Discussion of the mechanism-independent structure is also simpi ed by starting with naked RNA becase the possibiity spaces associated with living organisms can become very numeros as the complexity of their part-to-whole organization increases the case of nicelular organisms like bacteria, for example, we must consider at least three ifferent spaces: the space of possible genes; te space of possible structural proteins and ezymes that these genes code for; and the space of possible spatial structures and metabolic pathways that the structral proteins and enzymes can form I oter words, we need one possibiity space for the hereditary information (the "genotype) and two possibility spaces for the bodiy traits produced from that information (the "phenotype) In multicelllar organisms several other possibiity spaces must be added since the phenotype also incldes the space of possible cell types (such as muscle bone nerve, blood); the space of possible tissues and organs these cels can form; and the space o possible organsms these tissues and organs can compose Moreover, in large animals the mapping of genotype into phenotype involves a complex embryological process that is oly imperfectly nderstood By contrast naked RNA has a very simple 9
PHLOOPHY ND MULATON
relation between genotype and phenotype: the unfolded poymer constitues he former whe he foded version and its caayic capac ities constitutes the atter. Tis means that we need to consider only two possibiity spaces the space of possibe sequences of the four nuceotides, the size of wich is the number four raised to the ength of the poymer and the space of foded forms a smaer space given that severa different sequences of nuceotides can end up with equiv aent foded forms3 Like other discrete combinatoria spaces the space of possibe A poymers is intrinsicay amorphous. But we can mpose an order on i as long as his is justied by reference to the mechanism of replica tion and as ong as it seves to revea the structure of the possibity space Thus owing that the variation in a popuation of moecuar replicators invoves copying errors at specic poins in a sequence we can arrang them so that each moecule has as neighbors other moecues differing from it by only one muaion Tis space is muti dimensiona because it must incude a he variants that can be cr ated by varying one monomer aong the full ength of a given poymer and each of these one-mutant variants must be assigned a differen dimension of the space But wie the sheer number of dimensions makes the space very compex it aso grety smpies the distribu tion of possibe moecues every possibe poymer is in direct contact wth al its one-mutant neighbors and a series of neighbors forms a conected path for evoution to foow I other words given this spatia arrangemen moecuar evoution can be visuaized as a con inuou wa from one neighbor to the next driven b events produc ing one mutation at a tme.4 To compete the characerization of the possibiity space we must assign a tness vaue to each of the possibe poyers Fitness refers in tis case to the consequences for reproduc tive success of the cataytic capacities that each poymer woud have if it were foded into a threedimensiona shape We showed in the previous chaptr how caaytic capacities exercised in cooperative networks can ead to the focusing of resources on a particular set of moecues so w can now take tis for granted to concentrate on the effects of cataysis on rates of replicaion A possibiity space of sef-replicating entities to wich tness vaues have been assigned is called a ne andcape. The strucure of a given andscape depends both on the distribuion of sguarities, 50
GENETC ALGORTHM AND THE PREBOTC OUP
sequences of nuceotides with optmum tness and on how ness varies in he neighborhood of hose singuariies. For exampe if one mutant neighbors have smiar tness hen the singuarities i form a distribuion of ocal optma resemblng the "peaks of genty soping hils If tness varies great between neighbors on the other hand hen the andscape be more rugged with loca optma peaks surrounded by steep cliffs Fitness andscapes are importan because they compement the traditiona mode of evoution I that mode scarce environmenta resources cause some repiators to be seected at he expense of others, a seecion process expained by the capacity of the envronment to affect the repicators But we aso need to consider he capacity of the replica ors to b e affected and this is what a tness andscape captures. More specicay in a andscape with many oca opma separated by ong distances distance being mea sured in tis case by the number of mutations needed to convert one sequence nto another seecion pressures aone canno account for particuar outcomes5 A evoving popuation may, for exampe be trapped in a oca optmum if the path to a singuarity with greater tness passes through points of much esser tness In tis case seec tion pressures aone canot disodge the popuation from the trap. Ths, o the extent tha moecuar evoution can be modeed as a wak aong a continuous path of one-mutant neighbors the "topogra phy of the tness andscape must be part of the mode since it fai tates or obstructs those waks A precise characterization of tis topography woud mpy know edge of the exact tness vaue of each foded poymer But even with out tis quantitative information we can get a sense of the quaitative characeristics of a andscape whether it has a sge oba optimum or many oca optma, for exampe, or whether the neighborhood of those singuarities is smooh or rugged A mportant factor deter minng is qualitative structure is the degree to wich the different components of a polymer interact o determine tness When repro ductive success is determined by cataytic capacity nteractions between components mater because as an RNA poymer fods two or more nuceotides from different parts of the sequence may end up together forming he par of the catayst that recognizes its target causing the ness of the pomer to depend on severa components at once. By systematicay varing he degree of interaction among 51
PHLOOPHY AND MULATON
componens we can deerne he effecs ha his facor has on he opography of he landscape A one exreme here are no inerac ions and he landscape is srucured by a single global opimum one bes caalys surrounded by muans declinng smoohly in ness A he oher exreme when he ness of he polymer involves iner acions beween all is componens, he landscape is srucured by many local opima around which ness decreases seeply The rs exreme makes evoluion enirely predicable while he second one makes i enirely unfeasible The neresing cases are, herefore, hose corresponding o inermediae degrees of ineracion A paricularly imporan case is obained when he number of componens ha inerac is small relaive o he lengh of he polymer in ha case here are many local opima wih relaively high ness heir sides slope smoohly and h e highes peaks end o b grouped ogeher in paricular areas of he landscape6 To undersand how he srucure of he possibiliy space affecs evoluionary walks we also need o deermine wha exacly performs he walk Is i for example, individual replicaors compeing agains each oher or is i enire groups compeing wih oher groups? When considering he evoluion of large animals and plans he rs alea ive is more plausible while he second one is more adeque o deal wih molecular populaions he members of which can be convered no each oher by a single muaion e's imagine a paricular RNA polymer wih very high ness locaed a a local opimum wih smooh slopes Sponaneous muaions will end o produce many one-muan neighbors wih simar ness These neighbors can also muae o produce slighy less wo-muan neighbors ha, in u can generae hree-muan neighbors and so on. As we move away fro he es polymer he number of viable muans d ecreases unil heir number fades o zero he enire se forming a coheren cloud of muan Ahough he cloud conains a es member his donan ype may no be he one ha reproduces he mos as i would if he compeiion was among single replcaors The reason is ha a non donan mun ha is surrounded by very ones can ou repro duce he dominan one, no because of is own capaciy o make copies of isef bu because is neighbors can by a single muaion cre ae copies of i This amplcaion of reproducive success by nearby
52
GENETC ALGORTHM AND THE PREBOTC OUP
muans means ha he cloud as a whole a s ocaled "quasi species becomes he arge of selecion7 e's imagine a muan cloud moving over a ness landscape in which local opima of differen heighs are disribued in relaive proximiy he cloud iniialy clusered around a local opimum wih he donan sequence a he peak As evoluion proceeds par of he cloud may sumble upon anoher nearby opimum of higher ness A ha momen he niial cloud wl mel and recondense around he new sngulariy This means ha when a whole cloud of muans moves n a ness landscape he overall movemen is no a random walk s one would expec given he randomness of muaions, bu a earch: no a search wih a goal of course, bu a groping n he dark ha is neverheless beer a nding local opima of higher ness han a random walk A crucial facor behind his searching capaciy is he rae a which copying errors are made, ha is he muaion rae If he muaion rae is zero he cloud does no form while if i is oo high he inernal coherence of he cloud is desroyed As before i is he inermediae values ha yield he ineresing cases I addiion, he engh of he polmers mus be aken ino accoun since he lon ger he molecular chain he more possible muans i can have These wo facors ogeher deerne wha is caled he error hrehold a singulariy similar o a phase ransiion because beyond he hreshold he "liquid cloud ends o "vaporize The heory of quasispecies predics ha clouds wih he capaciy o search a ness landscape wil end o lie ear he error hreshold ha is o have as many mua ions as possible wihou suffering disinegraion Evidence ha his is in fac he case comes from aboraory sudies of he only conempo rar orgsms ha use RNA o sore geneic nformaion, viruses Sudies of viral popuaions as wel as of populaions of naked RNA exraced from hose viruses conrm ha hey end o form quasi species esing in he viciniy of he error hreshold 8 As in previous chapers we cn derive furher philosophical insigh if in addiion o considering he heoreica reamen of evoluionary walks on ess andscapes we could acually folow hose walks in a simulaion The replicaors in hese simulaions could be equaly "naked ha is, exis in a disembodied sae and have he simples relaionship beween genoype and phenoype Simulaions using he
53
GENET C ALGORITHMS AND THE PRE BIOTIC SOUP
PHILOSOPHY AND SMULATON
(a "secondary structure) ca be reliably produced. And while we cannot calculate the catalytic capacity from this intermediate form we can calculate the energetic stability that the folded form would have12 This measure of tness is justied because an unstable shape would not last long enough to exercise its catalytic capacties. If we hnk of a folded RA polymer as the form produced by a gradient as it dissipates then the stable folded forms would be those that managed to achieve a combination of bonds that fully dissipate the gradient This stability can be calculated from the secondary structure and used to evauate tness Given this simplied relation between genotype and phenotype a genetc algorithm can be used to folow a popUlation of simulated RA molecules across may generations and check whether stable forms emerge that can become the centers of quasi-species One sim ulation for example started with a population of 3000 strings, each 0 components long, but in which all components were identcal Since such a homogenous string would be incapable of folding the population started its evolutionary walk at a zone in the landscape that was a "lat plain of very low tnss and had to nd and climb a "hill representing a local optmum The most important parameter that could be varied in the smulation was the mutation rate, the fre quency with which copying errors were made since as we saw before this partly determines the coherence of the mutant cloud The muta tion rate is normally kept xed in most applications of genetic algo rithms but in this one it had to b e varied from o ne run of the smulation to another to test the existence of the error threshold The results of the smulation conformed to the theoretical prediction: the popula tion of strings eventually found a local optmum and became capable of folding (as inferred from the secondary structure) but only if the error rate was below the critical threshold 13 This is a relatively imited result particularly if we consider what the eventual goal of this kind of smulations would have to be: to show how the searching capacity of molecular replicaors could have led to he emergence of the genetic code d not nly the code itself that is the rigid correspondence between three nucleotides in one type of polymer (genes) and a single amino acid in another typ (proteins) but the means to reliably and routinely translate one into the other
56
While the naked replicators of genetic algorithms are idea to model the prmordia soup prior to the emergence of the genetic code a dif ferent approach is needed to track evolution after that singular event In particular, we need the means to increase the level of embodiment of the replicators given that the coupling of genes and proteins could no have occurred without the encapsulation of these interacting polymers within prmitive membranes And as we consider more embodied replicators the relation between genotype and phenotype must be modeled in more detail ven the earliest cellular organisms had to solve metabolic problems by using their lmited repertoire of proteins and enzymes as recombnable components to form primitive chemical circuits So simulations of celular evolution must have the necessary resources to give rise to simple part-to-whole relations between different levels of the phenotype And if these relations are to be inheritable the genotype tself must be modeled by symbol strings capable of incremental compositionality n the previous chap ter we saw that smple forms of chemical circuitry (autocatalytic loops) could emerge in a soup of symbol strings belonging to a recur sive function language that is in a Tring gas We also saw that those languages alow for complex computer programs to be generated using a part-to-whole relation: the smplest functions can be used to create more complex functions that in tu can enter into even more complex ones The computer programs so generated can be expressed as symbol strings in which the nested functions at different levels are separated by parentheses The branch of evolutionary computa tion using these more sophisticated symbol strings is called genetic programming. I fact in genetic programming the smulated chromosomes are not mplemented as strings but as graphs having the form of a tree This eliminates the pleasing similarity between a chain of molecules and a string of symbols but compensates for it by clearly displaying the part-to-whole relation between problems and subproblems: the over all metabolic task forms the root of the tree while progressively more detailed sub-tasks form the branches of the tree I industria applica tions the power of genetic programming has been clearly demon strated by going beyond the solution to optimization problems ike the gas pipeline discussed above In particular genetic programming
57
PHLOOPHY AND MULATON
can tackle more challengng design problems One example is the design o analog electrical crcuits, an aea n which human expertise has traditionally been required and n which the degee o orignality o a given design can be oughy assessed by the act that a patent oce has accepted it Usng genetic programming populations o replcators searching the space o possible circuits have red scoveed several circut designs that had been previously patented have ds covered novel esigns that match the unctionality o patented designs and n at least one case they have ound an entirely new design that is, one not logically deducible rom a previously patented invention 1 4 This abiity to generate nonobvious designs in which components are interconnected in complex ways can be very useul to explore the evolutionary origin o the metabolic circuitry o unicel ular organisms To show how ths could be done let's begin by descibing how genetic programming works in appications not involving the model ing o biological processes First o all, given that complex programs are evolved by recursion, that is by unctions that result rom the composition o simpler ones, we need a set o elementary unctions to get the process started. These pmtive unctions orm the nodes or branching points o the tree We also need to speciy the variables and constants that those elementary unctions use as their inputs (the "leaves" or terminals o the tree) These basic components must be chosen by a designer to match the type o problem to be soved5 the case o analog electrical circuits or example the eementay operations include unctions that insert a new component (a resisto a capacitor, an inductor) unctions that alter the connectivity o those components (the topoogy o the circuit) and unctions that set the intensity (or sizng) o a component that is the degree o esistance o a resistor the capacitance o a capacito and so on the case o dgital circuit design the elementary unctions shoud be operators like "nd "O and "Not the case o robotic motion that is when the poblem to be solved is the synthesis o motion paths the elementary reertoire must contain unctions ike "rn Let, "rn Rght, and "Move Forwad These elementay unctions are the component o the evolutionay process that is not emergent snce their behavior (the way the unctions aect their inputs) is dictated by rues ny complex unctin that can be created om the elementary 58
GENETC ALGORTHM AND THE PRE BOTC OUP
ones by recursion on the other hand can be considered emergent i it evolves spontaneously as a speci solution to a design problem 16 The nal step is to create a tness unction dening the problem · to be solved Snce the evolvng entities in this case are computer programs evaluating their tness demands actually running them because what is asses sed is the behavior o the program or the con struction capacities with which that behavior endows the program This evaluation is pero rmed by comparing the running program with a set o test cases that already possess the required capacities the degree o tness expressed as a distance rom that target the case o analog electrical circuits the evaluation o tness is in act, more complex because we need to check not ony that a unning pogam can constuct the circuit but also that the circuit itsel behaves as it is intended to: as a lter as an amplier as a sensor To do this an "embryo crcuit (an electrical substructure with modable wires and components) is placed into a larger circuit in which no component is modable Only the embryo evolves but its placement into a larger unctional setting allows it to be checked or viabity 7 s it happens the nsights and techniques developed in the appication o genetic programming to electrical circuits can be used to model cellular metaboisms because in many cases thee is an isomorphism between the two. Lets examine a concrete example The target o a simulation o this type is an actual piece o biochemical machinery that has been studied in the aboratory This typically includes a network o chemi cal reactins a set o enzymes acting as catalsts and an initial set o concentration gradents o the substances serving as substrates tra dtional biochemistry dierential equations ae used to model the rates at which the substrates are transomed into eithe intemediate or nal products as well as the way in which the catalysts aect those rates by acceerating them or decelerating them These equations are then used to pedict the concentration o the nal poducts a predction that can be tested by carrying out actual experiments The question or genetic programming is this given data about the initial and nal substance concentrations can we revese enginee the chemical network that ks the two? This would involve discovering the topoogy an sizing o the network that is the connectivty o the dieent reactions and the values o the rates at which they proceed 59
PHILOSO PHY AND SIMULATION
one simulation, for example, an actual metabolic pathway usng four different chemical reactions was used as a target. The topology of the pathway was complex: it contaned an internal feedback loop; a bifurcation point at which one substrate was used for two different reactions and an accumulation point at which the concentration of a substance was affected by two different sources18 The primitive functios used in this simulation included mathmat ical equations describing chemical reactions wle the terminals were the inputs to those equations, such as substrates and catalysts Both of these came from models used by biochemists ad were therefore not emerget On the other hand, the initial population of random programs had no nformatio about how many reactions the meta bolic pathway was supposed to have, how these reactios should be coupled to one another, or what substrates each reaction got as inputs and what products it produced as outputs The only information the evolutioary process had about its target was built into the tess function: the concentration of the nal product of the metabolic pathway The results of the simulation were encouraging A initial population of 100,000 replicators was able to discover the topology of the chemical network (including the feedback loop, and the bifurca tion and concentration points) in about 1 2 0 geerations, while the sizing of the network (the values for rates of concentration and rates of catalysis) was closely approximated n an additional 100 genera tions.19 The evaluation of the tness of each generation was facilitated by exploitig the isomorphism of chemical and electrical circuits voltage gradiets played the role of concentration gradients, while circuits that performed addition and subtraction represented the pro duction ad consumption of a given substance2 Once an isomorphic electrica circuit was created could be tested for tness using the same approach discussed above: placing an evolving embryo circui ino a larger non-evolving one ad using standard software to check for viability The coclusion we may draw from this and the previous biological simulation is tt the capacity of a population of variable replicators to search a pos sibility space is ideed real On the other hand, demon strating the existence of a capaty is ot the same thng as explaining it The simplest explanation, the one applyng to naed RA, would rely on two assumptions: that the space of possible polymers does in I
60
GENETIC A LGORITH MS AND THE PRE BIOTIC SOUP
fact possess t he order we imposed on it, that is, that the sequenc es are arranged so that onemutant variants occupy neighboring positions, and that the search is carried out in parallel by an entire cloud of mutants But once the relation between genotype and phenotype is taken into account this explaation ceases to be sufcient Even in the case of naked RA the topology of the possible becomes more complicated In particular, to understand how one folded catalyst that targets a certain chemical reactio could evolve ito another one ith a different target, the path from one phenotype to the other must be also represented in the way the space is ordered Oe way of achieving this is to impose an order that esures that between two molecular sequences, each one occupying a local tness optimum, there are "ridges joiing the two "mountai peaks, ridges in which the sequences all have the same tness Tis way a mutation could transform one sequence into a neighboring one ith the same tess while keeping the transformation invisible to selection pressures A series of such neutral mutations could therefore lead from one folded catalyst to a very different one by walking across the ridge even if there are zones of lower tness betwee the mountain peaksY A differet ld of problem wth the order we imp ose on the other ise amorphous space of possible polymer sequences is that in addi tion to mutation some unicellular organisms use sexual recombination. When two differet chromosomes are sexually combined each paret contributes half of the genes composing the offspring 's chromosome If we imagine that the ancestors of each parent had clmbed a differ ent local tess optimum the resulting offspring would fall between the two peaks, that is, in an area of lower tness, and would therefore be eliminated by atural selection Ts egative effect may be ame liorated if the landscape has the right topography, that is, if all local optima are grouped together forming a tight cluster of peaks22 But a more general solution may force us to rethink how the space of possible polymers should be ordered so that its connectivity is com patible ith both mutation ad sexual recombination This is particu larly importnt whe giving a accout of the emerget searcing capacity of simulations based on geetic algorithms or geetic pro graming because in most industrial applications these favor the use of a crossoer operator over a mutation operator as a meas to gen erate variation The crossover operator works this way: it takes two 61
PHLOOPHY AND MULATON
separate smulated chromosomes and picks a particular point in each one of them; it then breaks both at that point crossing oe of the two halves from one parent over to the other; nally it reattaches the two parental halves before releasing the offspring into the population. n other words, the crossover operator mimics the effect of sexual recombInation. One plausible explanation for the superiority of crossover over mutation starts with the assumption that the surival problem posed to an evolving population can be decomposed ito subproblems each of which can be solved separately n other words, the assuption is that complex adaptive problems that have a modular structure can be solved one buiding bock at a time, ith sexual recombination playng the role of bringing separate building blocks together into a full solution To explore this scenario a special ethod has been developed to keep track of the propagation of buiding blocks in simu lations using genetic algorithms f we imagine the simulated chromo somes to be ten bits long the population contaning strngs ke "1101100000 or " 0 0 1 1 0 1 1 1 1 0 we can identify the similarities among strings with a notation using wld cards The two strings just mentioned for example have similar genes at positions four and ten a slarity that can be represented by a string using a wld card symbol ke "# The string "###1#####0 is called a schema23 Once every generation each member of the population of strings is evalu ated by the tness function for its performance as a solutio to a prob lem f the population contained say 50 strings that would cover a relatively sall area of the possibility space But if in addition we evaluated the average performance of each schema we would expand the area searched For exampe, if the two strings above belonged to the population but not the strings "1001100010 or " 0 1 1 1 0 1 1 1 0 0 the the latter two would not be explicitly evaluated. But because these two strings are also instances of "###1#####0 we would get information about them wen evaluating the average tness of the schema. f we tnk o each strng in the population as a point in a multi imensional possibility space then a schema with many instances is like a slice (or hyperplane) of that space24 Thus a population in wch strngs share buing blocks woud not be like a cloud of points mov ing through the possbility space but rather like a set of slices covering 62
GENETC ALGORTHM AND THE PREB OTC OUP
a much wider area and therefore increasin the parallelism of the search The crossover operator, on the other hand, can easy destroy building blocks if the cut point happens to separate its componets This mplies that a instance of a schema in which the xed posi tions are close to one another has a better chance o preserving its buiding blocks after many recombinations: instances of the schema "#10#######, for example, have a better chance of preserving the buiding block "10 than those belonging to "###1#####0 t can be shown that a schea with any wld cards (representing bigger slices of the possibility space) and in which the few xed positions are clustered together ca propagate through a population as long as the tness of the schema is greater than the aerage tness o the population Ts result is known as the schema theorem. The schema theorem, however says nothing about how building blocks are brought together over many generations to be combied into a full solution to the problem n other words the schema theorem explins the propagation of good buing blocks but not their accre tion.5 But the latter is equally important because as builing blocks come together the evolutionary search is conducted on spaces of ever decreasing imensionality and ever increasing tness that is, each schema represents an increasingly larger slice of the space and contains several partial solutions to the overall problem s there any eidence for the exstence of bung blocks in real populations of replicators? o separate but related iscoveries in biochemistry and genetics bear on ts question On one had biochemists studying protens and their folded structures have identi ed those parts of their surfaces that play a role in enzymatic actiity: their bining sites and their catalytic sites Because these sites are relatively compact and exhibit specic folded motifs (sheets helces ribbons) they are named domains Smilar domans in ifferet proteins tend to have simlar functions a fact that suggests that they are indeed reusable bung blocks6 On the other hand, geneticsts have iscovered that th genes of unicellular orgaisms in which the genetic materials are encapsulated into a nucleus are not structured as one continuous unit as they are in bacteria that lack a nucleus nstead their genes have a mosaic structure in wch sequences of nucleotides that code for proteins (called exons) are interspersed with sequences that do not code for aytng (called introns") 63
PHLOOPHY AND MULATON I many
cases one or more exons code for a specic domain in a pro tein, that is, for a buling block, so exons can be reshufed through sexual recombination to produce novel proteins at relatively low cost, a hypothesis made more plausible by the fact that each protein domain has its own foling capabilities so that new combinations can be expected to fold together neatly27 Additionally, the presence ointrons should lower the risk of disruption of exons if the cut point used in sexual recombination lies on an intron then the coevolved compo nents of a building block will not be disrupted Despite its more elaborate mapping between genotype and pheno type the replicators used in genetic programming remain disembod ied and are not situated in space And the fact that the evaluation of tness is performed exogenously, that is, that the target of the search is set by test cases provided by the designer, implies that the process is more like the controlled breeding of farm animals and plants by human beings than an evolutionary process in which some species acts as selection pressures on other species This means that questions involving the coevolution of species or the emergence of food webs in ecosystems cannot be explored using these simulations n the follow ing chapter we wil remove these lmtations by embodying the repli cators, situating them in space, and making their reproductive success depend on their capacity to meet their metabolic needs by tapping into resource graients distributed in that space
CHAPER FIVE Genetc Algortms and Ancent Organisms
With the emergence of the rst living creatures the aqueous environ ment we have been exploring acquired a new feature a gelatinous stratum of colonial bacteria at the interface between bottom sedi ments and water forming the earliest biosphere This new stratum remained deceptively smple for over two billion years because the motionless microorganisms that composed it hardly changed in their anatomy] But contrary to what a supercial look at their external appearance may suggest those ancient organisms had managed to dis cover over that period of time all the biochemical processes of energy extraction that exist today After that, evouton produced many new mechanisms to use energy to perform work, for complex locomotion or for neural control, for instance, but no new major ways of extractin enery from radients2 Roughly, the earliest bacteria appeared on this planet three and a half billion years ago scavenging the products of non-biological chemical processes; a billion years later they evolved the capacity to tap into the solar gradient, producng oxygen as a toxic byproduct; and one bon years after that they evolved the capacity to use oxygen to greatly increase the efciency of energy and material consumpon By contrast, the great diversity of multicelular orgasms that populate the planet today was generated in about six hundred million years Thus the history of the earliest biosphere is a narrative of how the capacity of the environment to sustain ife was extended by the discovery through evoutionary search of the biochemical cir cuitry needed for fermentation photosynthesis and respiratio n
5
PHLOOPHY AND MU LATON
To give an iea of the increased efciency represented by these met abolic landmarks we can use some numbers obtaned from studying contemporary microorganisms. Using the least efcient process, fer mentation, 180 grams of sugar can be broken down to yield 20000 calories of energy Te sugar used as raw material for this chemical reaction was originally taken ready-made from the environment by the earliest bacteria but with the discovery of photosynthesis those organisms coud now produce it using 264 grams of carbon dioxide, 108 grams of water, and 00000 calories taken from sunlght, they coud produce the same 180 grams of sugar, plus 192 grams of oxygen as waste product With the advent of respiration, n turn, that waste product coud be used to burn the 180 grams of sugar to produce 00000 calories of energy3 Thus, adg photosynthesis to fermenta tion made the growth of the earliest populations of livng creatures self-sustaining, while adding respiration produced a nt surpus of bacterial lesh (or "biomass ) Given that surpus it became possible to go from mere population growth to increas ed species diversity through the complexication of food chains other words, bacterial biomass itself became a gradient that could be tapped into by newly evolved predatory speci es, the ancestors of contemporary unicellular organ isms le paramecia or amoebae Whether predatory or not, more complex iving creatures did nt have to confront the same metaboic problems that had taken bacteria billions of years to solve. Rather, ancient animals incorporated as a whole microorganisms tha� had already mastered respiration (the ancestors of mitochondria) while ancient plants absorbed those that had acquired the capacity for photosynthesis ( chloropasts) other words, the plants and animas that would eventualy become multi ceular came into being by using bacteria and their metaboic cir cuitry as builg bocks.4 The interactions between the new creatures and the older bacteria may have started as parasitic and only later developed into a more mutualistic relation as the intimate knowledge that parasites had abut their hosts was used for their common enet The crucial eleent in the relation was simultaneous r�plicat ion: the former parasites had to be replicated in synchrony with their hosts so that their reproductive interests did not diverge.5 I many contempo rary organisms there are enduring symbioses in which bacteria in the guts of larger animals allow the latter to digest food they could not 66
GENETC ALGORTHM AND ANCENT ORGANM
otherwise process, but the microorganisms must be reingested every generation The decreased ntimacy this implies means that mutations may arise that switch the symbiosis back nto parasitism. I the case of ancient amals and their mitochondria, or ancient plants and their chloroplasts, on the other hand, any retained mutation in the absorbed bacteria had to also be benecia to the host. This deeper form of mutualistic relation is called endosymbiosis Endosymbiosis is an exotic illustration of the part-to-whoe relation in ecology The more ordinary examples form severa eves of organi zation each one studied by its own sub-discipine behavioral, popula tion' community, and ecosystem ecology At the smallest scale the object of study is the behavior of individua organisms n their envi ronment: motle bacteria sensng nutrient gradients in their vicinity and moving in the direction of higher concentratio n, for example, or ancestral amoebae using owing protoplasmic extensions to encirce bacterial prey prior to ingesting it. The next scale focuses on the popu lation of which these organisms are component parts and studies the way it grows imited ony by the capacity of ocal gradients to sustain it Details about the behaVior of individua organisms become causally redundant in an explanation of this growth. At the next scale entire communities composed of several interacting populations of different specis becom he target of anaysis. nd nally, at the largest scale, the interactions between several of these communities are studied, disregarding the distinctions between species and deaing only with the function they perform, as producers or consumers, in the overall low of matter and energy in an ecosystem. Although each of these spatial scales is distinct enough to be studied by a different ed it is important to emphasize that we are deaing here with differences in relatie scale that is, with scale as is generated by the relation of part-to-whoe, not with absolute scale: a single contemporary large plant, for exampe, may house an entire ecosystem of microorganisms displaying a four evels of organization. A good starting point for a philosophical discussion of ecological issues is the scae studied by popuation ecoogy, a level of organiza tion characterized by emergent properties
1
PHLOOPHY AND MULATON
immigration also affect the rapidity or slowness of population growth On the other hand there is the available space, matter, and energy in the location inhabited by the population. These two factors together dtermine the caing capaci, the maxmum population that can be sustained indentely wthout depleting avalable resources Since one and the same environment may house many dfferent species the carryng capacity varies depending on the energetic, material, and spatial requirements of each species: a population that can survive on small amounts of energy for example, grow to a much larger size than one whose energy needs are hgher Because growth patterns may be affected by overcrowdng it is preferable to measure the out come of growth not in absolute numbers but in densities that is by the average population size in a given amount of space Carrying capacity can then be treated as a singularity in the space of possibil ties for the interactions between population densities and resources, the special value of densty that ensures a population be at long term equilbrium wth its environment Real populations o f course, need not exist at the singularity but lctuate around it or periodically overshoot it How a real population will behave relative to the singularity depends on which of the two factors that determnes carrying capacity dominates its repoductive stratgy A strategy based on growth rate mples that a species will tend to produce many offspring quickly, offspring that will tend to be small and short lved A strategy targeted at existing resources, on the other hand produce fewer larger longer lived offspring more able to lea to control their environment more energy efcient and more capable of ne-tuning its numbers to environmental condtions Insects and weeds are often used as examples of the rst strategy whle large mammals and trees exemplfy the secon d one Given the small size and rapid reprouction of mcroorgansms one would think that they should always use the rst strategy but studes of microbial ecologies have shown that both strategies exist: some populations ignore carryng capacity overshoot it then crashperhaps going dormant as spoes or cysts to wait for new resourceswhle others grow fast at rst but then show densitydependent saturation effects and gently converge on the singularity6 In addition changes in the availablity of resources can contrbute to shaping growth pattes as when seasonal variations affect resource abundances makng the 68
GENETC ALGORTHM AND ANCENT ORGANM
