Biasing Monte-Carlo Rollouts with Potential Field in General Video Game Playing
Chun Yin Chu, Tomohiro Harada, Ruck Thawonmas · 2015
This paper proposes the use of potential field and biased Monte Carlo rollout in General Video Game Playing (GVGP). Monte-Carlo Tree Search is a famous technique for General Video Game Playing, thanks to its adaptability. However, since the rollouts are performed randomly, it may not be able to search the game state efficiently. Existing research has attempted to bias the rollout by using Euclidean distances to the closest sprites as features, and training the bias weights with Evolutionary Strategy. In this paper, we propose the use of potential field features instead of Euclidean distances as the rollout bias, so as to further improve the performance of Monte-Carlo Tree Search in General Video Game Playing.