Coevolving Probabilistic Game Playing Agents using Particle Swarm Optimization Algorithm

Coevolutionary techniques in combination with particle swarm optimization algorithms and neural networks have shown to be very successful in finding strong game playing agents for a number of deterministic games. This paper investigates the applicability of a PSO coevolutionary approach to probabilistic games. For the purposes of this paper, a probabilistic variation of the tic-tac-toe game is used. Initially, the technique is applied to a simple deterministic game (tictac-toe), proving its effectiveness with such games. The technique is then applied to a probabilistic 4x4x4 tic-tactoe game, illustrating scalability to more complex, probabilistic games. The performance of the probabilistic game agent is compared against agents that move randomly. To determine how these game agents compete against strong non-random game playing agents, coevolved solutions are also compared against agents that utilize a strong hand-crafted static evaluation function. Particle swarm optimization parameters/topologies and neural network architectures are experimentally optimized for the probabilistic tic-tac-toe game.