density of a population vary in a periodc way These cyclic pattes can ecome even more complex when we move to the level of community ecology because at that point the density of a prey species becomes the main factor determng the carryng capacity for a predatory species Using density instead of absolute numbers to characterize a popula tion means that we can think of the latter in terms of gradents: like temperature or pressure density is an intensive property In this case the gradient is a concentration of biomass from whch energy can be extracted by othr species Ecological relations can, in fact be dened by the effect that the density of one species has on the density of another f an increase in the density of one species decreases that of another then the relation is either predatory or parasitic the densi ties of both species decrease there is competition f they both increase the relation is one of mutualsm or symbiosis Ad nally, if the densty of one population grows without affecting that of the other then their relation is one of commensalsm To study these relations communty ecology must add to the mechanisms determning the growth rate of a population a rate couplig mechansm that is a enduring nteraction making the rates of growth of the two popula tions depend on each other7 Lets use the interaction between preda tors and prey as an example n standard models the prey population is assumed to have access o unlimited resources so the primary pro ducers are not explcitly modeled In the case of ancient food chains the prey population can be magined to be bacteria usng fermenta tion to consume the organic products of photosynthetic producers n the absence of predators such a bacterial populaton would grow exponentially but in their presence its density will depend on the rate of predation The density of the predator population, in tu, depends on its own growth rate as well as on the rate of prey capture Fnally the densities of each poplation ffect the probablty that they wll encounter each other and interact, the higher the densities the more likely the nteraction n such a senario it is relatively easy to obtain population densities that do not settle into a steady state but instead cycle repeatey: as the predator population grows it reduces the population of prey up to the point where there is not enough prey to hunt; ths makes the predator poplation crash allowing the prey population to recover 6
PHLOOPHY AND MULATON
and start the cycle again On the other hand a rhythmic pattern o boom and bst in the respective densities does not imply that the space o possibilities or the copled system predatorprey contains a cyclc snglarity (a periodic attractor) This will be the case only i the oscillation n the densities is asymptotically stable that is only i the period o the oscillation has a tendency to retrn to its original vale ater an exteal shock Mathematical models with periodic attractors can be generated i to the previos model we add the assmption that predators can become satiated that is that as the prey poplation becomes more abndt rates o predation slow down becase predators are already g all the prey they need Data rom actal poplations in their natral habitats as well as rom poplations stdied nder controlled laboratory conditions tends to conrm that stable cycles in density are ineed common I some cases even more complex attractors cycles on cycles or even deter ministic chaos have been docmented9 Ths the interplay o ormal analysis d laboratory experimentation is giving commnity ecolo gists the means to explore the mechanisms and the mechanism independent strctre o their sbject matter To this it mst be added the role that simlations can play as intermediaries between theory and experment allowg s to ollow poplation densities as they move toward a singlarity To play this role a simlation mst embody its smlated organsms by giving them a metabolism and sitate them in a space with a cer tan distribtion o resorces Unlke the ses o genetic algorithms discssed in the previos chapter, n which disemboded replicators were capable o nding a sngle optimal soltion, embodied and sitated creatres tend to nd a range o acceptable soltions given that resorces can vary greatly n their spatial distribtion and that the soltions ond by evoltionary search need not be optmal or the entire poplation I addition tness evalations mst not be per ormed rom the otside bt need to be endogenou: reprodctive sccess mst depend on the ability o the smlated organisms to meet their metabolic reqirements Whe reas with exogeno s tness evoltionary search nds soltions to problems posed by the experi menter in these smlations the problems are posed by the environ ment in the case o poplation ecology or by other organisms in the case o commnity ecology Let's begin with the rst case smlations 70
GENETC ALGORTHM AND ANCENT ORGANM
o poplation ecology in wch the target o the stdy is emergent growth pattes and tness is dened by carryng capacity On one hand the growth may take place becase the environment has been enriched with new gradents njection o mneral resorces cased by an exteal event like a volcanic erption, or example On the other the growth may be cased by the evoltionary discov ery o novel metabolc strategies, lke photosynthesis or respiration That is we want to model the capacity o the environment to aect the organisms as well as the capacity o the orgisms to aect their environment I one approach called "Latent Energy Envronments (LEE) a spatial dimension is added to genetic algorithms n two dierent ways the simlated organisms nhabit an exteal space with a certain distribtion o resorces d they have an nteal space (or "gt) n which they can carry resorces ater ingesting them I addition nlke standard genetic algorithms in which a new generation com pletely replaces the older one and n which poplation nmbers remain xed, in LEE dierent generations can coexist and poplation size c vary allowg the exploration o densitydependent eects on growth patterns Bt the most mportt eatre o LEE is the way in which it enables s to distingish the eects o resorce availability rom those o evoltionary adaptations in the ate o a given popla tion To this end LEE does not distribte resorces as readymade ood bt rather as bits and pieces that mst be combined in specic ways or their "latent energy to become "sel energy The possible metabolic combinations o these bit d pieces are predened in a table that is they are not emergent bt they are rich enogh to yield not only dierent amonts o energy bt also byprodcts that can enter into rther combnations Ths an experimenter can control the problem that the environment poses to an organism in two ways by mag the possible metabolc combinations more or less complex (changing the entries in the table) or by determnng how spatially concentrated or dispersed the ood components are Since the table o reactions is not evolvable and the creatres gt is a simple container the soltions they mst discover are not new metabolic processes bt new behaviors More specically, they mst evolve novel foraging trategie. Searching or ood in an ecient way is important becase movement nvolves an energy cost This means that the simlated 71
PHLOOPHY AND MULATON
between predators and prey. Symbiotic species, for example, may come to depend on each other so closely hat their evolution becomes tightly coupled, as illustrated by the way n which many contempo- rary plants have developed an obligatory relation with the insects that pollinate them. Evolutionary interactions are different from ecologi cal ones not only because the latter take place in relatively short te scalesthe periods of the oscilations in density, for example are measured n yearswhile the former take place in much longer time scales but also because they are associated with different possibility spaces I the preous chapter we saw that an evolutionary possibility space can be conceptualized as a tness landscape: a space containing possible gene sequences on top of which a valuation n terms of expected reproductive success has been superposed. The distribu tion of singularities (tness opta) in this space denes the complex ity of the survival problem that has to be solved: a space ith a sngle global optimum surrounded by areas of mnum tness is a tough problem (a needle in a haystac) while one with many local optima grouped together denes a relatively easy problem In the case of predators and their prey (or of any other ecological relation) we need to couple two of these spaces because although predator and prey species have their own genotypes thy nluence each other's tness: when evolutionary search happens to nd a new combination of genes that proves the prey's capacity to evade its predator, for example ts will lower the tness of some existing predator genes, and conversely, a new gene sequence that improves huntng behaor il lower the tness of some existing prey genes I other words, the coevolution of predators and prey implies that their tness landscapes are constanty beng deformed with some "hills beng transformed nto "valleys3 Extending ts lne of thought to entire food chains implies that many tness andscapes may become serially couped raisng the question of how long food chans can be before the cou pled possibility spaces contain singularity distributions that mae survival prob6ms impossible to solve These considerations have led theorists to expect that evoution itsef may lit the ength of food chans to avoid a dead end, hypothesis made pausibe by empirical ndings of reatively short food chains in actu ecosystems. 14 74
GENETC ALGORTHM AND ANCENT ORGANM
When designing sulations to study coevolution having control of the details of metabolic processes or of the spatial distribution of resources is not as necessary as being able to deploy embodied crea tures of different species that can recognize and nteract ith one another I one approach, called Echo, the internal metabolic machin ery of the simulated creatures is made so simple that they n effect "eat genes. In Echo evolution is simulated through a genetic algo rithm n hich the standard symbol string of ones and zeroes acting as a chromosome has been changed to one using a larger alphabet. The creatures live and move in a spe space containing resources that are symbols from that alphabet and reproduction taes place only hen the creatures have accumulated enough symbols to create a copy of their chromosomes. Constructing energetic resources and chromosomes from the same raw materials allos Echo to disregard metabolic processing and nclude only storage in an nteal "gut le sultaneously making tness evaluations endogenous To simulate ecological nteractions symbols that are in the guts of other creatures can be captured in the case of predation or exchanged in the cas� of symbiosis Echo's creatures are ess embodied than those of LEE not only becase of the identity of energy and genes but also due to the lac of a transformation of genotype into phenotype: wle in LEE the simulated chromosome encodes for a neural net, the inputs and outputs of which act as sensors and motors, in Echo the chromo some acts both as inheritable material and as bodiy traits More specically, the chromosome's symbols play the roe of tags that are "visible to other creatures In real microorganisms these tags would be membranebound proteins allowing for molecular rec ognition using a locandey mechanism, or even chemical byprod ucts of an organism's metabolism through wch it can be identied by others, but these details are ignored in Echo. Tags wor not only as a means of identcation but also as the means to specify the capaci ties of a creature to affect and be affected by other creatures Ts has the consequence that Echo's creatures do not interact ith each other directly but need an external program to compare their tags nd determne the outcome of an "interaction. The tags are of to types: offensive and defensive. The basic procedure is t randomly pick pairs of creatures inhabiting the sme site and then compare the offense tag 75
PHLOOPHY AND MULATON
of on to the defense tag of the other, other, and vce versa The comparison aims only at establishing he number of symbols of each tag that match those of another tag at the same position in the chromosome f the match is asymmetricthe asymmetricthe offense tag of one creature matching the defense tag of another but not the other way aroundthe rst creature is taken to be a predator and the second one its prey. f the match is symmetric, that is, if both the offense and defense tags of each creature match each other, then the two are considered symbionts The number of matches determines the aount of resources that are captured by the predator (it can "eat the entire prey or siply take a "bite from it) or the amount of resources that are mutually exchanged in symbiosis. This allows for complex ecological relations to emerge: one species may prey n another that is symbiotic with a tird one that in turn, preys on the rst. 1 5 I a more complex version of Echo the chromosome is vided into two parts one acting as tags the other performing the role of con tionaction rules that is, rules specifying that if a given contio obtains a particular action is taken. I this more comple version the interactions are still not performed by the creatures but the tags are compared through the contionaction rules allowig for outcome that are more lexible: the creatures can "refuse to interact if the contios specied in their rules are not met for example 16 The rela tive lack of embodiment of Echo's creatures relects the emphasis of tis approach on the longterm outcomes of interactions (structure of food chans relatve relatve species abundances, arms races) rather than the interactions hemselves As in LEE the rst task to be performed is to calibrate these emergent outcomes to those of either matematcal mdes or data from communty ecology For instance, in many cases the relative abundance of fferent species has a dstinct statistical shape: many speces have a few representatives whe a few have many representatives Experients with Echo designed to test whether it can produce these abundance patterns have been conducted ad the initial results are ecouraging The experiments attempted to match not ony a targ�t statistical stribution but also tred to answer the question of whether the the outcome was prod uced by coevolution To test tis hypothesis two versions of Echo were ran for each experiment one in wich interactons were mediated by tags and contions (hence subject to coevolution) and another in which interactions 6
GENETC ALGORTHM AND ANCENT ORGANM
took place randomy The tagmeated siulation t the data better than the purely random one 1 7 From the point of view of simulating the ancient biosphere Echo provides a means to explore the very different ways in which the descendants of the earliest microorgasms evolved There were two stinct evolutionary evolutionary dynamics depending on whether or not a microorganism possessed an inner membrane encapsulatng its genetic materials within a nucleus Lacking suc a nuclear membrane allows the free sharing of genes by entirely different organisms. t is well known, for example, that transferable sets of genes called plasmids are responsible for the spread of antibiotic resistance in bacteria in the past 0 years, but this genetic promiscuity is nothing new. Ancient bacteria may have engaged in this form of orzontal gene transfer from the start using plasmids and other movable genes as combinato rial builng blocks to rapidly acquire capacities they previosly lacked The ancient predators of those bacteria on the other hand dd possess a separate nucleus and were not capable of sharng genes that way They compensated for that by acquring the capacty for reproducive reproducive isolaion. I larger organisms the formation of new species occurs largely through the isolation of their gene pool from lows of genetic materials from other species. Ths mechanism of speciation (the birth of a new species) is a major evolutionary force promoting divergence and anatomical dfferentiation The evlutionary hstores of microorganiss wth and without a nucleus (eukryots and prokaryotes) testfy to the importance of this stincton: while the former went on to form all the different plants and animals that exist today, the latter followed not a divergent but a convergent path. That is instead of yielding a plurality of gene pools more or less isolated from each other they enerated what is basically a giant gene pool spang the entire planet 1 8 I adtion to the presence or absence of speciation the othr major determinan of early evolutionary outcomes was endosymbiosis, the already mentioned absorption of entire microorganisms and their metabolic capacities. Simulating endosymbiosis and speciation can be done in Echo simply by adding new tags or contion action rule s The rst mocation s to add a new tag medating adhesion interac tions and a means to manage the boundaries or membranes of the interacting creatures I a sulation of ancient microbial ecosystems
PHLOOPHY AND MULATON
adhesion tags could play several roles depending on the outcomes of the interactions and the resulting boundaries. Allowing adhesion between similar orgaisms can easily yield colonies arranged in lay ers much like the matlike structures that nonmotile bacteria formed in the rst two biion years of evolution With this spatial organiza tion the evolution of individual creatures may be inluenced by the environment provided by similar creatures on the same layer or by different creatures in a neighboring layer promoting specialization. Endosymbiosis on the other hand can be simulated if the result of adhesion interactions is not the creation of a single boundary around many creatures but the absorption of one creature by another in which the absorbed creature keeps its own boundaryJ9 To simulate speciation the obvious choice would be to add a new tag to simulate expressions of species identity This choice would be justied by the fact that many organisms achieve reproductive isolation through the possession of special markings (colors odors pigmentation pattes) But in Echo selective mating is achieved by adding a new conition action rule that u ses an existing tag (the offense tag) Whenever two creatures occupying the same site are ready for reproduction that is when they have ingested enough symbols to make a copy of their chromosomes the condition part of their mating rules are compared to their offese tags and if both conditions are are satised copies of their chromosomes chromosomes are made, crossed over and two new offspring are added t the population This proce dure resembles that used by paramecia te ancestors of which were among the earliest eukaryotes By varying the specicity of the mat ing condition lower degrees of reproductive reproductive isolatin can be achieved to yield the euivalent of prkaryotes2 The distinction between the types of evolutin that result with and without the possibiity f speciation must be kept in mind when testing Echos evolutionary outcomes One experiment for example attempted to check whether Echos sulated evolution displayed the same open-ended character as real evolutin r whether it remained bounded within limits. The results failed t match the degree f divergence that is evident in the fssil record after the great burst f differentiation known as the "Cambrian explsion.21 But the experent did not cntrol for the degree of reproductive isolation Since the fssils of eukaryotes pre dominate in that record a sulatin of evolution that matches its 8
GENETC ALGORTHM AND ANCENT ORGANM
openended character should be conducted with populations that can speciate through strong reproductive isolation Once our simulations reach th postCambrian period we need to enrich them with new resources So far the simulated uniceular organisms have been able to learn how to solve survival problems only as a species That is the solutions are reached by an entire popu lation over many generations But after eukaryotes underent their their explosive divergence changes in their composition created the condi tions for the emergence of problem solving abilities that could be exercised during a single organisms lifetime LEEs creatures in fact can be used in this mode by allowing the strength of the connections of their neural nets to be established not by a genetic algorithm but by the training the creatures get as they move around searching for food The next step in this investigation investigation should therefore therefore be a discussion of the nervous system and of the emergent leaing capacities with which it endows living creatures
9
NEURAL NET AND NECT NTELLGENCE
CHAPER SIX Neual Nets and Insect Intelligence
The explosive divergence that multicellular orgasms underwent sx hundred million years ago had many consequences. The diversity of the survival problems that these organisms had to solve, for example, increased enormously as their environments led with other living creatures capable of affecting and being affected by them To confront this diversity organisms began to develop inteal models to guide their behavior Even humble motile bacteria, as they swim up a gradi ent of nutrients toward the point of maximum concentration, can be said to have an nteal model of their environment. This inteal model is not, of course, based on representations of any kind and it is not a model of the world at large but only of the concrete opportuni ties and risks afforded to bacteria by their immediate surroundings I other words, from the beginng of life the inteal models mediating the interaction between a primitive sensory system and a motor appa ratus evolved in relation to what was directly relevant or signicant to living beings With the advent of multicellular organisms and the progressive differentiation of their cells into multiple kinds the scope for internal models was greatly enlarged I partcular, a new kind of biological material, neuronal material, began to grow and interpen etrate the living jelly accumulating at the bottom of the ocean. Ad with the availability of neurons the capacity to distinguish the rele vant from the ielevant, the ablity to foreground only the opportu nities and risks pushing everything else into an undifferentiated background, was vastly increased. Smple multicellular organisms, like the hydra or the jellysh, have a few separate neurons linked together nto a network I the hydra 80
this network is concentrated around the mouth while in the jellysh it is linked to simple receptors for light and chemical substances These assemblies of neurons, n turn, sustain the emergent capacity to modulate inherted behavioral responses during the lifetme of an organsm The simplest of these capacities is called habituation: it operates on sensorymotor activities that are innate but it alows organisms to control their intensity A organism like a hydra, for instance, can gradually decrease its response to a stimulus-can become habituated to itas experience nds it harmless or avoidable. I other words, habituation transforms a signicant stimulus into an insigcant one. The opposite capacity, sensitization, alows these organisms to behave toward a previously irrelevant stimulus as a potential source of an opportuty or a risk, that is, it makes the stimulus behaviorally relevant I a sense these two ancient forms of learning mark the emergence of subjective gradients: before a hydra becomes habituated to a novel stmulus, for example, it may be thought of as beng "surprised by it, a protosubjective state that slowly inishes until it disappears as it reaches equilibrium Behind these subjective gradients, on the other hand, there are objective ones: cncentration gradients of electricaly charged substances constantly produced and maintained in the uids inside and outside neurons The capacity to become habituated to exteal stimuli can serve as a useful starting point for a philosophical investigation of amal lea ing because explaining it demands looking only at what happens inside a single sensory neuron as it interacts with a single motor neuron The standard model of neurons places the manipulation of gradi ents of metac ions, potassium or sodum ions, for example, at the heart of their functiong. Manipulatng these chemical gradients keepng potassium at high concentrations inside the cell and at low concentrations outside of itproduces yet another gradient, one of electrical potential The key to the conversion of one gradient into another is the selective permeablty of a neuron's membrane to metallic ions: as membrane permeability changes ins low in and out of the cel creating variations in the gradient of electrical potential. Neurons use these smooth lowleve variations to record the electrical signals acting as their input Producing an output nvolves changing the electrical gradients from continuous to discrete, transformng 8
PHLOOPHY AND MULATON
them ito a chai or trai of practically idetical electrical spikes This spike trai is used to produce a chemical sigal that, whe received by other euros icreases the permeability of their membraes if the chemical substaces are excitatory or decreases it if they are ihibitory 2 To provide a mechaism for habituatio i simple creatures all we eed to assume is that the supply of a excitatory substace i a sesory euro becomes progressively depleted after successive stimulatios, so that less ad less of this substace crosses the gap to affect the permeability of a motor euro's membrae leadig to a decrease i its respo se A siilar mechaism ca accout for sesitizatio if we add a secod sesory euro iteractig with the rst causig it to icreae productio of the excitatory substace ad therefore the amout of the latter that reaches the motor euro 3 The capacity to become habituated or sesitized implies oly the possessio of iate iteral models But with the advet of more complex assemblies of euros ew forms of learig emerged that made possible the creatio of iteral models durig the lifetime of a orgaism I particular, aciet isects were capable of the kid of learig referred to as cassica conditioning i which a iherited associatio betwee a stimulus ad a resose provides the basis to form a ovel associatio with a previously isigicat stimulus tudyig this form of learig i a laboratory ivolves idetifyig a particular behavior (salivatio i a dog fo example) that is always triggered by give stimulus (fo od odor ) f the trigger is preseted immediately preceded by a eutral stimulus (the soud of a bell) the latter becomes capable after may presettios to elicit the desired respose aloe Although mammals were origially used to prove the existece of this associative capacity it became clear later that isects also possess it Whe the ateas of hoey bees, for example, are touched by a drop of sugar i solutio the bees relexively exted a part of their mouth toward the sugar ad lick it f a odor or a color which was previously irrelevat to the bees is made to coicide with the presece of sugar it will after some traiig trigger the lickig behavior aloe Techically, the stimulus to which the aimal is geetically predisposed to respod is called the "ucoditioed stim ulus while the oes that may be associated with it are referred to as "coditioed stimuli 82
NEURAL NET AND NECT NTELLGENCE
Outside the laboratory this associative capacity has survival value oly if the coditi oed stimulus is i fact predictive of the presece of the ucoditioed oe Parasitic wasps for example have hutig relexes triggered by the odor of the caterpillars o which they feed but they ca lear to display the same behavior whe stimulated by the odor of the plat the caterpillar feeds o.4 other words the wasps ca lear to predict the presece o f their prey through a codi tioed stiulus The evolutioary relevace of coditioed stimuli is sometimes igored by psychologists who are iterested oly i the betwee a stimulus ad a respose But behavioral ecologists kow that evolutio ca build biases as to the kids of stimuli that may or may ot be successfully associated, ad that ecologically se sible explaatios ca be g ive for those biases Food search i hoey bees, for example, implies learig to recogize differet lowers that is, objects that are colorful, have a special aroma ad have geometric arragemets of petals This suggests that they shold be able to asso ciate ot sods but odors, colors, ad maypoited shapes with the presece of food We ca also expect a orderig of these stimuli rela ive to their usefuless to make predictios: lorl odors should be easier to lear tha the shape of lowers sice the latter varies with the agle of observatio ad is thus less redictive of the presece of food Experimets with hoeybee s corm these two hypothes es 5 Classical coditioig i isects illustrates the begig of a · true iteral model while the behavior of a oey bee may ot be affe cted by a lower as a whole but oly by a set of stimuli that ted to co occur i the presece of lowers the fact that there is a causal lik etwee each o f these stimul ad their source suggests that a simpli ed model of the lower (a set of likd stimul) is ideed created i the bee's brai Ulike the hydra ad the jellysh isects possess a more elaborate ervous system i which may idividual euros form the compo et parts of larger wholes like a primitive brai or the gaglia that coordiate their articulated lb movemets Nevertheless, we eed ot cosider their etire ervous system to explai their capacity for classical coditioig t is well kow, for example, that may sects ca lear to avoid certa limb positios by associatig them with a electrical shock but the same associatio ca be achieved by a isolated isect leg together with its gaglio6 Thus, it is the properties 83
PHLOOPHY AND MULATON
of local assembies of neurons, not entire nervous systems, that sus tan this capacity, properties lke their connectivty and the strength of their connections. These two properties are related because a con nection or synapse between two neurons becomes stronger the more often the neurons on either sde of it are actve simultaneously Or more exactly, simultaneous activity inluences a connection's strength although not necessarly in a constant way. Classical conitoning, for example, is characterized by a learning curve that isplays a clear deceleration: while at the begning of the training the joint presenta tion of the conitioned and uncondtoned stimulus has a large effect on behavior the effect becomes less and less pronounced with time. This suggests the existence of a state of expectation in the anmal's mind so that the more surising the stimulus the larger the coni tonng effect.7 I aditon to this subjective graient a new kind of entity may be needed to fully explain classcal conitioning a mental representation Theorists have traditionally been ambivalent about what exactly becomes assocated when an anmal s trained s the anal nking the conditioned stiulus with a patte of behavior (the relexve respose) or is it associating t wth a representation of the unconditoned stimulus? Whle behaviorist psychologists favor the former explanation laboratory experiments suggest that the latter is the correct one: if an anal is articially paralyzed durng traing so that the relexive behavior cannot occur, an association stll forms and affects the aal's behavior once its ablity to move has been recovered8 Given that we don't know what a representation in an insects mind is lke other than the fact that it cannot possibly be a symbolic representation, it will be useful to explore the question of insect intel igence at the mechanism-ndependent level That is we need simula tions that display an emergent capacity for classcal conditoning even if the mechanism of emergence is entirely different from hat of real neurons On the other hand, to make sure that we are exploring a space of possibilites that overlaps that of real neurons we need simulations in hch the computing units interact with one another through excitaton and inhibition; in which some form of learning takes place as these nteractions modify the strength of ther ntercon nections; and in whih non-symboic representations of relatively simple patterns emerge spontaneously from the dynamics The type 84
NEURAL NET AND NECT NTELLGENCE
of simulation that satises these requirements is caed an articial neural net A typical neural net consists of many computng units connected to each other n a certain patte each computing unit being much less sophsticated than the central processing unit nside a desktop computer: a it must be able to do is calculate its own degree of actvation given the exctatory or inhibitory signals it receves from other units The computing units nteract through connections that can vary in their ablity to transmit exctation or inhibton increasng in capacity if the interconnected units are active together or decreasing if they are simultaneously inactive. The variable capac ity to transmit exctation or nhbition is called the strength or weght of the conection The smplest neural net has computing units arranged in two layers, the input layer and the output layer, a design known as a "perceptron. Simply put a perceptron s a device that maps an input patte into an output patte9 One patte may be for example a simple sen sory stmulus while the other may be a smple motor response, although in practice neural nets are stuied in a isembodied state ignorng the details of the sensors producing the input patte and those of the actuators that convert the output pattern into a motor action The association between input and output pattes s created by training the neural net. The tranng process consists in repeately presenting a patte to the input layer, actvating some unts but not others, while xng a desred actvation patte at the output layer Or in more embodied terms the trainng involves stimulating the neural nets sensors while simultaneously showng it the desired motor response Once both activaton pattes are set the computing units in each layer begn to interact with each other, increasing or decreasing the weight of their connections depending on whether the units at both ends of a onnection are active or inactive The smulta neous presentation of iput and output pattes is performed many times untl a stable conguration of weights in the conectons is produced. After the traiing s over the xed output patte is removed and the neural net is tested to check whether it i able to reproduce it whenever it is presented with the original input patte This emergent capacity s explained by the fact that the connectons have now specic weights and wil transmit activation from the input layer to the output layer n just the way needed to reproduce the 85
PHLOOPHY AND MULATON
experience that is the weihts of its connections must be found throuh radient descent Finally the hidden units of each neural net should be connected to each other laterally in such a way that their non symboc represent ations can interac t with one another Du rn trainin, as each neural net is stimulated with the pattes corre sponn to the two stimuli the emerent representations that form n the hidden units will increase the weit of some of the lateral connections and decrease the weiht of othes After trainin the lateral connections l possess a conuration of weihts able to produce the representation of the unconditioned stimulus if iven that of the conditione one and vice versa. This means that when the input layer of the second neural net is presented with the conditioned stimulus it will create in its hidden units a patte of activation capable of producin the extracted prototype o the unconditioned stimulus in the rst neural net which in tu, will trier the nheited motor response The concept of a prototype extracted from experience and stored as a non symbolic representatio n will play such a crucial role in the explanation of animal behavior n the followin two chapters that its nature should be made very clear Firs of all an emerent represen tation is not explicitly stored as such the product of the leain process bein a conuration of connection weihts that can recreate it when presented with the riht input I other words what is stored is not a static representation but the means to dynamically reproduce it Second, unlike a photoraph these representations are dispersed or dstibuted in all the hidden units and are thus closer to a holoram. This means that they can be superimposed on one another so that the same conuration of weihts can serve to reproduce several repre sentations dependin on its input, simulatin the ability of insects to associate several colors or odors with the presence of food The ds persed way in which extracted prototypes are represented is so impor tant for the capacity of a neural net to eneralize that these emerent representations are usually referred to as distributed representations 1 8 Finally, unlike he conventional between a symbol and what the symbol stands for distibuted representations ae connected to the world in a non arbitrary way becaus e the process throuh which they emere is a direct accommodation or adaptation to the demands of an external reality Thus multilayer perceptrons offer a plausible account 90
NEURAL NET AND NECT NTELLGENCE
of the intentionaty of mental states 1 9 When neural nets are studied in a disemboded way, that is, when their input is preselected and pre structured by the expeimenter and their output i s simply an arbitrary pattern of activation that has no effect on an externa eniron ment, this emerent intentionality is not displayed But the moment we embody a neural net and situate he simulated body in a space that can affect it and be affected by it, the creatues behave n a way that one feels compelled o characterize as ntentional that is, as oriented toward exteal opportunties and risks There is one more question that needs to be addressed when considern the applicability of neural nets to simulate ins ect intellience When a simple perceptron is traned the desired motor behavior is iven to them by the experimente. This impes that, unlike ins ects n thei natua envionment, neural nets do ot tran themselves Tis exoenous tranin is even more problematic in the multilayer case because the neural net is supplie not only with the desired output pattern but also with the deee to which its current output fails to match the taret patten at any point durn tranin Moreover, n both cases it is the experimenter who decides what patterns to include in the tanin set and this may introduce statistical biases in the nput that aticially facilitate a lea task. Any of these probems can cast dubt on the extent to which the neural nets themselves are creatin a linl between their intena repesentations and the external world The situation is similar to that of exoenous tness evaluations in the case of enetic alorithms. n that case too the fact that the problem that evolution must sove is posed by the experimenter (throuh the desin of the tness funct on) may introduce unwanted biases into the search pocess When neura nets or enetic alorithms are used in ndustrial appications these biases need not be a problem and may even be desirable if they shoten the amount of time needed to perfom the trann o to conduct a search. But when used as bioloical models exoenous trainn and exoenous tness can be problematic A way out of tis dilemma has aleady been suested embody the imuated enes or neurons and situate them in a space that provides hem with opportunities and risks One of the two simulations di scussed in the previous chapter, latent enery envronments (LEE), illustrates how humans can be replaced not only as animal breeders but aso as animal trainers When descibin 9
PHLOOPHY AND MULATON
EEs simulations of population ecology we focused on only one of the two ways in which it can use neual nets: link them to genetic algoithms and let evolution nd te conguation of weights that impoes foaging behavio. But EEs neual nes can also be congued to allow its simulated oganisms to lea to foage fo ood duing thei lifetimes EE endogenizes the taining pocess though a cleve adaptation of backpopagation the output laye of the neual nets poduces not only motion instuctions (move fowad, tu left, tu ight) but also pections egadig the possible location of food. hose pections ae compaed with the esults obtained fom futhe exploation of the envionment and the degee of mismatch is popagated back to the othe layes deally, these two uses of neual netssimulating leaing by the species and leag by the ogan ismshould be used togethe so that something like the classical contioning design outlned above could emege he st of the two neual nets, the one that ceates a stibuted epesentation of the unconditioned stimulus, should be slowly evolved within an envion ment that can constai and enable action to yield elexes lining sensoy stimuli to adequate moto esponses he second neual net, the one wth distibued epesentations of the conditioned stuli, must aive at a nal weight conguation duing the lifete of the ceatues, but the est of its achitectue could be evolved using pedictive success as pat of the endogenous evaluation of tness his way the biases built by evolution on what can play the ole of a con ditioned stuluscolo, loal aoma, petal aangement, but not sound i the case of honey beescould emege sontaneously in a simulation Moving beyond the leaing capacities of insects demands a moe adical stategy that simply embodying multilaye peceptons and situating them in space. n paticula, lage anmals with moe elaboate nevous systems ae capable of leaing to deVelop novel behavios Unlie classical conditioning in which the existence of an inheited elex to ancho an association is necessay, new leaed behavios can e developed though thei association with ewads o punishments Some insects may, i n fact, have access to ts moe sophisticated leaning capacity Honey bees, fo example, use classical contionng to lea to pect the pesence of food but they must also lea by tial and eo how to exploit a lowe blossom once it 92
' ( !
