Offline Equilibrium Finding in Extensive-form Games: Datasets, Methods, and Analysis

ICLR 2026 Conference Submission17471 Authors

Keywords: Game Theory, Equilibrium Finding, Offline Learning

Abstract: Recently, offline reinforcement learning (Offline RL) has emerged as a promising paradigm for solving real-world decision-making problems using pre-collected datasets. However, its application in game theory remains largely unexplored. To bridge this gap, we introduce ***offline equilibrium finding*** (Offline EF) in extensive-form games (EFGs), which aims to compute equilibrium strategies from offline datasets. Offline EF faces three key challenges: the lack of benchmark datasets, the difficulty of deriving equilibrium strategies without access to all action profiles, and the impact of dataset quality on effectiveness. To tackle these challenges, we first construct diverse offline datasets covering a wide range of games to support algorithm evaluation. Then, we propose BOMB, a novel framework that integrates behavior cloning within a model-based method, enabling seamless adaptation of online equilibrium-finding algorithms to the offline setting. Furthermore, we provide a comprehensive theoretical analysis of BOMB, offering performance guarantees across various offline datasets. Extensive experimental results show that BOMB not only outperforms traditional offline RL methods but also achieves highly efficient equilibrium computation in offline settings.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 17471