| Pac-Man is a popular chasing and evading game in the world. The Non-Player Character（NPC）Ghost in this game is controlled by script. Generally, the NPCs controlled by scriptoften have fixed actions, which are considered unintelligent and will affect the playability ofthe games. In order to obtain flexible agents, evolutionary neural network can be used to trainthe ghosts to get self-learning ones. The related researches in literature on this game aremostly focused on enhancing the evading ability to achieve higher scores, while there is littlework on the NPCs in this game.Yannakakis has developed the method of evolutionary neural network on the study ofGhost. There are two main shortcomings in his work. The first one is that the trained Ghostonly has one action style; the other is the used simulation environment is quite different fromthe original Pac-Man game, and his fitness function with parameters are not suit for realenvironment. Based on these analyses, we have done the following work in this research.Firstly, we preserve the settings of different Ghost characters in original game and designcorresponding fitness functions for Ghosts in order to generate different actions. Secondly, weuse the real simulation environment through game engine in WCCI2008, and design newneural network input, appropriate parameters and fitness function to describe the game state.Finally, through setting a penalty in the learning process, the activity range of the Ghost isexpanded. Experimental results have shown that the agent learns well and the proposedmethod is feasible and effective. |
