Abstract— General Game playing, a relatively new field in game research, presents new frontiers in building intelligent game players. The traditional premise for building a good artificially intelligent player is that the game is known to the player and pre-programmed to play accordingly. General game players challenge game programmers by not identifying the game until the beginning of game play. In this paper we explore a new approach to intelligent general game playing employing a self-organizing, multiple-agent evolutionary learning strategy. In order to decide on an intelligent move, specialized agents interact with each other and evolve competitive solutions to decide on the best move, sharing the learnt experience and using it to train themselves in a social environment. In an experimental setup using a simple board game, the evolutionary agents employing a learning strategy by training themselves from their own experiences, and without prior knowledge of the game, demonstrate to be as effective as other strong dedicated heuristics. This approach provides a potential for new intelligent game playing program design in the absence of prior knowledge of the game at hand. I. INTRODUCTION n game playing, one of the most important aspect is the ability of the player to make intelligent, legal moves during game play. Many different approaches have been explored in this area, and much research and potential still exists to develop intelligent game players. A. General Game Playing The field of General Game Playing (GGP) is an important part of Artificial Intelligence (AI) research, and provides an important leap in the direction and approach of the construction of intelligent agent systems. In the past, much of the emphasis in the creation of intelligent systems was on the system being intelligent in its behaviour only for the task it was constructed to perform well in. GGP systems, as the name implies, are far more general. They are able to accept descriptions of any game, and are able to play them. The importance of this research lies in the fact that GGP systems provide a step from intelligent systems giving an illusion of intelligence to intelligent systems that act in an intelligent manner. Though pure General Game Playing capabilities have not entirely been implemented, systems have been designed This work was supported in part by an NSERC Discovery grant. Z. Kobti is with the School of Computer Science at the University of Windsor, Windsor, ONT, Canada N9B-3P4. (Phone: +1-519-253-3000; fax: +1-519-973-7093; e-mail: kobti@uwindsor.ca). S. Sharma, is with the School of Computer Science at the University of Windsor, Windsor, ONT, Canada N9B-3P4. (e-mail: sharmaw@uwindsor.ca). which display a general behaviour with respect to a specific class of games. B. Early attempts at General Game Playing: Positional Games One class of games where general game playing has been investigated are positional games. These type of games were formalised by Koffman [6]. Banerji [1], Citrenbaum, Pitrat [3], and Banerji and Ernst [2] have studied these class of games. Some examples of position games include Tic-Tac- Toe, Hex, the Shannon switching games. A position game can be defined by three sets, P, A, B. Set P is a set of positions; with set A and B both containing subsets of P. In other words, sets A and B represent a collection of subsets of P, with each subset representing a specific positional situation of the game. The game is played with two players, with each player alternating in moves, which consist of choosing an element from P. The chosen element cannot be chosen again. The aim for the first player is to construct one of the sets belonging to A, whereas the aim for the second player is to construct one of the sets belonging to B. Programs that are capable of accepting rules of positional games, and, with practice, learn how to play the game have been developed. Koffman constructed a program that is able to learn important board configurations in a 4 X 4 X 4 Tic- Tac-Toe game. This program plays about 12 times before it learns and is effectively able to play and start defeating opponents. A set of board configurations are described by means of a weighted graph. C. General Game Playing Architecture For our purposes, we use the GGP architecture developed at Stanford University [9]. A GGP system consists of an agent designated as the Game Player (GP) and a Game Manager (GM). The GM is responsible for sending to the GP, initially, the rules of the game, and subsequently, the moves being made at each stage, and upon termination of the game, a termination message. The responsibility of the GP is to accept all the messages sent by the GM and take the appropriate action. Currently, Stanford University maintains at their website for GGP a GM to which GP’s can connect and play games. They also maintain a rich resource base detailing the model of communication between GP’s and the GM, the types of games that are playable in GGP and a set of game descriptions. The game descriptions are written in prefix Knowledge Interchange Format (KIF) [4]. They are written in such a manner that it is possible to use them and generate a set of legal moves from a given game state. In A Multi-Agent Architecture for Game Playing Ziad Kobti, Shiven Sharma I 276 Proceedings of the 2007 IEEE Symposium on Computational Intelligence and Games (CIG 2007) 1-4244-0709-5/07/$20.00 ©2007 IEEE