Reactive control of Ms. Pac Man using information retrieval based on Genetic Programming

Abstract

During the last years the well-known Ms. Pac Man video game has been - and still is - an interesting test bed for the research on various concepts from the broad area of Artificial Intelligence (AI). Among these concepts is the use of Genetic Programming (GP) to control the game from a human player's perspective. Several GP-based approaches have been examined so far, where traditionally they define two types of GP terminals: one type for information retrieval, the second type for issuing actions (commands) to the game world. However, by using these action terminals the controller has to manage actions issued to the game during their runtime and to monitor their outcome. In order to avoid the need for active task management this paper presents a novel approach for the design of a GP-based Ms. Pac Man controller: the proposed approach solely relies on information retrieval terminals in order to rate all possible directions of movement at every time step during a running game. Based on these rating values the controller can move the agent through the mazes of the the game world of Ms. Pac Man. With this design, which forms the main contribution of our work, we decrease the overall GP solution complexity by removing all action control management tasks from the system. It is demonstrated that by following the proposed approach such a system can successfully control an autonomous agent in a computer game environment on the level of an amateur human player.