Evolving a Minimum Input Neural Network Based Controller for the Pac-Man Agent

Abstract

In this paper we present the development and implementation of a neural network-based controller for the Pac-Man agent which is fed with minimal information about the environment. The implementation of the game we used differs from the original. In this regard some key aspects were changed in order to provide new challenges for the agent andthus better test our controller. A neural network is used to compute a desirability value for the locations to which the agent can move. The non-deterministic nature of the game does not allow for fast and accurate feedback, thus a different method for training was developed. Accordingly, we considered a genetic algorithm to evolve weights for the neural network. To conclude, the controller adapts very well to the environment, resulting in well-trained agents that can complete several consecutive levels of the game.