Training Bao Game-Playing Agents using Coevolutionary Particle Swarm Optimization

Bao, an African board game of the Mancala family, is a complex two-player game with a very large search space and complex rule set. The success of game tree approaches to create game-playing agents rests heavily on the usually handcrafted, static evaluation function. One of the first steps towards using a game tree is to design an appropriate, efficient evaluation function. This paper investigates the effectiveness of a revolutionary particle swarm optimization (PSO) approach to evolve the evaluation function for the game of Bao. This approach uses a PSO algorithm to evolve a neural network as evaluation function, using an unsupervised, competitive learning approach. The revolutionary approach to evolving game-playing agents assumes no prior knowledge of game strategies. The only domain specific information used by the model are the rules of the game, and the outcomes of games played. The performance of the evolved game-playing agents is compared to a game tree-based agent using a handcrafted evaluation function, as well as a player that makes random moves. Results show that the coevolutionary PSO approach succeeded in learning playing strategies for Bao