Improvement of Monte Carlo Tree Search AI using TD(λ) in Ms. Pac-Man
Hiuchi Akihiko, Miwa Makoto, Tsuruoka Yoshimasa, Chikayama Takashi · 2013
Due to its simplicity of the rules and high degree of difficulty, Ms. Pac-Man has become an at- tractive subject of research on digital game AI. In Ms. Mac-Man, Monte Carlo Tree Search-based, especially UCT-based players tend to give better performance than rule-based systems. In this study, we propose the improvement of Monte-Carlo tree search using the evaluation function that have been learned in TD ( ). At the use of the evaluation function in Monte Carlo tree search, we adopt the Progressive bias. Progressive bias makes the selection follow the evaluate function in opening phase of Monte-Carlo Tree Search and increase the search of favorable states. Experimental results obtained on a simulator show that the proposed method gives better scores than an existing UCT-based method.