Evolution versus Temporal Difference Learning for learning to play Ms. Pac-Man

摘要

This paper investigates various factors that affect the ability of a system to learn to play Ms. Pac-Man. For this study Ms. Pac-Man provides a game of appropriate complexity, and has the advantage that in recent years there have been many other papers published on systems that learn to play this game. The results indicate that Temporal Difference Learning (TDL) performs most reliably with a tabular function approximator, and that the reward structure chosen can have a dramatic impact on performance. When using a multi-layer perceptron as a function approximator, evolution outperforms TDL by a significant margin. Overall, the best results were obtained by evolving multi-layer perceptrons.