Machine learning algorithm design for playing Ms. Pacman using Java

Abstract

The Ms. Pac-Man competition aims to develop an intelligent computer controlled Ms. Pac-Man. This project is to design an algorithm to decide the next move performed by the agent in Ms. Pac-Man game as quick as hard-coded controller does, but in an adaptive way that the mechanism can also learn to achieve better performance as time goes. A Situation-Strategy Mapping (SSM) algorithm is designed to provide a desirable strategy, given a list of game parameters. To validate the approach, both a simple and severe environment is used. Initially, a SSM controller plays against a perfect player in Tic-Tac-Toe game. Then the SSM controller is placed inside Ms. Pac-Man game. The experimental results indicate that the optimization of feed-in parameters produces higher performance, given the same set of strategies. Moreover, the result suggests that the SSM algorithm is capable of making a decision even for unseen situation and steadily improving the game score of Ms. Pac-Man game during training.