Keywords: Monte Carlo tree search, Context-specific independence
Abstract: Monte Carlo Tree Search (MCTS) has showcased its efficacy across a broad spectrum of decision-making problems. However, its performance often degrades under vast combinatorial action space, especially where an action is composed of multiple sub-actions. In this work, we propose an action abstraction based on the compositional structure between a state and sub-actions for improving the efficiency of MCTS under a factored action space. Our method learns a latent dynamics model with an auxiliary network that captures sub-actions relevant to the transition on the current state, which we call state-conditioned action abstraction. Notably, it infers such compositional relationships from high-dimensional observations without the known environment model. During the tree traversal, our method constructs the state-conditioned action abstraction for each node on-the-fly, reducing the search space by discarding the exploration of redundant sub-actions. Experimental results demonstrate the superior sample efficiency of our method compared to vanilla MuZero, which suffers from expansive action space.
List Of Authors: Kwak, Yunhyeok and Hwang, Inwoo and Kim, Dooyoung and Lee, Sanghack and Zhang, Byoung-Tak
Latex Source Code: zip
Signed License Agreement: pdf
Code Url: https://github.com/yun-kwak/efficient-mcts
Submission Number: 65
Loading