情報処理学会 第78回全国大会講演要旨

This paper presents a comparative analysis between Reinforcement Learning (RL) and Evolutionary Strategy (ES) for training rollout bias in General Video Game Playing (GVGP). GVGP has become an emerging research field, where researchers attempt to develop AI programs that can play various types of video game without prior knowledge. Monte-Carlo Tree Search, which does not require explicit evaluation function, has been a popular technique in GVGP, and existing research has succeeded in improving its performance by biasing the rollouts with weights vector, which is trained by ES. This paper compares RL and ES, and investigates the advantages and disadvantages of both techniques as rollout bias training mechanism in the GVGP domain.