Development of Ghost Controller for Ms Pac-Man Versus Ghost Team with Grammatical Evolution

These days, artificial intelligence (AI) has been used in game AI. Additionally, video game AI is studied actively in late years, for example, application of commercial game or competition etc. In many video games of recent years, real-time action and non-player characters have been required to attract players. This paper describes how to develop a ghost team controller using evolutionary system to play the video game, Ms Pac-Man. Ms Pac-Man has been used as a testbed of AI, especially multi-agent system. We propose a method to generate the ghost team controller with Grammatical Evolution. In case of developing Ms Pac-man agent with Evolutionary Computation using fitness function, the criterion of the fitness is used its obtained high score in many cases. In contrast, ghost team has to prevent Ms Pac-man to get high score, namely hold score in check. However, if Ms Pac-man is captured in low score by accident, its ghost strategy have a possibility to survive next generation, and if the ghosts pursue Ms Pac-man in a line, agent isn't captured for all time. Therefore developing ghost team agent is required to avoid these issues, and we introduced a penalty to the fitness, grammar like instinct and to attack Ms Pac-Man on both sides. This paper introduces experimental data about the ghost team controller for Ms Pac-Man versus ghost team, we used ghost team agents and tested them Ms Pac-Man agents. The experimental results showed that proposed system could catch Ms Pac-Man agent compare with simple hand-coded ghost teams, and the evolved controller we made worked effectively. These results are concluded that proposed method works effectively for generating ghost controller.