Systematic Improvement of Monte-Carlo Tree Search with Self-generated Neural-Networks Controllers

摘要

In UCT algorithm, a large number of Monte-Carlo simulations are performed and their rewards are averaged to evaluate a specified action. In this paper, we propose a general approach to enhance the UCT algorithm with knowledge-based neural controllers by adjusting the probability distribution of UCT simulations. Experimental results on Dead End, the classical predator/prey game, show that our approach improves the performance of UCT significantly.