Reproduce “Look-ahead Reasoning with a Learned Model in Imperfect Information Games” 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
Ondřej Kubíček, Viliam Lisý
LAMIR, an algorithm developed at the Artificial Intelligence Center, Czech Technical University in Prague, learns an abstract model of game dynamics directly from interaction in imperfect information games (IIGs). This enables scalable depth-limited look-ahead reasoning, which significantly outperforms strong model-free baselines like RNaD in large IIGs, achieving win rates up to 80.49% in II Goofspiel 15.
Key insights
- It is possible to successfully learn an abstract model of imperfect information game dynamics from interaction, enabling principled look-ahead reasoning without explicit game rules.
- Learning and applying information set abstraction within public states is an effective method for scaling depth-limited look-ahead search to previously intractable large IIGs.
- Integrating a learned abstract model with continual resolving significantly boosts the strategic strength and robustness of pre-trained model-free policies in complex IIGs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.