Design and Implementation of Pac-Man Strategies with Embedded Markov Decision Process in a Dynamic, Non-Deterministic, Fully Observable Environment

Abstract

This project designs and implements Pac-Man strategies, whose decision-making protocol is solely based on Markov Decision Process, without the support of pathfinding algorithms nor heuristic functions, in a dynamic, non-deterministic, fully observable environment. This project provides the rare opportunity to refine the understanding of, and practice the application of Markov Decision Process in a classic arcade game setting. After significant number of hours of parameter tuning, my design achieved a win rate ranging between 50% and 60%. With the proven effectiveness of embedded Markov Decision Process, a spin-off Pac-Man AI project that incorporates the advantages of pathfinding algorithms, heuristic functions, and Markov Decision Process shall be on the agenda.

Design Document

Design of Pac-Man Strategies with Embedded Markov Decision Process in a Dynamic, Non-Deterministic, Fully Observable Environment

Style of Work

• Coding style complies with industry standard
• Design paper complies with IEEE Conference format

How to run

This version of Intelligent Pac-Man is running onPython 2.7 environment. The visibility of the agent in this Pac-Man project is extremely limited, just like someone is running through the maze for real.

Run the following command to play the game:

python pacman.py --pacman MDPAgent --layout mediumClassic --numGames 50