Monte-Carlo Tree Search by Best Arm Identification
This work addresses game tree search for applications like AI gaming or decision-making, presenting an incremental improvement over prior methods.
The paper tackled the problem of efficiently identifying the optimal move in game trees with stochastic payoffs by developing new algorithms that summarize deeper tree levels into confidence intervals and apply best arm identification at the root, achieving improved sample complexity and matching or outperforming existing methods in experiments.
Recent advances in bandit tools and techniques for sequential learning are steadily enabling new applications and are promising the resolution of a range of challenging related problems. We study the game tree search problem, where the goal is to quickly identify the optimal move in a given game tree by sequentially sampling its stochastic payoffs. We develop new algorithms for trees of arbitrary depth, that operate by summarizing all deeper levels of the tree into confidence intervals at depth one, and applying a best arm identification procedure at the root. We prove new sample complexity guarantees with a refined dependence on the problem instance. We show experimentally that our algorithms outperform existing elimination-based algorithms and match previous special-purpose methods for depth-two trees.