An Agent that Learns to Play Pac Man

Exhibitor: Donald Shepherd

Supervisor: Marcus Gallagher

Research Group: Complex and Intelligent Systems

Industry Sector: Media / Entertainment

An Agent that Learns to Play Pacman

The aim of this thesis project is to produce an agent capable of playing Pacman on its own, using only the information available to a normal human player. This is to be done using evolutionary techniques, in this case, the use of a genetic algorithm on the parameters defining the behaviour of the system. This system consists of ten separate yet identical genetic algorithms, working on twenty Pacman "chromosomes" across one hundred generations.

Within the system, the agent views each object (ghost, food, power pill and so on) as a centre of positive or negative influence, with each of these centres interacting to negate or complement each other. This essentially forms a potential field of sorts, based around the "goodness" of each object and the distance it is away from Pacman. The agent then heads towards areas of higher potential rather than low, recalculating the direction in which to head in after each movement.

Preliminary results indicate a high level of suitability of this model as a solution, including a marked improvement over previous efforts at this problem. However, modifications may be necessary to achieve an optimal genetic makeup, for instance an increased awareness of the state of the maze and the ghosts, or more specific implementations of the rules of the game rather than hoping that the agent will learn these itself through its playing.

- Donald Shepherd
s363224@student.uq.edu.au

Thesis Document (PDF)