An Agent that Learns to Play Pac Man

Exhibitor: Mark Ledwich

Supervisor: Marcus Gallagher

Research Group: Complex and Intelligent Systems

Industry Sector:

The aim of this project is to develop an agent that learns to play Pacman. It will determine whether an Evolutionary Neural Network can be applied as a complete solution to the Pacman AI problem. The effect of different sized network inputs on the agents learning ability will also be examined.

The plane centred around Pacman is used as input to the network which output the direction of movement. Three populations with different sized inputs of 30 Pacman are "Evolved" separately. Each Pacman has a corresponding neural network who's weights are stored as a genome. At each generation 15 agents are discarded, random parents are selected from the remaining population for "breeding" of new genomes. This process repeats until the agents have not improved for many generations.

Evolving Neural networks is not an exact science, there are many constants that need to be chosen individually for a problem. Better choices for these parameters will allow the agents to reach higher fitness values. A simplified version of the Pacman game is used with one ghost and no power pills. If successful in this reduced game, power pills and extra ghosts can be added and learning continued. The results of this experiment will show how well this approach works as a total solution to an AI problem. A successful outcome is important because the same software and methodology can be applied to other problems with very little implementation.

Thesis Document (PDF)