The goal of getting computers to automatically solve problems is central to artificial intelligence, machine learning, and the broad area encompassed by what Turing called “machine intelligence ” [161, 162].

Abstract. We apply genetic programming to the evolution of strategies for playing the game of backgammon. We explore two different strategies of learning: using a fixed external opponent as teacher, and letting the individuals play against each other. We conclude that the second approach is better and leads to excellent results: Pitted in a 1000-game tournament against a standard benchmark player—Pubeval— our best evolved program wins 62.4 % of the games, the highest result to date. Moreover, several other evolved programs attain win percentages not far behind the champion, evidencing the repeatability of our approach.

Abstract. We propose an approach for developing efficient search algorithms through genetic programming. Focusing on the game of chess we evolve entire game-tree search algorithms to solve the Mate-In-N problem: find a key move such that even with the best possible counterplays, the opponent cannot avoid being mated in (or before) move N.Weshow that our evolved search algorithms successfully solve several instances of the Mate-In-N problem, for the hardest ones developing 47 % less gametree nodes than CRAFTY—a state-of-the-art chess engine with a ranking of 2614 points. Improvement is thus not over the basic alpha-beta algorithm, but over a world-class program using all standard enhancements. 1

In this paper we demonstrate how genetic algorithms can be used to reverse engineer an evaluation function’s parameters for computer chess. Our results show that using an appropriate mentor, we can evolve a program that is on par with top tournament-playing chess programs, outperforming a two-time World Computer Chess Champion. This performance gain is achieved by evolving a program with a smaller number of parameters in its evaluation function to mimic the behavior of a superior mentor which uses a more extensive evaluation function. In principle, our mentor-assisted approach could be used in a wide range of problems for which appropriate mentors are available.

www.doc.ic.ac.uk/ccg Abstract. Behaviour trees provide the possibility of improving on existing Artificial Intelligence techniques in games by being simple to implement, scalable, able to handle the complexity of games, and modular to improve reusability. This ultimately improves the development process for designing automated game players. We cover here the use of behaviour trees to design and develop an AI-controlled player for the commercial real-time strategy game DEFCON. In particular, we evolved behaviour trees to develop a competitive player which was able to outperform the game’s original AI-bot more than 50 % of the time. We aim to highlight the potential for evolving behaviour trees as a practical approach to developing AI-bots in games. 1

Abstract — In order for computer Poker agents to play the game well, they must analyse their current quality despite imperfect information, predict the likelihood of future game states dependent upon random outcomes, model opponents who are deliberately trying to mislead them, and manage finances to improve their current condition. This leads to a game space that is large compared to other classic games such as Chess and Backgammon. Evolutionary methods have been shown to find relatively good results in large state spaces, and neural networks have been shown to be able to find solutions to non-linear search problems such as Poker. In this paper, we develop No-Limit Texas Hold’em Poker agents using a hybrid method known as evolving neural networks. We also investigate the appropriateness of evolving these agents using evolutionary heuristics such as co-evolution and halls of fame. Our agents were experimentally evaluated against several benchmark agents as well as agents previously developed in other work. Experimental results show the overall best performance was obtained by an agent evolved from a single population (i.e., no co-evolution) using a large hall of fame. These results demonstrate an effective use of evolving neural networks to create competitive No-Limit Texas Hold’em Poker agents. I.

This thesis focuses on the fields of real-time strategy games, evolutionary computation, distributed machine learning and multi-agent systems. In general, the problem is to automatically learn the best strategy to play a real time strategy game, more precisely-a two-player combat of marines and tanks. The idea was inspired by ORTS RTS Game AI Competition held annually at University of Alberta. The given problem is very complex and multicriterial, thus final solutions presented here are the result of a constant development and countless improvements. In the paper we try to underline the iterative nature of this process and propose a methodology that could be used for different problems in the real-time games field. We show how to model the strategy as a multi-agent system and how to fine-tune the evolutionary process of searching best players. We also explore the subject of distributed learning, focusing on using a computation cluster for evaluating solutions. The methods of evaluation are also elaborated in the context of co-evolution, we compare two different methods that use competitive fitness- single elimination tournament and hall of fame. In order to

Computers have difficulty learning how to play Texas Hold’em Poker. The game contains a high degree of stochasticity, hidden information, and opponents that are deliberately trying to mis-represent their current state. Poker has a much larger game space than classic parlour games such as Chess and Backgammon. Evolutionary methods have been shown to find relatively good results in large state spaces, and neural networks have been shown to be able to find solutions to non-linear search problems. In this paper, we present several algorithms for teaching agents how to play No-Limit Texas Hold’em Poker using a hybrid method known as evolving neural networks. Furthermore, we adapt heuristics such as halls of fame and co-evolution to be able to handle populations of Poker agents, which can sometimes contain several hundred opponents, instead of a single opponent. Our agents were evaluated against several benchmark agents. Experimental results show the overall best performance was obtained by an agent evolved from a single population (i.e., with no co-evolution) using a large hall of fame. These results demonstrate the effectiveness of our algorithms in creating competitive No-Limit Texas Hold’em Poker agents. 1

Abstract — We present the application of genetic programming (GP) to the zero-sum, deterministic, full-knowledge board game of Lose Checkers. Our system implements strongly typed GP trees, explicitly defined introns, local mutations, and multitree individuals. Explicitly defined introns in the genome allow for information selected out of the population to be kept as a reservoir for possible future use. Multi-tree individuals are implemented by a method inspired by structural genes in living organisms, whereby we take a single tree describing a state evaluator and split it. I.

This paper demonstrates the use of genetic algorithms for evolving a grandmaster-level evaluation function for a chess program. This is achieved by combining supervised and unsupervised learning. In the supervised learning phase the organisms are evolved to mimic the behavior of human grandmasters, and in the unsupervised learning phase these evolved organisms are further improved upon by means of coevolution. While past attempts succeeded in creating a grandmasterlevel program by mimicking the behavior of existing computer chess programs, this paper presents the first successful attempt at evolving a state-of-the-art evaluation function by learning only from databases of games played by humans. Despite the underlying difficulty involved (in comparison to learning from chess programs), our results demonstrate that the evolved program outperforms a two-time World Computer