NEURAL NET AND NECT NTELLGENCE
has been identied Nevetheless, the capacity to le an novel behav ios is geatly amplied in animals like bids and mammals because in thei case peception and memoy go beyond the wold of mee stimuli Poving evidence fo this claim and examining the kind of neual nets that can splay these peceptual and memoy capacities be the task of the followig chapte
93
NEURAL NET AND MAMMALAN MEMORY
CHAPER SEVEN Neura Nets and Mammalan Memory
As popultions of neurons grew and proferated inside ving creatures, gathering into layers and foling into elaborate threeimensional structures, they provided an ever richer substratum for the growth and proferation of more ethereal entities memories. The memores that birds and mamals form of actuay ved episodes, for example, are more or less vivid reexperiences of events or situations in which objects play specic roles (such as agents or patients) and in whch their nteractions make sense to the animal. The content of autobio graphical memories in anmas must be thought of as endowed with signcance not with signication, whch is a linguistic notion. The signicance of a scene or event is related to its capacity to make a difference in an animal's fe, to its capacity to affect and be affected by the anmal's actions, whle signication is a semantic notion refer ring to the meaning of words or sentences Birds and nonhuman mammals may be ncapable of dealing wth signication but they surely can attribute signcance to the opportunities and rsks that their enviroent affords them Evidence for the existence of different types of memory comes mostly from their dissociation in amnesic human patients or in ani ms that have undergone surgery to produce specic lesions in the bran. The type of memory involved n classical conditioning, whether in insects, birds, or mamals, is referred to as "procedural because of its content's lack of accessibility outside of the sensorymotor asso ciations that it enables. Memories whose content is accessible are referred to as "declarative, a category further subdivided depening 94
on whether the content of the memory is events and situations that have actually been experience, a type of memory referred to as episodic memo, or whether it consists of nguisticaly expressed facts, that is, semantic memory. 1 Damaging certain areas in anmls' brains can seriously disable their capacity for declarative memory while leaving procedural memory ntact. Smiarly, umans suffering from certain kinds of amnesia can retain their capacity to remember n guistic information but be ncapable of recaling autobiographicl events, and vice versa, they can recal actualy ved scenes but are unable to remember what has been said to them.2 The very fact that episodic nd semantic memory can become dissociated suggests that human sensory experience and the memories we form of it are nde pendent of language and strongly related to one another. But for the purpose of ths chapter we do not have to settle the controversy over the thesis of the lnguisticaity of experience: non-human mamls and birds do not posses language so their sensory experience and autobiographcal memories cannot possibly be structured by it. Unlike the perception of atomzed stmuli in insects the perception of objects with an endurng identity is a complex task, involving the synthesis of features from different externl as wel as nternl senses Predatory mammals, for example, must guide their searching behav ior by the nformation they get from dfferent senses, using smel to bring them nto close proximity to their target then switching to sight or sound d they must constantly cabrate the nformatio they receive from the external world to tht generated by their own prop roceptive senses so they can distinguish between a change n their visual eld produced by a prey's movements and the apparent motion created by the movement of their own head or body.3 The synthesis of nformation from dfferent sensory modaities is thought to be performed in an anml's hppocampus and related areas, an organ also involved in the formation of episodc memories. Damage to the hppocampus, for example, pairs performance on tasks that involve comparing one memory to another or using the content of a memory in a novel context 4 Until the mechanisms behnd these capacities are fuly understood a phlosophcal exploration of avian and am malian memory is best conducted at the mechanismindependent level Much as the multilayer perceptrons discussed in the previous chapter helped us develop a better idea of what a representation in an 95
PHLOOPHY AND MLATON
insect's brain could be likedespite the fact that the simulated mechanisms like the backpropagation of error have no anatomical counterpart in real brainsthe capacity of neural nets to recognize enduring objects and make sense of interactins between objects can be a great aid in netuning our philosophical intuitions about what goes on in the minds of larger animals. Given the complexity of the subject it will be useful to break it down into two ifferent capacities object recognition and scene analysis give examples of the evidence that animal psychologists have gathered for their existence and then iscuss the currently existing neural net designs that can approximate them he earest labora tory evidence for the existence of the rst of these two capaciies comes from experiments called "twoway classications oway classication experiments were rst conducted wth pgeons but were later extended to monkeys and other mammals Pigeons were shown projected photographs ad trained to peck at a switch that activated a food dispenser if certain kinds of objects appeared in the pictures n one experiment for example oer a thousand slides of urban and countryside scenes were used half of which contained humans in if ferentpositions (sitting standing) with ifferent appearances (clothed naked) and shot from ifferent angles his variety eliminated the possibility that the pigeons were reacting to simple stimuli such as color or brightness After a few weeks the pigeons had acquired the capacity to classify pictures into those that contained one or more humans and those that id n. his result strongly suggests that the pigeons were able to extract a prototype from their perception of human gures and then use it to assess the d egree of typicality that ifferent images had relative to the prototype Pigeons were also shown capable of performing toway classications of pictures with pigeons versus pictures with birds other than pigeons; of pictures with and without trees; and of pictures with and without a body of water5 Before showing how multilayer perceptrons can isplay similar object recogniton capacities one aspect of their design should be claried. Unlike the neural nets iscussed in the previous chapter in which the input layer received sensory stimulation while the output layer conrolled motor behavior we are dealing here with partial models of a more complex nervous system. his implies that a neural 6
NEURAL NET AND MAMMALAN MEMORY
nets input may not come directly from sen sors but from other neural nets while its output may be subject to further processing before it is converted into a bodily action by actuators I many cases the neural nets irectly connected to either sensors or actuators are not incorpo rated into the design but just assumed to be there the neural net that receives sensory information is not included for example then the one whose patte recognition behavior is being studied can have its input units labeled with words words that stand for the objcts that the latter would receive if it were actually connected to the for mer Similarly if the neural net that controls ovement is not part of the desgn the output units of the one being stuied can be labeled with words for motor actions. I the present case using words to label the input units would defeat the purpose of the experiment given that what we need to siulate is a bird's capacity to recognize objects from the visual information provided by images he output units on the other hand can be labeled linguistically since in a r eal bird this output would be the input to many other assemblies of neurons before becoming a motor action. A god illustration of the use of neural net for this purpose is a simulation designed to recognize human faces he input layer con sisted of a twoimensional array of units 64 units on each side into which an image could be projected Each input unit was connected to a onedimensional hidden layer consisting of 80 units that in turn were connected to an output layer containng eight units labeled with words During training the input layer was prese nted with ictures of 1 1 ifferent human faces (as well as of objects that were not faces) while its output layer was given an activation pattern the meaning of which was given by the labels: one unit was labeled "face being active if the input was indeed a face and inactive if it was not; two units were labeled "female and "male their activity signifying gender recognition; and the remaining ve units were labeled with a name and a number ssigned to each face. After being trained the neural net was not only abl to perform a twoway classication of pictures into those that contained and did not contain faces but also to correctly assgn face pictures to the right gender category and even to identify faces belonging to specic persons. o test its ability to gener alize the neural net was tested with pictures of the same people that were not part of the training set and it performed almost lawlessly 7
PHLOOPHY AND MULATON
Moeove, it could coctly identify faces in which 20 pecent of the content was deibeately obscued by completing the fal paten wit nfoaton fo he distibuted epesentations in its hidde unitS6 he pesence of labels in the outut units designatng ingustic ctegoies lke me" nd feme" had n effect on the abiity of the ne net to ecognize faces The expnation fo this capacity ies istead in the 80 units of the hidden aye Befoe tainng al the pssibe pattens of acivity that coud occu in ths aye foed n stucue space of 80 mensions As the neu net was tined it st subvided his intenal space of possibilities into two egions, one coesponig to faces d the othe to nonfaces. Futhe tining ed the neua net to patition the face egion into two subegions one fo femae the othe fo mae faces, which wee then futhe subivided into even smale egions fo inividu faces. The cente" of each egion coesponded to a pototypethe easiest to ecognie faces in the tang input o the easiest to ecognize views of an invdua facewhe the bodes between egions coesponded to mbiguous inputs, such as pictues in whch the content coud not be detemined to be a f�ce o in which a face was isplayed in such a way that it was mpossibe to decide whethe it beoged to a mae o a feme This simulation povides a poweful insight into how n objecive categoy can be captued wihout using any nguistic esouces The secet is the mapping of relations ofsimilari into rela tions of proximi in the possibiity space of activation pattes of the hidden aye That is, obects that esemle each othe become neigh boing points in the inte possibty space, nd vice vesa obects with a high degee of issimaity (faes an nonfaes) end up as points that ae fa away fom each othe n he spae of p ossible activation pattens One itaion of this simuatin is that een though the linguistic labes of the output units do not pay any eplanatoy oe thei pes ence inicates that we ae deang with a patly isemboied neal net. We coud elminate the ned f labels y addng othe neua nets using the otput laye's activity atten to geneate oto beavio hs behavio coud be elatively smpe, the smated equivaet of pecking on a food ispense In aboatoy anms ths kind of behavio is not achieved though cassical conitoning but though a diffeent tng pocess cale instrumental conditioning Instead of 98
NEURAL NET AND MAMMALAN MEMORY
using a eex as its stating point instument conitionng stats ith a behavio that tends to occu spontneously using ewads to incease its feuency o punishents to decease it The positive o negative enfocemet must be contingent on changes of behavio: if the fequency of occuene dos not chnge the einfocement is not given. In adtion, labotoy expements must include a sign a smple sound o ght, to help the anm tel which one of its possibly mny behvios shoud be assoiated with the positive o negative einfocement. Ths sign is called a scmnative stimuus." Outside of te laboatoy einfoceent eaning cn ead to the development of oine behavio that helps an m cope ith the copexity of a ituation but it can so lead to the fomation of supe stitious habits beaing no elatin to eal oppotuities and isks The oe of isnative stii is pecisey to allow n nm to istin guish the behavo that is bing ewaded fom cooccuing behav ios that ae ielent. The typic aboatoy outcome of instuent conitioning n anm that les to pess a eve to obtain a pelet of fod, gives a vey poo idea of the types of behavo that c be eaned though ths kind of tainng This sple behavo is used because of the ease with whih inceases o deceases in the fequency of evepessing c be measued But with caeful nipulation of einfoceent schedules d chges in iscminative stmuli complex pattens of behavo cn e geneated as exepied by the elaboate ticks that c be taught to cicus anis. The taget behavio ay be, fo exape a at stnding on its ind egs in a given coe of the cage The expeimente ust st ewad going to that coe, a behavio that tkes pace spontaneously once n a while this chance behavio is ewaded nd a iscative stmuus is used to le the at know wat is being ewade it wil incease in fequency t that point it sos being ewaded and a ffeent lowfuency behavo, such as moving the head to an pght position, is ewaded d nked to anothe iscminative stmuus unt its fquency inceases The poedue continues one behavioal patte afte the next (ising the left paw, sing the ight pa) untl the desied nal behavio is obtined Ths pocess is called successve apoximation."9 The main obstacle to e desig of ne nets abe to ea this way is not captuing the capacty of enfoces to affec but the capacity of 99
PHLOOPHY AND ULATON
a ama to be affected by them. Usig food as a reiforcer, for exa pe, impies that a aima is hugry or at east ot fuy satiated. But suatig huger woud seem o ioe reproducig aspects of its pheomeogy, such as the pags of a empty stomach Eary expa atios o huger, i fact, took stomach cotractios to be art of the mechaism through which huger motiates food seekig behaior This hypothesis was show to be fase by experimets i which te stomachs of rats were remoed, attachig their esophagi to their sma testies eaig the behaor associated with huger itact. 0 I the preious chapter we argued that the degree to which a aima is surprised or started by a oe stimuus is a subjectie graet that ca be ssipated by habituatio but its existece presupposes that of a objectie graet, a cocetratio of excitatory substaces i the eura substrate. Ad simiary for huger cotractios of the stomach ad other compoets of the experiece costitte a subjec tie grat but oe that is aso uderpied by a objectie oe, such as cocetratios of gucose i the ama's bood stream. Remoig a rat's stomach eimiates the frmer but ot the atter. This meas that a a eura et must capture is the motiatioa effect of huger ot aythig about its pheomeoogy. Impemetig the capacity for istrumeta coditiog ca be doe by adg to the eura ets hadig sesory ad motor actii ties thers actig as a source of motiatio for behaior I oe rea tiey spe pemetatio four eura ets were ked together, two for itera reiforcemet (peasure ad pai) ad two for motor beaior (approach ad aoid). Latera ihbitory coectios esured that oy oe member of each pair was actiated at ay oe time. Extera reiforcemet was coected to the itera reiforce met eura ets through xed coectios to capture the fact that aimas do ot hae to ea to respod to extea sources of pea sure ad pai. Simuated aimas equipped with these four eura ets were abe to master some tasks, such as aigatio through a maze, that ormay reuire traiig through istrmeta coditio ig aborator aas. 1 I oter simuatios the positie reiforce met ceters were further differetiated so that their output actiatio pattes deoted a gie degree of huger or thirst. This icresed the degree to which the behaor of the simuated orgaisms was 100
NEURAL NET AND MAMMALAN MEMORY
goaorieted, seekig ot just positie reiforcemet geera but food whe they were hugry or water whe they were thirsty 12 Oce we embody ad situate eura ets with the capacity to recogize edurig etities the ext step woud be giig them the abiity to aayze scees i which those etities iteract payig the roes of aget or patiet. Laboratory eidece for the capacity of birds ad mammas to perform scee aaysis is more difcut to obtai because of our compete ack of access to their pheomeoogy. I some cases we ca compesate for tis imitatio by teachig a aima how to express its experiece through extea meas It is we kow for istace, that chimpazees ca be taught to use a reatiey arge ocabuary i sig aguage. They ca aso be taught to use ot had gestures but coored pastic objects of differet shapes arbitrariy paired with referets, that is paired i such a way that either the shape or the coor of the pastic object bears ay resem bace to its refereta bue triage for istace associated with a appe Chimpazees are ot oy abe o associate te pastic objects with rea etities ad actios but ca aso arrage them i sequeces that gie us iformatio about how they make sese of a scee. Gie the pastic objects associated with the traier with itsef, with the actio "to gie," ad with the etity " appe" for exampe, a chimpa zee ca arrage them i what seems to e requests ike "aier Gie Appe Chimp" The chimpazee ca aso be traied to uderstad requests ike "Chimp Gie Appe aier."3 These resuts do ot show that chimpazees ca ea to use aguage because the sequeces of pastic objects do ot hae a rea sytax, that is they do ot hae the combiatoria productiity of huma words. But the outcome does show that they ca pace etities ad actios i reatios that are sig icat to them they ca uderstad ad memorize scees i which oe participat acts as a aget aother as a patiet ad i which certai actios ik the two together. Gie the scarcity of eidece about the eura substratum ad pheomeoogy uderyig the capacity for scee aaysis i ai mas ad about the reated abiity to memorize actuay experieced scees, the subject must be approached idirecty. We wi rst scuss the kids of modes that hae bee created for huma episodic mem ory a case i which we do hae access to the pheomeoogy ad 101
NEURAL NETS AND MAMMALIAN M EMORY
PHILOSOPHY AND SIMULATION
then work our way back to the anl case Current simlations of human episodic memory use proposiions, the meaning of sentences tha express (trly or flsely) facs about he world. Propositions are used as component parts of ore complex whoes called scripts, strc red repesentaions capring the reguarities of rouine siuaions Scrips can be sed, for expe, o represen he regulariies in activ iies lke gong o a estaurant The space of possible actions in resta ran scenes has a srcture-wih some acions being casa precondiions for he occurrence of ohersa srcure hat allows s to make inferences like "I the csomer does no ow wha food is available she mst ask for the men" or "If the csomer is nished eating he mst pay the check" The hypothesis behind scripts is that this strcre mst be captured by our minds if we are o sccessflly and routine anticpate acton possibliies sing our epsodic mem 0ry of past vists to restaurants The restarantgoing scip was cre ated b rst subdividing the possiblity space into types of restarant (facy restaran, cafeteria, fastfood joint) since the action opportu niies in each of these types is quite diferent Then the different res taurant types ( calld "tracks" ) were sbdivded into scenes "Entering," "Ordering," "Eating," "Exiting," each of which was dened by a sequence of simpe acons Eah track was lso given a set of charac teristic objects (tabes, menus, food) and roles (cstomer, waiter cook, cashier) The typcl conditons of entry to the sitation (hungy cs tomer wit money) as wel as its typica outcome (satiated cstomer with ess money) were lso incuded Scripts an other strcture representaions of scenes were creaed in the symboc schoo of Aia Integence so it is not surprising that there is a nguistic bias blt into them Specicay, the conten of epsdc memory was assued to be anguageike, an assumption hat may e defended on a varety of grounds We may argue, for instance that hman experence is strctured ingstcally a thesis that has lite experimental suppor but that has been very inuentia in pilosophicl crc�s in the past 2 0 0 years A weaker but more defensbe argument woud be that we have no accss to other peo pe 's epsodic emores except throgh verbl reports This s indeed true bt t ay be arged that those repors are ony a pae shadow of the more or ess vivid reexperiences constitting autobiographica recolecions Finay, his way of modeling episodic memory may be 14
15
102
defended in terms of the goal of a particular simulation The above script, for exampe, was created to ive simulated agents the capaciy to understand short stories aking pace in restaurans, the agents displaying their understanding by answering questions abou, or paraphrasing passages from, those short sories B ecase short sories typically leave o many deails the ability o paraphrase hem or to answer quesions abou hem invoves inferring hose missingpieces of informaion using as an aid something lie a scrip Given he goa of endowing a simated agen with this emergent capacity the con cept of a scrip may be accepted as being sefu for expanations of episodic memory at the mechanismindependent level The rst step toward adaptng his expanatory strategy to the case of avian or mammalian memory involves creating a neura net impe mentation of scipts, one hat displays the same capacities to para phrase and answer questions abot short sories Unle he scipts created by the symbolic school those using nerl nets do not have to be handcrafted by their creators: the scripts are leaed by exampe extraced from statistica reguarities in aca stores mch ike mu tiayer perceptrons extract prototypes from sensory stimuation In other words, unlike the symboc version of scripts those impemented with nera net are emergen The modues that compose ths impe menaion differ in important ways from the nera nets discssed so far so a desciption of their design wi be necessary before expaining how he entire assemby works The two nove architecres are recurrent neural nets and se-organizing maps. Recrrent nera nets are ke mltlyer perceptrons augmented with feedback connec tions In regar mulayer perceptrons there is backward movement of information abot he degree of discrepancy between current and desred opt pattes bt this feedback operaes only dring train ing After the weights of the connections have converged on their na values the nerl net ceases to use any feedback Recurrent neural nets, on the other hand, use feedback dring actua operation The simplest version adds o a mutiayer percepion an extra se of nits, caed "context nits" inked o the hidden ayer by connec tons wih nchangeabe weights and operating at f strength This means ha the context nits receive an exact copy of the activation pattern that the hidden layer had mmedatey before ts current one This copy is hen fed back to the hidden layer (long with the activation 16
103
PHOOPHY AND MUATON
coming from th nput units) causing its prsnt stat to b rlatd to its immdiat past Th rsult is an mrgnt capacity to xtract temporal regularities from snsory stimulation a capacity that can b usd, for xampl, to analyz spokn sntncs and prdict th nxt word in a sris17 Th scond typ of nural nt ndd to implmnt scripts is th slforganizing map I its simplst vrsion this dsign uss only two layrs on of which, th input layr, functions just lik in multilayr prcptrons Th scond layr, th map itslf, is mad of computing units arrangd in a spcic spatial pattrn, such as a twodmnsional rctangular array Tis dsgn is said to slforganiz bcaus unlk multilayr prcptrons it is not givn a targt output during trainng or th dgr of mismatch that ts currnt output has rlativ to th targt In othr words th traning of a slforganizing map is unsu prvisd Th way in which th trainng procds is as follows Evry unit in th input layr is connctd to vry unit in th map layr, th wights of th connctions st to random valus at th bginnng of th procss n activation pattrn from a st of training xampls is thn prsntd to th nput layr Bcaus an activation patt is simply a list of intnsitis that can b xprssd as numbrs, and a conguration of wights in ach st of connctions is a list of strngths that can also b numrically xprssd, th activation patt of th nput units can b compard for its dgr of similarity to th wight conguration of ach st of connctions to th map layr Aftr stimu atng th input layr a comparison is prformd for ach st of con nctions and th wight conguration that happns to b mor similar to th input pattrn is sctd That is, at th start of th trainng whatvr smilarity thr may b is compltly by chanc but it nv rthss alows us to pick a particular unit in th map layr as th winnr Aftr slcting th winnr th wight conguration of th st of connctions to th winnng map unit is changd so that it rsmbls th input actvation pattrn a littl bit mor Th connctions of th spatial nighbor of that map unit ar also adjustd to bcom mor lk th input, th dgr of adjustmnt dminishing th furthr away th map units ar from th winnr Thn th procss is rpatd with th sam input pattrn, th winnr now bing th sam ma unit as th prvious itration sinc th wights of its connctions hav alrady 104
NEURA NET AND MAMMAAN MEMORY
bn mad mor similar to th input Th wights of th connctions to this unit ar mad vn mor simlar as ar thos of its nighbors xcpt that no th siz of th nighborhood is dcrsd This pro css is rpatd until th siz of th nighborhood has bn rducd to a singl map unit At that pont th conguration of wights associ atd with th wining map unit ar a lmost idntical to th input acti vation pattrn whil thos of its nighbors ar vry smilar to it This mans that if w activatd th wing unit and alowd ths activa tion to travl backwars through its wightd connctions it would rcrat th original pattrn at th input layr nd f a nighboring unt was activatd it would rcrat a pattrn similar to th original on Through ts unsuprvisd training procdur th activation pattrn in th nput layr is ffctvy stord in a rgion of th map layr, or mor xactly, th mans to rproduc t ar stord in th conguration of wights btwn th winning map unit and th input layr 18 This trainng procdur can b rpatd many tims for iffr nt input pattrns ach on stord in a iffrnt rgion of th map Whn iscussing th fac rcognition nural nt abov w saw that smilar facs wr mappd nto narby rgions in th spac of possibl activation pattrns of th hiddn unts and that this transformation of smlarity rlations into rlatios of spatial proxmity was crucial to th ability of th nural nt to rcogniz objcts A simlar ffct taks plac n slforganizing aps· xcpt that in this cas th similarity btwn nputs is mappd nto proximty rlations in th map layr itslf I othr words, th rlations of proxmity ar now btwn rgions o a concrt actual spac not an abstract possibiity spac This changs th dgr to which th stord distributd rprsntation is accssibl by othr nura nts Th prototps that a multlayr pr cption xtracts ar availabl only to itsf, a limtd accss that maks thm usful to mod procdural mmory but not dclarativ mm 0ry, whthr pisodic or smantic But onc ths prototps ar mad xplicitly spatial thy bcom publicly availabl: othr nural nts that can activat th winnng map unit in th slforganizing map and that can rad th rsulting activation patt at its input layr will b abl to rtriv th distributd rprsntation I addi tion, th stord distributd rprsntation is mad simplr and mor compact bcaus a slforganizing map can prform a rduction of th numbr of dmnsions of th possibty spac of activation pattrns 105
PHLOOPHY AND MULATON
the number of hidden units in a neual net is, say, eight, then its possibility space has eight dimensions. This number can be reduced to only two when the activation pate is projected into the two spatial imensions of the map The selforganizing map itself selects the two imensions that have the most variablity, that is, the two that are most likely to be relevant for the ability to recognze patterns. The other six imensions are approximated by the way in which ifferent istributed representations are placed in the map, a placement that may develop an ntertwned form resembling the way in which a onedimensional fracta curve tries to approximate a twomensional surface y repeatedly foling on itself 19 Combining recurrent neural nets and selforganizing maps to mplement scripts has been done in a simulation called Discern." Disce uses eight separate modules, six of which are used for processng propositions and two for memory storage The rst processng module uses a recurrent neural net to parse sentences, that is, it inputs a temporal sequence of words (subject, verb, object) and outputs a series of roles agent, patient, action performed instrument used in the action. A second recurrent neural net uses those roles as an input to parse stories, that is, it maps the series of roles into the categores of a script. o other processing modules perform the reverse operations one takes scripts as inputs and outputs a series of roles, the other inputs the roles and outputs a sentence. Finally, two additional processing modules are used for the specialized tasks of answering questions and producing paraphrases 2 The two memory modules are selforganizng maps used for the storage and sharing of istributed representatins, one corresponding to semantic memory the other to episodic memory The former is needed because the ablity to understand short stories presupposes a previously acquired lexicon while the latter is where the structure of scripts is memorized The episodic memory uses several maps arranged n three ifferent layers one layer for scripts (going to a restaurant, shopping, traveling); another for tracks ( different places to eat or shop, different means of transportation) ; and one to store the pattes corresponding to the incumbents of roles (who is doing the eatng, shopping, or traveling where these activities are performed what is being eaten or bought) The mltiple maps form a pyraidal structur in the sense that each unt of the top map is connected to a separate track map, and each
106
NEURAL NET AND MAMMALAN MEMORY
unit of a tack map is connected to a larger map in which roles are bound to their incumbents. This pyramdal structure greatly facilitates the extraction of the inherently erarchical structure of scripts from the ifferent short stories included in the trainn set 21 The end result is a modular neural net that exhibits the emergent capacity to answer questions about short stories, or paraphrase parts of them, by inferrng missing pieces of information using its episodic memory Using Discern to guide philosophical tnking about mammalian and avian memory involves simplifying some aspects of its design as well as complexifying others On one hand, the design must be made simler because the animal capacities we need to reproduce are not as sopsticated as those that Discern simulates. we tnk of scripts as internal models that explain actually lived scenes then the needed simplication would boil down t this whereas human beings use as explanatory resources not only phsical causes but also reasons (social conventions, cultural values) and motives (personal goals), birds and mammals need only frame their expanations in terms of causes since they are not capable of ascribing mental states to others nor understand how shar ed conventions or values shape behavior. When, for example a chimpanzee uses colored shapes to make a request (ainer Give Appe Chimp") it is merely trying to causally affect the behavior of the trainer not to inluence his or her mind On the other hand, we would have to make Discern more complex because in its current form it makes use of inguistic nformation that would not be available to aimals Its nterface to the externa world is mplemented through mages of letters transformed into activation patterns, that is, it extracts scripts from printed texts not from sensory experience Ts greatly facilitates the task of assigning the roles of agent and patient to objects in a scene using the syntactic and semantic structure cotained in printed sentences, a structure that is not automaticaly provided by visual experience The changes necessary to break the inguistic dependence of scripts should be easier to implement in Discern than in the orginal symbolic version becase the task of making sense of situations is carrie out in the former with emergent distributed representations used as buiding blocks to compose equaly emergent erarccal taxonomies. Those istributed representations could, for example, be supplied by a video camera using the face recognition neural net iscussed above 107
PHLOOPHY AND MULATON
But the real problem is not so much the kind of informatio used as sensory input as the fact that Discern is entirely disembodied The lack of a body situated in space leaves Discern without the resources needed to distinguish what is relevant in an mage for the purpose of assigning perceived objects to the roles of agent and patient, or to understand the connection between a cause and its effect. As we sug gested in the previous chapter possession of a body gives anmals a variety of ways to lter out irrelevancies. First of all, learning at the level of the species builds into an animal's body a variety of inherited behavioral tendencies that have evolved in connection with a real environment and that are therefore relevant to operating in that environment Habituation, in turn allows animals not to spend cog nitive resources on stimuli that required an active response prior to getting used to it, in effect transforming a stimulus that was rele vant into an irrelevant one. Classical conditioning permits inherited relexes to be triggered not only by a stmulus that has survival value but also by a previously irrelevant stimulus tha is nevertheless pre dictive of the presence of the rst These smple forms of procedural memory represet assessments of signicance made by the species as a whole over many generations, assessments that can be further modulated or adapted during an ani mals lifetime To these irrelevancy lters already avalable to insects we must add the habits acquired through instrumental conditioning in larger animals habits that are less constrained by genetics The lexibility of habits comes at a price since as it was argued above learning through reinforcement can lead to the formation of supersti tious habits bearing no relation to the capacity of an environment to affect and be affected by the animal Nevertheless, when animals learn good habits they can not only automate a series of actions to be executed without any conscious planning but they can also learn to anticipate the routine consequences of their actions Considering that much of what makes up a script is routine and repetitive procedural memory in the form of good habits can considerably narrow down the range of experience that must be deliberately considered by the animal to make sense of situations I other words, an embodied agent can let different forms of procedural memory take care of lter ing out insignicant details whle it focuses its attention on the most iportant aspects of a complex scene 108
NEURAL NET AND MAMMALAN MEMORY
I addition to giving Discern a body we must also situate that body in space. When simulated agents are modeled without a body they are treated as smart planners but dumb executioners planning every thing in their heads prior to performing an action When they are simulated without situating them in space they are treated as if they had o information about their mmediate surroundings22 This information is referred to as indeical knowledge Knowledge about an external object's position, for example, can be expressed ndexically, the object is in front of the agent and to its left, or in a context independent form by givng its latitude and longitude Whle indexical knowledge comes from an agent's situated point of viewa ground level view of reality, as it werete source of contextindependent knowledge in simulations is the godlike aerial point of view of the designer23 I the particular case of scripts indexical knowledge is nec essary because at the lowest level of the taxonomy there are specic roles that must be bound to their incumbents, the incumbents being in most cases entities known only indexicaly A possible obstacle in this regard is that neural nets are not good at solvng binng prob lems, whether the problem is binding pronouns to proper names in a linguistic context variables to their values n a athematical context or roles to their incumbents in a sceneanalysis problem One way of overcoming this lmitation is suggested by the experimental observa tion that groups of neurons in real brains, each of which extracts different features from the same object, tend to re in synchrony with one another when perceiving that object I other words, temporal simultaneity in the rng behavior of neurons can act as a dynamic binder. This insight has already been successfully exploited in some modular neural net designs that perform scriptlike inferences.24 Even though we are still far from understanng how avian and mammalian memory and perception work the ideas we have derived from simulations suggest that the key ingredient in the explanation of those capacities continues to be the extraction of prototypes from sensory exposure and the "storage of those prototypes in the form of distributed representations. I the case of scene analysis the proto types must be extracted not only from smilarities in the spatial con guration of the dening features of sensed objects but also from patterns in their temporal behavior, such as the regular cooccurrence of lnear causes and their effects. The more enduring form of memory 109
PHLOOPHY AND MULATON
needed for autobograhcal recollectons, n turn, must be le mented through the use of selforganzng mas or smlar desgns n whch the number of dmensons of the ma s not xed n advance2 5 Fnally, we must rovde these neural nets wth the necessary rrele vancy lters wth whch evoluton sules emboded anmals These are artcularly necessary for the oeraton of selforganzg mas because the mechansm through whch they acheve the storage of dstrbuted reresentatons resuoses a crteron of slarty rovded exogenously, and smlarty s a very roblematc relaton two objects that are dssmlar to each other vsually may be smlar n ther audtory, olfactory, or hatc roertes Even more roblematc s the fact that any two enttes may be sar to each other n an nnte number of ercetually rrelevant ways n eg nhabtants of lanet Earth, for nstance, or n movng slower than the seed of ght26 Evoluton has ltered out many rrelevant dmensons of sm larty from the neural machnery used by anmal brans but dsem boded neural nets do not have access to these automatc lters It s tme now to move beyond sulatons of ndvdual agents and ther ndvdual minds to consder entre communtes of agents Whle n the former case we are conceed with how an emboded and stuated agent makes sense of the oortuntes and rsks ro vded to t by ts envroent, n the latter we must add the oortu ntes and rsks suled by ts nteractons wth members of the community of whch t tself s a comonent art Snce we wll now be conceed wth larger wholes wth ther own emergent roertes the detals of how ndvdual agents erceve and memorze can be taken for granted Ths means that smulatons can use ngustcally exressed rules and rocedures to generate agent behavor wthout beggng the queston The followng chater wll tackle ths subject startng with multagent smulatons of the soal behavor of our closest relatves n the anmal world and movng on to consder how the more comlex soal behavor of our huntergatherer ancestors could have emerged
110
CHAPER EIGH Mutiagents and Prmate Strateges
As the sychologcal lfe of anmals became rcher and ther behavor ncreasngly determned by ther membersh n a community ther nteracons rovded the medum for the rse of a new knd of entty, soal strateges, and generated new ossblty saces n whch the sngulartes dened sets of mutually stablzng strateges Although some nsect sees have an elaborate soal lfe wthn colones or hves ther nteractons do not nvolve recognton of ndvdual con secs and take lace by ndrect communcaton, that s, by altera ton of the hyscal and chemcal roertes of a shared enroent Monkeys and aes, on the other hand, are caable of recognzng each other not only as unque ndvdual organsms but also as belo ng ng a certan ksh grou and occyng a gven rank n a dom nance herarchy The lkelhood tha t a young moey or ae wll start a ght wth another, for nstance, deends not only on memores of the outcomes of revous aggressve encounters but also on the res ence or absence of an older relatve actng as a soal ally. Other examles of the kd of soal roblemsolvng erformed by these anm�ls nclude coaton formatona lower rankng anmal may be able to challenge a rval f he or she has revously establshed an allance wth a hgher rankng one-as well as reconclatory behavor to undo the damage of revous aggressve encounters The observed soal caates of contemorary nonhuman rmates rovde us wth the best gude to seculate about how our las t common ances tors could have affected (and be affected by) one another More se cally, they rovde us wth nsght nto the mechansms behnd the 111
PHLOOPHY AND MULATON
emerence of reciprocal altrism, a social nteraction involvin the exchane of favors n wch there is a temporal delay between the provision of a service and its reciprocation Reprocal altrism, in trn, may have been the point of departre for a lon process that wold eventaly lead to the emerence of comnal solidarity in ancient hnteratherer comnities Soldarity in contemporary hman commnities acts as a social radient that varies amon different rops endowin them with dif ferent capacities for political mobilization and other forms of colec tive social action. The mechanism of emerence of solidarity in these commnities depends on the existence o f a dense networ of connec tions amon their members In a comnity n which everyone nows everyone else word of moth travels fast, particlarly if the sbject of conversation is the transression of a local norm: an nfl lled promise an npaid bet, an nreciprocated favor This rapid transmission of linistic information allows the commnity to act as a reptationstorae mechanism If n addition ridicle and ostracism are sed to colectively pnish those who dishonor commitments then the comnity has the capacity to act as an enforcement mech anism The qestion to be explored in this chapter is how the transi tion between the reciprocal altrism of the common ancestors of hmans and chimpanzees and the comnal solidarity of the eariest hnteratherers cold have occrred withot the se of inistic resorces Some experts beleve that the rst emerence of lanae (lottoenesis) too place between thirty and fty thosand years ao when brial rites, sophisticated stone tools, and relar tade were already wel estabshed Others believe that lottoenesis occrred in an earler period, the middle paleolithic, that bean abot two hndred and fty thosand years ao2 Eithr way, by the tie lan ae appeared the ancestors of mode hmans had already spent hndreds of thosands of years as members of comnities in which a dal economy of hntin and atherin nvolved new role assin ments, pbic monitorin of relations, and colective evalation of resorce alocations I other words, a complex social life preexisted the emerence of lanae Reciprocal altrism amon nonhman primates invoves strateic interactions between two animals spplyn one another with an 112
