CONTENT 01 H OW D OES AI W ORK ? 01 I T ’ S A LL A BOUT L EARNING , M ACHINE L EARNING 02 M AJOR D RIVERS OF THE AI R EVOLUTION 04 E MERGING AI E CONOMY 09 U SES OF AI IN C OMPUTER 09 W HAT
IS
A RTIFICIAL I NTELLIGENCE ?
A NIMATION
AND
G AMING
10 AI IN THE S IMULATION OF C ROWDS 12 AI I N THE S IMULATION OF P HYSICAL P ROCESSES 13 P ROCEDURAL C ONTENT G ENERATION (PCG) 14 AI
IN THE
S IMULATION
OF THE
WITH
H UMAN M OTION
M ACHINE L EARNING
15 AI AND N EW G AME P ATTERNS 16 AI I N P LAYER M ANAGEMENT 17 B ENEFITS OF AI- BASED G AME D ESIGN 18
AI T O C REATE G AME L EVELS
AND
S TY LES
Issue 01
What is A R T I F I C I A L INTELLIGENCE
?
From ancient times, scientists dreamed to create machines that can reason, learn, plan, see, and move in much the same way as humans do naturally. The prospect of working Artificial Intelligence (AI) became tangible in the twentieth century with revolutionary breakthroughs made in computer technology, neuroscience, and mechanical engineering. A l t h o u g h A I t e c h n o l o g y i s n o t n e w, i t r e c e i v e d a powerful momentum in the 2010s when global digital infrastructure and science became ripe for incorporations of AI magic into business. For many observers, 2015 was a landmark year for AI technology . Funding for AI projects grew exponentially with major tech companies, such as Google, IBM, and Facebook, investing vigorously in the new
technology. The AI revolution spilled over into the consumer market, making AI systems, such as Apple’s Siri, Google Assistant, Facebook’s image recognition s o f t w a r e , a n d G o o g l e Vo i c e a n inherent part of our daily lives. Rather than pursuing an ambitious agenda of General Artificial Intelligence aimed to reconstruct the human mind in full scope, modern AI is much more specialized, which means it is business-oriented and consumer-centric. Contemporary AI no longer needs robots and machines to run; it may be embedded into various desktop, mobile, or SaaS (Software-as-a-Service) applications.
How does AI WORK? AI uses advanced mathematical, statistical, computational, and logical approaches to model the human thinking and planning process, find hidden patterns in data and solve optimization problems, predict trends and make decisions under uncertainty. Although humans excel here, machines powered by thousands of CPUs, and joined in complex networks are capable of finding solutions to some problems faster and more efficiently than them. AI cracks the most complex puzzles by using heuristic and optimized solutions that are suitable for unstructured data which is hard to visualize in human mind and language. AI’s ability to retrieve patterns and build models is useful in diagnosing deceases, making data-driven predictions of creditworthiness, volatility of financial markets, weather, or consumer behaviour. AI has already been used in network security to identify patterns of DDoS (Distributed Denial of Service) attacks, or in data mining to predict consumer behavior, financial performance, probability of earthquakes, or running out of stock. However, AI is not just a version of human intelligence on steroids. Its power depends more to cold calculation than intuition; although using AI to produce intuitive insights is not that far off either.
MAGA ZINE
01
Issue 01
IT’S ALL ABOUT LEARNING, MACHINE LEARNING AI ecosystem is enormously complex and diverse, but Machine Learning (ML) plays a pivotal role in the new generation of AI. In 1959 Arthur Samuel defined ML
as a method that gives “computers the ability to learn without being explicitly programmed” . Unlike standard programming that uses a series of ‘if/else’ statements, ML algorithms make predictions and devise complex analytical models from raw samples of data. In a typical ML setup, a human operator lets the computer optimize some initial random hypothesis by making it fit real data, and fine-tuning model parameters until a proper solution is found. To m a k e c o m p u t e r s l e a r n , s c i e n t i s t s c a m e u p w i t h a n idea of artificial neural networks (ANN). ANN is a digital approximation of the human brain that consists of a network of ‘neurons’ that compute and exchange information, extract features from data, and construct optimized solutions to complex problems. Unlike the human brain though, ANN may include just one hidden layer of neurons, as in the perceptron algorithm first implemented in the late 1 9 5 0 s . To s o l v e c o m p l e x p r o b l e m s , s u c h a s i m a g e or speech recognition, ANNs are designed with multiple layers of artificial neurons (‘hidden’ layers)
MAGA ZINE
02
Issue 01
that process data in complex non linear ways, which even their creators are struggling to understand.
