Pacman using genetic algorithms and neural networks

Abstract

Pacman is an arcade game where a group of "ghosts" try to catch a Pacman as he attempts to eat all the dots in a maze in order to progress to the next level. The standard Pacman game uses traditional AI search techniques for the ghosts and does not utilize the concept of machine learning. We propose to use an evolutionary computational method instead that implements a combination of genetic algorithms and neural networks in order to improve the level of play for the ghosts. We use a unique genetic structure for each ghost, a metric for determining the fittest ghosts of the current generation, and define parameters such as crossover probability and mutation percentage in order to create genetic offspring for the new population. Furthermore, we use a backprop neural network to optimize the crossover probability and mutation percentage values in order to minimize search time. The final test simulation successfully demonstrates the combined application of GA and NN in Pacman and its potential for machine learning in similar games.