MULTAGENT AND PRMATE TRATEGE
opportnity to cooperate-to roo m parts of the bod y that are inac cessible withot help, for examplebt also with the ris of bein cheated, since the delay before reciprocation means that there is no mediate assrance the service provided will be repaid This is an example of a ocial dilemma a social interaction that can have collec tive benets bt that is endanered by the possibility of reater indi vidal ains I the simplest social dilemmas interactions between pairs of aents are strctred by two opportnities an two riss: the temptin opportnity to cheat a coopertor ainin at its expense (called temptation or simply T); the opportnity of mtal an by cooperators (a reward or R); the ris of mtal cheatin (a pnishment or P ); and the ris for a cooperator of bein cheated (bein made a scer or S ) The eld of mathematics that stdies strateic interactions models them as ames so these opportnities and riss are referred to as the payoffs of the ame What maes a strateic interaction a social dilemma is the orderin of the payoffs I the sitation referred to as the Prione r' dilemma the payoffs are arraned lie ths: T > R > P > S If R > T then there is no conict between individal and colective ains so all dilemas mst have T > R Bt this particlar arranement also mplies that the worst thin that can happen is to be a cooperator that is cheated (P > S) 3 A different dilemma can be enerated by the arranement T > R > S > P This is caled Chicen-after the ame in wich two drivers on a colsion corse play to see who swerves ot of the way rst-becase nlie the previos case the worst otcome is when neither driver swerves and both crash (S > P) Lie all opportnities and riss the payoffs of strateic interactions strctre a space of possibiities, an objective possibiity space shared by the interactin aents In addition to this objective space we mst also model the sbjec tive space of possibilities, that is, the strateic choices available to the aents I the simplest case these are smply the choices to cooperate or cheat bt in more complex sitations the sbjective possiblity space contains not only choices bt strateies. Becase in the latter case the vale of one aent 's stratey depends on its relation to the strateies of other aents what we need to explore is the srctre of an intersbjective possibility space This intersbjective space is strc tred by sinlarties caled Nah equilibria denin combinations of 113
PHLOOPHY AND MULATON
strategies that stablize one another because no agent can ge a beter payoff by unlaerally changng to a ifferen srategy. hen he opporuniies and risks are arranged as in the Prisoners ilemma and he ineraction occurs just once he only Nash equlibrium is for boh agens to chea: they ould be better off if hey cooperated, since R > P, bu neiher one can choose cooperaion unlaerally ihou risking being cheated and geing S f he teracion is repeaed many times then mutual cooperation becomes a Nash equilibrium except in he very las encounter in hich mutual cheating pays best But if he agens o hen the las interacion ill occur hen it pays to cheat in he previous one as ell, and for that reason in the one prior to hat one, and so on all the ay o the very rst one So the nteresing case from the point of vie of modeling he emergence of coopera tion in primate communities is a repeated (or ierated) strategic nter acion in hich neither agen knos hen he las encouner ill ake place This is caed the indenitely iterated Prisoners dilemma, a version characerized by he fac that he "shado of the fuure (the probability of future interacions) helps to stablize cooperaive out comes4 hle social ilemmas involving only a single interacion represen he conict beteen nivdual and collective nterests, he iteraed version adds a emporal mension that captures the tension beeen shor and longterm inerests The simplest srategies are those that do not ake hisory nto accoun: Dove the srategy of alays cooperating and Hak he sra egy of alays cheating Nex in complexiy are srategies n hich an agents choices take into account only the previous acion of the other agent, cooperating if it cooperated or cheating if i cheated, for exam pe If e add to this tha in the very rs encouner, hen here is no history on hich o base a choice, an agent alay cooperates e ge he famous srategy on as Tit-fr-tat, iner of the rs simulaed competiion beteen strategies5 Next in complexiy are strategies tha nvolve remeering not only he other agens previous choice bu also an agens on previous choice example ould be a sraegy insrucing an agent o cooperae in the next encounter is previous coperaion as reciprocated; to shif back o cooperae if boh previously cheaed to chea agan if i cheaed hie the oher one cooperated; and nally, o shift back to cheat if it cooperated hile he oher one cheaed This srategy is referred o as Pavlv 114
MULTAGENT AND PRMATE TRATEGE
(or inSay, LooseShift) and dffers from Tiforat in hat i takes advanage of unconiional cooperaors (Doves) More complex stra egies can be generated by increasing he amoun of memory ha agens mus bring to bear remembering, for instance, no us he previous action bu he prevous o actions example ould be a more "forgivng version of Titfora (called "Titforoats) in hich an agen mus be cheaed tice in a ro before cheaing back These ideas can be used o model the social straegies of nonhuman primates in the case of dyac ineractions, such as the reciprocation of grooming or food sharing beteen to animals, and only hen the interacin< pair is not related by kinship, since kdred anmals may be genetially preisposed ord cooperation Some triaic interac ions, lile sarting a ht in the presence of an ally, may be broken don into o dyac ones: he ineracion ith he ay, ho may have been groomed in advance and is being present at the ght is its ay of reciprocaing, and he agonisic interacion iself, a ritualized contest in hich opponents assess each ohers srengh In this con est using the Hak straegy means to escalae he level of violence untl injured hile using Dove means to make aggressive isplays (such as bearing ones teeh) unil escalaion sars then ithdra6 I adtion to restrictng he application of he model to pairs of unreated anmals there is the quesion of heher nonhuman pri mates can n fac analyze each others behavior and base heir on responses on the results of hose analyses. Evidence in favor of this hypothesis comes from experimens using dyads of unrelated tamarin monkeys placed in a siuation in hich one of hem has an opporu niy to provide food for he oher ithout iself geing any reard In one experiment some amarin monkeys ere previously trained to use he Hak and Dove srategiessome ould never pull he lever hat provided he other one ith food hle ohers alays pulled iand then hey ere paired ih unraed monkeys The resuls ere clear: he untraned monleys puled signcantly more for the Doves han for he Haks This result, hoever, does no necessarily mply ha he untrained monkeys ere reciprocaing pas behavior tha as menally ana 1yzed and determined o be alruisic, so a ifferent experimen as carried out to rule out the possibility ha mere satiation as he moivaing factor o unrained monkeys ere placed in he above 115
PHLOOPHY AND MULATON
situation but alternating in roles each having several chances to deliver food for the other without mmeate benets to itself. The rate at which they cooperated in these interactions was recorded and then they were made to interact in a situation in which each pull of the lever resulted in food for both of them Finally they engaged another session of alteating altruistic food exchange. their rate of cooperation depended on mere satiation then the intervening situation in which both got food should have increased the rate of cooperation. on the contrary the rate depended on their memory of the others past behavior then the intervening situation should have had no effect. The results were compatible with the latter hypothesis the monkeys seemed to be responng to prior intentional acts of food giving by their conspecics.7 Finally there is the question of whether monkeys and apes have the cognitive capacity to use behavioral strategies lke Titfortat or Pavlov as part of a larger com munity in which each animal must remember the last outcome of every interaction with other members iterally keeping score of all its strategic interacions Laboratory stues incate that only chim panzees have the cognitive sophistication to do this while other species may be using smpler behavioral strategies such as mirroring each others behavior or favoring those with whom they spend time in close proximity. 8 Gien that all we need in this chapter is a way of modeling the social strategies of the common ances tors of humans and chmpan zees this evidence is enough to alow us to assume that they could memorize partnerspecic contingencies punish cheaters and perform score keeping The next step is to nd a way to explore the emergent consequences for an entire community of many ndenitely iterated dyac interactions I paricular we need to be able to track a chang ing composition of strategies over many generations. some strate gies come to dominate the community in the long run and if these strategies involve some form of reciprocal altruism (like Titfortat or Pavlov) then social cooperation can be considered a stable outcome and thus one lkely to have actualy occurred among our ancestors The smplest way of exploring this question is by coupling models of strategic interaction to genetic algorithms Coding a given strategy into the smulated chromosome of a genetic algorithm is relatively straightforward because the choices that make up a strategy are binary 116
MULTAGENT AND PRMATE TRATEGE
(cooperate or cheat) as are the bits that compose the genes f our bits can ode for the binary responses to an agents previous move while an added two bits can tell the agent what to do in the very rst move. 9 The outcome of the smulations on the other hand can be quite complex because the tness of a particular strategy is frequencydependent in the sense that as a strategy becomes more frequent in a population it can create the contions for its eventual dese. I other words, the relative success of a strategy depends not only on its intrinsic properies but on the proportion of agents that are using as well as on the proportion of agents using other strategies. Frequencydependency creates such complex dynamics that the results from one simulation may change when one of the parame tersinitial proportions of strategies, initial x of strategies number of generations mutation rateis changed. This is clearly illustrated by the fate of the very rst smulatio n of this kind in which Titfortat came to dominate the population after 50 generations in a large number of runs This result was viewed as a watershed in the study of reciprocal altruism and also carried some moral comfor because, to use psychological metaphors, the winnng strategy was "nice since it always cooperated in the rst move; it was "forgiving renewing cooperation if a former cheater cooperated; and it was not liable to be ausd because it was "retaiatory, punishing cheaters by cheating I the original simulation the initial mix contained only Hawk Do ve, ad Titfortat as alternatives. Hawks performed well at rst by taking advantage of Doves but as the Doves began to disappear from the population so did the Hawks leaving only Titfortat But when the same smulation was alowed to run longer than 50 generations frequencydependent tendencies became manifest a mutation transformed one copy of Titfortat into Dove and because al Hawk strategies were gone Doves were able to slowly spread as more mutations accumulated at random; and once the Dove strtegy had acquired a critical mass a mutation producing a copy of Hawk had again a chance to get a foothold and then invade the population. is now wel known that Pavlov will do better in this scenario because uike Titfortat can take advantage of Doves and prevent them from creating the conditions that lead to the spread of Hawks. Pavlov on the other hand, does not do wel against Hawks so to spread through the population it needs to use Titfortat as a stepping stoneY 117
PHLOOPHY AND MULATON
I ths simulations th bhavioral stratgis wr modld as if thy wr inhritd in a rigid way that is as if th agnts did not larn during thir liftims Whn larning is rformd by th s cis ovr many gnrations a Nash quilibrium rfrs to a oulation of stratgis that is stabl against invasion from mutant stratgis Ths ar rfrrd to as evolutona stable strateges, stratgis that rform succssfully whn intracting with othrs as wll as whn intracting with cois of thmslvs I som cass it is asy to stab lish whthr a strtgy is volutionary stabl th Hawk stratgy for instanc is clarly not stabl bcaus vn though it is good at xloit ing othrs it dos not do wll against itslf But as th cas of Tit fortat shows a stratgy rviously thought to b volutionary stabl may tu out not to b so in th long run and this comlxity cals for a comlmntary aroach using both comutr simulations and mathmatical analysisY Whil mathmatical analysis rvals th sin gularitis that structur th intrsubjctiv ossibility sac comutr simulations can actually folow th ath toward th singularitis W hav alrady ncountrd this comlmntarity btwn th two dif frnt aroachs in arlir chatrs but in th cas of simulations of intracting agnts bing abl to track a historical trajctory is vn mor imortant bcaus many intrsting vnts occur during tran sents, that is during th riod of tim bfor a singularity is rachd. I som cass a frquncydndnt dynamic may snd most of its tim in long transints This imlis that giving an xlanation of th longtrm consquncs of stratgic intraction involvs showing not only that a Nash quilibrium xists and that it can b rachd but also roviding an stimat of th tim scal ncssary to rach it 13 Simulations can also comlmnt mathmatical analysis by hling to clarify som conctual assumtions Thr ar two ascts of th conct of a Nash quilibrium that nd clarication On has to do with th maintnanc of quilibrium that is with th mchanism that rstors stabilty aftr an xtrnal shock whil th othr rlats to th slction of on Nash quilibrium ovr anothr in th cas of ossibility sacs structurd by multil singularitis Th slction roblm can b solvd through a mchanism that rforms a sarch of th ossibility sac to convrg on on of th svral availabl Nash qulbria. In arlir chatrs w hav sn that biological volu tion constituts a owrful sarch mchanism and th simulations 118
MULTAGENT AND PRMATE TRATEGE
just discussd shows how usful thy can b to ntun our intu itions about th slction rocss.14 Th maintnanc roblm on th othr hand nds a diffrnt aroach. In traditional mathmatical modls th roblm is solvd via an assumtion of full common knowldg on th art of all agnts it is not nough that a stratgy is such that no on can switch unilatrally to a diffrnt on and b bttr off but in addition vry agnt must know that al othr agnts know this Th assumtion of common knowldg howvr dmands ostulating cognitiv caacitis that th ancstors of humans and cmanzs could not hav had So th solution to th maintnanc roblm must b sought in th dvlomnt of good habits that cv th sam rsult.15 In th cas of our rimat ancstors thos good habits may hav bn dvlod through rinforcmnt larn ing in tims of scarcity tims in which it was in vryons slfintrst to coorat by sharing food but bing habits as oosd to rational dcisions thy could hav survivd into tims of lnty 1 Using comutr simulations to clarify th maintnanc roblm imlis th us of agnts that can acquir habits in thir liftim and this in turn involvs making svral changs in th assumtions built into th modls. In articular unlik th cas of dismbodid stratgis in which th ayoffs can simly b assumd to act as incn tivs rinforcmnt larning imis rlating to th oortunitis and risks of stratgic intraction as rwards and unishmnts Thus th rst chang that nds to mad is to introduc a cutoff oint in th sris of ayoffs to mak som unambiguously rwarding and th othrs unishing Placing th cutoff oint in th midl so that T < R < 0 < P < S yilds th most intrsting rsults Nxt w nd to crat nural nts that can rlat to ths ayoffs as ositiv and ngativ rinforcmnt As w saw in th rvious chatr this involvs a crtain dgr of mbomnt in th form of nural nts rforming motivational rols such as simulatd hungr or thirst but in th simlst cas all w nd to assum is that th agnts hav a (nonmrgnt) tndncy to avoid th risk of unrcirocatd coo ration This boils down to th unroblmatic assumtion that agnts dislik bing chatd but it also carris th imlication that vry mmbr of a dyad has an incntiv to try to gt th othr mmbr to coorat. Thus what is ndd is a ouation of nural nts that train on anothr on a continuous basis as thy attmt to icit 119
PHLOOPHY AND MULATON
the food patches were sma and widey spaced23 Kinship is another way of generating cooperating groups for two fferent reasons First, bioogica evoution can give rise to atruistic behavior between cosey reated animas by a mechanism known as inclusive ness any gene presposing an anma toward atruism can propagate if the anima itsef reproduces or if cose reatives with which it shares those genes reproduce, so atruistic gestures toward reatives can have simiar evoutionary effects as sefinterested behavior Second, kinship can inuence the formation of groups because the benets of knrected atruism can ony be enjoyed if reatives stay in cose spatia proxm ity Both of these possibe effects have been tested in simations using genetic agorithms couped to neura nets n which evoution fuy determined the na weight conguration In one simuation a popuation of 100 mutiayer perceptrons was evoved, using information about kinship as an input and the decision to share or hoard food as an output The popuation was broken down into 20 grops of reatives and tness was measured by the amount of food gathered during a ifetime The outcome conformed to the incusive tness hypothesis whie atruistic behavior toward strang ers was weeded ot atruism between reatives was not Then a sec ond smuation was performed in which the food was given a spatia stribution and the neura nets were provided with adtiona units to aow their behavior to be shaped by spatia position: for young" agents two input units gave the position of their parents (one for ange, the other for distance) whie for adut" agents the input was the position of food items F or both age g roups adtiona output units aowed the agents to perform simpe movement decisions (to move forward, to turn right, or to turn eft) Parents had to decide whether to feed themseves, feed their existing offspring, or save food for future offspring since they coud reproduce severa tmes during their ife The resuts were that in young" agents a tendency to stay n ever coser proxity to their patterns evoved, and with it the emergence of custers of cooperators For aduts the resuts were more compex because their choices invoved compromises particuar, if they were too atruistic and d not sufcienty feed themseves they had shorer ife spans and ess overa offspring if on the contrary, they were not atruistic enough their offspring tended to e So reproduc tive success beonged to those that stroke the right compromise24 122
MULTAGENT AND PRMATE TRATEGE
Let's pause to consider what has been argued up to this point Existing aboratory evidence suggests that chimpanzees can not ony remember the outcomes of past encounters with invidua members of their community but aso anayze scenes invoving socia interac tions to detect tendencies to cooperate or cheat The scssion of episoc memory in the previous chapter makes the existence of these capacities rather unsurprising but a new twist has been added in this one the responses of chimpanzees in future interactions tend to be correated with the resuts of those anayses making their behavior an instance of a socia strategy Using these capacities for socia probem soving as a gide the qestion we set out to answer was whether some form of reciproca atrism coud have become estabished in the communities of the ast common ancestors of humans and chim panzees though the concusion we reached was not unambiguous it seems pausibe to assume that sets of cooperative strateies of changng composition coud have become estabished in prehuman communities through an nteraction of severa factors genetic pres positions couped to good habits formation of custers of cooperators ivng n cose proxmty and the effects of incusive tness I other words stabe sotions to socia emmas nvoving indenitey repated dyac interactions coud have been arrived at coectivey and then maintained in a community The next step is to use the ideas deveoped so far to mode the transition from our primate ancestors to preinguistic huntergatherers as a historica event that invoved arrivng at soutions to new socia demmas One way of approaching the emergence of cooperative huntng (or gatherng) is to mode it as a transition from the twoperson Prisoner's diemman which a whoe community may be invoved but nter actions are dyacto the mutiperson Prisoner's emma n which each member strategicay interacts with the rest of the community nother name for socia diemmas invovng a puraity of agents is pubic good demmas I the case of cooperative huntng the pubic good is the meat obtained at the end of the hunt a pubic good invov ng a cost to ooperators but not to those who refran frm huntng but enjoy its rewards when the meat is restributed I this case the opporunities and risks structuring the objective space of possibiities become more compex to cacuate since they must be a function of the number of cooperators, the costs of cooperating, and the degree to 123
PHLOOPHY AND MULATON
wich coopertion mplies (or mltiplies) indiidl efforts These three fctors cn be sed to ccte the pbic good tht is prodced, nmber tht mst then be diided by the nmber of people into which the spos re redistribted ince cheter gets the same mont of met s someone who priciptes in the hnt its pyoff is smply the rest of tht diision, while the pyoff for coopertor is tht mont mins the costs of cooperting For there to be diemm the mtipiction fctor mst be lrger thn one else there re no benets to colbortie hnting oer nd boe those of indiidls hnting lonebt it mst lso be less thn the nmber of persons into wich the spos be redistribted The reson is tht while cheters lwys do better thn coopertors becse their pyoff neer incldes costs, if the mtipliction fctor is lrge enogh cheters cn mproe their pyoff by switching to coopertion2 5 A mportnt difference between dydic nd commn interctions is tht strtegies ike tforTt re not prt of the intersbjectie possibity spce becse of the difcty of focsing retition A good illstrtion of tis difcty is the soced tragedy of te commons sittion inoling commnity of gents expoitng common resorce tht is nerbe to depletion by oereoittion 2 6 If single cheter oerexploits nd former coopertor ttempts to retite by lso oerexploiting, s tfortt wod do, tis not ony nfiry penizes those who refrined from selsh behior bt it my lso trigger we of frther retlitions n which eeryone oerexploits the resorce with predictbly trgic otcome Ths ible strteges mst nd wy of replcing the indiid pnish ment of cheters by the collectie enforcement of copertion In ddition, while in the cse of dydic nterctions it is p to he mem bers of dyd to detect nd remember deprtres from coopertion, when n entire commnity is inoed sch indiid memories re no enogh Tis mens tht we need wy to replce the indiidl recognition of cheters by the coectie storge of reptions As mentioned boe hese two mechnisms re mpemented in contemporry commities by the rpid spred of word of moth bot ioltions o loc norms nd by the se of ostrcism nd ridice to pish iotors Bt in the present contex we mst mode the emergence of both collectie mechisms withot prespposing the existence of ngge 124
MULTAGENT AND PRMATE TRATEGE
The problem of storing repttions cn be so led by replcing word of moth with the shred experience of ioltions of coopertion For entire commnities it wold be nreistic to ssme tht eery ioltion cn be witnessed by members Bt if commnity tended to sort itself ot into distinc grops nd if members of these sm grops tended to sty in cose proxmity to one nother then the episodic memories of grop members wold sfce to keep record of the pst behior of other grop members As we jst sw, both ecoogic nd genetic fctors cn bring bo the formtion of sch grops so we cn tke those reslts s indictions tht there is soltion to this problem The problem of collectie enforcement, on the other hnd, is not so esy · to soe becse of the existence of freqencydependent effects This cn b e illstrted with simple exmple Let's ssme the existence of certin inherited tendencies tht wold in ence the otcome of pbic goods dilemm behiorl tendencies like engeflness, the tendency to pnish those tht re cght, nd bodness, the tendency to chet reltie to the chnce of beng cght We cn code degrees of boldness nd engeflness ino the simlted chromosomes of genetic lgorit nd trck the reslting colectie dynmic In one simtion popation of gents strted with genes determning intermedite leels of both tendencies After few genertions the degree of boldness begn to f becse e en withot ery high engefness being cght chting ws costly As the proportion of bod gents decr esed genes for engeflness so begn to dispper becse with only few cheters enforcement costs become libility Bt s the erge degree of engefness in the commnity fel nd with it the costs of being cght boldness mde comebck 2 7 imlions like tis show tt he existence of certin person chrcterisics is not enogh to yied the desired otcome becse the ctiity of punising ceaters is itse a pu blic goo d if some cooper tors refrn from pnishng becse of the coss ssocited with it then they re cheting those th do pnish In oher words, ny behior tht promotes pbic good is itself second order pbic good Tis sggests tht the sotion to pblic good dlemms mst inole both primitie normsiherited or hbi ptterns of behior leding to the pisent of cheterss we s metanorms, the endency to pish those ht do not pnish2 Tis possible soltion 125
PHLOOPHY AND MULATON
mu be coupled wh h for cheer deecon, h , wh he for mon of mll group whn lrger communy n one mulon for exmple, gen were mched no group ung he norm do no chooe cheer member nd he menorm do no chooe hoe h chooe cheer oh degree of coopervene nd degree of vengefulne were coded no genec lgorhm nd populon w llowed o evolve dfferen rege he gen mde decon bou group formon nd choce o coopere or che The oucome depended on hree condon ffecng group formon The r condon reled o he wy n wch member hp w decded, h , wheher unlerl choce by n gen who wned o jon w enough or wheher ccepnce or rejecon of memberp depended lo on he group (muul choce) Wh un lerl choce here w no room for menorm o he oucome w predcbly h low cooperor cheed on hgh cooperor The econd condon nvolved he pobly for vengeful member o pl from group o expre her dfcon wh hoe who refued o punh Whou h pobly gene for gh vengeful ne could no ge eblhed nd low cooperor connued o che Fnlly, he hrd condon nvolved he pobly h gen who my hve depred from prevou group could reener noher group llowng relen group of cooperor o recru new mem ber nd grow he e xpene of oher group The regy combnng muul choce, plng, nd rejonng performed be oluon o he mulperon Proner ' Dlemm29 We my conclude h when he emergence of menorm llowed commne o olve he publc good dlemm nvolved n collecve hunng nd gherng he rnon from nml o humn comm ne w complee A menoned he r of h chper hee ovel commune emboded new reource grden, grden of oldry From h pon on he furher developmen of humn ocly begn o depend no only on he nercon of mll group whn cmmuny bu on he nercon beween dnc com mune' he oucome of whch w deermned n pr by he ol dry grden h ech of hem could explo More pecclly, when he prcce of erly humn commune begn o nclude he producon of one ool, chngng he wy n wch hunng nd gherng were conduced new Nh equlbr h were mprovemen 12
MULTAGENT AND PRMATE TRATEGE
over he l d one ppered n her pobly pce A greer degree of oldry n communy could hve mde he rnon o he new ngulry mooher by ncreng he degree of ocl coordn on needed o move hrough he pobly pce whle yng n he cooperve pyoff regon of h pc Commune h were ble o uccefully me h move would hve herefore ncreed her chnce of urvvl relve o commune h fled o do S30 Thu, wh fuure mulon mu llow u o do o rc many commune of gen hey move from one Nh eUlbrum o noher n n ever chngng pce of poble n he followng chper we wll exmne one of he ource of new Nh eqlbr, one ool echnology well new ocl grden h emerge from he nercon beween commune: he grden repreened by he gn from re
12
MULTAGENT AND TONE AGE ECONOMC
CHAPER NINE Muliagens and Sone Age Economcs
As ommuities o eary humans began to spread over the panet their ativities added to its surae one more ayer o ompexity. On one hand their members on whose bodies anient ommuna habits had aready sedimented aquired the apaity to transorm stone into manuatured toos On the other the toos themseves and the debris produed durng their manuature sowy aumuated eaving behind a reord not unike that o ossis embedded in geoogia strata. Most o the knowedge we have about the uture o those eary ommunities omes rom those deposits o stone artias Next to them areoogists have aso ound the bones o animas that ded at dierent ages giving us inormation about hunting tehniques as we as ashes rom wood or bone providing us wth evidene or the routine use o re Nevertheess it is the stone artiats themseves that are our rihest soure o data about the way o ie o our anes tors One o the saest onusions that an be drawn rom the strata in whih these deposits are arranged is their great ongevity the od est stratigraphi sequene (the Oduvai sequene) is about one and a ha miion years od. Muh o the ontent o this and other sequenes is reativey unior but as we approah more reent times the too and debris deosits begin to diversiy the variabiity in too types inreases in the midde paeoithi (two hundred and teen thousand to orty thousand years ago) and reahes a peak in the upper paeo ithi (orty thousand to ten thousand years ago) Toos rom these more reent per iods dispay a ompexity · that bears witness to the inreased sophistiation o the manua operations 128
used to produed them rom the odest method o perussion the deiberate breaking o a stone nto piees by a ew strong bows; to the more reent tehque o retouhing using ighter but more reuent bows; to the use o pressure aking rst or the reation o sharp bades then or the preision shaping o a too by ontrong the thikness o its dierent parts the sharpenng o its ends and the eimination o urves habits promoting ooperation onstituted the rst orm o human utura ontent transmitted aross genera tions the manua skis invoved in the manuature o stone toos greaty enrihed this utura heritage The distintion between habits and skis is ike that between tendeies and apaities habits are repetitive and imited in variation whie skis are more exibe and open ended. Like a apaities to aet skis must be adapted to a apaity to be aeted perussion bows ust aim or raks or ra tures; edges must be reated in a rigid stone apabe o hoding on to a shae; aking operatios must be appied to negraned stone; pro dution o arge toos must use roks o the right dmensions Both manua skis and the ompex proedures to whih they gave rise are ertainy oder than spoken anguage suggesting that the hand may hae taught the mouth to speak, that is that ordered series o manua operations may have ormed the bakground against wh ordered series o voaizations rst emerged. Variabiity in too and debris deposits is evidene not ony o the progressive dierentiation o skis but aso o utura enrihment wthin ommunities and even o ontat between dierent ommuni ties The atter oows rom the at that good quaity stone was not aways avaabe oay so that it had to be imported Reiabe meth ods to determne the provenane o dierent materias have been deveoped by arheoogists who have shown that a great amount o movement o highquaty int took pae in the midde paeoithi Roughy i a given materia has traveed ess than ve kometers it is assumed to be o oa origin i the distane is between ve and twenty kiometers it is said to be o regiona origin; and i the materia has been transported over dstanes greater than 30 kometers it is deemed exoti or imported Whie in the midde paeoithi the on gest movements seem to be o the order o a hundred kiometers by the upper paeothi they have inreased to 400 kiometers. hat is not ear is whether the ongdstane movement o int and other 129
PHLOOPHY AND MULATON
materials was effected through some form of trade or by another mechanism One possibility is that the movement began with a trav eng delegation carryig an offering that was then entrusted to a variety of intermediary communities none of which ever left its own territory4 Ts would inply that some form of cooperation between different comunities had become stable enough to support such a multistaged transport mechanism but, as was arged in the previous chapter, the stability of cooperation is hard to achieve within a single comunity let alone between several different ones So trade remais a possibility even if it ivolved only a very primitive form of barter with no clear conception of prots Evidence from contemporary hunterga therers indicates that witn communities the nly type of stable exchange is i the form of trans actions internal to insp groups mutual aid or gifts not immediately reciprocated The interactions between such groups within a tribe can become less personal but it is only in transactions between different tribes that social distance icreases enough for impersonal exchange to exist 5 This evidence can serve only as a rough approximation to thin about middle paleolithic commuties because today 's hunter gatherers already possess spoen language ut this is less of a mita tion than it may seem because the existence of a complex language has never been a prerequisite for trade even i relatively recent times in areas of itense commercial activity between communities of dif ferent linguistic bacgrounds simped languages called "trade pid gins, are often created in a relatively shot time and they disappear as the contact situation ceases to exist For simpler forms of barter a few gestures and unaticulated sounds could have sufced to carry out a transaction The main question then is not so much about lin guistics but about economics, that is, about situations in which choices must be made due to the scarcity of raw materials and the alternative uses to which the may be put Scarcity is a concept that must be handle carefuy because it is easy to introduce ethnocentric preju dices in an analysis Judging contemporary huntergatherers as suffer ing from res
MULTAGENT AND TONE AGE ECONOMC
The concept of scarcity, therefore, must always be used relative to specic resources The typological diversity of tools i middle paleo litc deposits and the differential intensity with which the tools were wored is evidence that their producers had an awareness of the lim ited availability of ghquality materials and were forced to mae choices between possible uses if the raw material was imported it was used to create a wider variety of tools and to create more complex forms (such as doubleedged tools) than if it was of local origin7 Incentives to trade were also created by environmental factors that limited the mobility of early humans constraining their access to geo logical sources of raw materials The middle paleolitc lasted so long and witnessed such dramatic climatic luctuations (from full temper ate to full glacial) that in the cold periods such incentives to trade must have been common Finally, there were plenty of distict com munities to provie the degree of social distance necessary for imper sonal exchange Although archeologists label similar deposits of stone tools recovered from different sites with names lie "Mousterian or "Aurignacian these terms should not be taen to refer to selfidentied ethc groups occupyig large territories Rather, the transition from the middle to the upper paleotc ivolved many small bands ihab itig favorable areas with large empty zones between them, silari ties in tool design being the result of cultural contact8 It is plausible that some of these contacts ivolved trade but this hypothesis is normally resisted by anthropologists because orthodox economic models mae certain assumptions-personal selfinterest optimal decisionmaing, costfree information-that are clearly iadequate in the context of contemporary huntergatherers let alone middle paleolithic humans9 On the other hand several simulations of barter have been created that dispense with these assumptions removing at least one source of objections To the extent that thse simulations are used to model huntergatherers the agents deployed in them should not be consid ered to be pers ons but bands or tribes since trade does not tae place within commuties but between them Simulated agents can be used to model entire communities because, lie the persons composing them, they are individual entities The term "idividual has come to be used as synonymous with "person but that is misleading As an 131
MULTAGENT AND TONE AGE ECONOMC
PHLOOPHY AND MULATON
denes an indierence curve 1 O Each agent may have several indiffer ence crves iffering in desirabity While desires or preferences cannot be compared qantitativelywe cannot jdge by exactly how mch more one object is desired over anotherwe can nevertheless arrange preferences in an order maling rigoros jdgments abot an object's greater or lesser desirabiity Using these qaitative jdgments to rank many inifference crves ields an indifference map captring the strctre of the sbjective possibility space for each agent he objective possibility space can aso be dened by different com binations of the two trade objects bt now the actal amonts of high-qaity lin and partially worked stone tools that th agents have available to them mst be taken into accont That is, nlike the sbjective space is which only desires are mapped in the objective space we mst inclde the nitia endowments of each of the traders since these dene what can actally be traded n a space of all com binations of trade objects the inial endowment of both agents are each a single point: one agent may have all the partially worked tools while the other has all the highqaty lint for example or they may each bring to the traing sitation a given amont of each object The other objective component is the "price at which the two objects are exchanged ince there is no money involved these prces are sply the amont of one object that is cstomarily exchanged by an amont of another object From the endowment point (the point representing the initial state of the interaction) price lines may be drawn showing the ratios of the two goods that cold be traded at a given price Finally becase both objective and sbjective spaces dene possible combinations of qantities of the same objects the endowment points and price ines of the former can be integrated wth the indifference maps of the latter to yield a single composite iagram caled an "Edgeworth Box a beatif geometrica constrc tion condensing in two ensions p to 1 2 ifferent concepts from economics Becase this composite diagram incldes sbjective and objective factors it allows s to dene the two graients that drive trag inter acions n one hand the sets of indifference crves arranged by increased desirability dene two sbjective gradients of satisfaction one for each agent An important feate of these graients is pairs of
adjective it can legitimately be sed to speak of individal orgaza tions, inividal cities, individal nation-states or any other entity that is singlar and historically niqe Another reason why comm ties can be modeled as agents is that once they have acqired eer gent properties of their own properties ike solidarity or inteal cltral homogeneity they can play an explanatory role indepen dently of the persons tht compose them Whe two conities interact for example the actal interactions may be performed by inividal persons bt the details of their personal interests and pref erences will be explanatorily irrelevant if any other two commnity members cold replace them and leave the otcome invariant. n other words the otcome of an inter-conity interaction is not altered by changes in the identity of the pesons that performed it then the dening featres of their personal identity can be left ot of an explanation Becase we are modeling individal commnities notions like that of a commnitys self-interest or of its cltral and material preferences are perfectly vad The persons composing these commnities can still be conceived as possessing their own self-interest bt one that is sborinated to that of the entire grop wit denite costs associated with the breaking of that bond Ths in what follows the term "agent will refer to an entire com mnity and the term "sbjective to the shar ed preferences character izing that commty Lets rst describe the objective and sbjective possibiity spaces involved in a trag interaction between two com mnities The sbjective space contains all the possible combinations of qantities of the objects to be traded the sple case of paleo ithic barter we can assme there wold be only two sch objects: high qaty lint and partially worked stone toos A twodimensiona possibiity space in which every point is a combination of a certain qantity of lint and a certain qantity of tools can be sed to express the preferences of each agent the sbjective preference for one of the objects in terms of a nmber of nits of the other object one agent may be willing to let go of one piece of high qaty lint in exchange for two partially worked tools for example while another may prefer a ifferent combination f an agent cannot choose between say two pieces of lint and three tools becase both qantities are eqally desirabe then that point in the possibiity space is a combination that makes no difference to the agent Joining several sch points by a line 132
. !