Figure # 1 Architecture of a Neural Network
Example of Cancer Classification Problem Using Machine Learning
SCENE
To i l l u s t r a t e h o w A N N s w o r k , let us take an example of a cancer classification problem. Let’s assume we have a training set of 1000 patients with malignant and benign tumors. Our task is to teach AI how to find patterns and features that differentiate one type of tumor from another, and, hopefully, construct a model that would help doctors automatically predict the likelihood of cancer in p a t i e n t s . To p r o c e e d w i t h t h i s t a s k , we construct a hypothesis (cost function) and feed tumor data labeled with 1 for malign tumor and 0 for benign tumor to the ML algorithm, which optimizes our initial hypothesis to make it fit tumor data. Afterwards, the computer runs thousands or even millions of iterations and gradually adjusts parameters (‘weights’) of our hypothesis until the right model is found. At the end of the day, AI manages to infer function that best
MAGA ZINE
distinguishes medical data of patients with malignant and benign tumors, which may be fruitfully used in medical diagnostics. For example, using a similar procedure, researchers at Stanford University have created a learning algorithm to identify skin cancer. The program trained on nearly 130,000 images of rashes, lesions, and moles, could perform at least 91% as good as human doctors .
03
Issue 01
MAJOR DRIVERS OF AI
REVOLUTION
The rise in relevance of AI in recent years owes to increasing synergies between electronic engineering and computer technoloy, Big Data, and Cloud infrastructure, which simplify the task of running complex ML algorithms, making M L- e n a b l e d infrastructure more affordable, and providing data for training AI on the internet.
MAGA ZINE
D IGITAL
INFRASTRUCTURE PAVES THE WAY FOR FASTER AND CHEAPER COMPUTATION
Unlike their predecessors, modern computers have multi-core CPUs with immense processing power, large volumes of memory, and high bandwidth. Contemporary processors produce millions of operations per second (MIPS) required by neural networks and ML algorithms in order to learn efficiently. Unprecedented progress in computing is largely due to Moore’s law, according to which a number of transistors in CPUs doubles
04
Issue 01 approximately every two years. Similarly, recent advances in GPUs’ productivity (Graphics Processing Units) have been shown to speed up training algorithms by orders of magnitude decreasing time of learning dramatically. Today, leading GPU manufacturers, such as NVIDIA, sell GPUs with in-situ AI architectures (e.g NVIDIA Jetson TXI) that simplify running complex artificial networks in the enterprise setting. Other companies, such as Google, are ready to provide ML-enabled computation in their cloud offerings (Google Cloud Platform) . Cloud IaaS (Infrastructure-as-a-Service) and PaaS (Platform-as-a-Service) solutions make computing power cheaper via pay-as-you-go model that allows companies and researchers to use huge server farms of cloud providers for AI research and production. With such cloud providers as Amazon AWS or Google Cloud, enterprises can easily launch hundreds of computer instances to train their neural networks without making any upfront investments in the in-house IT infrastructure.
‘B IG D ATA ’ F UELS AI ML R ESEARCH
M ACHINEL EARNING A LGORITHMS A RE B ECOMING B ETTER mobile AND
Internet, applications, and IoT (Internet of Things) generate huge volumes of structured and unstructured data that fuels AI research, predictive analytics, and strategic management. Data farms and databases provide training sets of data as diverse as medical records, stock market prices, and images that are mined for patterns and business insights. Big data solves the problem of data shortage that was a barrier to earlier research into AI. For example, neural networks for image recognition require millions of labeled images to learn how to tell a cat from a dog. This is no longer a problem with such visual databases as ImageNet providing over ten million images hand-annotated to indicate what objects are pictured . Today, data providers report a surge of requests for big data for AI research and analytics. For example, in 2015 a major structured data provider CloudFlower sold a hundred million spreadsheet rows of data compared to only two millions in 2012 . Also, it becomes easier for companies to obtain data. In 2015 only 51% of data decision-makers were able to easily obtain data, whereas in 2017 this figure is expected to rise to around 66%.
