MAPLE: New Tree Search Method for Imperfect-Information Games
A new approach called Multi-State Aggregated Policy Evaluation (MAPLE) has been introduced by researchers. This tree search technique is designed for imperfect-information games (IIGs) and consolidates policy and value assessments from various sampled world states within one search tree. MAPLE leverages the strengths of both Perfect Information Monte Carlo (PIMC) and Information Set Monte Carlo Tree Search (IS-MCTS), all while keeping computational costs manageable. To identify informative world states, a Siamese-based sampling strategy is employed. Tests conducted on Phantom Go and Dark Hex showcase the effectiveness of MAPLE. The findings are available on arXiv under the identifier 2605.24139.
Key facts
- MAPLE aggregates policy and value evaluations from multiple sampled world states.
- It combines PIMC and IS-MCTS advantages with controllable computational cost.
- A Siamese-based sampling strategy selects informative world states.
- Experiments were conducted on Phantom Go and Dark Hex.
- The paper is available on arXiv (2605.24139).
- Imperfect-information games require decisions without full game state observation.
- AlphaZero succeeded in perfect-information games but faces challenges in IIGs.
- PIMC suffers from strategy fusion; IS-MCTS has high computational cost with neural networks.
Entities
Institutions
- arXiv