133
PHLOOPHY AND MULATON
indifference curves that just touch each other At these points of tangency the amount of satisfaction that each agent would get from an additional unit of the traded objct are identical and hence there is no further incentive to trade hese are the points at which the subjective gradients have been "dissipated On the other hand an objective gradient dening the potential gains from trade is captured by the diagram as a region containing combinations of the two goods that are more desirable than the initial endowments If we could exactly quantify desire we would be able to estabish that in a given transaction the loses of one agent were more than compensated by the gains of the other and that the trading interaction had achieved unambiguous gains. But since we caot quantify it the only unam biguous judgments we can make about interactions that improve satisfaction are outcomes involving gains for one gent and no losses for the other Using these ualitative judgments we can divide the diagram into two areas, one containing allocations that make at least one agent worse off and another containing those that do not and that therefore act as an incentive to trade Within the second area there is a singularity, called a "competitive equiibrium, dening the point at which all potential gains from trade have been exhausted, that is, the pont at which th objective gradient has disappeared and cannot longer drive interactions. Like any other diagram capturing the mechanismindependent component of an explanation the Edgeworth Box says nothing about the actual process that would drive two agents toward the singularity When modeing this dynamic process economists usually make two assumptions which are unjustied in the present context. First it is assumed that the agents have the calculational capacity and the nec essary nfoation t make optimal choices between alternative com binations of the two goods Second the interactions between the traders are assumed to be guided by an auctioneer who does not allow the transaction to poceed until it can take place at the singularity: the auctioneer roposes a price and the traders answer by giving the quantities of a good they ould be wiling to trade at that price; if these quantities leave gains o trade unrealized (i there are wasteul excesses or decits) then a second price is proposed quantities oered, and the process is iterated I this way the autioneer gropes its way toward the competitive equilibrium ad authorizes the trade only 134
MULTAGENT AND TONE AGE ECONOM C
when the singularity has been reached. his mechanism is early not realistic in most trading situations even in contemporary ones, and it does not show how prices can emerge endogenously from actual negotiations So from the point of view of modeing paleolithic trade we would have to get rid of the auctioneer and of agents possessing unrealistic calculational capacities (optmizing rationality). 12 Both of hese problematic assumptions have been addressed in multiagent computer simulations In one of these the assumption of optimzing rationaity was tackled directly and radically: the agents had no awareness of the trading situation no memory of past transac tions no notion of prots; and only enough inteligence to avoid losses. hese agents were appropriately called "zerointelligence trad ers Instead of an auctioneer the simulation used a double auction agents did not respond to exogenous prices by providing information about the quantities they were willing to trade but suppied prices directly: "bids if they wanted to buy or "asks if they wanted to sell. he bids and asks were matched continuously even if the singularity had not been reached In other words, unike the process directed by the auctioneer in which transactions take place ony at te com petitive equiibrium in a simulated double auction they can occur anywhere withn the area dening the gains from trade that is trans actions can take place away from equiibrium When zerointeligence traders were placed in an Edgeworth Box they were able to ove toward the singularity and come ose t it using only local informa tion to cimb the gradients of satisfaction. What the outcome of this simulation showed was that the crucial ingredient in the mechanism was not the calculational abiities of the traders but the structure of the interaction an interaction that came ose to the haggling that takes place in marketplaces or bazaars 1 Simar results were obtained in another smulation called "Sug arscape in which the totally disembodied zerointelligence traders were replaced by agents that had a body (that is, a metabolism) and that were situated in a space with a certain distribution of resources he behavior of both agents and resources was entirely determined by rules and was thereore not emergent but the use of rules is an advantage in this case because i the actions o the agents had been produced by a neural net it would be hard not to view those actions as mplying personal agency With rules on the other hand no 135
PHILOSOPHY AND S IMULATION
assumption about underlying mechanisms is made so agency can be attributed to larger units like entire h ouseholds or comunities 14 Nevertheless some care must be taken when interpreting the rules since these specied the range of vsion of agentsdened as the number of grid squares of the space they inhabited within �hich they could detect the presence of foodas well as their metabolism the amount of resources they had to consume to continue moving and prevent death I our case "vision and "metabolism must be taken to refer to the perceptual capacities and almentary needs of enire comunities: the consensus view reached after many personal obser vations and the subsistence activities needed to keep the entire group alive Rules also speced the behavor of resources n the simulation at each point in time each square in the two-imensional space had both a given quantity of sugar, a rate at which it replenished itself and a lt on the maximum amount that could exist there Agents used their vision to inspect all squares within their range identied those unoccupied squares with the most sugar and moved to the nearest one to gather the resource Through ts smple rule they acquired the emergent ability to climb resource gradients.15 Although the agents n this smulation had (non-emergent) prefer ences to move to the nearest square containing more sugar, for example, they did not yet have to make choices between alternative resources To allow for this to happen a second resource together with a second metabolism was intrduced: spice Now as the agents moved around they confronted choices between squares more or less rich in sugar and those with a greater or lesser abundance of spice Since the resources were dened by rules it would be relatively simple to change them to simulate choices between "high -quality lnt and "partially worked tools but this would not make a difference in a smulation designed to explore the mechanism of emergence of prices Let's therefore stick to the original formulation To determine an agents decisions its preferences for different combinations of quantities of sugar and spice }ad to be charted, that is each agent had to be assigned an indifference map Several factors entered into the determination of the indifference curves in Sugarscape: the agents metabolisms for sugar and spice; the current amounts of each good they currently pos sessed; and the current degree to which their metabolic needs were satised These factors determined the relative desirability of sugar l36
. MULTIAGENTS AND STONE AGE ECONOM ICS
and spice at any given moment for each of the agents as a function of their relative internal scarcity and their current endowments Finally the agents were given the (non-emergent) capacity to trade by adding a new rule The trade rule can be descrbed in Edgworth Box terms: if the indifference curves of a given pair of agents were tangent to each other there were no incentives to trade so the couple did not interact; if the curves were not tangent then a direction of trade was established one agent may have needed sugar and the other spice so sugar moved in one direction and spice in the other. Then a ratio of sugar to spice was established (via a bargaining rule) and the quanti ties of each good that would be traded at that price were checked to see i no agent was made worse off, that is, if the transaction was within the area dening the gains from trade in which case the trans action was carried out16 To see whether after many such tradg interactions an entire population could collectively reach the point of competiive equilib rium 200 agents were ueashed in a space with a certain distribution of both sugar and spice At the start of the smulation there was sig nicant variation in prices across the population but as it proceeded there was a tendency to converge toward the equlibrum price, never quite reaching the singularity but cong quite close and staying within a small neighborhood around it I other words, prices emerged spontaneously away from equilibrium without the need for an auctioneerY The mechanism of emergence depended on the fact that as agents repeately interacted their internal valuations for the two resources became more and more alike; and the more similar their valuations became the closer they all got to a single price On the other hand, some degree of price dispersion survived until the end of the simulation a degree of dispersion that was increased by any factor that interfered with the intensity of the interaction: reducing the range of vision of the population (since this made agents less able to locate potential traders); allowing agents to die even if they met their metabolic needs (shorter lifetmes reduced number of interactions); and allowing agent preferences to change in res ponse to noneconomic factors ( since this increased the time needed for internal valuations to become smilar) . 18 In the case of paleolithic trade a large degree of price dispersion is sometng to be expected so the outcome of the simulation supports the thesis that the gains from trade could have l37
PHLOOPHY AND MULATON
been realized in that context without making any of the usual assump tions that anthropologists nd so objectionable. Let's draw some conclusions from what has een argued up to this point. Multiagent simulations can guide us in the search for solutions to archeological problems as dened by the available eidence: depos its of stone tools and the debris left behid during their production I some cases the problem is to account for a particular distribution of these deposits, a problem to be discussed below, but there is also the question of explaning how rare imported materials got to the place where they were found Because we do not know the method of transport that was used multiagent simulations can help ne-tune our intuitions about those unknown transport mechanisms While the currently preferred explanation is some kind of ceremonial trans port, an account that draws its plausibity from the fact that instances of it have een observed in some contemporary huntergatherers, there is also the simpler mechism of trade between communities. This option has been underestimated by archeologists partly because of the presuppositions bult into economic models But if the latter are taken to address only the mechanismindependent component of explation, that is, the singularities structuring the possiility sace of a tradig situation, d not the actual mechanisms involved in the making of decisions or in price negotiations then the assumptions become less problematic. Athough simulations also operate at a rela tively mechanism-independent level the capacity of computational process to track historical paths toward a singularity, and the added advantage of using embodied and spatially situated agents, makes them ideal to explore potential cddates for the transport mecha nism that accounts for the available evidence To go beyond this relatively humble role and tackle archeological problems e explaing the acual distribution and composition of deposits of stone artifacts in a specic geographcal area, we need a more expicit way of modeling the different elements of a multiagent simuation th resource landscape, the behavioral rues, and the tem poral scale of the subsistence activities This is a highy interdiscipin ary task but archeology itsef has ong reied on other eds (chemistry, geology, ecology, ethnology) to reconstruct subsistence activities and settlement patterns from stone tools, fosslized polen, anima bones, and hearth remains. Smuated reconstructions of the subsistence 138
MULTAGENT AND TONE AGE ECONOMC
economy of huntergatherers ary in the degree of detail of both behaioral patterns and resource gradients I the earliest siulations the type of plants and animal prey assumed to be the target of forag ing actiities were guessed either from contemporary hunter-gather ers in the same region or from the fossiized remains found in a cave or another site Then the con�tions affecting the relative abundance of those resources today (rainfall for plants, population dynamics for animals) were modeled to create statistical series of good, bad, and average years to be used in the simulations Athough the resource distributions generated this way were clearly more realistic than those of the Sugarscape simulations just discussed, extrapolating from prs ent conditions can be problematic if the landscape is not teporally stable, if target species are now xtinct, or if ecological diversity has been greatly reduced 19 To avoid these potential sources of error data d techniques fro ultiple disciplines must be used to reconstruct ancient landscapes: using the information stored in ice sheets, tree rings, sediment, and rocks paleoclimatology can help to recreate the temperature d pre cipitation condtions of a given time period; geomorphology can e used to understand the history and dynamics of land forms, and to recreate the effect of rivers, volcanoes, and glaciers on the distriution and quality of soils; ecology can furnish information about the succes sion dynamics of different plants and this, together wth the output from palynology (the study of pollen, spores, and particulate organic matter), can guide the reconstruction of cient distributions of plants of a given species20 Once the spatial component of resource dstribu tions has been worked out we need to add a temporal dmension to the simulations This is smplied y the fac that the subsistence activities of huntergatherers tend to have a cyclical structure I soe cases the plants that are the target of foraging acivities ripe in differ ent seasons and this gives the activities a yearly cycle, sometimes accompanied by settlement patterns that display a simiar periodicity In other cases a much shorter period may be used such as a daly cycle divided into dual and nocual activities21 This periodic repetition of activities can be used to give tempora structure to a simuation and to correlate its output with time series for temperature and pre cipitation changes, the rise and fall of available groundwater, and the changes in sol quaity created by the deposition of sediments 139
PHILOSOPHY AND SMULATION
The next step is to make more realistic the behavioral and etabolic rules that capture the decision-makng and consumption activities of agents. Etnological studies of contemporary huntergatherers can be a valuable source o f information in this regard but we must be careful not to reify particular cases into a timeless set of properties In the 1960s, for example, the consensus view of contemporary hunter gatherers characterized them as living in groups of 25 to 50 members, with no material poss essio ns to constrain their high mobility, a sexual division of labor lexible membership, and very egalitarian authority structure But a decade later other tribes were found that id not match this list of properties, being sedentary part of the year, living in larger groups, and storing some possessions These two cases can be used to dene the two extremes of a continuous possibility space and the behavioral rules for the sulations can be derived by making educated guesses about the likely position of paleolithic hunter gathrers in this continuum Several sulations may be performed with rules derived from ifferent parts of the continuum to check the sensitivity of the results to ifferen t assumptions Finally to be able to use these sulations to make specic predictions about the istribu tion of deposits of stone artifacts the latter must be available in a well preserved state I general the further back in te we go the more isturbed the archeological record is so most detiled simulations have focused on more recent periods, such as the mesolithic the tran sitional period between the upper paleolithic and the already agricul tural neolithic. This period differs from the ones iscussed so far in one crucial respect the availability of language a change relected in the sulations by the need to incorporate the low of semantic infor mation among the agents In one smulation called "Magical the goal was to reproduce the istribution ad composition of stone artifacts on an island of Western Scotland where mesolithic hunter-gatherers once foraged for hazelnuts I this sulation resource graients were modeled in detail using te paleoenvironmental approach just discussed includ ing competition among typical temperate forest trees (oak birch) with hazel Yearly monthly and daily te scales were included The activiies on a yearly te scale were determned by the life cycle of the trees themselves: one season with a hazelnut harvest and one without This in turn was assmed to affect ·human activities like 140
MULTIAGENTS AND STONE AGE ECONOMIS
gathering on a base camp followed by dispersing around a given area of the island The monthly time scale was determined by decisions reached by consensus about the location of the base camp: whether or not to move it to a new place with better chances of increasing the productivity of foragng activties Finally, on a daily basis agents foraged or explored during the day and got together at night to share the results of their activitie3 Unike the previously icussed simula tions in which the agents were entire communties here the aents were taken to be inividual persons so the nightly reports ha to be modeled explicitly The solution to this problem was very ingenous Since the activities o f the agents had to take place in a simlated space with the same characteristics as the original island a standard format for the storage o f spatial data called a Geographic nformation System (or GS) was used . A GS is basically a map on which multiple layers are superimposed each containng nformation about a particular landscape property (soil composition distributions of plant species water availability) . n Magical these maps were given an aditional use: to represent the knowledge that the agents acquired about the landscape n other words the cognitive maps that the agents formed of resource istributions maps that would have been created by real huntergatherers through a combination of episoic and semantic memory were each represented by a GIS map That way when the agents gathered at night to report their daily ndings the linguistic exchange among them could be modeled by merging all the ndivid ual GS maps to form a composite one that was then redistributed among all the members of the group.4 Ulike the agents in Sugarscape the use of smulated vsual infor mation did not form the basis on which movement decisions were made the agents of Magcal using instead their inteal GIS maps for that purpose But as in Sugarscape Magical's agents were heterge neous in this capacity: while in the former they possessed iffrent ranges of vision in the latter they analyzed ifferently sized portions of the map around their current position before moving to a neigh boring location. Agents who made decisions based on a smaller por tion of the map were being " conservative in the sense that they were tryng to reduce risk and prove their shortterm gains; those usng larger portions took more risk by moving to locations with unknown hazelnut abundance but could benet from longer term gains if they 141
PHLOOHY AND MULATON
discovered new sources of food Finay, the stone artifacts were divided into different categories (waste from production processes, small tools, larger scrapers and probabilities that the agents would dscard them were assigned according to their position relative to the base camp and the stage of the yearly season.5 So far the main results of usng Magical have been to reduce the likelhood of some previous scenarios proposed to account for the existing archeological data In particular, the hypothesis that the only determinant of and use was the gathering of hazelnuts, an amittedly imporant resource, has lost some credbility. Many runs of the simulation were performed each run representing about ten years of foraging actvity, using different landing sites around the island as well as different istributions of risk-taking behavior. Yet, none of the outcomes managed to repro duce the existing distribution of deposits of stone artifacts.6 though the simulation's results were negative its contributon was clearly positve: it showed that other spatial constraints must have been in operation, such as the avalability of highquaity int near the coast that may have been an incentive to create speciaist base camps used for the production of tools. I other words, the simulation's results pointed to useful alternatve hypotheses Smilarly negative but highly informative results have been obtained in a Sugarscape simulation n which resource distributions were reconstructed using paleoenvironmental models The simula tion tackled an even more recent period the Neolithic, a period in which agriculture was already practced and the nomaic ifestyle of huntergatherers had been abandoned While stone tools continued to be used and ther increased sophistication and functional differ en tiation left behind a richer archeological record, the Neolithic intro duce a more sensitve medium: pottery The plasticty of clay gives pottery a greater range of styistc varation and this together with the fact that it was typcay produced locally supplies us with evidence about regional groupings and about the interactions between neigh boring comunities.7 nother source of information from this period is settlement patterns Settlements leave mineralized remains arranged in separate strata: some layers may show that for a while most of the population was concentrated n one large village, for example while other layers may reveal that at other times the settlers dispersed and ived in any small villages. For archeologists trying to explain these 142
MULTAGENT AND TONE AGE ECONOMC
variations the question is whether they can al be attributed to changes in resource gradients (soil productivity availabiity of water or whether cultural changes were also involved Multiagent simulations can be used to tackle this question for specic historical cases. One particular case is the nasazi culture that inhabited the Long House Vley in northeaste rzona rst as huntergatherers (before 1800 B) then as agriculturalists surviving on a economy based on the cultiva tion of maize until they abandoned the region in A 1300 for reasons stil unknown. Unlike the static resource landscapes of the Sugarscape simulations discussed above those of the Long House Valley had to be modeled dynamically because the productivity of maize crops the main gradi ent driving the process, undergoes change at several time scales: it is inluenced by relatively shortterm variations in the moisture and temperature of the atmosphere as well as by longer term variations in the dynaics of rivers, affecting the availabiity of groundwater and the creation and destruction of new fertile soil The time scale for the simulation was determined by the agrcultural cycle that is the simulation proceeded in yearly periods. The agents in this simulation were not individual persons but individual households, composed of ve members each whose metabolic requirements were calculated using anthropological studies of other subsistence agriculturalists each member of a family was assumed to consume 160 kilograms of maize per year storing hatever was left for up to two years. Families were given behavioral rules the most important of which was a reset tlement rule: at the end of each year each family calculated on the basis of the current harvest and their stored grain the chances that there would be enough food nex year; if it decided that there would not be enough it made a decision to relocate constrained only by the availabilty of unoccupied farm land and proximity to drinking water. If the wrong decision was made and not enough food was produced the family disappeared from the simulation.8 Each run of the simula tion generated as the emergent cumulative effect of the movement decisions by many families year after year a possible settlement hs tory between the years A 800 and 1300 To capture statistical trends many such possible histories were generated starting from differen initial condtions and varying certain parameters. The output was both a geographical representation of setlement locations hat could 143
PHLOOPHY AND MULATON
be compared with maps crea ted by archeologists, as w e as statistical series on population growth and aggregation pattes that could be checked against those in the archeological record. The comparisons, however were not intended to be exat and quantitative: it would be unrealistic to expect agreement on overa population sizes number of households, or size of settlements In fact, the simulation consistently generated quantities that were too large packing many more househods in a sigle residential zone· than is consistent wth the evidence But the qualtative agreement was striking: for popuaton growth for instance, the qualtative features of the statistical series the periods of growth and decline as well as those in which numbers remained constant, were mostly repro duced29 Simiarly the responses of the simulated agents to environ metal changes-to aggregate in a large viage to perform intensive agriculture when groundwater was scarce and to disperse into many small ones when plentiful groundwater lowed cultivation in many places-generated settlement pattes qualatively similar to those in the data The main discrepancy between the simulation and the archeological record was itsef highly informative Around the year 2 5 0 long and shortter variations in environmental factors coin cided to produce a subsistence crisis i the Long House Valley and by the year 300 the Anasazi had abandoned the are The simulation failed to reproduce this event suggesting that the decision to emigrate was not necessary linked to gradient degradation fact the simu lation indicated tht th available resources could have supported a smaler population dispersed into smaer communities That the real Anasazi did not take advantage of this opportunity suggests that the cultual and social factors that were not included in the simulation must have played an imporant role in their decision D particular when deang with agents that already possess language we must assume not only that they can exchange factual information but aso that they an frm and share magical belefs that can play an equally important role in their own explanations about resource availability and hence on their dcisions to stay or to emigrate abandoing the area altogether We wil return in Chapte 1 to the role that magical beliefs play i the determination and explanation of human behavior, and to the way in wch the effect of those belefs can be icorporated ito 144
MULTAGENT AND TONE AGE ECONOMC
multiagent simulations But before we tackle that problem we must confront the one that ultimately gives rise to it: the emergence of language Although huntergatherr communities during the hn dreds of thousands of yeas they spent skillfully working stone may have ready developed simpler forms of communcation, based per haps on hand and facial gestures the gradua development of articu lated language marked a threshold beyond wch an entirely new world became possible The crucial property of language is that from a nite set of words an inite number of sentences can be produced Thus what needs to be expained is how ts inite combinatorial productivity could have emerged starting with a few monolitc symboc artifacts incapable of entering into complex parttowhole relations It is to ts task that we now tu
145
MULTAGENT AND PRM TVE LANGUAGE
CHAPER EN Muiagens and Pimiive Language
When humans rst began to shape lows of air with ther tongues and palates th aousti matter they reated introdued yet another layer of omplexity into the world The prodution of sound reled on the same mehanism used by other animals-a soure of osillating air shaped by a voal avity ating as a lter-but in humans ths meha nism underwent several iportant evolutionary hanges The soure the larynx and its voal hords lowered its position relative to the bone at the base of the skull extending the range of vowels that ould be produed even at the ost of losing our ability to swallow and breathe at the same time The lters in tu onstituted by all the different organs that an hange the form of the resonating avity (tongue lips teeth) beame apable of manipulating pressure gradi ents with ever greater preision resonanes in the voal avity on entrated the energy of the airlow around partiular frequenies and these onentrations (alled "formants) beame the building bloks that n different ombinations would eventually yield all the different sounds of language Sine evolution has no foresight neither of these evolutionary hanges was amed at making language possible The desent of the larynx may have ourred as it has in other speies due to sexual seetion a lower larynx an produe sounds that mimi the effet of a larger body size an effet that anials exploit during territorial disputes And the greater ontrol of the organs produing formants was most likely a byprodut of the inreased size of the brain Life in the savana had long ago hanged dietary patterns adding new nutrient soures (tubers meat) and new tehnques for 146
food preparation (the use of tools to rush tubers and re to ook meat)inreasing the amount of energy onsumed nd making its digestion more efient The extra metaboli energy available in turn was invested in a longer embryologial development wth a more elaborate folding of neuronal layers paking more brain matter into a given volume2 But if there was nothing uniquely human about the biologial evo lution of the sound prodution mehanism the aousti matter itself was unique beause as soon as it emerged it beame part of an evlu tion that was separate from genetis Prelinguisti hutergaterers possessed learned habits and skls that were transmitted aross generations and that ould therefore play the role of repliators in a ultural evolutionary proess a proess ino whih the newly born linguisti sounds ould be instantly grafted The details of what hap pened to those early sounds may never be known but we must assume that linguisti evolution began with monolithi symboli artifats That is we must assume that the rst "words were not made out of parts that ould be reombined to form other words and tha the words themselves were not parts that ould enter into larger sentenes. The only ommuniative funtion that those monolithi symboli artifats ould ave had is that of labels that eferred to diretly experiened objets and ations Smple onatenations of those artifats ould have later on been used to express the observed onstant onjuntion of linear auses and their effets an observation that had been made routine by the praties assoiated with stone tool prodution The next evolutionary step going from a set of monolithi sounds to one of words that ould be ombined into an innite set of possible sen tenes is muh more difult to explain We need not only to aount for the apaity to ombine a set of words like "the prey and "runs away into a sentene like "The prey runs away. We also need to aount for the abilty to feed this simple sentene bak as an input together with another set of words like "the hunters and "hase to produe a omplex sentene like "The prey the hunters hase runs away In other words we need to explain the unique human ability to use reurson T get a sense of the problem that had to be solved by ultural evo lution it will be useful to onsider the spae of all possible languages. The struture of this spae an be explored using automata theory 147
'T PHLOOPHY AND MULATON
a theory in which memory, recursion and language complexity are closely intertwned Automata can be classied into a series of ncreased computational capacity determined by their degree of access to memory resources: nite state automata, pushdown auto mata, linearbounded automata and Tring machines This series in turn, can be related to a series of languages of icreased complexity Fiite state automata possess no memory and can handle only lan guages without recursion regular languages. The next type has access only to the last item stored in memory (the others being "pushed down a stack) but this gives it the ability to handle simple recursion and master context-free languages. I particular pushdown automata have the ability to keep track not only of embedded clauses ("the hunters chase) but to match verbs and nouns separated in a sentence that is, to match "the prey to "runs away The next level of compu tational capacity involves a more powerful memory in the form of an inite tape on which symbols can be stored but access only to a small portion of it This is enough, however, for linear-bounded automata to handle the stored symbols together with their iediate context and master context-sensitive languages. Examples of the more complex recursive constructions that this allows are rare in natural languages but do exist as illustrated by crossdependencies in which verbs are not matched to nouns in the order they appear Finally the most powerful automata Tring machines have unlimted access to the memory tape, can take as much context into account as needed can handle any form of recursion and master phrase-structure languages3 A tpical desktop computer lies somewhere between a linearbounded automaton and a Tring machine approximating the latter ever more closely as memory gets cheaper and more abundant As mentioned in Chapter automata theory is important philo sophically because it provides for capaities what state space and its attractors do for tendencies a clear and rigorous way of exploring the structure of a possibility space The singularities in this case are constituted by the four automata with nite state automata and Tr ing machines constitutig a minimum and a maxum of computa tional capacity respectively The term "langage in automata theory efers not oly to the approximately 6,000 currently existig natural languages but also to l the programng languages used to create 48
MLTAGENT AND PRM TVE LANGUAGE
software for the different automata as well as many languages that are possible but have never been used. Lke many of the combnato rial possibility spaces that have been discussed so far this space is in nite in size but for the puose of exploring the evolution of natural languages we can concentrate on the region contag context sensitive languages This implies the assumptio that the part of the human brain that handles language must have the computational capacity of a linear-bounded automaton Using neural nets to explore how this capacity is implemented i our brains would seem at rst to be an ipossible task because the kind of neural net that can process the temporal sequences of sounds that make up human speech, recur rent neural nets has only the computational capacity of a nite state automaton. We will retu to this important question at the end of this chapter. Having dened the nal state of the evolutionary process let's go back to its beginnng the emergence of reference Having the ability to refer implies the creation of a conventional association between monolithic symbolic artifacts and real entities mediated by meanings But what could "meanngs be in a world of pre-linguistic hunter gatherers? In Chapter 6 it was argued that in both animals and early humans "meanings are the prototpes that neural nets extract from sensory experience nonlinguistic categories or generalizations stored as distributed patterns of activation These distributed representations are associated to their referents directly because they are extracted from repeated encounters with many particular instances of a class of referents In addition to giving us an account of the association between meanings and referents neural nets can also explain the association between meanings and sounds because the process of training that generates the distributed representations is often per formed with the goal of producing a particular motor output If we think of ths motor output as involving the larynx as a source and the vocal cavity as a lter then the nal product is a vocalization that is automatically associated with the distributed representation Beig in possession of an explanation for the k between meanings and referents and meanings and sounds all that is left to explain is the emergence of the link between sounds and refeents This is harder than it seems because prelinguistic huter-gatherers could not rea minds, that is, they did not have access to each other'S distributed 49
PHLOOPHY AND MULATON
reresentations. Ths mlies that early humans had no choice but to use ublicly available information to converge on a shared set of soundreferent associations. To elore the emergence of ths shared set we need to deloy a oulation of neural nets situated in a sace containing simle refer ents and to give them a minimum of embodiment to ensure that the agents have no access to each other's minds I fact neural nets are not strictly necessary: we can use any other imlementation of rototyeetraction, using video cameras and standrd image rocessing software, for eamle, as is done in a simulation called semiotic dynamics4 I ths simulation airs of agent take tus laying the roles of seaker and lstenr The agents inhabit a simle world of geometric gures (the otential referents) and ossess three smle caacities: they can erceive and categorize the geometrical gures they can oint at a gure, and they can observe in what drection other agents are ointing They are also able to roduce arbitrary symboic artifacs (random syllable combinations) to use as "words artifacts that can either be invented or be already in use in the com munity. Thus using semiotic dynamics to model the emergence of a shared set of sound-referent associations mlies only the assum tions that relinguistic huntergatherers could use gestures to achieve shared attenton and had the wlingness to articiate in ritualized tasks, such as taking tus to vocaize semiotic dynamic simulation unfolds in the followng way gents are aired at rando to attemt to communicate The agent whose tu it is to seak selects a geometric object as the toic of conversa tion, checks its memory to see what revious symbolic artifacts have been assocated wth it, and icks one; since at the beginnng of the simulation no artifact is available it invents a new one The agent layng the role of istener then checks its own memory to see if a revious association between that artifact and some referent has been made and uses tat as its rst guess I the very rst encounter it can not nd anything so t invents its own sound Thus, the starting oint for the smulation is a situation in which each agent has its own idiosyncratic vocabulary and the isteners always fail to guess the seakers referential ntentions But seakers can also hel listeners by onting to the inteded referent and this can sometimes result in a
MULTAGENT AND PR MTVE LNGUAGE
correct guess after several trials Ths random guessi ng cannot by itself converge to a single shared vocabulary for all the otential referents but convergence can be acheved if we give agents the caacity to kee score of what airs of symbolic artifacts and referents have led to communicative success in the ast.5 When a seaker searches its memory for a revious artifact associated with its intended referent for eamle, it may discover that there are several choices with differ ent scores. The history of the different alteative associations, some of which have led in the ast to more correct guessings is releced in those scores so all the seaker has to do is ik the one with the highest score The listener too searches among cometing associations for the least ambiguous one (the one with a single referent) and if there are several it icks the one with the highest score as its rst guess Iterating a rocess in which agents are attemting to guess each others referential intentions and heling each other reair break downs in communication eventualy leads o a stable reertoire of shared soundreferent associations.6 The outcome of this simulation is valuable but limited: the symboic artifacts reresent only one class of linguistic entities, labels or roer names whle the set of meanngs is restricted to objects that have actually been eerienced, not to meanings that can be dened com ositionally by other words. Nevertheless, within its limited goals the simulation is a success because it shows how the rocess of lin guistic evolution could have got stared. To simulate the net ste, the concatenation of labels into simle monolithic sentences, the agents must be situated in a world that demands from them engaging in causal reasoning In the revious chater we saw that relinguistic hunter-gatherers were caable of masterng comle sequences of manual oerations as art of the roduction of stone tools This task resented them with the oortunity to learn about linear causa relations by the direct observation of the constant conjunction of a cause and its effect. Eressing this constant conjunction symboicaly would demand no more than concatenating the labels for the actions or events acting as cause and effect. The resultng symbolic artifact would not be a real sentence since its meaning would not result from the comosition of he meanings of the comonent labels, but it would nevertheless be caable of bein true or false. On the other 151
PHLOOHY AND MULATON
hand if shared personal observations were sufcient to lea about causal relations what would have been the motivation for expressng them symbolically? To answer ts question the next simulation poses the following challenge what if there were causal regularities that could not be easily observed by persons but only by entire communities? Would the production of artifacts expressing guesses about OOccurrences of events improve the chances that over several generations all members of the community would make the corrct association? The situation confronting the simulated agents was inspired by a real example from contemporary huntergatherers When the ocean tide is ow a gradent of concentrated protein (shellsh) forms at the beach. If the agents lived near the beach they could drectly observe low tides and know when to collect the shellsh But if they lived far away thy would have to use anothr observable event that corre lated with a low de, such as the phase of the moon As it happens when the moon is new or full the tide reaches both is hghest and its lowest ponts, maximizing the chances of ndng shellsh So a com munity could benet if collective observations of the conjunction of tide states and moon phases could be encapsulated into a monolithic sentence7 The members of this community would not have to pos sess any knowledge o f real mechanisms that is th ey need not know that when the moon is new or full its gravitational force is in phase with that of the sun and its effect on the state of the tide is greatest n the other hand the results of Chapter allow s to assume that prelin guistic huntergatherers could perform perceptual analyses of situa tions using riptlike structures, assign nteracting objects to the roles of causal agent and causal patient and remember those experiences using episodc memory While the symbolic artifacts must have the capacity to be true or false, a capacty that propositions have, they need not necessarily have a propositional form snce the simulated agents have to solve a single causal problem n their lives a simple numerical vocabulary specialized to refer to tide states and moon phases suce Never theless it be useful for us to imagne the symbolic artifacts as monolithic sentences like this FullMoonCausesLowTide 52
MULTAGENT AND PRMTVE LANGUAGE
I this precursor of a real sentenc the absence of spaces between words indcates that the labels for "full moon and "low tide cannot be used compositionally to generate other sentences while the verb "causes simply marks the fact that the cooccurrence of the two labels stands for the cooccurrence of the events playng the roles of cause and effect At any rate, because in this simulation the agents are entirely dsemboded and are not situated in a space in which the relevant events could take place; the terms "label and " event refer to activation patterns in neural nets suppled to the agents by the experimenter Each agent consists of three neural nets two used to handle the association between labels and the events to which they refer and one used to create an association between the events them selves The design of the two neural nets that handle the linguistic part of the task can be very simple beaus the emergence of a lexicon to refer to events has already been explained by simulations lke semiotic dynamics. More specically the neural nets that handle labels and their referents do not need to have hidden layers capable of extracting prototypes they need to do is produce a label as an output if given an event as an input and vice versa produce an event if given a label. The third neural net is a regular multilayer perceptron that takes events as its input and produces a causal association as its output At any one time in this simulation a comunity of 20 agents exists each spendng its life learning aout the lnk between moon phases and tide states from "direct eerience and from the mono lithic sentences produced by pri or generations At the end of its lif e an agent des leaving behnd a monolitc sentence expressing its best guess about the causal link and a new untrained agent to replace it8 Because the agents in this simulation are dsembodied the responsi bity for recreatng the desired situation falls on he training process itself if the multilayer perceptron was trained using only information about events then culture would not matter and vice versa if the only traning examples were past symbolic artifacts then the events and their objective causal connections would not make a dfference. So the set of training examples needs to nd a good balance between the two n one hand, it must make it hard for any agent to learn the association between events from drect experience mimicking the effects of cloudy weather, dstance and other interferences that n real life would preclude personal observatons of the constant 53
PHLOOPHY AND MULATON
conjunction between u or new moon and extreme tide variation. On the other hand, it must reect the accumulated knowedge o past generations so that a bias toward true artiacts must be a dded to the traning, a bias capturing the act that in a rea commuity the reative success o agents from previous generations woud inuence the choice o guesses used as a starting point by a new generation. With this bias acting as a seection pressure cutura evoution arrives at a situation in which the entire communty can reiaby ea about te reation between moon phases and tide states.9 The outcomes o this and the previous suation suggest that monolithic smbolic artiacts coud have emerged n a community o hunter-gatherers to perorm spe reerentia nctions and cpture simpe causa reations This is not in itse surprisng because as we argued n Chapter chmpanzees possess the abiity to assocate abes (coored shapes) and their reerents and to chain severa abes together to express simpe causa reations, athough they need to be traned by humans to deveop this abity and have never been observed exercisng it in their natura habitat But as was aready pointed out this consttutes te part o the probem that is relativey easy to sove The hard part is to suate a process that can break down a set o monolithic sentences into recursivey recombinabe components, a process reerred to as gammaticalization. Beore dis cussing a suation that tackles this probem it wi be useu to describe the traditiona account. Instead o expaining the uman capacity to use and understand recursion by postul atng an evoution ary process based on replicatng acoustic matter, as we are doing ere, the tradtiona expanation assumes tat this capacity was produced by bioogica evoution This account starts by creating a orma mode that can produce sentences throug a recursive process. The mode is based on rewrit ing rues, operators that tae as their input strngs o inscriptions, rewrte them, and produce a new string as their output. The strngs are not random cains o sounds (or etters) but sequences aready assigned to grammatica categories, ie noun, verb, sentence, and so on Both the rewriting rues and the grammatica categories are assumed to be innate constituting a universa grammar common to a anguages The innate hypothesis is deended on the grounds that when chidren earn a anguage they acquire a capacty to produce an IS4
MULTAGENT AND P RMTVE LANGUAGE
innite number o sentences even though tey are ony exposed to the very smal number o sentences actuay produced by their parents In other words, the poverty o stuation characterizing the earng situation is taken to be evidence that chldren do not earn a grammar by makng step-bystep nerences rom adut sentences but rather that the atter act as mere triggers or rues and categories inherited bioogicay To put this dierenty, i as automata theory shows the space o possibe anguages is innite in size how can chidren earn a anguage given the poverty o the stimuation they receive rom their parents? It woud seem tat we woud need to postuate a way o constraing this space to mae anguage acquisi tion not just possibe but highy probabe A unversa grammar is precsey such a constraint: it eminates rom the space o possibili ties that the chid must conront a except the humany earnab� anguages. The traditiona expanation runs into diculies wen tryng to speciy te evoutionary process that ed to the deveopment o nnate rues and grammatica categories This is not just a question o expain ing the surviva vaue o recursion given tat inguistic communica tion can be acheved without it. is aso the probem o showing that genetic variation or the neura substrate o categories e "noun and "verb existed The atative account, on the other hand, must not ony expain how anguage evoved externay driven by its capac ity to encode propositiona normation about cause and eect rea tions' but in addition it must expain how i was ater intealized2 The second requirement is necessary because we need to account or the existence o brain organs specazed or anguage, organs that can act as neritabe constraints on the space o possibe anguages. The two aspects o this aternative account have been successuly investigated using computer suations: grammaticaization as been shown to emerge through cutura evoution in groups o inter acting agents in which one generation pays the roe o teacher and the other one o student; and the mechaism o internalization, the so-caed Baldwin eect has been shown to be viabe in simuations o evoving neura nets The rst simuation does not have to produce te compex type o recursion that characterzes contextsensitive languages but ony the sper type o contextree languages. We can agne that once such a spe recursive anguage emerged in ISS
MUAGENT AND PRMTVE LANGUAGE
PHLOOPHY AND MULATON
ancient huntergatherer comunities the Badwin eect began to operate and that the emainder o the journey though the space o possibe anguages took pace by evoutionary search based on both genetic and inguistic repicators The agents used in the simuation o gr aticaiation ae even moe disembodied than thoe o the simuation just discussed. pa ticua, the agents have access to the intended meanings o monoithic atiacts Since agents are supposed to earn rom obseving each othe 's behavior not by reag each others inds this is a majo imitation addition, the simuation does not use neura nets but stays within the word o automata, pushdown automata in the case o contextree anguages, and the notation they understand. Thus the meaning o a monothic sentence ike: FuMoonCausesLowTde is a proposition that is expressed in automaton notation as: Causes (Fu Moon, Low de) This choice is justied on the basis that using symboic epresenta tions aciitates tracking the progressive evoution o compositionaty and ecursion: checking that a new syntactica eature has emerged is smpy a matter o ooking at the current repertoire o propositions in the popuation o agents. 3 A crucia component o the process though which compositionaty emeges is the abiity o the agents to genera ie rom experience In a simuation using neura nets this woud be impemented by the extaction o prototypes rom sesoy stimua tion In the present case on the other hand, it is easie to impement this capacity through an inductive agorithm speciaied to opeate in the word o pushdown automata Finay the agents invoved must be abe to interact paying the roe o teacher or studnt an interac tion that in th case o preinguistic hunter-gatherers woud have taken ace between parents and osping, or between aduts beong ing to an extened amiy group and kindred chidren During this interaction the students have direct access to the meaning o decara tive sentences that is, the teacher pesents them with pairs o popo sitions and their monoithic symboic expressions To justiy the use o mind reading we woud have to assume that in another version o 156
this simuation a version in which agents woud be embodied and situated in a space containing the reerents o the abes, the proposi tions woud b e r econstructions o intended meanings that the agents woud arrive at as they observed each other 's interactions with those reerents We wi retun to the question o how to eiminate mind reading beow. The word in which the agents ive is xed and simpe I a typica run or exampe the simuation incudes as reeents or the abes ve objects and ve actions Give that each proposition has two arguments such as "u moon and "ow tide, and that these must be dierent om each other this yieds 1 00 possibe propositions Agents spend ha their ie as students o teachers beonging to the previous geneation the other ha as teachers to a new geneation To incude the eect o the poverty o stimuation each agent teaches ony ha o the possibe propositions in its ietime 4 When the simu ation starts a symboic artiacts are monoithic and teaching invoves making the student memorie a ong st o associations From the act that these ssociations ae taught as customay, however, students can ine the simpe rue that a proposition i
PHOOPHY AN MULATON