MAGAZINE
Modern ML algorithms use sophisticated mathematical and statistical models that allow process complex data structures and solve non-trivial problems under uncertainty. Neural networks based on reinforcement learning teach machines to take right decisions in uncontrolled environments, such as road traffic. Self-driving cars using this technology are successfully tested these days, and the time of their commercial production is not far off. In the same vein, new architectures for neural networks are becoming more suitable for handling language generation and processing tasks that were beyond the realm of possibility until very recently. Modern neural networks are able to retrieve the most complex patterns from data. Based on the novel technique of deep learning, they are organized in hundreds of layers that create a long chain of data transformations known as credit assignment path (CAP). Passing data through multiple layers allows us to construct better models and identify the most opaque patterns in data. By virtue of these advantages, deep learning has became a state-of-the-art technology in computer vision and speech recognition. Deep learning
05
Issue 01 algorithms are fruitfully applied to teach intuition, planning, and decision-making in complex combinatorial decision spaces with astronomical amounts of solutions. For instance, combining intuition and calculation has helped AlphaGo AI, a program designed by Deep Mind beat the best human players in the old Chinese game Go which has been considered an impossible feat for a computer to master anytime soon.
https://deepmind.com/blog/exploring-mysteries-alphago/
AI
REACHES WIDER DEVELOPER AND BUSINESS COMMUNITIES
AI is entering a syndication stage with more AI solutions and applications included in B2B (Business-to-Business) and consumer offerings. Popularization of AI technology is supported by major tech companies who spread it across wider developer and business communities by open-sourcing ML libraries and APIs. Such ML libraries as TensorFlow, originally developed by the Google Brain project, provide basic ML functionality, such as neural networks, classification, and optimization algorithms out of the box. Open-source ML software with optimized algorithms and standardized solutions makes the threshold of entering into the booming AI market much lower.
MAGA ZINE
06
Issue 01
Emerging AI Economy Ascent of AI technology creates numerous economic and investment opportunities in the emerging AI economy. AI adoption across a broad range of industries will increase global revenues from approximately $8.0 billion in 2016 to over $47 billion in 2020 . The annual growth rate of 55.1% over the 2016-2020 period is promising to attract billions of investments in industries as diverse as pharmaceutical research, medical diagnostics, recommendation systems, fleet management, and logistics. Big investments into AI have been already made by banking and retail, manufacturing, and healthcare sectors. The emerging trend is that the growing amount of AI investments goes to ML software and applications (neural nets, filtering, visualization, data mining), which facilitates the development of data-driven and intelligent systems that support businesses. ML software will generate the largest share of revenue, which is projected to reach $18.2 billion by 2020. AI is beginning to gain traction among the leading tech companies. At its recent Google I/O annual developer meeting, Google announced that the company is moving toward the AI-first approach integrating many AI features in the company’s major software products, such as Gmail and Google Assistant . Also, Google is one of the leaders of AI research with such projects as Google Brain and Deep Mind standing behind some of the most exciting innovations in General Artificial Intelligence and natural language processing. IBM and other tech giants have managed to keep pace with these rapid development of AI. In 2011, IBM’s question answering supercomputer Watson managed to win the quiz show Jeopardy! competing against the former winners Brad Rutter and Ken Jennings. Apart from revolutionary AI R&D in major tech companies, AI adoption grows among non-tech enterprises as well. According to a Narrative Science report based on the survey of 235 businesses, 38% of them are already using AI solutions, and 62% will be using AI by 2018 . Growing AI adoption across the broad range of industries will drive a 300% increase of investments into AI in 2017 compared to 2016 . Standardization and popularization of AI make it easier for businesses to integrate AI solutions into their operations and products. As AI is gradually entering all spheres of life, its economic potential becomes almost boundless.
MAGA ZINE
Use Cases of AI in Computer Animation and Gaming Unfortunately, traditional animation and gaming have failed to appreciate the power of Artificial Intelligence. In the opinion of Dave Mark, the founder of AI consultancy Intrinsic Algorithm, standard game AI has more to do with ‘artificial behaviour’ than ‘artificial intelligence’ , because it makes characters be exactly as intelligent as it is needed to be predictable and provide fun, drama, or enjoyment for players. Nowadays, however, the trend is for a deeper convergence between game AI techniques and state-of-the-art AI and ML methods in academia and business. Pure AI is coming to the foreground of animation, game design, and player experience. AI and ML technologies have a strong potential to improve simulation of human motion, physical processes, and crowds. With AI technologies, we can also create new game patterns, rich PCG (Procedural Content Generation), unique game styles, and human-computer interaction. In what follows, we describe these use cases of AI in animation and game development.
