Fuzzy Q-learning in a nondeterministic environment: developing an intelligent Ms. Pac-Man agent

Abstract

This paper reports the results from training an intelligent agent to play the Ms. Pac-Man video game using variations of a fuzzy Q-learning algorithm. This approach allows us to address the nondeterministic aspects of the game as well as finding a successful self-learning or adaptive playing strategy. The strategy presented is a table based learning strategy, in which the intelligent agent analyzes the current situation of the game, stores various membership values for each of the several contributors to the situation (distance to closest pill, distance to closest power pill, and distance to closest ghost), and makes decisions based on these values.