But then the generazing abities o the students come into action: the common part o these two setences is detected and the two propositions are repaced b a more genera one: Cases (Fu Moon, X) together with an assignment o the category "no un to possibe vaues o the variabe X. When the student that made the generazation becomes a teacher it does not teach the two origna associations but the new one S/auses (Fu Moo, NX) � FuMoonCausesX At this point the povert o stimuation characterizing the earning situation starts to work in avor o the more genera orms because smboic ariacts that are used more requent tend to be overrepre sented in the impoverished sampe used or teaching In other words, a setence ke "FuoonCausesX is used twice as oten as the two smboic artiacts it repaces since it can eress the meang o both, so the probabit that it wi be seected as a teaching exampe is twice as high Iroica, the same povert o the stimuus invoked as a jus tication or a uiversa grammar works here to eiminate the need or nnate rues b acting as a botteneck avoring the transmission o more genera orms to uture generations.1 5 But it aso eads to the start o the segmentation o the monotic artiacts since the more genera orms are aso the ones that spa compositionait. I other words, the botteneck acts as a sorting device, avoring some repca tors at the expense o others, whe aso creating the contions or the urther segmentation o monotic sentences I act, the simuations scussed in this chapter, as we as other reated ones, operate b using a botteneck as a seection mechaism16 In semiotic dnamics, or exampe, the origina purat o iosncratic vocabu aries is sow narrowed down to a singe shared exicon b the tendenc o agnts to choose smboc artiacts accorng to how re erentia successu the have been in the past And simiar or the simuation o abe concatenation as a means to express causait: ever generation has a tendenc to start with those mooithic sen tences that were more predictive o the rea ink between moon phase and tide state, and this acts a botteneck acitating the passage o some but not other artiacts to the next generation 158
MULTAGENT AND PRM TVE LANGUAGE
As a exampe o how the grammaticazation botteneck coud ead to urther segmentation et's imagine that the sae process just described was apped to the monothic sentences FuoonCausesLowTide NewMoonCausesLowTide and that the resuting more genera proposition, Causes (Y, Low Tide), was compared by a student with the previous one: Causes (Fu Moon X) Causes Low de) The student may then reaze that the abe in the rst proposition is a possibe vaue or the variabe in the second one and vice versa, and produce by generaization the abstrct proposition: Causes (y, X) At this point, a monothic smboic ariact has been eecive bro ken down into pieces that are now read to enter into partto-whoe reations to compose other sentences Not on have the abes or events bee assigned to the categor "noun but the previous impicit precate "Causes has been isoated and can be assigned to th categor "verb I other words, compositionait has been achieved and grammatica categories have emerged in a pure cu tura evoutionar process But what about the other required com ponent o a graar, rewriting rues that can be used recursive? A simpe orm o recursio coud emerge i to the primitive precate "Cause s that takes events as its inputs we add a precate ke " Sas taking propositions themseves as inputsY The new precate can generate propositions ike these Sas(ation Causes(Fu Moon, Low Tide)) Sas(M Teacher, Ca uses (Fu Moon, Low de) ) Or n Engsh, "ation sas that u moon causes ow tide or "M teacher sas that u moon causes ow tide These are the knd o prop ositions that woud arise spontaneous n a teacing situation when ever the content o what is beng taught must be justied b ascbng it to a tadition or to the ow dead teacher o the current teacher 159
PHLOOPHY AND MULATON
The type of recursion involved is of course very simple but that is all we needed in the present contet. Once an eteal language had acquired this minimal ompleity it could start co evolving with its biological substratum through the Baldwin effect The Baldwin effect is not a general mechanism for the transforma tion of leaed capacities into inheritable ones. Such a general mech anism does not eist Rather it is a means to give genetic adaptations that have little survival value in themselves a chance to increase in value by getting partial credit for their contribution to the tness of a given generation A leaed capacity can become inherited only in very special circumstances it must already have a genetic basis but one that does not fully determine it; it must directly increase repro ductive success to help its associated genes propagate; and this propa gation must be such that without the acual eercise of the leaed capacity the genes would be eliminated by natural selection One way in which the eercise of a capacity can affect biological evolution is if it is culturally transmitted from one generation to another giving it the necessary temporal stability to hav relevant effects.8 The capacity in question must also preeist the emergence of linguistic communication and be coopted for this purpose. The perfect candi date would be the ability to eploit sequential dependencies in the operations involved in stone tool manufacture This ability did have a partial genetic basis it was transmitted across generations and it had to be actually eercised to impinge on reproductive success. Its eten sion to sequences of sounds on the other hand mst have been at rst a leaed capacity How this etension took place is suggested by the operation of recurrent neural nets These can be used to etract temporal dependencies in sequences of manual operations as well as in series of vocalizations In fact a simple recurrent neural net, one with the same computational capacity as the simplest automaton, can learn how to predict the net word in a sentence even if the sentence has embedded lauses. This does ot contradict the idea that nite state automata cannot handle recursion because unlike a pushdown automaton the neural net does not have to be able to tell whether or not a sequence belongs to a contet-free language ll it has to do is detect statistical pattes in a sample of sequences and etract the probability that a given word is a likely continuation of a particular sequence.9 160
MULTAGENT AND P RMTVE LANGUAGE
The interaction of learning and inheritance involved in the Baldwin effec can be understood through the joint use of genetic algorithms and neural nets There are two etreme cases of this combined use a genetic algorithm can be used to nd the nal weighs of a neural nets connections in which case there is no room for learning or used to specify only certain properties of the neural net (like its connectiv ity or the initial values of the weights) allowing the nal setting of the weights to be done by training.20 A simulation of the Baldwin effect must start with the second case, in which a leaed ability is only partly genetically determined and end with the rst case, a fully genetically specied neural net. In one simulation etremely simple neural nets were used, much too simple to be useful in language processing but the problem that had to be solved through the com bined use of learning and genetics had only one solution This means that the structure of the search space had the form of a "needle in a haystack with the only good solution represented by a high tness peak surrounded by low tness points Without the availability to learn from eperience the tness peak coresponding to the solution was so steep that it was basically impossible to climb Introducing learnng on the other hand changed the shape of the search space urrounding the peak with a gentle slope that facilitated evolutionary clmbing.2 In our case the needle in a haystack would be the problem of generating a full language together with specialized brain organs for producing and parsing its sentences, using as a starting point an eter nally evolved language possessing only a minimum of compositional ity and recursion Before discussing how this process could have unfolded we need to address a limitation of the previous simulation the assumption of direct access to propositional content That is, we need to replace the capacity to read minds with the inference of com municative intentions from the observation of behavior We argued in previous chpters that prelinguistic humans shared with birds and other mammals the ability to analyze sensory eperience in terms of comple scenes in which objects with enduring identities play roles of agents and patients and that they shared with other primates the ability to detect tendencies to cooperate or cheat in social interactions Understanding behavioral roles and detecting behavioral tendencies is not of course, the same thing as guessing intentions but the abilities 161
PHLOOPHY AND MULATON
nvolved could have been further developed n that directon once simple reference and causal predication emerged The abity to aceve shared attention by the use of a common set of soundreferent associations could have led to the ability to infer intentions if the objects referred to had a capacity to affect and be affected by the agents, if the referent was a predator affording them danger or a prey affording them nutrition, for example In other words, the ntention to warn others of potential opportunities and risks would be easy to guess from the production of a sound if the lat ter took place in a situation to which all agents attributed signicance And simarly for the symboc artfacts denotng the conjunction of causes and effets If the production of the sound for "Fuloon CausesLowde" had acquired the stats of an urgent call to go to the beach to forage for shell sh, agents would come to nfer the cus tomary ntention behnd the producton of that sound These abilities, pritive as they are, could have sufced to replace the elicit pre sentation of propositional content n a leaing sitution if the teach ing of lnguistic associations had taken place s part of the performance of noinguistic tasks, such as huntng, gatherg, or orkg stone n the other hand, once monotc artifacts began to be broken down a new type of observable behavior became avalable: word choice behavior. I particular, words ke "full moon" and "low tide" that tended to occur together more often than n the company of other words would have created patterns n the frequency of co occurrence that were detectable in the choices of words made by agents These are exactly the kind of statistical pattes that can be extracted from a population of sentences by even the simplest recur rent neural nets From then on the history of lguage can be magined to have pro ceeded on two levels at the same time: biological evolution continued to operate through the Baldwin efect helping to propagate the genes underlying the caacity to detect pattes of word cooccurrence wle cultural evolution began to transform customary patterns nto obgatory consraints on wordchoice behavior While early on n the cultural evolutionary process the teacherstudent nteraction may have volved only the transmission of pattes that had become abitual, later on the pattes could have been transmitted as if they 162
MULTAGNT AND PRMTV LANGUAG
consttuted enforceable social obgation That is, a word that used to require another one simply because the two tended to cooccur very often now obged its user to provide the second word. This way cultural evolution could have provided a mechism for the emergence of lnguistic rules, that is, a mechaism to institutionalize habitual usage nto obgatory convention The basis for such a mechaism would have already been available if, as was argued n Chapter 8, the habit of punisng those that do not punish departures from cooperative behavior (metanorms) had previously emerged nce metanorms had become part of the cultural content that was transmitted from one generation to the next their use could have been extended t patterns of behavior that distnguished one com munity from another I other words, metanorms could have become a mechaism not oy for the emergence of sodarity and a sense of communal dentity but also for the enforcement of behavioral pattes that were uique to a commuity d consolidated that identity A mechanism to transform habit nto convention is an importt component of theories of nonbiological nguistic evolution at he level of both syntax and semantics The way syntax could have been affected by stitutionazation c be shown by gog back to the previous example of the proposition "Causes Y X). nce such an abstract proposition had emerged and the customay cooccurrence of a predicate and its arguments had become a conventional con straint, the pattern itself could have been used as a template (the operatorargument template) for any other eressions I other words, the patte could have been used by alogy to constrin the creation of new pattes the result beg distnguishable from sen tences produced b y the application of gratical rules3 The effect of institutionazation on semantics, in turn, is related to the fact that when two words tend to co occur with very gh frequency, so that a stener can aready tell what the second word will be once the rst one is heard, the second word can be omitted from a sentence That is, very gh frequency of cooccurrence could have made some words redundant for a speaker since they could be supplied by the stener I some cases this could have allowed entire phrases or even sen tences to be compacted into a single word while preserving the mean ing of the phrase or sentence When the product of packing the 163
PHLOOPHY AND MULATON
composite meaning of series of wo rds was m ade obligatory a powerful mechanism for the extension of vocabulary, and even for the produc tion of new categories of words, would have come into existence24 this scenario is correct we can conclude that the possiblity space that was searched through evolution, both biological and cultural, was isomorphic with but not identical to the space studied by autom ata theory Automata thory charts the space of possible languages and of the computational capacities needed to master them in an entirely mechanismindependent way involvng no commitment to any particular mplementation of the automata themselves This point is often overlooked by those who think our brains must embody a genetically coded lnearbounded automaton usng an explicit con textsensitive "language of thought When it comes to the design of future simulations of nguistic evolution keeping the distinction between mechanisms and mechanismindependent structure rmly in mind can have mportant consequences particular, useful as the just discussed simulation of grammaticalization is to study the decom position of monolithic symbolic artifacts, performing it in a format that automata can understand reduces the degree to which the pat terns that result can be considered to be truly emergent More pre cisel, while comositionality did indeed emerge in that smulation the grammatical rules and categories did not emerge they were partially built n by the choice of symbol sequences and nference algorit25 What we need are smulations that can yield rules and categories as an emergent product by using recurrent neural nets to extract pattes of cooccurrence ong words and by situating the agents in a social space in which the transformation of habitual practices into obigatory conventions can take place is time now to retu to a subject that was left unnished at the end of the prevous chapter There we saw that detaled archeological smulations of mesolithic and neoithic communties failed at tmes to match the avalable evidence ad that a possible explanation for this is that once �omplex meanings could be formed by the comb ina tion of words, and once the referents of those wrds ceased to be directly observble objects and actions, magical beliefs havng a pow erful effect on behavior could have emerged the next chapter we wll tackle this kind of beliefs but without separatng the imaginary words they give rise to from the material world of resource gradients 164
MULTAGENT AND PRMTE LANGUAGE
To keep these two worlds together magical beliefs w be scussed as a means to generate a new kind of social gradient a graent of egitimacy, that goverment organizatios can use to manipulate physical graents to an extent that early humans would have found mystifyng I particular, the construction of huge funerary monu ments' like the pyramids of Egypt, w be shown to have depended on the couplng of magical and practical beliefs, the pyramids them selves constituting a powerful mechansm for the production and mantenance of legitmate authority
165
MULTAGENT AND ARCHAC TATE
CHAPER ELEVEN Mut iagents an d Archaic States
Neolitc communitis, like their hunter-gatherer predecessors, were relatively egatarian. Certain personal qualities such as charisma oratory prowess, or the abity to lead may have created fferences among their members but these tended to be transient and had ittle impact in the long run For this reason these communities had access to ony one social graent a graent of soidarity But when ffer ences in prestige were not ssipated quicky and were allowed to accumuate they began to form another gradient a gradit of status marking the beginning of the process of social stratication Govern ment organizations with a well-dened authority structure, on the other hand needed more that status fferences to emerge Authority can always be eorced in a material way through physical punish ment or connement but it is much ess costly and more stable if it is beeved to be legitimate nd ts implies that a gradit oflgitima has been created a concentration of the capacity to command justi ed by a regious tration g ete members to supernatural forces or some cases, justied by the successful practical reasong of specialized bureaucracies 1 The storical process that gave rise to these two new exploitable resources used to be conceptuazed as a simpe linear eoution: egalitaian agricultural tribes were super seded by chiefdoms organized around status fferences which were in turn, replaced by archaic states and their centralized authority But archeologists have moed ts simple model in severa ways First, it has become clear that there is more than one transitional entity between rural and urban forms of social organization At the 166
very least we must make a stinction between simpe and complex cefdoms the latter already splayg rigid social stratication but not centrazed decision-makng Second the focus of research has shifted from looking for simlarities to exploring the wide range of variation of each of these social forms nd nally the idea that these social forms followed each other in time has been replaced by the study of their coexistence in space: an archaic state may have formed the core of a arge region dominating a few compex cefdoms with simpe cefdoms and agricultura vilages forming an exploited periphery2 The concept of a linear evoution usually leads to modes in wch a single entity "society as a whole, develops from one "stage of deveopment to another But once we replace that concep tion with one involvng the coexstence and complex interaction of agricultura communities, cey lneages and institutional orgaiza tions, we need several models In some cases we can take for granted the emergent properies of these larger invidual entitiessodarity status or legitimacyand model the entities themselves as agents Invidual orgaizations, for example, can be modeled as agents because they have their own goals (of wch most of their staff may be unaware); are capable of affecting and being affected by their environment (as when an environmental crisis affects an entire orgaization) possess institutiona knowledge and memory (written records routines); have their own egal stanng rights and obga tions; and are capable of coordinating the activity of many workers and managers the performance of complex tasks3 Multiagent simulations in wch the behavior of agents is entirely determined by formal rules are adequate to deal with ts case as we saw in Chapter 9 where invidual communities were modeled as single agents But if we want to explain how properties ike he status of a ineage or the legitimacy of an organization emerge we must simuate the interactions between persons expcity and model their behavior as determined by beiefs and desires: the beef that a sacred text justies the authority of a supreme leader for example, or the desire to be part of the mission of an orgaization The agents we have scussed so far do not meet ts requirement: the rules that guide their behavior may mimic the effect of having beefs or desires but the latter are not expcitly modeled and cannot interact to produce emergent effects The simplest way of modeng beefs and 167
PHLOOPHY AND MULATON
desies are as atitudes or sances aken by persons toward pro osi ions that is as proposiiona attiudes Using propositions as he objecs o mena attitudes disregards he ac that pre-guisic humans aready possessed beies and desires reative to he content o their episodic memories: the beie that something happened because one witnessed it happeng o insance or he desire or a avor aroma sound or sigh that one has peiousy experienced. But he beie on supenatura orces or the desire not to oend those oces caot exis without anguage To acke more compex cases o pe sona agency ecen appoaches o the design o simuaed agents have repaced rigid ues wih moe exibe proposiiona atitudes. This is the stategy oowed in the socaed BeieDesirentenion (BD) appoach4 BD agents ae no in act designed o dea with reigious o ideoogica bees bu to eason about practica maters As in the case o other agent designs his invoves mapping sensoy inputs o appropriate behavioa outpus but insead o using ues to mediae between perception and action a more eaborate pocess is empoyed: ormation about he word about a cetain disribution o resources or exampe is rs used to updae an agent 's beies; hese are then used to generate a set o options concerng he avaabe opportuni ties and risks; he options ae hen checked or how desrabe and achevabe ey are; opions ha are desred and can be acheved become goas and hese combed with a commitment to a couse o acion become t he intentions driving e behavio o e agent5 Usg BD agens or e siuation o supeatura bees woud vove two simpe modicaions. First we woud ave to give agents te capacity to ascribe propositona atitudes o oer agents in oder to make sense o ei acons a capacity eered to as takng "e tentiona stance toward te expanaton o beaior6 Second agents soud be abe to treat resources as agents7 Togeer ese two modications woud give BD agents te capacity to adopt he inen tiona sance owar maeria gradiens attibuting beies and desires to enities ke e sun te rain te soi and so on To transorm a ew magica beies into an ideoogy one more gredient woud be necessary: debates among agents aimed at persuasion n e accoun o he emergence o anguage given n e pevous capter it was imporant o assume ta agens did not ave access to e conents 168
MULTAGENT AND ARCHAC TATE
o each ohes minds and ha hey coud ony orm intenions abou changing each othes behaor But once anguage emerged and coud exress he conent o poposiiona attitudes humans became abe to orm tenions in which the goa was to modiy he very con tent o those beies and desies. te subjec mtter o such pesua sion attempts voved expanations o eso urce scarcity in intentiona terms e The so has become barren because the eiity goddess is angry then om he ensug discussions a coheren set o super natura beies coud emerge. Proocos or negotiaion through which BD agens can estabish their positions on an issue make conces sions and reach ageements are currenty being deveoped8 The BD approach to agent design is in its inancy and the kind o appicaion o archeoogica simuations just suggested does not yet exst But simper approaches using rues to geneae beavio at mimics he eec o proposiiona aitudes are aready yieding inteesting esuts. One o them caed "Emergence o Society ( EOS ) expores he rst step in the evoution o organizaiona auhority: the rise among egaitarian communies o enduring leadership n a ypica EOS simuation agens propose pans abou specic ways o avesting resources: what resource to arget how many oher agens ae needed how o aocate hei abor othe agents beieve that his plan is bette tan hei own hey adopt it When an agent has paicipaed in anoher agens pan on sevea occasions it becomes a oower whe those whose pans ae oowed become eaders. When a group as ormed aound a eader cerain tendencies become mani est tha srengthen e bonds hoding he goup togethe: its mem bers end to communcate with eac oter more oten about te ocaion and concentaion o resources and tend to ake eac oe into account in e ormation o uture pans9 Once his dynamic o group ormaion as been estabised e agens can be cononed wit envonmenta crises o ceck wa eects ese ave on e sructure o te groups n oe wods wat ths simuaion expores is te reation between endurg eadership and a space o possible resource distributions Tis space may be picured as conssing o wo mensions one eatg to reative abundances and scaciies e ote to reaive concenrations and dispersions Abundance weakens eadershp sce mos agens can meet tei meaboic needs by woking aone whie scarciy acs an incenive to work ogeer. 169
PHLOOPHY AND MULATON
Dispersion of resources in turn make creating successfu pans harder since it increases the difcuty of coordinating the actions of a group whe concentration faciitates panning The outcome of EOS simuations has increased the pausibiity of an eisting archeoogica hypothesis that the hstorica emergence of endurng eadershp was triggered by environmenta changes eading to both scarcity and concentration of resources This simuation captures the materia aspect of the transition from a strict egaitarian community to one in which an accumuation of prestige around a eader is not rapidy dissipated The maintenance of such a gradient over onger tme scaes however is probematic because in huntergatherer and agricutura communities departures from an egaitarian organization are activey prevented by interna devices Leaders are demanded for eampe to be generous and redistrbute in ceremonia feasts whatever food surpuses have accu muated n their hands or ese to rituay burn them Customs ike this may act as an unconscious mechanism to guard against the crys tazation of forma authority Moreover war heroes prophets or promoters of agricutura productivity may nd their eadership com promised if any misfortune (a defeat in war the faiure of a pan a natura disaster) is beieved by their foowers to be a supernatura sign that the eaders do not in fact possess the admired charismatic vrtue Magca beiefs on the other hand can in specia circumstances stabize rather than dssipate a status gradient The catayst for such a change can be a succession crisis ensuing after the death of a cher ished eader These crises can ead to the transformation of charisma from a persona characteristic into one that can be transmitte aong a bood le a transformation typicay justied by the nvocation of supeatura beiefs1 And once eadership can be inherited it can pay a crucia roe in the generation of enduring concentrations of prestige around particuar kinship groups Simuang this critca transition woud invove augmenting simuatios ike the one just described with BDI agents capabe of having magica beiefs and of engaging in dbates about the sources and potentia transmissibiity of charismatic powers Once status becomes a property that can accumuate it opens the way for the differentiation of communities or of groups within com munities into socia strata rcheoogica evid�nce for stratication 170
MULTAGENT AND ARCHAC TATE
comes from a variety of sources Snce compe chiefdoms organize entre geographica regons the remans of settements in a regon may be ranked by size the number of ranks counted and the distri bution of settement sizes per rank studied This regiona hierarchy can then be used as an indication of the number of hierarchica eves of decisionmag that eisted in those communities At a smaer scae buria sites offer another window nto socia strata Systematic buria was aready practiced by humans in the midde paeoithic and by the upper paeoithic it had become ceremonia the dead were bured cothed and wearing persona adornments 3 The buria sites of chiefs contain in addition to persona information vauabe indica tions about the steepness of status gradients the size of a tomb its demarcation from other graves and more importanty the richness of its contents Of paricuar signicance are the presence of adorn ments made from eotic materias ranging from obsidian imestone and basat to copper apis azui turquoise and aabaster These rare objects are signicant because they constituted a different form of weath than agricutura producs they coud be used not ony to epress status differences but aso as a nd of oitica currency to confer status to others as when the eader of a compe chiefdom distributed them to those of simpe chiefdoms to cement aiances and dene oca rankings4 This movement of objects of status from a pace of high concentration to one of ow concentration conrms that prestige can form a gradient Keeping it from dissipating invoved the strict contro of the ongdistance trade of rare raw materias and of the speciaized crafts needed to produce eite adornments Indeed burying these rare objects together with their former owners may have been a way to prevent "nationary pressures from deveoping by permanenty taking them out of circuation The eistence of an enduring gradient of status can ead to the cre ation of rank differences that are uid and contnuous as n simpe chiefdoms or ed and sharpy discontinuous as in compe chief doms To epore how these different outcomes may be produced we do not need simuations in whch persona agency is modeed n detai but rather ones that capture the interaction between arger scae socia entities taking pace at onger tempora scaes One eam pe is a simuation (caed TongaSm) inspired by a contemporary case the Tonga chiefdom n Poynesia This chiefdom is organized into 171
PHLOOPHY AND MULATON
knsip groups called "leages, each with its own cief or chiefs, that are ranked in terms of status These rankings can change depending on the relations between lineages Specically, marriages among the offspring of titled chiefs can increase or decrease the rank of a lineage: if the titled heir marries someone from a igher status in eage is or her own leage will gai in status and vice versa, heirs marrying down in status bring down their own lineag Signicant canges in the ranking of lineages can take up to four to ve generations so only a series of bad marriage decisions can have this effect Each leage is involved economic activity, its agricultural wealth determined by number of kn, the land available to it, and the personal ability of its chief to motivate both kin and commoners to work harder This agricultural wealth, i tu, lows within lineages in the form of redistribution to comoners, and between lineages from wifereceiving to wifegiving ones Agricultural wealth also lows in the form of tribute paid to lineages of igher rank 15 Thus in chiefdoms material and status gadients are intertwined, their teraction pivoting on marriage decisions as well as on decisions affecting the selection of titled heirs The question that te simulation set out to answer is how the rigid social stratication observed in complex chiefdoms could emerge from the dynamics of these interacting gradients One possible answer is the relaxation of prohibitions to marry close relatives Incest taboos come in different strengths: they may prohibit the marrying of brothers or sisters; te maryg of rst cousins; or the marrying of any relative however distant The stronger the prohibitions the more likely it is that an heir belonging to a high status lineage will be forced to marry someone with lower status It is therefore in the terest of those with the highest status to lower te strength of te taboo Historically this could have been acieved hroug political strategies by chiefs but with smple simulated agents this level of scale cannot be captured The relaxation of taboos, on the other hand, can be performed exogenously to check whether it has the ostulated effect on degree of stratication In the smulation ten chiely lineages started the process As population increased and ome lineages ran out of agricultural land they split into two increasig te overall number (The smulation coul handle up to SO lineages) The smulation went through a series of cycles as each lineage ranked potential spouses and selected heirs, hundreds o f 172
MULTAGENT AND ARCHA C TATE
smulated mariages took place, agricultural wealth lowe among leages, and status was transmitted Once this basic dynamic was established experments on the effect of taboos of different strengths could be conducted Measures of stratication cluded the steepness of the status gradient (how large the differences in status between leages were) and the mobility of the lineages (how often changes the raing occurred) The outcome of te smulation conrmed the basic logic behind marriage strategies: te lower the strength of incest prohibitions the higer the degree of stratication 6 If the burial sites of titled chiefs can be said to be our window into social stratication then architectural monuments like pyramids, temples, and palaces are our main source of evidence for the exis tence of centralized decisionmaking in arcaic states The central decisions leading to the erection of monumental structures were based on both technical and magical beliefs and informed both mate rial and ceremonial practices Let's take the example of te famous pyramid of Giza in Egypt On one hand, this funerary monument was clearly tended to have magical ffects since it was supposed to function like a resurrection machine Needless to say, the pyramids internal mechanism did not allow it to actually transmute a king into a god but it nevertheless functioned like a machine for the production of legitmacy, both by its powerful effect on the religious beliefs of those who contemplated it and by serving as a ritual center for the continuous performance of traditional ceremonies associated wit the cult of a dead kg On te other hand, the design of te monument volved higly technical knowledge such as the astronomical knowledge neede for is proper orientation toward the heavensit is very precisely orientd true northand the applied matematical knowledge involved calculating its sape Its construction, in tu, demanded wellorganized material practices The pyramid is made of more ta two mllion blocks of stone eac weighing over two tons on average These blocks were originally covered by polished white limestone that, unlike the core stone, was an exotic material transported from far away Building such a structure in just over two decades is evidence of the relative efciency of the bureaucratic organizations involved, the legitimacy of wich derived from their pragmatic ability to overcome the cognitive, physical, temporal, and stitutional ltations of personal agency In short, te pyramids 173
PHLOOPHY AND MULATON
were both ceremonial and material entities, the rst aspect related to their capacity to express n their solid permanence the validity d contnuity of authority and the second to the fact that they consti tted a frozen indication of the total amount of energy that a given eite was able to haessY Before the great pyraid of Giza there was only a century of con struction experience starting with the very rst Egyptian monument built entirely out of stone, the Step pyraid at Saqqara. between these two there were a few experiments in the building of smooth sided pyraids one of wch seems to show that the architects were stil leang since they changed the slope of the monument as it grew taller (the B ent pyraid at Dashur) On the other hand, prior to ts relatively short burst of creative activity there was nearly a millen nium of experience in the construction of tombs and other funerary monuments going back to the late pre-dynastic era when ceftan's tombs built out of brick with some stone elements seemed already intnded to last for an eternity18 These two different time scales call for different explanations Modeling a leang process lasting many centuries should use as agents the organizations themselves: royal workshops and the parts of the rising central goveent on wch the workshops depended. Modeling a process involving two or three generations on the other hd a period of time in wch projects grew so large so fast that bureaucrats had to lea to solve novel logis tical problems during their lifetmes, should use simulated personal agents One way to approach the smulation of organizational lea ing taking place at lennial tme scales is to model it as an evolu tionary process. Organizations can be modeled as entities capable of evolution if their daily routines and procedures, and their division of labor and coordination mechisms are passed from one generation to the next or even from one organization to another witn the same generation. 19 We can agine the activity of tomb building early on in Egyptian story as divided into separate tasks each performed by a separate artisan in which organizational learning involved the elimination of inefciencies though further subdivision into subtasks More speci cally if there were subtasks that had to be exected as part of several tasks then a certain degree of overlap existed among the acivities If a single artisan were assigned not one but two overlapping tasks 14
MULTAGENT AND ARCHAC TATE
I
the the shared subtasks would have had to be performed ony once reducing the overal amont of effort nvolved On the other hand, the effort of coorinatng the different activities by overseers would have increased Ths there was a tradeoff between the advtages of clstering overlappng tasks and the coorination costs ncured by such clusteg To smulate organizational evolution using genetic algorithms the routine task assignments must be coded into a chro mosome n wh the ones and zeroes stand fo agents d subtasks, whe the tness function measures the overal amount of effort taken to complete the task Simulations like these have been carried out and have shown that clustering of tasks does occur but not in a way that takes ful advtage of functonal complementarities among sub tasks20 One way of mproving organizational learning would be to model it not as a blind evolutionary search but rather as a hybrid of genetic algorithms d reinforcement leaning These hybrids are caled asser systemsY a classier system what evolves are rules that guide the perfomance of tasks, rules that specify that if a condi tion arises then a certain action must be taken Fitness is evaluated not by the amount of effort spent perforng a task but by the ability to get a reward from a performane. Moreover, tness is assigned not to individual rles but to groups of related rules a shared credit assignment that tends to produce coadapted sets of rules rather th a single ttest rule22 the present case the conitionaction rules would represent either customary acivities in wch behavioral responses to specic conitions have become rule-like or written regulations specifying the connection between a condition d an action Royal workshops would be the evolving entities while the source of rewards would be other organizations on wch the work shops depended for legitmacy and nancing addition to the slow organizational larning behind the practice of tomb builing there is the much faster one leading to pyraid builing. To get an idea of the challenge ts task posed to arctects, ofcials, and bureaucrats we need to model the construction process itself, a process involving a m of the technical and the ceremonial. On one hand the workforce was motivated by the intense religous feelings geneated by its involvement in a ivine projec Athough in he past this wokforce was assumed to be composed of unskiled slaves blinly following comands from above it is now clear that 15
MULTAGENT AND ARCHAC TATE
PHLOOPHY AND MULATON
to create wooden sledges and metal tools, to potters, bakers and brew ers to produce containers food and drk. Overall this experiment showed that a force of 25,000 workers could have nished the task in the 23 years that the builn of the pyramid originally took.5 n addition to these physical experiments to duplicate some of the original tasks multiagent simulations can be used to explore the part to-whole relations between tasks and subtasks While smple projects can be subdivided in such a way that they are no more than a sum of their parts if there are interdeendencies between component tasks then the components cannot simply be added but must be carefully articulated he objective space of possibilities of different task envi ronments can be explored using a formal language caled ask Analysis Environment Modeling and Smulation ( or aems aems can be used· to model the relation between the main task and its subtasks and to capture the interdependencies between individual tasks: performance of one task, for example, may enable the perfor mance of another one so it must take temporal precedence; or it may merely facilitate its performance so the rst task need not be fuly completed before the second one can start; one task may replenish the resource needed by another task so the two may proceed smulta neously; or the two tasks may consume the same resource in which case there can be conicts between them6 hese relationsenables facilitates rplenishes and consumesallow aems to capture the structure of a task environment in the form of a hierarchica graph the termnal ends of which are directy executable operations. When used as a tool to explore the objective possibility space of task environments these directly executable operations must be left unspecied and the denition of what they produce must be made very abstract capturing only the contribution of an operation to the solution of a problem Using aems this way is facilitated by the fact tat its formalism was designed to be entirey domain-independent: a graph representing a given task decomposition may capture the structure of a task environment that is common to many domains (a factory a hospita an airpor, a bureaucratic agency explainng why human agents in those domains face similar organizational challenges he mensions of the possibility space are typicay prob abilistic because the properties of large graphs can only be given statistically: the average number of subtasks nto which tasks are
most of the operations involved in pyramid builing demanded high skl and motivation Recently excavated cemeteries in the town that used to house ofcials overseers artists and artisans show that even the workers were buried respectfuy, strong evidence that they were not slavesY On the other hand both the religiously motivated workforce and the religiously legitimated organizational structure above them had to take into account the capacity of the material world to be affected by human activity as we as the capacity of humans to affect the world at that scale More specicay to under stand the material chaenge posed by the activity of pyramid building we need to dene both the objective nature of the task environment as well as the kinds of coornatng mechanisms needed to organize a large workforce to meet that challenge. n what foows we wll concentrate on the technical component of the construction process but it should be kept in ind that without the ceremonal component the explanation would remain incomplete he way the Egyptians dealt with the construction problem was through a recursive decomposition approach, grouping together related tasks and assembling them as parts of a larger whole he overall task of pyramid building was divided nto at least three main task groups: quarrying th stone; transporting it to the construction site; and erectng it Quarrying, in turn, was subdivided nto several tasks carving deep channels into a lmestone bed nearby; carving narrow channels to subdivide it into blocks; and using wooden levers to detach them from the bed ansporting the blocks was subdivided into the buing of tracks and ramps on top of which they could be movd; loading them into wooden sledges ; and pulling or roing the sedges over cyndrical beams Finay, erecting the stones nvolved setting them using wooden evers; carving marks on the stones to control the slope; and trimmng the stones to their nal shape.4 he efforts of expermental archeoogists in the past few decades have given us a sense of the size of the workforce and the amount of tme taken to execute each task I one experiment a team of 1 2 quarry man was able o produce eight stone blocks a day; 20 men could hau the blocks from the quarry to the builng site at a rate of ten blocks a day; and a team of four men two to push and adjust and two masons to perform the tg was needed to erect each block o this we must add a variety of support workers from carpenters and blacksiths
177
176
I
PHLOOPHY AND MULATON
decomposed; the average duration of operations and tasks; the statis tical istribution of hard interdependencies (like "enables) and soft interdependencies (lke facilitates). On the other hand, the graph produced by Taems can be coupled to a multiagent system to investi gate a specic task enviroment in which case the operations as wel as their products and durations must be fuly specied. In the case o pyramid builing the irectly executable operations would be activities lke measuring, aligning, ciselng, levering, cutting, polish ing These operations can be assigned the average time they take to be performed and the potential value of their product using the results rom experimental archeology Tis iscipline can also provide insight into the subivision of tasks and sub tasks to be ncoorated into a Taems graph But n adition we would need another graph speciy ing the authority stucture o the relevant parts o the Egyptian gov erment. Ths can be reconstructed usng existing textul evdence related to the buing o the actual pyramids Control of the overal task was the hands o the Vizier in his role of Overseer o Royal Wors Evidence for this comes rom a text nscribed n hieroglyphs on the wals of several tombs low as "The Duties o the Vizier. From tis text it is clear that authority roles had not yet ierentiated into the plurality of unctions ith which we are amiliar today I particular, the Vizier played several roles: head o the civil aministration with uicial and amistrative responsi bilities; managng director o the royal palace getting daily reports about security and logistic matters; and deputy to the king, personaly hearing his commands and ensuring their enactment.27 The Vizier controlled the conglomerate o central departments constituting the Egyptian bureaucracy as wel as the local authorities o the iferent urban centers and their rural areas The text also reveals that the different departments were relatively autonomous: although detailed written procedures existed to regulate their activities the Vizier does not seem to have initiated any o the departments' actions by irect command but only to have supervised the results ater the act 2 8 This suggests a hierrchical coorinatng mechanism but one n which oy strategic gals were given as explicit orders while their tactical implementation was let to the perso�el ithin the departments. Other texts give us the part-to-whole relations at the bottom end o the hierarchy: the workforce consisted of several crews each composed 178
MULTAGENT AND ARCHAC TATE
o two gangs o ,000 men ivided into ve groups of 200, each fur ther subivided into ten teams of 20 men called "ivisions29 The overall hierarchical structure may be modeled by a graph in which nodes represent institutional roles, edges represent relationships between those roles, and arrows represent asymmetrical relations, such as those that exist between superiors and suborinates. The unity of command under the Vizier implies a graph with a single node at the top, while an unambiguous chain o command implies a graph in wich all nodes but the top one have a single arrow ponting to them, representing the fact that every subordinate had a single irect boss30 These two graphs, one describng the command hierarchy the other deing the specication and decomposition o the overal task, give us an objective view of the organization and the problem it must solve But to animate them ith agents the graphs must be supple mented by a plurality o partial views. The subjective view of the problem varies rom agent to agent and it is deterined in large part by the way in which the workorce is organized this organization is achieved n a rigid hierarchical way n wich commands ow down wards while reports ow upwards the agent occupying the top posi tion must be modeled as hving access to the objective tas structure hile suborinate agents should have o a ragmentary view, the more lmited the lowe their ra the organization included teams and i their members were alowed to commuicate with one another then agents may orm in adition to their local vews a no-local view through the nformation provided by their peers. This distinction also deterines how interdependencies between tasks are handled by agents: in a rigid hierarchy schedules and deadlines for every subtask are centrally decided and implemented by coand, while ithin teams agents must be able to negotiate dealnes and make commit ments to one another.31 I the case of pyramid builing teams are needed to model what happens inside each o the hierarchical levels The Viziers oce, or example, consisted of several people workng together: the Vizier himsel his council, an ofcial o nteal aairs, and several scribes Teams also existed inside each relatively autono mous goveent department and at the level o worker ivisions where most physical operations were perormed This complexity cals or agents that are not ust boundedly rational (as in most multi agent simulations) but also task oriented and socialy situated32 179
PHOOPH AND MUATON
Finally populating the graphs with agnts and settng them nto motion demands simulating a range o coornating mechanisms. The simplest and etter explored o these are prohiitions that is simu ated socia aws that generate commitments to avoid certain actions without incrring the costs o communication exaple o prohi itions are trac laws In a simulation o the uilng o a pyramid the need to regulate the trac o agents would mmediately ecome evident i the agents were situated in a twodimensional space: er ent gangs would have o e assigned o erent sides o the pyramid and withn each side agens ringing stones mortar tools ood and drin would have to avoid colisions and creae relaively ecien circulaion paths Moreover i we added a hird dimension o this space then as the pyramid grew aller and the room to maneuver decreased spatial assignmens allowing the simultaneous perormance o asks would ecome even more importan Circulaion pates could e rigiy specied y a central conroler ut rac laws can e a less expensive coornation mechanism i they are crated o ake advantage o the act ha agents can perceive one anoher within certain stance and can hereore conriue to he soluion o circulaion prolems y making some local decisions hemselves Ideally a social law should give agens enough reedom o action to construct a plan o actionnd an opimal pah rom one poin o another or exam peu no so much tha hey can inerere with the acions o ohers33 The dynamics eween erent levels o he herarchy can e capured y augmening he ineracions etween agens wih speech acts. These are actions in which he very uterance o a sentence carries nding social impicaions with eren degrees o orce o promise to reques to demand and most importanly in the present case o command34 Usng the graph ha captures he auhority structure he inng orce o these commands can e made o low rom their poin o uterance o those tha mus ulimatey execute them The dynami wihin eams in turn can e expored using an approach called "General Partial Goal Planning (or GPGP) In this approach agents use heir partia eies aout a ask to orm inen tions o act and to sequence hose actions o create a oca schedue Because he conten o an agents eies do es no ncude inormation aout oher agents' asks is oca schedue may ener ino conlict 180
MUTAGENT AND ARCHAC TATE
with their schedules Thus the rst coornating mechanism in GPGP is used to create nonlocal v�ewpoints This mechanism alows an agent to get inormation aout other agents' eies and to detect coornation prolems etween its tasks and those o others. A sec ond mechanism allows agents to communicate the results o their operations to others ut only those that are relevant since communi cation itsel is costly Other mechanisms are then used to aid agents in the solution o prolems ike avoing redundancies and haning inerdependencies. The end result is that agents can modulate heir local schedulng o include commitments o others: a commiment o mee a deaine; o stop an action n order to ree a resource; or o start an action a the earlies possile time ater anoher agent's action is nished35 Armed wih hese erent coornating mechanisms (rac laws cenral commands negoiated comitmens) it should e possile o use a Taems graph and a GPGP multiagent sysem o simulae he complex process o uilding gian unerary monumens: each set o coornating mechanisms ogeher wih a proposed decomposition o he overall task ino suasks would consiue an archeological hypothesis aou wha really wen on during the construcon o he grea Giza pyramid Such a simulaion would capure only a rie episode n the history o ancien Egyp and would include only a segmen o Egyptian society: the Vizier as Overseer o Royal Works the relevan departmen under h and he crews o workers under the command o tha department. This means hat much o what made the workings o this social segmen possile would e included only implicitly The provinces ha suppied ood and laor to he cenral government or example would e ncluded only as resource inpus. Bu as we have argued repeatey in previous chapters there is nohing wrong with unexplained presuppositions as long as a simulation explains somethng else such as the emergen capacity o an organization to pan and ud large archiectural structures and as long as there are oher models hat accoun or the emergence o what is eing presupposed Finly here is he quesion o how o simuae the organizaiona earnng process ha led to a specic set o coornang mechanisms The cenury o pyramid udng experience eore Giza is oo short or an evoutionary process o have yielded any signican resus 181
MULTAGENT AND ARCHAC TATE
PHLOOPHY AND MULATON
ig rank would not earn to improve te sequencing of operations lke stone cutting or polising but te seduling of entire task groups like quarring auling and erectig Te modications to coordina tion mecanisms woud be still be loca and oriented to specic tasks but te iger te rank te more general in scope tee local solu tions would be Like genetic algoritms and neural nets multiagent sstems ave a variet of idustrial uses tat ave noting to do wit modelng social reait Tese range from creating user terfaces for new software applications tat can learn about te abits of uman users to increas ing te efcienc of te aocation of resources i te iteet by allowing autonomous agents to engage in decentralzed trade.3 8 Tese and oter uses of multiagent sstems wi contiue to drive teir development regardless of te rate at wic social scientists adopt tis tecnolog for modeing purposes But it as been te underlying message of tis capter tat social simulations as enacted tougt experiments can greatl contribute to deveop insigt ito te work ngs of te most complex emergent woles on tis planet Weter its rate of adoption accelerates or not wl depend on te imaginative use tat social scientists make of tese new resources as well as on teir wilingness to abandon te idea of "societ as a wole and replace it wit a set of more concrete entities (communities organizations cities) tat lend temselves to partial modelng in a wa tat vague totalities do not I sort te future of multiagent simuations as models of social reality wl depend on ow social scientists can affect tis tecnology b depoyig creativel and on ow te can be affected b it troug te possession of te rigt social ontolog.
