Learning Strategies in Games by Anticipation

Abstract

Game Theory is mainly seen as a mathematical theory which tries to replace pure chance and intuitive behavior in a competitive situation by calculations. This theory has been widely used to define computer programs. The aim of the researc h described here is to design an artificial system which is able to play efficiently certain ga me s to which Game Theory cannot be applied satisfactorily (such as games with incomplete or imperfect information) . When it cannot find a winning so lution , the system is able to play through a process of anticipation. This is done by building and refining a model of the adversary's behavior in real time during the game. The architecture proposed here relies on two genetic classifiers, one of which models the adversaries' behaviors while the other uses the models thus built in order to play. The system's strategy learning abi lity has been tested on a simple strategic game. The results show the advantages of thi s apprOach over human and traditional artificial adversaries (simple probabilistic and adaptive probabilistic) and illustrate how the system learns the strategies used by its adversa ries.