09
Issue 01
AI IN SIMULATION of
HUMAN MOTION
Since the 1970s, motion capture has been the main technique for motion animation in computer graphics and video games. In motion capture, motion data is obtained via specialized electromagnetic or electronic devices placed on human actors to detect motion features and geometry. Motion data retrieved this way has to be manually edited by animators to create believable 3D models of movement. This technique has obvious disadvantages, such as high price of equipment, complicated operating procedures, constraints in actor movement, and more. Consequently, alternative motion capture methods were designed to address these difficulties. In video-based motion capture, for example, a single camera taking a series of shots of human characters and applying automatic feature tracking to extract movement patterns is used. Similarly, in silhouette analysis a video space is searched for human gestures and correspondence is built between extracted image features and a 3D human model. Despite their advantages vis-a-vis traditional methods, both video-based approach and silhouette analysis require individual characters to do a motion capture and thorough hand-made fine-tuning of resultant animations. Unlike traditional motion capture, AI methods allow for algorithmic learning of human movement patterns with the end result of more precise, realistic and flexible human motion. In fact, high-fidelity motion and facial animation may be created without any motion capture at all. For example, researchers from the University of Edinburgh and Method Studios created a machine learning system that trains on a large data set of motion capture clips to learn various types of locomotion, such as walking, running, jumping, and climbing in virtual environments. The system uses Phase-Functioned Neural Networks for Character Control that models motion based on numerous parameters, such as the previous state of the character and geometry of the scene . As a result, AI is able to produce smooth motions where the character automatically adapts to various geometric .
MAGA ZINE
10
Issue 01
environments, including walking and running over rough terrain, jumping over obstacles, and crouching under low ceilings. The neural network makes flowing transitions between these types of movements. Another benefit is that a training process has minimal computing requirements: it takes the system milliseconds of execution time and several megabytes of memory to produce these excellent results. Modern AI is also powerful in simulation of highly dynamic and rich motion patterns, such as dancing. In the early days of computer graphics, animators struggled with animation of dancing. However, with the arrival of advanced machine learning, computers can learn to generate dance movements that look realistic to an average human observer. This ability was demonstrated by the neural network software l a b Pe l t a r i o n a n d t h e L u l u A r t G r o u p t h a t created a self-taught dancing human skeleton named “Chor-rnn”. Spending two days learning from dancing clips featuring dancing people shot with depth-sensing Kinect cameras for Xbox, AI learned the overall choreographic style of motions and even managed to create new improvisational dance styles. By virtue of new ML solutions, modern AI systems solve many problems usually faced by game designers and animators. The problem with traditional animation techniques is that they have little flexibility. If the animator discovers that the right motions were not captured in the studio, it will be needed to retrack and capture some more. For video games the situation is even worse. Here, all motions that might conceivably be needed must be captured beforehand. With AI technologies, animators no longer need to envisage all potential movements working with libraries of motion. Starting with a given amount of captured motions, they may procedurally generate more data that fits the style of the original. At the same time, specific aesthetic requirements of animators can be also taken into consideration
MAGA ZINE
11
Issue 01
AI in CROWD SIMULATION In the early days of computer graphics and VFX (Visual Effects), directors used thousands of extras in battle and sport event scenes. The world record belongs to Richard Attenborough’s 1982 film Gandhi in which 300,000 extras were used in the f u n e r a l s c e n e . To d a y , w i t h n e w A I c r o w d s i m u l a t i o n technologies available we can’t afford the luxury of drawing crowds in games, or hiring extras. On a large scale, crowd simulation AI was for t h e f i r s t t i m e i n Pe t e r J a c k m a n ’ s L o r d o f t h e R i n g s premiered in 2001. Crowd simulation software MASSIVE (Multiple Agent Simulation System in Virtual Environments) used in the film allows to generate thousands of extras that act as individuals without being explicitly programmed, or animated. The system uses fuzzy logic to make characters realistically respond to surroundings and other characters. In MASSIVE, crowd behaviour patterns are retrieved from pre-recorded animation clips that are made in motion-capture sessions, or drawn by hand. Modern ML techniques push the limits of crowd simulation technology even further by using learning to track crowd behavior patterns from real crowds. In the state-of-the-art crowd simulation, motion of a real human crowd is first recorded from an aerial view to form a training set of data. Then, two-dimensional moving trajectories of each character in the crowd are learned by powerful neural networks. By studying real crowd footage, AI system can assess the impact of environment and nearby agents more precisely than in hand-made animations and motion capture. As a result, AI-made crowds behave much the same way as real crowds.
