KrwEmd: Revising the Imperfect Recall Abstraction from Forgetting Everything

FU Yanchang, Junge Zhang, Bai dongdong, Zhaolingyun, Kaiqi Huang

Keywords: game theory, imperfect-information games, games with ordered signals, computer poker, imperfect-recall abstraction

Abstract: Recent research has shown that an extreme interpretation of imperfect recall abstraction —- completely forgetting all past information —- has led to excessive abstraction issues. Currently, there are no hand abstraction algorithms that effectively integrate historical information. This paper aims to develop the first such algorithm. Initially, we introduce the KRWI abstraction for Texas Hold'em-style games, which categorizes hands based on K-recall winrate features that incorporate historical information. Statistical results indicate that, in terms of the number of distinct infosets identified, KRWI significantly outperforms POI, an abstraction that identifies the most abstracted infosets that forget all historical information. Following this, we introduce the KrwEmd algorithm, the first hand abstraction algorithm to effectively use historical information by combining K-recall win rate features and earth mover's distance for hand classification. Empirical studies conducted in the Numeral211 Hold'em environment show that under identical abstracted infoset sizes, KrwEmd not only surpasses POI but also outperforms state-of-the-art hand abstraction algorithms such as EHS and PAEMD. These findings suggest that incorporating historical information can significantly enhance the performance of hand abstraction algorithms, positioning KrwEmd as a promising approach for advancing strategic computation in large-scale adversarial games.

Primary Area: Algorithmic game theory

Submission Number: 18392