Evolutionary computation and games
Simon Lucas2006, IEEE Computational Intelligence Magazine
Sign up for access to the world's latest research
Abstract
and Games beat the best human players. Board games usually succumb to brute force 1 methods of search (mini-max search, alpha-beta pruning, parallel architectures, etc.) to produce the very best players. Go is an exception, and has so far resisted machine attack. The best Go computer players now play at the level of a good novice (see [3], [4] for review papers and [5]-[8] for some recent research). Go strategy seems to rely as much on pattern recognition as it does on logical analysis, and the large branching factor severely restricts the look-ahead that can be used within a game-tree search. Games also provide interesting abstractions of real-world situations, a classic example being Axelrod's Prisoner's Dilemma [9]. Of particular interest to the computational intelligence community, is the iterated version of this game (IPD), where players can devise strategies that depend upon previous behavior. An updated competition [10], celebrating the 20th anniversary of Axelrod's competition, was held at the 2004 IEEE Congress on Evolutionary Computation (Portland, Oregon, June 2004) and at the IEEE Symposium on Computational Intelligence and Games (Essex, UK, April 2005), and this still remains an extremely active area of research in areas as diverse as biology, economics and bargaining, as well as EC. In recent years, researchers have been applying EC methods to evolve all kinds of game-players, including real-time arcade and console games (e.g., Quake, Pac-Man). There are many goals of this research, and one emerging theme is using EC to generate opponents that are more interesting and fun to play against, rather than being necessarily superior. Before discussing possible future research directions, it is interesting to note some of the achievements during the past 50 years or so, during which time games have held a fascination for researchers. Games of Perfect Information Games of perfect information are those in which all the available information is known by all the players at all times. Chess is the best-known example and has received particular interest culminating with Deep Blue beating Kasparov in 1997, albeit with specialized hardware [11] and brute force search, rather than with AI/EC techniques. However, chess still receives research interest as scientists turn to learning techniques that allow a computer to 'learn' to play chess, rather than being 'told' how it should play (e.g., [12]-[14]). Learning techniques were being used for checkers as far back as the 1950s with Samuel's seminal work ([15], which was reproduced in [16]). This would ultimately lead to Jonathan Schaeffer developing Chinook, which won the world checkers title in 1994 [17], [18]. As was the case with Deep Blue, the question of whether Chinook used AI techniques is open to debate. Chinook had an opening and end game database. In certain games, it was able to play the entire game from these two databases. If this could not be achieved, then a form of mini-max search with alpha-beta pruning and a parallel architecture was used. Chinook is still the recognized world champion, a situation that is likely to remain for the foreseeable future. If Chinook is finally defeated, then it is almost certain that it will be by another computer. Even this is unlikely. On the Chinook Web site [19], there is a report of a tentative proof that the White Doctor opening is a draw. This means that any program using this opening, whether playing black or white, will never lose. Of course, if this proof is shown to be incorrect, then it is possible that Chinook can be beaten; but the team at the University of Alberta has just produced (May 14, 2005) a 10-piece endgame database that, combined with its opening game database, makes it a formidable opponent. Despite the undoubted success of Chinook, the search has continued for a checkers player that is built using "true" AI techniques (e.g., [20]-[25]), where the playing strategy is learned through experience rather than being pre-programmed. Chellapilla and Fogel [20]-[22] developed Anaconda, named due to the strangle hold it places on its opponent. It is also known as Blondie24 [22], which is the name it uses when playing on the Internet. This name was chosen in a successful attempt to attract players on the assumption they were playing against a blonde 24-year-old female. Blondie24 utilizes an artificial neural network with 5,046 weights, which are evolved by an evolutionary strategy. The inputs to the network are the current FEBRUARY 2006 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 11 © DIGITALVISION In natural evolution, the fitness of an individual is defined with respect to its competitors and collaborators, as well as to the environment.