Crowd Simulation. Source: http://gamma.cs.unc.edu/research/crowds/)
With crowd behaviour patterns learned via ML learning algorithms, game designers can blend existing crowd data into new crowd animations, selecting their own virtual environments and characters. This makes crowd animation in games more flexible and creative. Characters in AI-based crowds may also express social and emotional intelligence reflected in a rich palette of facial expressions, emotions, and gestures. Using reinforcement learning, crowd agents can learn from their mistakes and alter behavior in response to rewards and punishments received from the environment. Thus, incorporation of learning into crowd simulation allows to combine group and individual realism and introduce ‘fuzziness’ of behavior one may find in real crowds. AI crowd as a living and changing organism becomes a powerful tool of experimentation in social dynamics and creates opportunities for new interactive methods in gaming.
MAGA ZINE
12
Issue 01
AI In Simulation of Physical Processes Physically based computer animation aims to simulate physical laws of a system for graphical purposes. It is widely used in water, smoke, explosion, and character animation. Real-time simulation of physical substances and processes, such as fluid or smoke, is one of the hardest problems in computer graphics. High-fidelity simulation of physics requires immense computational resources, which makes real-time applications impractical. With AI, however, we may use a data-driven approach that combines deep learning methods with classical fluid and gas equations to obtain highly realistic and computationally cheap results. For example, Jonathan Tompson from Google and researchers from New York University proposed a semi-supervised learning approach in which a Convolutional Network (ConvNet) is trained on a data set of smoke and fluid simulations. At the end of the training, AI learns to produce relatively fast simulations of smoke and fluid that outperform standard non-AI simulation. Using statistical data of fluid and smoke dynamics, new AI-based simulations can create high-resolution animation that accounts for numerous unseen properties of gases and liquids and complex interrelations between substance particles. Not just simulations of physical objects and processes, but game characters, as well, should obey Newton’s laws. This problem is addressed by the physically based character animation. Since 1999, AI played a crucial role in emulation of physical properties of characters. For example, a NeuroAnimator approach used a neural network to observe and interpret physics-made models in action in order to synthesize motion that suits animation goals. ML used in NeuroAnimator allowed to produce faster computations than traditional numerical methods . In pair with simulation of physical processes, physically based animation of characters allows for more realistic rendering of human response to physical shocks, such as explosions, and simulation of other dynamic environments that experience the impact of physical forces (wind, earthquakes, floods). Flexibility of ML approach also allows simulation of motion under new perturbations, different from those in the training sample, which opens opportunities for experimentation in sci-fi games.
MAGA ZINE
13
Issue 01
PROCEDURAL CONTENT GENERATION (PCG) WITH MACHINE LEARNING
Procedural Content Generation (PCG) is widely employed in game design for algorithmic generation of content in the runtime. PCG may be effectively used to increase replay value, save storage space, reduce game production efforts and costs, personalize narrative for players, and create infinite diversity of quests and game environments. In the early days of video games, memory shortage was the main reason behind using PCG. To reduce memory usage, conventional PCG employed pseudo random number (PRN) generators that initialized pre-defined parameters of game space size, texture, color, or character appearance. These days, with AI and ML offering innovative methods to create content in games, conventional PCG approach becomes outdated. Such ML techniques as generative adversarial networks are more efficient in generation of such game artifacts as music, speech, or animation. In the state-of-the-art ML methods, music and speech in games may be generated with ML algorithms that use raw waveforms instead of sound samples. Increase of computing power allows neural networks to process thousands of sound frames in milliseconds and output more realistic speech and music content that perfectly simulates physical properties of sound.
MAGA ZINE
One example of such approach is the WaveNet neural network developed by Deep Mind, an AI research company bought by Google in 2014. WaveNet generates human voices that sound more natural than existing Text-to-Speech systems (TTS) based on large datasets of small speech fragments. In contrast to TTS, WaveNet directly models the raw waveform yielding more natural-sounding speech. Speech generated by WaveNet is also more natural by virtue of non-speech sounds, such as breathing or mouth movements. Similarly, WaveNet is able to produce creative musical samples in styles of classical or popular music, which sound as though they were written by a human composer. With such a technique at their disposal, game designers will be able to generate more naturally-sounding NPCs (Non-Player Characters) and create rich sonic experiences for gamers. Today, AI solutions may also be fruitfully used in procedural video and animation generation. In one of experiments concerning this idea, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) used a deep learning algorithm that could learn to transform still images from a scene into a brief video that simulates the future of the scene. To achieve this, researchers trained the neural network on two million of unlabeled videos that contained a year’s worth of footage. In gaming, such approach could be used to simulate diversity and relationships between objects (e.g trees) in complex structures (e.g forest). It’s still hard to develop a good script that covers a variety of tree species (e.g birches, palm trees). It is even harder to generate a realistically looking forest by simply assembling trees, because forest trees tend to form complex ties and grow unevenly. AI-based PCG solves this problem by learning on large samples of real trees to identify real-world patterns and relations between them. As a result, AI-based PCG becomes a powerful alternative to hand-made animation that falls short of expressing the full complexity of our world.
