Reproduce “Re-determinizing Information Set Monte Carlo Tree Search in Hanabi” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
James Goodman
An extension to Information Set Monte Carlo Tree Search, Re-determinizing IS-MCTS (RIS-MCTS), addresses information leakage in imperfect-information cooperative games like Hanabi. This approach, which won the CIG 2018 Hanabi competition, combines game-specific heuristics and an offline training methodology for state evaluation to achieve high performance under strict time constraints.
Key insights
- Standard MO-ISMCTS exhibits a critical pathology in Hanabi, where information leakage and strategy fusion render it largely ineffective, scoring very low regardless of computational resources.
- Re-determinizing hidden information for each active player within the MCTS tree effectively mitigates information leakage, significantly improving opponent modeling and overall agent performance.
- The combination of domain-specific heuristic rules for action space restriction and the integration of shared conventions (like "playable now") dramatically improves MCTS performance by enhancing communication and search efficiency.
- Utilizing training data from deeper, partially explored MCTS nodes ("penumbra of exploration") improves the robustness and generalization of learned evaluation functions for real-time game AI.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.