Evolving Pac-Man Players: Can We Learn from Raw Input?

Pac-Man (and variant) computer games have received some recent attention in artificial intelligence research. One reason is that the game provides a platform that is both simple enough to conduct experimental research and complex enough to require non-trivial strategies for successful game-play. This paper describes an approach to developing Pac-Man playing agents that learn game-play based on minimal onscreen information. The agents are based on evolving neural network controllers using a simple evolutionary algorithm. The results show that neuroevolution is able to produce agents that display novice playing ability, with a minimal amount of onscreen information, no knowledge of the rules of the game and a minimally informative fitness function. The limitations of the approach are also discussed, together with possible directions for extending the work towards producing better Pac-Man playing agents