14
Issue 01
AI To Create Game Levels and Styles
The power of AI may be leveraged to generate realistic game levels and innovative game aesthetics. For example, research has been conducted into creation of new Super Mario Bros. levels by adapting functional scaffolding for musical composition (FSMC) learning approach used in the generation of musical samples. In the original FSMC technique, music is represented as a function of time and relations between voices in a musical piece. Amy Hoover and her colleagues creatively adapted this technique to break up each level of Super Mario Bros. into sequences of columns with a tile-sized width. Each column is treated as a unit of time, which allows us to represent a series of tiles in the game as a sequence of musical pitches and durations. This data is then fed to the ANN to generate new levels . The idea that AI can retrieve hidden patterns from human-authored levels based on just little starting data, such as the layout of a single tile-type, is, indeed, very intriguing for game designers. In addition to game level generation, machine learning makes it possible to transfer or imitate certain styles and game genres to create unique game aesthetics. MIT Nightmare machine is an interesting project that probes into this direction. In the Nightmare Machine, deep learning is used to transform famous a r c h i t e c t u r e s i t e s , s u c h a s B i g B e n , t h e U n i t e d S t a t e s C a p i t o l , o r To w e r o f L o n d o n into the haunted locations one would see in horror movies . To g e n e r a t e h o r r o r i m a g e r y , M I T n e u r a l n e t w o r k s t r a i n o n t h o u s a n d s o f images retrieved from horror movies and thrillers. Over time, AI learns horror aesthetics and develops strong neural connections that may later be excited to generate new horror images from any visual content.
MAGA ZINE
15
Issue 01
AI AND NEW GAME PATTERNS Shooters, real-time-strategies (RTS), RPGs (role-playing game), fighting and survival games are just a few game types that have earned high reputation among gamers. AI technologies promise to create new game patterns that enrich gaming experience and discover innovative ways of human-computer interaction . AI-based game design creates new intelligent NPCs who learn as you play, adjust game difficulty to the player’s skills and style, and interact with players and each other through the gameplay. O n e e m e r g i n g g a m e p a t t e r n i s u s i n g A I a s a r o l e m o d e l , l i ke i n S p y Pa r t y Game, where one player is a spy at the party whilst another is a sniper trying to identify him in the crowd of AI agents mimicking human behavioural cues. AI learns to emulate human behaviour to make it harder for players to identify the spy. Each new player leaves his mark on AI tactics that transforms as it gains new experience. Game AI may be also thought of as a trainee taught by the player to perform certain tasks. In this pattern, the player trains AI to generate game content in order to solve puzzles or develop skills needed to proceed in the game. For example, in Black & White god game that implements this pattern, the player trains a creature to act as his autonomous assistant in spatial regions with no direct access to the player. The AI is taught to obey his commands through positive or negative reinforcements; the player rewards it by petting and punishes by slapping. In Editable AI games, players can change internal model’s values until acceptable piece of content is generated. For example, the player may tweak weapons and ammunition to create a combination that helps destroy the enemy, or d e s i g n n e w t y p e s o f w o r k e r s w i t h https://www.thesims.com/media?page=3 properties that allow for more productive use of limited resources. Editable AI moves procedural content generation to the next level by allowing gamers to generate their own content. In this approach, a human player leverages the power of AI to boost his own creativity. In its turn, a guided AI game pattern reflects upon the fact that many AI algorithms are brittle and prone to collapse if not used in a highly limited environment. However, rather than avoidind placing AI in situations that could harm them, one may build games in which players act to protect AI in such situations. One may find this pattern in the The Sims game series published by Electronic Arts company. In The Sims, AI agents have a set of basic needs and wants that should be met by players through food, work, shelter, socialization, and self-actualization. I n t h i s w a y , T h e S i m s p u t s t h e c o n c e p t i n i t i a l l y i n t r o d u c e d i n Ta m a g o t c h i t o a qualitatively new level. Even more exciting scenario is viewing AI-based games as a spectacle where AI or a group of AI agents creates a complex social hierarchy which the player may observe, or interact with. Such approach would allow for more innovative interaction between AI and players, and promote autonomy of AI agents in game design.
