This is my final project for metaheuristic course. A detail report in IEEE format is at report.docx The idea of this work is originated from the paper Evolving Multimodal Behavior With Modular Neural Networks in Ms. Pac-Man The features in the paper end up to be linear combination. From my opinion, the combination function should have some other forms.

I try to use GA to delvelop my own agent. In my agent design, I use an evaluation function to evaluate every successor states, and take the one with the highest evaluation score. Fix features including distance to opponent agent, distance to food, distance to the middle line, etc. Each feature comes with a weight, which will be trained in GA. The setup of GA is stated in the section 3 of the report. My agents and trainning scripts are in "my_works/".

According to the result, all of the game scores is positive, indicating that our agent wins all games and does a decent job even on maps other than the one it is trained on. The result data is inside the “pacman_contest/data/”, the snapshots of result is in "result_snapshot/".