Scale-Adaptive Balancing of Exploration and Exploitation in Classical Planning

Stephen Wissow, Masataro Asai

Keywords: classical planning, multi-armed bandit, monte-carlo tree search

TL;DR: We present a new MCTS/THTS algorithm for classical planning, using a new Multi-Armed Bandit algorithm based on UCB1-Normal, that outperforms GBFS and MCTS/THTS using UCB1 and deepens our theoretical understanding of heuristic search in general.

Abstract: Balancing exploration and exploitation has been an important problem in both adversarial games and automated planning. While it has been extensively analyzed in the Multi-Armed Bandit (MAB) literature, and the game community has achieved great success with MAB-based Monte Carlo Tree Search (MCTS) methods, the symbolic planning community has struggled to advance in this area. We describe how Upper Confidence Bound 1’s (UCB1’s) assumption of reward distributions with known bounded support shared among siblings (arms) is violated when MCTS/Trial-based Heuristic Tree Search (THTS) in previous work uses heuristic values of search nodes in classical planning problems as rewards. To address this issue, we propose a new Gaussian bandit, UCB1-Normal2, and analyze its regret bound. It is variance-aware like UCB1-Normal and UCB-V, but has a distinct advantage: it neither shares UCB-V’s assumption of known bounded support nor relies on UCB1-Normal’s unfounded conjectures on Student’s t and χ2 distributions. Our theoretical analysis predicts that UCB1-Normal2 will perform well when the estimated variance is accurate, which can be expected in deterministic, discrete, finite state-space search, as in classical planning. Our empirical evaluation confirms that MCTS combined with UCB1-Normal2 outperforms Greedy Best First Search (traditional baseline) as well as MCTS with other bandits.