MAGA ZINE
16
Background image source: https://www.thesims.com
Issue 01
AI in PLAYER MANAGEMENT With rowing competition among video game companies, it becomes harder to attract and retain new customers. Unfortunately, quality of game content is not enough to create persistent attachment of players to games. In a new environment driven by multi-player online and social games, companies emphasize long-term gaming experience by offering products that evolve over time and support player communities to produce more game buzz. Married with this approach, AI technology may give another competitive edge for companies in attraction and retention of gamers. One of the most promising areas of innovation in this respect is game analytics. AI may be efficiently leveraged to capture, process, and understand player behaviour, model player responses to new game features, and manage player experience via customizable and adaptable AI features. This expertise may be also used to create innovative player assistance AI. This type of AI will play a role of lifelong agents who learn about players over time, recognize and adapt to their changes, and use history of interactions to improve their gaming experience. Lifelong agents can also assist players in overcoming game challenges that would otherwise make them quit playing a game. By knowing the players’ strengths and weaknesses, game AI may intervene in a timely manner to provide them with useful hints, automatically generated tutorials, and gameplay adjustments. In this way, game AI may serve as a coach to new players along their difficult mission. Using AI-powered Dynamic Difficult Adjustment (DDA), the AI coach will adapt game parameters, enemy behaviours, and other game features to the player’s skills and mood. Similarly, given the growing popularity of online games, new game AI may employ user data, such as GPS location, social profile, and user preferences to design a better gaming experience with more adaptable gameplay. In this way, big data generated by modern online and social games can become a powerful training sample to be used by AI lifelong agents to improve player experience and create lasting and pleasant connection of players to games.
MAGA ZINE
17
Issue 01
Benefits of AI-based Game Design Gaming is at the peak of its popularity. According to the ESA (Entertainment Software Association) study, the number of gamers in the world has reached 1.8 billion people in 2016 . To attract and retain customers, video game companies should not hesitate to adopt state-of-the-art AI solutions that enhance and broaden gameplay experiences. Advantages of AI for gaming and animation mentioned above are too obvious to be ignored.
AI CHARACTERS AND GRAPHICS ARE MORE REALISTIC
AI BASED CHARACTERS DO NOT MERELY RESEMBLE INTELLIGENCE, BUT ARE, ACTUALLY, INTELLIGENT
AI-based techniques for animation of human motion and physical processes outperform traditional approaches. Instead of fine-tuning animations manually using standard motion libraries, we can leverage the power of neural networks to learn patterns of human motion and physical processes from data. This allows us to produce animation models that account for a variety of use cases that are hard to envisage in the game design process.
Unlike traditional game characters, who only behave ‘intelligently’, AI is able to create NPCs able to learn from players, adapt, and interact with them in various innovative ways. Rich palette of behaviours, skills, and emotions displayed by AI characters will not only enhance player experiences, but create new product offerings for video game companies.
AI-BASED GAME DESIGN COMES WITH GAME DIVERSITY AND VERSATILITY
AI GAMES BRING HUMAN-COMPUTER INTERACTION TO THE NEXT LEVEL
With new AI-based game patterns, we may create an infinite number of games with creative characters, narratives, and player roles. Such diversity can not be easily achieved by hand-picking character traits and features. AI-based PCG (Procedural Content Generation) married with human creativity promises to produce unmatched abundance in video game
Being a playground for innovation, AI games can revolutionize the entire space of AI research. We may use AI-based games to test future relationships between AI and human society and explore them as prototypes of our digital future driven by new technologies and artificial intelligence.
