Nested Rollout Policy Adaptation for Monte Carlo Tree Search
Christopher D. Rosin
Monte Carlo tree search (MCTS) methods have had recent success in games, planning, and optimization. MCTS uses results from rollouts to guide search; a rollout is a path that descends the tree with a randomized decision at each ply until reaching a leaf. MCTS results can be strongly influenced by the choice of appropriate policy to bias the rollouts. Most previous work on MCTS uses static uniform random or domain-specific policies. We describe a new MCTS method that dynamically adapts the rollout policy during search, in deterministic optimization problems. Our starting point is Cazenave's original Nested Monte Carlo Search (NMCS), but rather than navigating the tree directly we instead use gradient ascent on the rollout policy at each level of the nested search. We benchmark this new Nested Rollout Policy Adaptation (NRPA) algorithm and examine its behavior. Our test problems are instances of Crossword Puzzle Construction and Morpion Solitaire. Over moderate time scales NRPA can substantially improve search efficiency compared to NMCS, and over longer time scales NRPA improves upon all previous published solutions for the test problems. Results include a new Morpion Solitaire solution that improves upon the previous human-generated record that had stood for over 30 years.