Dynamic resource allocation—adjusting the number of parallel simulations and how budget is split between them—makes MCTS stronger in games with hidden information, without requiring domain-specific tuning.
This paper improves Monte Carlo Tree Search (MCTS) for games with hidden information and randomness by dynamically adjusting how many parallel game simulations to run and how to distribute computational budget across them. The authors test their approach on three card games and show it plays stronger than the baseline method.