MAGA ZINE
18
Issue 01
References • Clark, Jack. Why 2015 Was A Breakthrough Year In Artificial Intelligence? Bloomberg. https://www.bloomberg.com/news/articles/2015-12-08/why-2015-was-a-breakthrough-year-in-artificial-intelligen ce • Puget, Jean Francois. (2016). What Is Machine Learning? https://www.ibm.com/developerworks/community/blogs/jfp/entry/What_Is_Machine_Learning?lang=en • Vincent, James (2017). Artificial Intelligence Can Spot Skin Cancer As Well As A Trained Doctor. https://www.theverge.com/2017/1/26/14396500/ai-skin-cancer-detection-stanford-university • Pardes, Arielle. (2017). The Slickest Things Google Debuted Today At Its Big Event. https://www.wired.com/2017/05/slickest-things-google-debuted-today-big-event/ • ImageNet. http://image-net.org/about-stats • Pres, Gil. (2016). Forrester Predicts Investment In Artificial Intelligence Will Grow 300% in 2017. https://www.forbes.com/sites/gilpress/2016/11/01/forrester-predicts-investment-in-artificial-intelligence-will-grow300-in-2017/#55093b115509. • McCormick, James (2016). Predictions 2017: Artificial Intelligence Will Drive The Insights Revolution. Forrester. https://go.forrester.com/wp-content/uploads/Forrester_Predictions_2017_-Artificial_Intelligence_Will_Drive_The_I nsights_Revolution.pdf • IDC (2016). Press Release. http://www.idc.com/getdoc.jsp?containerId=prUS41878616 • Pardes, Arielle. (2017). The Slickest Things Google Debuted Today At Its Big Event. https://www.wired.com/2017/05/slickest-things-google-debuted-today-big-event/ • Pres, Gil. (2016). Artificial Intelligence Rapidly Adopted By Enterprises, Survey Says. https://www.forbes.com/sites/gilpress/2016/07/20/artificial-intelligence-rapidly-adopted-by-enterprises-survey-sa ys/#b5b578c12dae • Pres, Gil. (2016). Forrester Predicts Investment In Artificial Intelligence Will Grow 300% in 2017. https://www.forbes.com/sites/gilpress/2016/11/01/forrester-predicts-investment-in-artificial-intelligence-will-grow300-in-2017/#55093b115509 • Graft, Kris (2015). When Artificial Intelligence In Video Games Becomes ... Artificially Intelligent. http://www.gamasutra.com/view/news/253974/When_artificial_intelligence_in_video_games_becomesartificially_i ntelligent.php • Phase-Functioned Neural Networks for Character Control. http://theorangeduck.com • Templeton, Graham (2016). Deep-learning Neural Network Creates Its Own Interpretive Dance. https://www.extremetech.com/extreme/227287-deep-learning-neural-network-creates-its-own-interpretive-dance • Yu, Jun and Tao, Dacheng (2013). Modern Machine Learning Techniques and Their Applications in Cartoon Animation Research. Hoboken, New Jersey, John Wiley & Sons, p.3. • Grzeszczuk, Radek, Terzopoulos, Demetri, Hinton, Geoffrey. (1998). NeuroAnimator: Fast Neural Network Emulation and Control of Physics-Based Models. http://web.cs.ucla.edu/~dt/papers/siggraph98/siggraph98.pdf • Deep Mind. WaveNet: A Generative Model for Raw Audio. https://deepmind.com/blog/wavenet-generative-model-raw-audio/ • Summerville et al. (2017). Procedural Content Generation via Machine Learning (PCGML). https://arxiv.org/abs/1702.00539 • Nightmare Machine. http://nightmare.mit.edu/ • Treanor, Mike et al. (2015). AI-Based Game Design Patterns. http://julian.togelius.com/Treanor2015AIBased.pdf • McKane, Jamie (2016). There Are 1.8 Billion Gamers In the World, And PC Gaming Dominates the Market. https://mygaming.co.za/news/features/89913-there-are-1-8-billion-gamers-in-the-world-and-pc-gaming-dominates -the-market.html
MAGA ZINE
19
About Team Absentia: Founded by three BITS Pilani drop outs, Absentia consists of a small, highly skilled and electric team partnering with knowledge partners from around the world. The team, with over 500 publications is developing Norah AI, an artificial intelligence capable of generating premium games in a few minutes. Chief Mentor – A.P. Parigi Ex-board member, Bennet Coleman and Co. Ltd. Ex-advisor at NEA, India Ex-CEO and MD at Radio Mirchi Ex-CEO at Network 18 Investors: Exfinity Venture Partners LLP is focused on funding Technology Ventures but with a difference. They seek to act as a strategic partner to portfolio companies through high impact interventions. Their impressive portfolio includes Mad Street Den, Uniken systems. TV Mohandas Pai Leadership team and ex-board member of Infosys Ltd (1995-2011). Chairman, Manipal Global Education Services. Board of Director National Stock Exchange of India (NSE). Prolific angel investor in India & US. V Balakrishnan Leadership team and ex board member of Infosys Ltd. Ex-CFO, Infosys Ltd. Head of BPO, Finacle & India Business Unit at Infosys.
Global Knowledge Partners Natural Language Processing: Soma Halder – InfoSec Specialist at Visa Deep Learning: Denis Tarasov – Co-founder at meanotek.io Gaming: Nebojsa Brindic – CTO at Bincode Entertainment
