Summary
Many two-player zero-sum games possess not a unique Nash equilibrium but a convex set of equilibria sharing the same minimax value. This paper asks whether standard solvers, often treated as interchangeable, systematically select different members of the Nash set as a function of the algorithm rather than the random seed. Using a tabular, exactly solvable testbed of six games with analytically known Nash sets — including a two-dimensional Nash polytope and Kuhn poker — the authors show that (i) equilibrium selection is determined by the algorithm, not the seed, with families differing only on asymmetric Nash sets; (ii) regularized last-iterate methods such as R-NaD and magnetic mirror descent select the maximum-entropy member, while regret-averaging methods such as CFR, CFR+, and fictitious play drift toward lower-entropy faces, a result confirmed on a stochastic set of 180 games; and (iii) the selected member has downstream consequences against suboptimal opponents. Two negative results correct common intuitions: removing CFR's orthogonal quadrant projection does not eliminate boundary drift, and R-NaD's selection is anchor-following rather than initialization-independent.
Paper: arXiv:2606.28308 — Luis Leal, ML, posted 2026-06-26.
Abstract (translated/condensed)
Many two-player zero-sum games admit not a unique Nash equilibrium but a convex set of them: a polytope of profiles that all share the minimax value V* yet prescribe different behaviour. Standard solvers each converge to some equilibrium and are treated as interchangeable. This work asks whether they instead select different members of the Nash set, systematically as a function of the algorithm rather than the seed.
Key findings
Using a tabular, exactly solvable testbed of six games with analytically known Nash sets — including a two-dimensional Nash polytope and Kuhn poker:
- Algorithm, not seed, decides. Equilibrium selection is determined by the algorithm rather than the random seed, though solver families differ only on asymmetric Nash sets.
- Regularized last-iterate methods select maximum entropy. R-NaD and magnetic mirror descent converge to the maximum-entropy member of the Nash set, while regret-averaging methods (CFR, CFR+, fictitious play) drift toward lower-entropy faces. This was confirmed on a stochastic set of 180 games.
- Downstream consequences. The selected equilibrium member matters against suboptimal opponents.
Negative results
- Removing CFR's orthogonal quadrant projection does not eliminate boundary drift.
- R-NaD's selection is anchor-following, not initialization-independent.
---
*Auto-collected on 2026-06-30.*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178208312