Extreme Value Monte Carlo Tree Search

원문 초록

Despite being successful in board games and reinforcement learning (RL), UCT, a Monte-Carlo Tree Search (MCTS) combined with UCB1 Multi-Armed Bandit (MAB), has had limited success in domain-independent planning until recently. Previous work showed that UCB1, designed for $[0,1]$-bounded rewards, is not appropriate for estimating the distance-to-go which are potentially unbounded in $\mathbb{R}$, such as heuristic functions used in classical planning, then proposed combining MCTS with MABs designed for Gaussian reward distributions and successfully improved the performance. In this paper, we further sharpen our understanding of ideal bandits for planning tasks. Existing work has two issues: First, while Gaussian MABs no longer over-specify the distances as $h\in [0,1]$, they under-specify them as $h\in [-\infty,\infty]$ while they are non-negative and can be further bounded in some cases. Second, there is no theoretical justifications for Full-Bellman backup (Schulte & Keller, 2014) that backpropagates minimum/maximum of samples. We identified \emph{extreme value} statistics as a theoretical framework that resolves both issues at once and propose two bandits, UCB1-Uniform/Power, and apply them to MCTS for classical planning. We formally prove their regret bounds and empirically demonstrate their performance in classical planning.