Efficient representations for team and imperfect-recall equilibrium computation

A team of researchers including Luca Carminati, Brian Hu Zhang, Federico Cacciamani, and Tuomas Sandholm from MIT, CMU, and Politecnico di Milano have published a breakthrough in computational game theory. Their paper, 'Efficient representations for team and imperfect-recall equilibrium computation,' tackles the NP-hard problem of computing mixed-strategy Nash equilibria in two-player zero-sum games with imperfect recall—equivalent to two-team zero-sum games. The key innovation is the 'belief game,' a perfect-recall construction that is equivalent to the original imperfect-recall game. While the belief game can be exponentially larger, the authors show its strategy spaces can be directly represented as a directed acyclic graph (DAG), called the Team Belief DAG (TB-DAG), which offers exponential speedups over naive approaches.

TB-DAG simultaneously achieves essentially optimal parameterized complexity bounds and integrates seamlessly with efficient regret minimization techniques like counterfactual regret minimization (CFR). The paper also establishes completeness results: finding Nash equilibria in mixed and behavioral strategies for these games is Δ₂ᴾ-complete and Σ₂ᴾ-complete, respectively. Experimentally, TB-DAG paired with existing learning algorithms yields state-of-the-art performance on a wide variety of benchmark team games, demonstrating practical viability. This work consolidates and supersedes four previous arXiv preprints, representing a major step toward practical equilibrium computation in complex multi-agent settings.