Monte Carlo Tree Search

Monte Carlo Tree Search builds a partial decision tree, repeatedly selecting promising branches, evaluating outcomes through simulations or learned models, and backing those evaluations up to guide later choices. Common approaches include UCT and PUCT exploration rules, neural policy and value guidance, progressive widening for large or continuous action spaces, and parallel search. MCTS is used in games, robotics and partially observable planning, as well as language-model tasks such as mathematical reasoning, coding, and tool use; evaluations span game-playing tests, POMDP benchmarks, and reasoning benchmarks. A central challenge is allocating limited computation effectively: search depends on the quality of its models and evaluators, and more search does not always improve results, so methods need reliable uncertainty handling and fair equal-budget comparisons.