even if augmented b reinforcement eaning Tis means tat we ave no coice but to use persona agents capable of montorig te execution of teir own tasks and to detect problems like te perfor mance of redundat subtasks mistakes in te sequencing of opera tions misallocated resources or broken commitments. More importantl once a problem s detected agents must be able to diag nose its cause and propose a soution To ave tis capacit agents must use te portion of te Taems grap to wic te ave access to ocate te detected problem i its proper contextto determne for example if iterdependencies among subtasks were one of te causesas we as use traces of past activit recording not only te series of decisions te took about wat operation to execute and wen to execute it but also te reasons w tose decisions were made36 I is sceme an agent as direct acess onl to its own memor traces but in te presence of task interdependencies it ma ave to request iformation about te traces of oer agents to be abe to reconstruct te process tat ed to te problem and tese requets can be costl To keep communcation costs low relative to te gains in efcienc agents sould attempt to diagnose onl situa tion specic problems. I oter words agents sould tr to gater nonlocal iformation but onl to te extent tat it can be exploited locall37 Troug a process like tis an organization can learn to improe coordinatig mecanisms but onl at te level of its teams Tat is once a correct diagnosis is made modications to coordination rules w be appled onl loca Tis approac on te oter and can be extended to te entire ierarc b appling it to all te different ranks Historically one of te rst organizaiona roles to differentiate between a cief and is folowers must ave been tat of overseer Early on ito Egyptian dnastic istory severa ranks of overseers alread existed But wat is an overseer if not an agent tat special izes i monitoring and detecting probems ocurring at te rank just below and creatig eports expaining problems addressed to tose in te rank just abe? I oter words te same diagnosing capabilities tat team members ave in is sceme could yied an agent plaig te roe of overseer. Te difference between te two cases is tat te iger te rank of an overseer te larger te portion of a Taems grap it woud ave access to Tis means tat an overseer of sufcientl 182
I
183
ANDX
APPENIX Links to Assemb lage Theoy
this history the restriction against totalities has not always been enforced, it w be useful to introduce a new word for wholes that are irreduble and decomposable We will refer to these wholes as assemblages 1
The variety of mechanisms of emergence discussed in this book show that the fears of the early emergentists that clockwork mechanisms could not possibly account for the irreducible properties of a whole, were mispl&ced. There is noting in the notion of a holistic property that emerges from the interactions between components that mits the potential complexity of those interactions On the other hand the concept of emergent property does place limits on the kinds of wholes tha can be legitimately accepted I particular it militates against the idea of a totality the parts of which are so inextricably related that their very identity is constituted by their relations witin the whole I a seamless totality the properties of components derive from the role they play in the whole so detaching them from it would auto matically deprive them from their dening characeristics. But none of the wholes we have discussed here, from thunderstorms to institu tional organizations, are seamless totalities The fac that the proper ties of a whole depend on the actual exercise of the capacities of its parts implies that removing one of them may indeed destroy the wholes identity but the part itself need not lose its own identity: pulling a live animals heart out will surely k it but the heart itself can be implanted into another animal and resume its regular function Thus wha we need is a concept that allows us to retain both irreducibility and decomposability, a concept that makes the explanation of synthesis and the possibility of analysis intelligible We cold, of course, simply build this requirement into the denition of the term "emergent whole but since this concept has a istory, and since in 18
The advantage of a new term is not limited to the fact that it is unblemished by past philosopical mistakes or that it is devoid of undesirable connotations It also signals a fresh startand acts as an incentive to get things right this time around So lts carefully dene what else should be included in the content of tis new concept First of all, the identity of an assemblage should always be conceived as te product of a istorical process, the process that brought its components together for the rst time as well as the process that maintains its integrity throug a regular interaction among its parts. Tis implies that the identity of an assemblage is always continent and it is not guaranteed by the existence of a necessary set of properties constitut ing an unchanging essence Or to put this differently assemblages are not particular members of a general category but unique and singular individuals. Even if two assemblages resemble each other so much that no one can tell them apart, each will still be unique due to the different details of its individual istory I the last few chapters it was argued that despite the fact that te word "individual has become synonymous with "person it can legitimately be applied to individual communities, individual organizations or individual cities And tis argument can be extended to the emergent entities of all other chapters: individual atoms and molecules, individual cells and organisms individual species, and ecosystems All hese different assemblages are born at a paricular time, live a life, and then die It follows that knowledge about an assemblage does not derive from a "botanical classication hat takes properies for granted but from an account of the origin and enduance of those properties Given the importance of the ontological status of assemblages we will need a techcal term to refer to it: every actual assemblage is an individual singulari A second ontological commitment must be built into the denition of the term "assemblage because these emergent wholes are dened not oy by their properes but also by their tendencies and capacities Tendencies can make the properties of a whle vary, sometimes even changing its identity as when an ice sculpture characterized by its solidity and its shape manifess its tendency to melt at a certain 185
APPENDX
tmpratur osing thos proprtis and much of its idntity Capaci tis, tu, mak whos xhbit aspcts of thir idntity that wr prviousy hiddn, as whn an innocnt ookng pant tus out to b poisonous or, on th contrary, to possss unxpctd haing powrs Wh tndncis mak a ist of ssnta proprtis ook fasy pr annt capacits xpod it sc for a givn who it woud b impossib to ist a th diffrnt ways in whch t can affct and b affctd by innumrab othr whos And vn f w coud ist a actuay manifstd tndncis and actuay xrcisd capactis that woud st not b nough sc nithr on has to b actua To tack this probm w usd th notion of th structur of a possibiity spac, a structur that xpains how tndncis and capacits can b ra vn whn thy ar not actua though w saw how varid this structur can b and acknowdgd many gaps in our concption of , it is saf to say that t is charactrzd by spcia or rmarkab trats th ontoogica status of whch is that of universal singularities Th trm "univrsa is usd for two ratd rasons rst, th vnts n whch tndncis ar manifstd and capacitis xrcisd may b ntiry diffr t n dtai and yt b shapd by th sam snguarits; and scond two sris of vnts ach constitutng a diffrnt mcha nism may nvrthss poss ss ovrappin ovrapping g possibiity possibiity spacs and ds pay common faturs that ar mchanismindpndnt2 mchanismindpndnt2 addtion to ths ontoogica commitmnts anothr rurmnt must b buit into th dniton of th trm "assmbag though ach assmbag is a uniu historica ntty it aways bongs to a popuation of mor or ss smiar assmbags In othr words, dspit th indvidua snguarity of ach assmbag th procss of assmby bhd it tnds to b rcurrnt so what is synthsizd is nvr a sing individua but many of thm Th rcurrnc itsf s xpaind by th fact that th assmby procss is govd by univrsa snguartis but th actuaization of th attr is aways subjct to contngnt vnts so what is gnratd s a popuaton n which variants ar distributd n a cain way Th fact that shard proprtis vary, and that th variation variation in a popuato n dspays a crta statistica dstribu dstribu tion, is anothr rason why th idntity of an assmbag shoud not b concivd in trms of a xd ist of proprts To facitat "popuation thinking w nd a mans to spcfy not ony t pos sib ways in which th mmbrs of a popuation can chang but aso 186
APPENDX
th stat of thir idntty at any paricuar pot in thir hstory Ths can b achvd by parametrizing th concpt of assmbag, that is, by providing it wth "knobs wth modiab sttings th vaus of which dtrmn th condition of th dntity of an mrgnt who at any givn tm. This mak t asir to pictur in our mnds an ntir popuaton of assmbags ach possssg a sghty diffrnt idntity dpnding on th sttgs of th paramtrs And it aso factat thiking of th popuation as an ntty that can undrgo an tns shock from th outsid that drasticay changs thos sttgs producng variants that go byond th norma rang of variaton short, a paramtrizd concpt can popuat our minds a way that brings variaton and th hstory of that varation wth On paramtr must spcify th ratv homognity or htrog nity of th componnts of an assmbag if in a givn popuaton a assmbags us th xact sam componnts thy shar a bttr dnd idntity than if thr is arg dgr of variaton thos com ponnts This paramtr must aso spcfy th stat of th boundaris of an assmbag sharp and xd or, on th contrary, fuzzy and uc tuating Th trm "boundary rfrs to vry diffrnt thngs dpnd ing on th assmbag th outr sh of ctrons in an atom th mmbran of a bactrium; th outr sk of a pant or anma; th gographica frontirs of an cosystm; th rproductv barrirs of a spcs Th paramtr nds a nam and sinc t party dtrmns dning boundaris w can ca it territorialization3 Th mor homo gnous th ntrna composition of an assmbag and th bttr dnd its outr boundaris th mor trritoriazd its idntty may b sad to b addition givn that tndncis and capacitis aso dtrmn dntity w must incud bhavora factors in ths param tr For xamp, on facor affctng th rang of capacits actuay xrcisd by an assmbag is mobty, snc a mob assmbag is mor ky to ncountr nov situations than an mmob on ths sns w may say that an ama that can mov around and that can affct and b affctd by a wdr rang of othr ntitis is mor dtrrtoriazd than a pant that is tid to th so For som assmbags w may nd mor than on paramtr sinc nw ways of dtrmining dntt hav mrgd hstoricay n nov mans of xing intity is th gntic cod, or mor prcisy, th compx machinry through which gntic nformation is rssd 187
APPENDX
Genes can exercise tight control over the mechanisms of emergence behind organic wholes, as when they coordinate the gradients and emergent forms of a developing embryo An obvious name fo this new parameter would be codin. An organism, for example, may be said to be highly coded if every detail of its anatomy is rigily deter mined by its genes and relatively decoded the environment also contributes to its anatomical detion And similarly for behavior: if the behavior of an organism is entirely inherited the coding param eter w have a high value and vice versa, if genes dene ony the connectivity of its neurons allowing much of its behavior to be deter mined by learning during its lifetime the parameter will have a low value Entities other than genes can affect this parameter Much as the territorialization parameter can apply to all sorts of boundaries so the coding parameter can be applied to other sources of information based constraints, such as those provided by language We can imag ine a community of paleolithic huntergatherers, for example, using its newly acquired power of speech to rigily categorize al the entities that are important for its continued existence: food could be coded as raw or cooked; sacred, taboo, or ordinary And similarly for dress, behavior, and other cultural expressions of its identity Another good example would be the institutional organizations belonging to an archaic state that code in written form al the material ows they control Both of these assemblages may be said to be highly coded Conversely, there may exist a few commercial organizations in the periphery of an archaic state that coordinate material ows using prices instead of ofcial categories and central comands, so as assem blages these orgazations may be said to be relatively decoded4 Let's summarize what has been said about the term "assemblage and then review the content of previous chapters in the light of this new concept First of all, the term must refer to concrete wholes that are irreducible ad decomposable, that is, it must accomodate the epistemological demands of both synthesis and analysis hese wholes must be assigned a clear ontological status Every assemblage must be treated as unique historical entity characterized both by a set of actual emergent properties (making it an individual singularity) as well as by the structure of possibility spaces dening its tendencies and capacities (a structure dened by universal singularities) his strucure strucure may be teed the diaram of the assemlage he historicay 188
APPENDX
contingent identity of assemblages must be allowed to change so the concept must be supplied with one or more parameters dening those changes And nally, assemblages must ways be thought as parts of populations in which their identities can chan ge within liits produc ing a certain statistical distribution of variation he potential range and possible lines of variation are determined by the diagram shared by all population members but their actual degree of similarity is determined by the values of the parameters at any given time the more territorialized or coded the members of the population are the more they wtend to resemble one another his way the similarities that tempt us to classify the population's members under a general category, and then to reify that category, can be given a concrete historical explanation Chapte 1 we examined the simplest way of creating an asse blage: placing two separate molecular populations (two bodies of air or water) at different temperature or pressure in contact with each other his, of course, creates a gradient ith the tendency to dissi pate and the capacity to act as fuel for other processes Because the two populations are identical in composition the ony heterogeneity involved is the intensive difference itself he territorialization para meter reduces in this simplest of al assemblages to the distance from thermodynamic equilibrium, that is to the degree of intensity of the gradient he diagram of this assemblage can be studied with the help of athematical tools like state space, its trajectories, and singulari ties he dimensions of the possibility space are the relevant ways of changing for an asseblage, that is, ts degrees of freedom there are interactions between these degrees of freedom the diagram is nonin ear possessing multiple singularities of different types, whereas if there are no interactions the diagram is linear structured by a single singularity of the steady-state type Finaly, some of the members of these molecular populations may be recruited to form the component parts of larger assemblages the identity of which will be determined in part by the linear or noninear nature of the diagram and in par by their own territorialization parameter Emergent wholes that form y dissipating a local gradient and reaching a lowintensity equilibrium such as a minimum of bonding energy in the case of ice crystals or a minimum of surface tension in the case of air bubbles, may be said to be highy territorialized hose that on the contrary, emerge when 189
APPENX
a local graent is nense enough and i is no allowed o ssipate, ike he pate of low denng a convecon cell, may be sad to be more deerritoriaized. Chapter 3 we generated the nex layer o complexty by ntro ducing chemical nteracions and he synthesis o novel molecules, grealy increasng the heerogeneiy heerogeneiy o the molecular populaons, as well as caalyss o manpulae the rapiity or slowness wih whch such syntheses ake place some cases the oucome o chemical nteracons s an equbrum concenration o he producs, a territo rialized isribuon o subsances wih a sable identy. others the caalyc control o reacon rates s perormed a loop-one sub sance accelerang he producon o anoher subsance tha n turn, acceleraes he rae o producion o he rsdrvng the molecular populaions away rom equbrium and leading o he emergence o more or less deerrtoriaized wholes: chemical clocks or more com plex auocaalyic loops Alhough he identity o an autocaalytic loop s sabzed by its coonens i s also subjec o change as new componens are added hat catalyze and are caalyzed by exsng ones. The iagra o hese assemblages can sill be suied using sae space but now the number o mensions o the space nd he ier enial relaions hat dene he sribuion o sngularites mus be allowed to change to reec the ac ha enirely new subsances can be produced and new chemical reacions made possible To keep track o he changng mensons o the possibty space we need a graph n whch he nodes represent subsances and he edges chemical reac ions. mporan sngularity · n he graph itse l is a percola ion hreshold, a crical degree o connecivity a whch the number o possible reacions changes rom ne to nnte Aer ths hreshold is reached the identy o he producs o chemical neracons can vary n an innite number o possble ways, or wha aounts o the sae thng the diagram becomes lled wih potenal es o deerritoriazatin. Chapter 4 added one more layer o complexty by ntroducng the capacity or sereplicaton nto molecular popuations A populaon o repicators such as molecules o nal
APPENX
mutans mutans ha evolve as a whole Too muc h accuracy in the replication replication process leads to a hgh value or the territoriaization parameer and a lower capacity to orm quasispeces But there s also a deerritor aizaton hreshold consiued by oo hgh a mutation rate ha ssolves he identiy o he muan cloud The agram o these assemblages mus capture the combnaorial possbes o he repi cators as well as the possble erences n repicaive success hat would resul i the population conroned lmted resources. pos sible RA sequences can be arranged so ha they have as neghbors other sequences ierng oy by one muation such a space a series o onemuant neighbors denes a possible pathway or iden tiy to change, a possble e o deerrioriaizaton that can take a molecule o RA and ransorm it step by sep nto an enirely ier en one Bu wheher such pahways can acually be ollowed depends on the isribuon o ness values. These orm a opography o val leys and hlls supermposed on the combnatorial space, a opography tha determines the elhood hat a given path be pursued paricular, paricular, scarcty o resour ces avor avor the clmbng o hills pnnng down a popuaton once is top has been reached, homogenizng its composon, and errioriaizng is dentity Escapng rom a ness peak can be done by genetc drit, hat s, by a series o changes hat are neural n ness and invisible to selecion pressures. Such an escape would constitue a relave deterrioriazaon populaons o semboed repicators here is only one param eter aecng he identity o assemblages Bu wih he emergence o he genetic code and the acquison o a mnimum o emboment through encapsulaion wihin a membrane, ideniy could now be dened n wo ierent ways As we saw n Chapter 5 he behavior o all ancent organisms was rgdly deermned by heir genes ha is, they could oy learn as a speces over many generaions, so he codng paraeer had a xed hgh value or several billion years But he erritoriazation parameer could change and lead o ierent ypes o behavior Early baceria lacked he abty o move and tended to orm colones tha accumulaed as layers a the nerace between waer and ottom sements To a casual observer hese baceria would hav looked just ike anoher sementary layer, moionless and uchanging n is anatomy, ha is, very erriorialied But enor mous changes were akg place wihn her membranes as he 191
APPENDX
APPENDX
retaliating against them, in the multiperson ase punishment is hard to fous and an have oletive unintended onsequenes, ike waves of retaliation that endanger solidarity and shared resoures Punishment is also ostly so there are inentives to let others do the punishing. Ts meas that an entirely new habit had to be added to the assemblage to stabilize its identity metanorms the habit of punishing those that do not punish heaters Although metanorms in prelinguisti huntergatherer ommunities do not yet onstitute a new type of oing, being siply a new territorializing devie, they do prepare the ground for the ventual rise of linguisti ategories transmitted from one generation to the next by enfored soial obligation. The iagram of these more omplex ssemblages of strategies a also be studied using payoff matries and Nash equilibria but as we saw applying these ideas to the multiperson ase is more omplated and less well understood One stable human ommunities emerged they beame apable of interating wth other ommunities ad form larger sale assemblages Some of these were relatively stable and enduring like the larger wholes onstituted by many ommunities assembled into a urban enter Others were relatively ephemeral but if they were reurrent they ould have interesting emergent properties of their own I Chapter 9 we explored a partiular reurrent interommunity interation, traing interations, the emergent propery of wh was pries These were not, of ourse, pries dened in monetary terms but in ustomary amounts of one objet in terms of another objet The rea son we neded to analyze trade was to explain the long distanes traveled by ghquality raw materials for the prodution stone tools One small urban enters speializing in longistane trade had ome into existene they an explain ts phenomenon but evdene for the gh mobilty of materials like lint predates the emergene of towns Either way in the absene of a powerful trade partner that an set pries or of a powerful government organization that an replae them wth entral ommands an assemblage of trading ommunities and the ow f goods between them is relatively deoded The diagram for two trading ommunities was stuied with the help of inifferene maps apturing the olletive preferenes of the ommu nities. These maps were ombined into an Edgeworth box to display the regon of the spae of possible ombinations of traded objets 196
i
.
