Learning game players by an evolutionary approach using pairwise comparison without prior knowledge

There are many studies to learn an artificial game player or game strategy automatically or by unsupervised learning. However, it is still difficult to make a good player without using prior knowledge such as a heuristic player to play with, game records, and so on. One of representative methods for unsupervised learning of players is evolutionary learning approaches. In this study, an evolutionary approach with using pairwise comparison of two players is proposed to learn Othello players under the condition that the players only know the rules of the game. In order to solve a highly dynamic and unstable problem of learning players, Differential Evolution (DE) is adopted because DE adopts pairwise comparison and has been successfully applied to a highly uncertain and unstable problem of optimizing human preference. In our proposed method, players are randomly generated and form a population. Each player in the population is perturbed by DE operations and a child player is created. As pairwise comparison, the child plays against the parent player. The winner becomes a survivor. The population is replaced by the survivors. Players are evolved by repeating these processes. Computer simulation of Othello games is performed and it is shown that the method can generate good players who can win a standard heuristic player.