Training a Pac-Man Player with Minimum Domain Knowledge and Basic Rationality

Abstract

In this paper, a Pac-Man player (agent) is trained based on four neural networks. The motivation is to demonstrate how computational intelligence techniques can be used to simulate the self-learning process of human players with minimum prior knowledge about the game and basic rationality in their behaviors. Experimental results show that, on a simplified version of the original Pac-Man game, the agent can achieve reasonable scores after only a handful of trials. This performance is in contrast to existing work on evolving Pac-Man agents where thousands of trials and huge amount of computational efforts are typically required.