ontaining the potential gains from trade I this region a singularity denes the optimal outome for the trading interation a singularity referred to as "ompetitive equibrium If atual trade always took plae at the singularity as eonomists sometimes assume, a assem blage of trading ommunities would be relatively trtorialized but as we saw this assumption is not neessary atual trading interations an take plae awa from equilibrium and still realze some of the gains from trade. I that hapter it was suggested that trade may have preeded the emergene of language but that temporal preedene is not stritly neessary language and trade may have oevolved not only wth eah other but with other praties like the eremonial burial of the dead As noted above the effet of language on human behavor (as well as on the emergent behavior of ommunities and organizations) is so unique that it must be assigned its own parameter in an assemblage words and sentenes just like genes have the apaity to ode identity But languages and genomes are themselves assemblages that an be studied independently of the role they play as a oding parameter Contemporary laguages, for example, an form a deter ritorialized ontinuum of dialets or a territorialized set of standard languages the boundaries of wh are arefully xed through the use of ofial iionaries, grammars and rules of pronuniation. I Chapter we approahed the mergene of language as a proess that began with monolit symboli artifats that were slowly transformed into the kinds of entities that ould form the parts of an assemblage a nite set of words that ould be syntatially ombined into an innite number of possible sentenes Going from a set of monolit arifats with a rigily dened identity to a set of reombinable omponents the identity and meaning of wh depends on what other words they oour with, is itself a proess of deterritorialization The iagram of these assemblages was stuied taking advan tage of the fat that the spae of all possible languages, as dened by their sntax, is losely linked to the spae of possible automata that an master that syntax The automaa an be onsidered singularities in the spae of omputational apaities, some exemplifying a mni mum others a maximum of omputational apaity Our nal theme was organizational assemblages possessing an authority struture was argued that the rise of suh assemblages 197
APPENDX
was dfcult to explain because of the active means that hunter gatherer and agricultural comuties had (and some stl have) to prevent the cystalization of authority: dssipating gadents of status or pestige to pevent leadership from becong pemanent, or dssi pating gradents of energy (suplus agricultural products by burng them or edstributing them to prevent the creation o a stock. But if these mechanisms o prevention wee necessary it was because the dagram of these comunal assemblages their social topoogy, already contained cental authority as a possibity. 6 Comunal assemblages were teritorialized to dferent degees more they had setted to practice agricuture and less if they remained mobile and anged ove a wider territory By neothic times oth possessed anguage and this gave them the means to code thei own identity. When the mcha sms o pevention ailed and chiefdoms and archaic states emeged these simpler comunities were not eft behind as a pimitive stage of development but coexsted with the more complex ones as pat of an expoited periphery. As such they were sujected to an ovrcoding as state organizations supeimposed thei own codes of conduct on them withou epacng the native ones7 other words comuna assembages became a component part of arge assemblages retanng some o thei own identity but being foced to exercise capacties in the sevice o the production and maintenance of the identty o a ager whoe The subject o centralized authority s compex because t involves several eves of the part-to-whoe reaton There s the argest assem bage in which a cental state occupes a core suounded by arge chiefdoms and a perphery of smal chiefdoms and agrcltural com munites Then there s the assemblage of the captal city tsef, made up o a varety of communites and oganizations (tempes, monuments, wokshops, esdences, and the oya paace tse) as well as the physcal nfastrcture ng a these together. Fnally, there s the nividual organizatons acting as nsttutona agents with their own goas and the resources to acheve those goals. The terrtoriazation paameter can ; have dierent values o each eve of the part-t o whoe relaton but in general the rse of a cental state can be conceved as the product o a powerfu deterrtorializaton. t the argest scale t invoved a change n the way the dentty o a social whoe s elated to ts physcal territoy: commnal assembages occpy a teritoy and 198
APPENDX
the piece of land on which they are settled or around which they roam is part of what makes them who they are state organizations controlng a vast territory, on the other hand, detach themsel ves from the land to be able to make objective comparisons about its agricultural productivty, assigng dffeent value to dffeent plots for adstrative puposes 8 At the smallest scale an indvdual orga zation can be sd to be deterritoriazed if the resources it controls are nked to an ofce as opposed to the incumbent o that ofc adtional legitimacy, what made the authoity of an Egyptian pharaoh ow down ecty om the gods, left plenty of room for personal caprice so oce and incumbent wee not separate. But the legitimacy derved from the capacity of an organization o sove problems o secuity, logistics monumental construction, and other actities in which the outcome depends on the successu matching of means to ends, was more impersonal. The coexistence o dffeent sources of egitimacy impes that there wee two poles deing orgazational assembages in archaic states, the magical and the egal-rational9 These two poles can be seen materiazed in the Giza pyid, a monument that was at once a resurrection machine and a eat o structual engineering Chapter 1 1 we concentrated on the second poe at the smalest scale: the laws, chain of comand, task detions and assignments, and contro o near and far resource gadents though which govenment organizations solved the poem of pyraid builing We saw that the iagam of these orgazational assemages coud e studed with the hep of graphs of dfeen types some to capture the hierarchica reatons in a chain of command others to ispay the dvison of abor within each of the aks The rst gaph reveas whether, for exampe, there are posse concts of athorty in the case of a rak whose members must obey commands from more than one iect boss The second graph dspays the intedependencies among tasks that speciazed workes must take nto account as they carry out those commands, nterdependencies that mark points of possbe conct in schedues or n the use of resources These poten tal authorty or perormane concts are ile snga events struc tung the space of possibites for the assemby process of giant unerary monments and other arge-scale goverment pojects. They party specy the probem that an ogazaton must s ove much 199
APPENDX
as othe diagams dene poblems fo othe assemblages: the state �pace of copled molecla poplations denes a poblem fo a gadi ent' the poblem of what low patten to geneate to dissipate as much enegy as possible; the coupled spaces fo possible genotypes and phenotypes dene poblems fo evolving oganisms, poblems solved by seaching these spaces fo soltions; the oppotunities and isks affoded by the envionment dene poblems fo leaning oganisms the soltions of which involve the development of associations of stimuli, good habits, o even skills; and the social dilemmas geneated by the choice to coopeate o cheat dene poblems fo animals o hmans involved in indenitely epeated stategic inteactions. O chaacteizatio of assemblage theoy needs only one moe featue to be complete The identity of an assemblage is not only embodied in its mateiality but also expessed by it. This distinction coesponds to that between matte-enegy on one hand and info mation on the othe, not the semantic infomation conveyed by the meaning of wods o sentences, but aw physical patten A live gadient, fo example, may contain the same amount of enegy as one that has become dissipated but it contains moe infomation because its comosing molecles ae odeed and have patten In addition to dening the degee of ode in an assemblage physical infomation can expess its identity To use a simple example, atoms can expess thei chemical identity because one of thei popeties (a cetain distibution of electons) gives them the capacity to inteact with adiation and leave a uniqe "ngepint in it: each electon shell absobs some wavelengths of the adiation bt not othes ceating an infomation patt that can be used to identify the chemical species of the atom Astophysicist use these pattes (the pattes that the ngepints leave in photogaphic plates) to identify the components of fa away stas, bt even if no one used this infomation the expes sive pattens wold still be thee It was becase the wold in which living ceates evolved was lled with infomationfom the nti ent gadients that a bacteium can climb; to the odos, colos, and shapes that a honeybee can use to nd necta; to the silhouette o gait patten that a pedato can se to identify its peythat they wee able to stat foming intenal models of thei envionment While the distinction between the mateial and the expessive, between matteenegy and infomation, is impotant to tack the 200
APPENDX
paallel histoies of bodies an minds, it is also elevant hee because the compute simulations discssed thoghout this book ae eme gent wholes composed of infomation existing above the compte hadwae that povides thei mateial and enegetic sbstatm. The technology that makes simlations possible had to undego seveal tansfomations that can also be explained within the famewok of assemblage theoy The most impotant deteitoialization was the tansfomation that conveted a special-ppose Ting machine, igidly dened by its inteal states and the symbols it ses to wite on its memoy tape, into a nivesal Ting machine capable of simulat ing any specialpupose one This simlating capacity deives fo the coespondence between atomata and langages, a coespon dence that allows a specialpupose atomaton to be given a symbolic expession and be placed on the memoy tape of the nivesal automaton. Although these automata ae conceptal entities they can be given a mateial fom that clealy displays the deteitoializa tion involved: when a specialpupose Ting machine is mateially embodied its identity is xed by its specialized tas bt a univesal Ting machine has an identity that is programmable dened by what eve specialppose machine it is cuently simulating Thus the st deteitoialization bings the metalevel (opeations on data) into diect contact with the object level (data) ceating the possibility of assemblages mae out o operators an ata The components of a genetic algoithm, fo example, can be divided into opeatosmutation and cossove, tness fnction, selection fnctionand data, the popula tion of symbol stings constituting the simlated chomosomes As with any assemblage these components must inteact, the inteaction edcing in this case to the application of an opeato to the data sev ing as its inpt These inteactions, on the othe hand, must be choeogaphed since the ode in which opeatos ae applied the nmbe of times they ae applied mst be caeflly specied This involves an additional componen fo the assemblage: contol stctes like loops and banching instctions All togethe, opeatos, contol stctues, and data allow an automaton with xed identity to be coded in the memoy tape of a univesal atomaton tansfoming it into a piece of softwae: an application like a wod pocesso, a spead sheet, o a web bowse This implies the possibility that the identity of a coded 201
APPED
APPENDX
automaton can undergo changes through decoding, a process that depends on the type of prograng language used to express it sybolcaly. Old prograng lnguages (lke Fortan, Pasc, or C) control coputaton processes n a rigid herarchcal way: a aster progrm, ebodying the basc dentity of an applcaton yields control whenever it cals a subroutne to perfor a particular task but t recovers it the oent the subroutine nshes ts task The latter ay in turn surrender control to an even more basc subprogra but eventuay control of the process moves back up the herarchy to the master prgra other prograng languages referred to as bjeriene there are no aster progras or subroutnes Control s always decentralzed as software objects encapsulatng a set o operators are caled nto acton by pattes n the very data they operate on Ths ples that the dentty o an applcation s dened dynaicaly n nteracton with the data: the latter changes so l withn ts the applcaton's own dentty If deterritoralzaton transforms a pece of hardware into software decoding elnates a rgid master progra in favor of a populaton of autonomous and exibe sotware objects These two transormatons provded the envronent in whch the sulatons scussed n ths book could be bo and thrive On the other hand, whie a deterritoralzaton may explain the existence of assemblages of operators and data t does not account for the act that the behavor of those operators as they act on data can c the behavor of ater processes Chapter t was argued that an explanaton of the isomorphism between a athematic odel and the process it models can be given the obective exis tence of the dagrams of assembla ges is accepted. The unvers singularities structuring these diagrms are by detion re bu not actu althogh they can become actualzed when tendencies are anfested or capactes exercised The unversality of the structure of possibilty spaces eans that two entirey different processes can share the sae diagra or that heir diagras can oerlap so the processes share some of the �ame sngularities In these terms, the isomorphsm between moels and what they model can be explained as a aualizain of the sae diagram or o different but overlapping a grams that chapter we went on to argue that the man danger o ths account is making uiersal singularties into transcendent 202
entites entities existng entirey dependenty o the materia wor But ths potentia pita can be avoided by aways treatng diagr as mmanent to matter, energy and oaton wie te objectie exstence o diagrms may not depend on any particuar atera energetic or ioration mechism does depen on the actu existence o some mechais or another ths account turns out to be correct then it w pot to an tmate between ontology and epistemoogy Ad the exstence o such a tu w constitute a poweru argument for breang th the ontoogy we inerted from the cassica Gree posophers an ontoogy based o the gen era and the partiar and an centive to deeop a new one base on te vidu sgar and te uvers sguar ,
203
NOTE
Notes
Introduction 1 John Start Mill A Sytm f Lgic Raticinativ and Inductiv (London: Longans Green and Co 1906) p 243 2 George Henry Lewes Prbm f L and Mind Vol 2 (London: bner & Co 1 875) p 415 3 Samel Alexander Spac, m and Di Vol 2 (London: Macan 1 920) p 4647 See also: C Lloyd Morgan Emrgnt Evutin (ew Yor Henry Holt 1931) p 8
Chapter One 1 Peter Atns Th Scnd Law (ew Yor: Scientic American Library 1984) p 73 2 Peter Atns Ibid p 38 3 Eric D Schneider and Dorion Sagan Int th C Enrgy Fw Thrmdynamic and if (Chicago: University of Chicago Press 2005) p 1 1213 4 Gregoire Nicolis and ya Prigogne Ering Cmpxi (ew Yor: H Freeman 1989) p 12 5 Joseph M Moran and Michael D Morgan Mtrgy Th Atmphr and th Scinc f Wathr (New Yor: Macan Pblishng Co 1986) p 287 6 Joseph M Mora and Michael D Morgan Ibid p 71 Joseph M oran and Michael D Morgan Ibid p 301 and 31 5 8 Gregoire Nicolis and ya Prigogine Expring Cmpxi Op Cit p 1820 9 lan Garnel Frm f Eanatin (ew Haven: Yale University Press 1981) p 5862 204
1 0 Peter Smith Eaining Cha (Cambridge: Cambridge University Press 1998) p 72 am J Kafan and Larry 1. Smarr Suprcmputing and th Tranfrmatin f Scinc (New Yor: Scientic merican Library 1 99 3) p 141 5 Credit for this landmar simlation is given to Robert ilhemson It made its debt at SIGGRAPH 89 and was noinated for an Academy Award for anmation Detals and graph ics can be fond n the website of the National Center for Spercom ptng Applications 12 am J Kaffman and Larry 1 Smarr Ibid p 1 3 1 3 Egene P igner Th Unranab Ectivn fMathmati in th Natura Scinc I Symmetries and Reections (oodbridge: Ox Bow Press 1979) p 22237 14 Morris Kline Mathmatica Thught m Ancint t Md m Vol 3 (ew Yor: Oxford University Press 1 972 ) p 8 82 15 Morris Kline Mathmatica Thught m Ancint t Md m Vol 2 Ibid p 732 16 Jne BarrowGreen Pincar and th Thr Bdy Prbm (Providence R: American Mathematical Society 1 997 ) p 323 1 Ian Stewart D Gd Pay Dic Th Mathmatic f Cha (Oxford: Basil Blacwe 1989) p 107 18 Alexander oodco and Monte Davis Catatrph Th (ew Yor: EP Dtton 1 978) p 42
Chapter o 1 am Pondstone Th Rcuriv Univr (New Yor: am Morrow 1985) p 2631 2 am Pondstone Ibid p 8 2 3 am Pondstone Ibid p 3840 4 am Pondstone Ibid p 105 5 am Pondstone Ibid p 202 6 The implementation of this machine was perfored n the year 2000 by Pal Rende and can be seen n operation at his website Stephen olfram Universality and Complety n Celllar Atomata I Cuar Autmata and Cmpxi (Readng: Addisonesley 1 994) p 14055 205
NOTE
OTE
Stephen Worm o Dimension Cear Automata n Cellular Automata and Complexi bid, p 2 1 3 Christopher G Langon e at the Edge o Chaos n Articial ife II
4
5
Edited by Cistopher G Langon, Chares Tayor, J Doyne Farmer and Steen Rasmussen (Redwood City: Addison -Wesey, 1992) p. 1 0 Cistopher G Langon bid, p 83 1 1 ' Stephen Worm A N Kind of Science (Champagne: Worm Media, 2002) p 845 1 Bros Hassacher Discrete Fds n From Cardinals to Chaos Reections on the e and ega ofStanislav am Edited by Necia Grant Cooper (New ork: Cambridge Uiversiy Pres 1989) p 1 8 1 1 Tomasso Tooi and Norman H. Margous nveribe Cear Automata a Reiew n Cellular Automata Theo and Eeriment Edited by oward Gutowitz (Cambdge: M Press, 1 9 9 1 ) p 2 3 ] 4 Bros Hassacher Discrete Fds Op Cit p 1889 ]5 Stephe Woram Theodynics and Hydrodyics o Cear Automata n Cellular Au tomata an d Complexi Collected Essays (Read ing: AddisonWesey, 1994) p 260 ]6 Tsumu Shimomura, Gary D Dooen, Bros assacher, and Castor Fu Cacuations Using Lattice Gas Techniques n From Cardinals to Chaos Op Cit p 20210 1 Tomasso Tooi and Noan H Margous Cellular Automata achines A N Environmentforodeng (Cambdge: Press, 1 9 87 ) p 1 5 3 1 Tomasso Tooi and Noan H. Margous nveribe Ceuar Automata Op Cit. p 2445 1 Peter Gaison Computer Simations and the ading Zone n The Disuni of Science Edited by Peter Gaison and David J Stump (Standord: Standord Uiversity Press 1996) p 1384
6
p 1 68 1 0 Water Fontana. bid p. 179 11 Water Fontana. bid. p. 184
Chapter Four 1 Manred Eigen. Steps Towards e (Oxord: Oxord University Press, 1992) p 28 Manred Eigen bid, p 657 Peter Schuster ow Do RNA Moecues and Viruses Eore Their
4
5
6
Chapter Three 1 Ronad E. Fox Energy and the Evolution ofe. (New ork: WH. Free man 1988) p 89 2 Stuar Kauan The rigins of rde Se-rganization and Selection in Evolutio (ew ork: Oxord Uiversity Press, 1993) p 2906 Daid Dresser and unngon Poter. Discovering Emes (New ork: Scientic American brary, 1 9 9 1 ) p. 16872 And: NC Veitch and
J.P Wiiams. The Moecuar Basis o Eecron anser in Redox
Enzyme Systems n Frontiers of Biotransformation Vo. 7 (Berin: Akademie Verag 1992) p 28390 Stuart Kauman. The Origins o Order. Op Cit p. 3023 Stuar Kauma. bid p. 30910 Stuar Kauman. bid p. 320 Richard J. Bagey and J Doyne Farmer. Spontaneous Emergence o a Metaboism. n Artcial ife Edited by Christopher G. Langton Chaes Tayor, Doyne Farmer, and Steen Rasmussen (Redwood City: Addison-Wesey, 1992) p 1 1 2 Richard J Bagey and J Doyne Farmer bid, p 11826 Water Fontana Agoritmic Chemistry n Articial e Op. Cit.
10 11
I .�
Words? n Complexi etaphors odels and Reali Edited by George A Cowan, David Pines Daid Metzer. (Reading: AddissonWesey 1994) p. 400 Manred Eigen Steps Towards e Op. Cit. p 925 Stuar Kauman The rigins of rder Serganization and Selection in Evolution (New ork: Oxord University Press, 1993) p 345 Stuar Kauman bid p 4562 Manred Eigen. Steps Towards Lie Op Cit p 289 Manred Eigen bid p. 826 Meanie itche and Stephanie Forrest Genetic Agorithms and Aricia Lie n Articial ife Edited by Christopher Langton. (Cambridge: M Press 1997) p 272 David E. Godberg The Design of nnovation essons om and for Competet Genetic Algorithms (Boston: Kuwer Academic Pubishers, 2002) p. 3 David E Godberg. Geneti Algorithms in Search ptimization ad achine eaing (Reading: AddissonWesey 1989) p 12530 20
NOTES
17
18 19 0 1
John H. Holland. Hidden Order How Adaptation Builds Compli (Readng Addison-Wesley 19 95 ) p 11 11 3 Peter T. Hraber Terry Jones and Stephanie Forrest The Ecology of Echo. In Artal Le, Vol 3 No 3 ( Cambridge: M Press 1997). p 1 1 9 . Sorin Sonea Bacterial Evolution Without Speciation In Symbiosis as a Source ofEvolutiona Innovation. Op Cit p 002 John H Holland Hidden Order Op Cit p 1 20 John H. Holland Ibid, p. 2223 Mark A. Bedau, Emle Snyder Norman H. Packard A Classication of LongTerm Evolutionary Dynaics In Artal Le V. Edited by Christoph Adai ichard K. Belew Hiroaki Kitano and Chales E Taylor (Cabridge: M Press 998) p 232
Chapter Si
3
4
5
6 7 9
10
210
John H Holland Hidden Order How Adaptation Builds Compli (Readng: AddisonWesley 1995) p 32 Ronald J MacGregor and Edwn R Lewis Neural Modeling (New York: Plenum Press 1977) p 230 John T Bonner The Evolution of Culture in Animals (Princeton: Princeton University Press 19 80) p 1 3 James L Gould Ethological and Comparative Perspectives on Honey Bee Leag In Insect Leaing Edited by Daniel R Papaj and lcinda C Lewis (New York: Chapman and Hal, 1 993 ) p 3 James L Gould Ibid p. 3 18 Eizabeth A Beays. Aversion Leag and Feedng. In Insect Leaing Op Cit p 0 J Mackntosh Conditioning and Assoative Leaing (Oxford: Clarendon Press 983) p 8992 N J Mackintosh Ibid. p 2 Wlliam Bechtel and Adele Abrahamsen Connectionism and the Mind An Introduction to Parallel Distributed Processing in Newor (Camridge: Basl Blackwell, 199) p 067 Karl Ss volving 3D Mohology and Behavior by Competition In Artal Life I Edited by Roey A Brooks and Pattie Maes. (Cambridge M Press 994) p 33
NOTES
13
14 15 1 17
18
19 20
1
James J Gibson The Ecological Approach to Visual Perception (Boston: Houghton Miflin 1979) p 16 The anmation called "Evolved Virtual Creatures may be watched at the author's website. David E Rumelhart and James L. McCleland PDP Models and Gen eral Issues Cognitive Science. In Parallel Distributed Processing Exlorations in the Microstructure of Cognition Vol Edited by David Rumelhart and James McCleland (C ambridge: M Press 1986) David E. Rumelhart, Geoffrey E nton and Ronald J . Wlliams. Leaing Inteal Representations by Error Propagation In Ibid. p 322. Nel Gershenfeld The Nature of Mathematical odeling. (Cambridge: University Press, Cambridge UK 9 99 ) p. 79 Wlliam Bechtel and Adele Abrahamsen Connectionism and the ind Op Cit. p 32 Paul F M J. Verschure and Anthony C C Coolen Adaptive Fields: Distributed Representations of Classically Conditioned Associations In Nework: Computation in Neural Systems Vol 2 No 2 (Infoa Healthcare Onie Journal 99) p 89206 Geoffrey nton James MClelland and David E Rumelhart Distributed Representations In Parallel DistributedProcessing. Op Cit. p 867 And: T Van Gelder. Denng "Distributed Representation In Connection Sence Vol. 4 No 3 (Oxfordsre: Carfax 992 ) p. 7687 Wlliam Bechtel and Adele Abrahamsen Connectionism and the ind Op Cit p 268 Filippo Menczer and ichard K Belew Latent Energy Environments. In Adaptive Individuals and Evolving Populations. Edited by ichard K Belew and Melaie Mitchell (Reang: Westview Press, 9 96) p. 19 6 James L Gould. Ethological and Comparative Perspectives on Honey Bee Leag In Insect Leaing Op Cit p 9
Chapter Seven I
Dael L Schacter and Endel lving What e the Memory Sys tems 1994 In Memo Systems 1994 Edited by Daniel L. Schacter and Endel lving (C ambridge: M Press 994). p 28 Daniel L Schacter and Endel lving. Ibid. p 9
211
NOTES
NOTES
3
4
6 7 8
9 10 1
12
13 14 15 16
17
18
19
20 21
- 2 12
Stephe Walker Animal Thought (Lodo: Routledge d Keg Paul, 1985 ) p 243 Howard Eichebaum The Hippocampal Syste and Declarative Memo in Humans and Animals In Memo Systems 1994 Op. Cit p 152-5. Stephe Wker Animal Thought Op Cit. p. 2578. Paul M. Churclad. The Engine of Reason the Seat of the So ul (Cambridge: M Press 1996) p 40-5 Paul M Churchland Ibid p. 47-5 1 . Derek Blackma Operant Conditioning A n Eerimental Ana lysis of Behavior (Lodo: Methue, 1974) p 48 Derek Blackma. Operant Conditioning Ibid p 35 ad 97. David G Mayers Eloring Pchology (New York: Worth Pubishers 2005). p. 345 A Harry Klopf, James S Morga ad Scott E. Weaver. Modeig Nervous System Fuctio with a Herarchical Network of Cotrol Systems that Lear I From Animals to Animats 2 Edited by JeaArcady Meyer Herbert 1 Roitblat, ad Stewart W Wilso (Cambridge: M Press 1993). p. 255 Federico Ceccoi ad Domeico Parisi Neural Networks with Motivatioal Uits. I From Animals to An imats 2 Ibid., p 3467 Stephe Wker. Animal Thought Op. Cit. p. 359-64 Roger C. Schak d Robert P Abelso Scripts, Plans Goals and Understanding. (Hillside: Lawrece Earlbaum 1977) p 18. Roger C Schak d Robert P belso Ibid p. 42. Risto Miikkulaie Submbolic Natural Language Processing: An Integrated Approach to Scripts Lexicon, and Memo (Cambridge: M Press, 1993) p. 5-6 Jeffrey 1. Ema. Finding Structure in me Cognitive Sence, ol 14. (Norwood: blex Pubishing 1990) p 179211 Teuvo Kohoe Se-Organizing Maps (Bern: Sprger 2001) p 0619. Bed Fritzke. Growig SelfOrgzig Networks I Kohonen Maps Eited by Ekki Oja ad Samuel Kaski (msterdam: Elsevier 1999). p. 1313 Rsto iulaie. Subsymbolic Natural Language Processing Op. Cit p. 1929. Risto Miulaie. Ibid. p. 125.
22
23
24
2
26
Phlip E. Agre. Computatioal Research o Iteractio d Agey I Computational Theories ofInteraction and Agen. Eited by Phiip A. Agre d Staley J Rosese (Cmbridge: Press 1996) p 12 Yves Lesperace ad Hector J. Levesgue Idexical Kowledge ad Robot Actio I Computational Theories ofInteraction and Agen Ibid p 443 Lokedra Shastri Advaces n Shruti A Neurally Motivated Model of Relatioal owledge Represetatio ad Rapid Iferece Usg Temporal Sychroy I Applied Intelligence ol. 1 1 No 1 (msterdam: Spriger Nethelads 1999) p 79108. Paola Baldasari Paolo Puiti Aa Motesato d Guido Tasci. SelfOrgaizg Maps ersus Growig Neural Gases n a Robotic Appicao I Computational Methods in Neural Modeling. Vol 2. Eited by Jose Mira d Jose Alvarez (Bern: SprigerVerlag 2003) p. 2013 Nelso Goodma Seve Strictures o Similarity. I Problem and Projects (Iiaapois: BobbsMerrl, 1972). p. 445
Chaper Eight I
2
3 4
5 6
7
Fras B M. De Waal. Social Sytax: The IfThe Structure of Social Problem Solvig I Animal Soal Cmplexi Eite d by Fras B. M De Waal ad Peter 1 ack. (Cambridge: Harvard Uversity Press 2003) p 237-46 Duae Quiatt. Laguage d Evolutio the Middle Stoe Age I Language Evolution Eited by Gabor Gyori (Frakfurt: eter Lg 2001) p 29-34 Samuel S Komorita ad Craig D Parks Soal Dilemmas (Bolder: Westeriew Press 19 96) p 12-1 3 Robert Axelrod The Evolutio of Cooperatio (New York: Basic Books, 1984). p 59. Robert Axelrod Ibid p 42 Sus E. Rechert. I Game Theo and Animal Behavior Le A. Dugatkn ad Hudso . Reeve (Oxford: Oxford Uversty Press, 1998) p 77. M. Keith Che ad Marc Hauser Modeng Reprocatio ad Coop eratio n Primates: Evidece for a Puishing Strategy. I Joual of Theoretical Bio logy Vol 235. (Amsterdam: Elsevier 2005) p 6-7 213
NOTES
NOTES 8
9 10 11 12 13
14 15 16 17 18
1 20 21 22 23
24
214
Sarah F Brosnan and Frans B de Waal. A Proxate Perspecive on Reciprocal Altruism. In Human Nature Vol 1 3 No 1 (New York: Walter de Gruyter, 2002 ) p. 148 Robert elrod The Complexi of Cooperaion (Prnceton Prnceton Unversity Press 1997 ) p 1 7-2 1 Robert Axelrod The Evolution of Cooperation Op. Cit p 54 Martn A Nowak EvolutionaDnam i (Cambridge: Belknap Press 2006) p 8690 Ken Bore Game eo and the Social Contract (Cambridge: Press, 1998) p 31421 Joshua M Epsten and Ross Hamond In Joshua M Epsten, Gene tiveSocial Science (Prnceton Prnceton niversity Press, 20 06 ) Chapter 3 Ken Binmore Game Theo and the Social Contract Op Cit p 319. Brian Skyrms The Dynamics of Rational Delberation (C ambrdge: Harvard niversity Press, 1 990) Chapter 6 Ken Binmore Game Theo and the Social Contract Op Cit p 403 David W Stephens and Kev C Cle ments. Game Theor and Leag In Game Theor and Amal Behavior. Op Cit p 244 Michael Macy Natural Selection and Socal Leg n Prisoner's Dlemma In Sociological Methods and Research Vol 25, No 1. (S age Publications, On Lne Joual, 19 96 ) p 12 24 Martn A Nowak Evolutiona Dnami Op Cit p 79 Michael Macy Natural Selecion and Social Leang n Prisoners Dlemma Op Cit p 125 Ken Bore Game Theo and the Social Contract Op Cit p 2946 Marn A Nowak. Evolutiona Dnamics Op Cit p. 14559 Jon W Pepper and Barbara B Smuts The Evolution of Coopera tion n an Ecological Context: an AgentBased Model In Dnamics in Human and Primate Societies Edited by othy A Kohler and George J. Gumerman (O xford: Oxford University Press, 1 99 8) p 634 Domenico Parisi Federico Cecconi, and tonio Cerni Kin Dreced Altruism and Attacment Behavor n an Evolving Popula tion of Neural Nets In Artcial Societies e Computer Simulation of Social Lze Edted by Nigel Gilbert and Rosaria Conte (London nversity College of London Press, 1 995 ) p 23 848
25
26 27 28 2
30
Chrstph Hauert Cooperaton Collecives Formation, and Special zation In Advances in Complex Systems, Vol 9, No 4 (Sngapore World Scientic Publisg 2006 ) p 3 16 Thomas C Schellg Micromoives and Macrobehavior (New York: W. W Norton 19 78) p. 1 104 Robert elrod. The Complexi of Cooperation. Op. Cit p. 52 Robert elrod Ibid, p 525 Tomohisa Yamashita, Koshi Izu, and Koii Kurumatani Effec of Mutual Choice Metanorm n Group Dynamcs for Solvng Social Dlemmas In Agent-Based Simulaion From Modeling Methodologies to Real World Applications Edited by Takao Terano Hajme Kta, Toshijuk Kaneda yoshi Arai and Hiroshi Deguchi (Tokyo: SprngerVerlag 2005). p 3846 Ken Bnmore Game Theo and the Social Contract Op Cit p 34458
Chaper Nine 1
2 3
4 5 6
7
Mar D Leakey Cultural Pattes i the Olduva Sequence In Aer the Australopithecines: Stratigraphy, Ecology, and Culture Change in the Middle Pleistocene Edited by Karl W Butzer and Glynn saac. (The Hague: Mouton Publishers, 1975) p 477 SA. Semenov Prehistoric Technology (Totowa Baes and Noble Books 1985) p 3955. Clve Gamble Exchange, Foragng, and Local Homid Networks In ade and Exchange in Prehistoric Europe Edited by Cris Scarre and Francis Healy (Oxford: Oxbow Books, 1993) p 36 Chris Scarre. Introduction In Trade and Exchange in Prehistoric Europe id, p 1 Marshall Sahns Stone Age Economi (Chicago: leAtherton, 1972) p 27780 Marshall Sahlns id, p 10-12 Harold 1. Dibble and Nicolas Rollard On Assemblage Variablity n the Midle Paleolithic of Weste Europe In The Middle Paleolithic Adaptation, Behavio and Variabili Edited by Harold 1 Dibble and Paul Melars (Philadelpha: niversity Museum, niversity of Pennsylvaa 1992 ) p 11 1
215
NOS
NOS 8 Lawrence Guy Straus Even the Notion of "ansitional dustry
23 Mark Winter Lake GICAL Computer Smulation of Mesolithic
Is a Suspec pological Construct I New Approahes to the Study of Upper Paleolithi Transitional Industries" in Weste Eurasia Edited by ulien RielSalvatore and Geoffrey A. Clark (Oxford Bar 2007) p. 11-15 Leonard Joy One Economst's View of the Relation Between Eco nomcs and nthropology I Themes in Eonomi Anthropology. Edted by Raymond Firth (London: Routledge 19 67) p 34 Paul Woacott and Ronald Woacott Eonomis (New York: ohn Wey Sons 1990) p 393 Thomas M Humphrey The Ealy Histor of the Box Diagram (Federal Resere Bank of Rihmond Eonomi Quarterly Vol 21 Winter 1996) p 37-5 Leigh Tesfatsion. AgentBased Computational Eonomics: A Co structive Approach to Economc Theory I Handbook of omputa tional Eonomis, Vol 2. (Amsterdam: Elsevier 2006) p 4550 Dan K Gode Stephe E Spear and Shyam Sunder Convergene of Double Autions to Pareto ptimal Alloations in the Edgeworth Box (Yale School of Management working paper 2 004) effrey S Dean George Gumerman oshua M. Epsten Robert L Axtell Alan C Swedlund Mles T Parker and Stephen McCarroll Understandng nasazi Culture Change Through AgentBased Modeling I oshua M Epstein Generative Soal Sene. (Pnceton: Pnceton Unversity Press 2 006 ) p 99 oshua M Epstein and Robert Axtell Growing Artal Soeties (Cambridge: MIT Press 1996) p 236 oshua M Epstein and Robert Axtell Ibid p 30-2. oshua M Epstein and Rober Axtell Ibid p 106. oshua M Epstein and Robert Axtell Ibid p 1 09- oshua M Epstein and Robert Axtell Ibid p 1 20. Steven Mithen Smulatng Prestoric HunterGatherer Societies I Simulating Soeties Edted by Nigel Gber and m Doran (Lodon: UCL Press 1994) p 113 effrey S Dea et al Understandng nasazi Culture Change Through AgetBased Modeling Op Cit. p 9 Steven Mithen Smulating Prestoc HuterGatherer Societies Op Cit p 179 Steven Mithen. Ibid p 1 6970.
Foraging I namis in Human and Primate Soeties AgentBased Modeling of Soal and Spatial Proesses. Edted by othy A. Kohler and George Gumerman (O xford Oxford University Press 20 00 ) p 1156 Mark Winter Le. Ibid p 1 1 Mark Winter Le Ibid p 1 30 Mark Winter Lae. Ibid, p 137-8 Grame Clark World Prehisto . (Cambridge: Cambridge University Press 1977) p 64 effrey S Dean et al Understandng nasazi Culture Change Through AgentBased Modeling Op Cit p. 94-100 effrey S Dean et al Ibd p 10 3 effrey S. Dean et al. Ibid p 1067
9 1 II 12 13 4
15 1 17 18 19 2 21 22
24 25 2 27 28 29 3
Chapter Ten 1 Mark D Hauser and W Tecumseh Fitch at Are the Uniquely
2
3 4
5 7
Human Components of the Language Faul? I Language Evolution Edted by Morten H Chrstiasen and Smon Krby (Oxford Oxford Unversity Press 2003 ) p 16 1- Sonia Ragir. Towards an Understanding of the Relationship beeen Bipedal Walking Enephalization and Language rigins. I Language Evolution Edted by Gabor Gyry. (Frakfurt: Peter Lang 200 1 ) p 5 Richard Y Kain Automata Theo (New York: McgrawH 1 972 ) . p 4-90 (Tring machnes) 1 22 (linearbounded automata) 142 (pushdown automata) Luc Steels and Frederic Kaplan Bootstrapping Grounded Word Semantics I Linguisti Evolution through Language Aquisition Edted by Ted Briscoe (C ambridge: Cambridge University Press 20 02 ) p 56-7 Luc Steels and Frederic Kaplan Ibid p 59 Luc Steels and Frederic Kaplan Ibid p 7 Edwin Hutchins and Brian Hazlehurst Leaing in the Cultural Proess. I Articial Life Edted by Christopher G Langton Charles Taylor Doyne Farmer and Steen Rasmussen (Redwood City: AddsonWesley 19 92) p 69 3
NOTES
NOTES 8 9 10 11
12 13
1 15 16
17 18
19
20
21
22 23 2 25
Edwn Hutcns and Ban Hazlehurst Ibid., p. 6978 Edwn Hutchns and Ba n Hazlehurst. Ibid p 7024 Noam Chomsky Aspects of the Theo of Synt (Cambdge: M Press 1 9 6 5 ) . p 6673 Natala L. Komarova and Marin Nowak Language, Leaing and Evolution In Language Evolution. Edited by Morten H. Chstiansen and Smon Kirby Op Cit p. 3268 Steven Pinker Language as an Adaptation to the Cognitive Niche In Ibid p 258 Simon Kirby. Leag Bottlenecks and the Evolution of Recursive Synta In Linguistic Evolution through Language Acquisition Op Cit. p 1778 Smon Krby Ibid p 184 Smon Kirby. Ibid p. 192 James R. Huord. Eression/Induction odels of Language Evolution: Dimensions and Issues In Linguistic Evolution through Language Acquisition Op. Cit. p 30611 Son Kirby Leag Bottlenecks and the Evolution of Recursive Syntax Op Cit p. 189 J Mark Baldwn A New Factor n Evolution Adaptive Individuals in Evo Populations odels and Algothms. Edited by Richard K Belew and Melaie Mtchell (Rea dng: Westview Press 1996) p 64 Morten H Christiansen and Ni Chatek Fte Models of Innite Language: A Connectionist Approach to Recursion. In Connectionist Psychoinguisti Edited by Moren H Chrstiansen and Nick Chatek (estport: Ablex Pubshng 2 0 0 1 ) p. 147 Richard K Belew John Mneey and Nicol N Schraudolph Evolvng Networks: Usng the Genetic Algothm with Connecion ist Leag In Artcial Le II Op Cit p 52730 Geoffrey E nton and Steven J Nowlan How Leang Can Guide Evolution In Adaptive Indiduals in Evolving Populations Op. Ci. p. 4 5 1 Zelg Hars. A Theo of Language and Infoation A Mathematical Approach (Oxford: larendon Press, 1 9 8 1 ) p 367 Zelig Hars Ibid. p. 3937 Zelg Hars Ib p. 339 James R. Hurford Eression/Induction odels of Language Evolution Op Cit. p 3201
Chapter Eleven 1
Max Weber The Theo ofSocial and Economic Organization (New York: Free Press, 1964) p. 12432 2 Kstian Kstiansen Chiefdoms States and Systems of Soal Strati cation In Chiefdoms Powe Economy, a nd Ideology Edited by mothy Eale (New York: Cambdge University Press 1 9 9 1 ) . p 24 3 Kathleen M Caley and Lee Gasser Computational Organizaton Teo Multiagent Systems Edited by Gerhard Weiss (C ambdge: M Press 2 0 0 1 ) . p 300 4 Michael Wooldridge telligent Agents. Ibid p 54-6. Michael Wooldridge Intelgent Agents Ibid, p. 8 6 Daniel C Dennet. ue Beievers The Intentional Stance (Cmbdge: M Press 1 9 9 0) p . 1 5 2 2 7 J E Doran. ajctoes to Complexity n Arial Soeties. In nami in Human and Primate Societies gent-Based Modeling of Social a nd Spatial Processes Edited by mothy A Kohler and George J Gumerman. (Oxford: Oxford University Press 2000) p 94-5 8 ichael N Huhns and Larry M Stephens Multiagent Systems and Societies ofents In Multiagent Systems. Op Cit p 8595. 9 J E Doran and Mike Pamer The EOS Projec: Integratng o Models of Paleoithic Soal Chnge In Artcial Societies The Com puter Simulaton ofSocial Le. Edted by Nigel Glber and Rosara Conte (London: University Colege of London Press 1995) p 10611 0 J E Doran and ke Pamer Ibid, p 1 1320 1 1 erre Clastres Socie Against the State (New York: Zone Books 1987) p 30 12 Ma Weber The Theo of Soal a nd Economic Organizaton Op. Cit p 35966 1 3 Grahame Clark Wold Preistory (Cambdge: Cambdge University Press 1977) p 1 0 5 1 Eizabeh M. Brumel and mothy K Earle Introducion In Speciazation, Exchange, and Complex Societies. Edited by Eizabeth M Brumeld and mothy K Eae ( Cambridge: Cambridge University Press 1987). p 34 15 Cathy A Smal The Poitical Impact o f Marriage in a Virual Polynesian Soety In nami in Human and Pmate Societies Op Ct p 22831 219
218
