A Methodology for Creating Generic Game Playing Agents for Board Games

Abstract

General Game Playing (GGP) consists in developing agents capable of playing different games. Normally these agents go through an initial learning process to gain some knowledge about the game and be able to play it well. In board games, this normally requires learning how to evaluate a great variety of states in a game tree. This work introduces a methodology called UCT-CCNN to generate value functions for evaluating states in generic board games. The UCT-CCNN method executes a large number of matches between Monte Carlo Tree Search (MCTS) agents using a tree policy known as Upper Confidence Bounds for Tree (UCT) in an off-line process that generates a database of state-utility examples. From those examples, a value function for the game states is learned through the use of constructive neural networks known as Cascade Correlation Neural Networks (CCNN). The UCT-CCNN method was tested with two classical board games: Othello and Nine Men's Morris, and the obtained agents were capable of winning matches against agents specifically developed for these games. Moreover, the UCT-CCNN method can control the strength of the obtained agent, ensuring a flexible method capable of generating intelligent agents with different levels of difficulty. Another set of experiments shows that the UCT-CCNN method can also be easily integrated into any algorithm such as the MCTS itself, leading to higher winning rates when compared to the standard UCT with the same number of simulations.