[arXiv]score: 0.58
MAPLE Aggregates Multi-State Policy Evaluations to Reduce Strategy Fusion in AlphaZero for Imperfect-Info Games
May 26, 2026
MAPLE combines PIMC and IS-MCTS by aggregating policy/value evaluations across sampled world states in a single search tree with Siamese-based state selection, outperforming baselines on Phantom Go and Dark Hex at controllable compute cost.
cs.AIcs.LG
HOW THIS AFFECTS YOU
●
researcherThe Siamese sampling strategy for informative world-state selection is the key technical contribution worth examining for scaling MCTS-based methods to larger imperfect-information games.