Model-based reinforcement learning (MBRL) has achieved remarkable results in single-agent domains, yet its extension to competitive imperfect information games (IIGs) remains underexplored. In multi-agent settings, opponent-induced non-stationarity complicates the learning process, and decentralized model learning faces severe identifiability barriers, which we argue make centralized model learning a mathematical necessity. Building on this analysis, we propose NashDreamer, a principled MBRL framework for two-player zero-sum IIGs. It introduces a centralized Multi-Agent Recurrent State-Space Model (MARSSM) that decouples environment dynamics from the effect of players' strategies on their individual observations. NashDreamer is designed to use arbitrary policy gradient algorithms and inherits their convergence guarantees towards Nash equilibria under an idealized model. Empirical evaluations across four benchmark games demonstrate that NashDreamer substantially improves sample efficiency over model-free baselines early in the training. Finally, we theoretically analyze the architecture's optimization landscape, identifying the vulnerability of the Dreamer family of algorithms to posterior collapse in stochastic environments. We leave it as an open challenge.
Chase McDonald, Nathan Tsang, Wesley N. Kerrcs.AI cs.GT
We present FootsiesGym, an open-source environment for learning in a non-trivial two-player, zero-sum, imperfect-information game. Built on HiFight's minimalist 2D fighting game Footsies, it isolates the cyclic, non-transitive strategic interactions of fighting game neutral play while remaining simple enough for efficient analysis. We provide a vectorized simulator that enables high-throughput training on standard hardware, making the environment accessible and reproducible. We describe the design of the environment, benchmark several reinforcement learning algorithms, and discuss open research directions it enables. The code is available at https://github.com/como-research/FootsiesGym.
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. We ask whether they instead select different members of the Nash set, systematically as a function of the algorithm rather than the seed. Using a tabular, exactly solvable testbed of six games with analytically known Nash sets -- including a two-dimensional Nash polytope and Kuhn poker -- we find that (i) selection is determined by the algorithm, not the seed, but families differ only on asymmetric Nash sets; (ii) regularized last-iterate methods (R-NaD, magnetic mirror descent) select the maximum-entropy member, the information projection of their uniform reference onto the Nash set -- exactly on the 2-D polytope and at 99.7% of maximum entropy in Kuhn -- while regret-averaging methods (CFR, CFR+, fictitious play) drift to a lower-entropy face; we confirm this on a randomized 180-game ensemble, where R-NaD attains the maximum-entropy member in 100% of converged games while CFR+ sits strictly below it in 94% (paired Wilcoxon p < 10^-27); (iii) the selected member has downstream consequences against sub-optimal opponents that scale with sequential/hidden-information structure but stay bounded -- in Kuhn the max-entropy member is a strictly better hedge, whereas on the matrix games the members differ without either dominating. We also report two negative results correcting common intuitions: removing CFR's positive-orthant (max(R,0)) projection does not eliminate boundary drift; and R-NaD's selection is anchor-following, not initialization-independent. We state the maximum-entropy / I-projection characterization as a strongly data-supported conjecture, checked throughout against analytic ground truth.
Self-play with naive gradient ascent cycles in two-player zero-sum games: the last iterate orbits the equilibrium. Modern methods restore last-iterate convergence by regularizing toward a reference policy -- MMD a fixed one (reaching only the regularized equilibrium), R-NaD a periodic snapshot (the engine of DeepNash). We study GARIP, which anchors to the running average, and isolate what the choice of reference controls. Our central result is a mechanism: collapse tracks the peak lag of the reference, and among causal convex averages of a fixed mean lag the running average (flat profile, peak $=$ mean) uniquely minimizes that peak, while a snapshot's sawtooth has peak $= 2\times$ mean (a one-line theorem). Two consequences follow. Convergence: we prove local last-iterate convergence at constant anchor strength -- the anchor scales the base map's rotation by $1-β$, crossing the stability boundary and turning a recurrent base into a contraction (global convergence is conjectured at small $β$; we characterize a large-$β$ consensus failure). Robustness: GARIP matches R-NaD's peak performance -- on matrix games, the Coin Game, and the board games Connect Four/Othello, both moving references are far more robust than fixed-magnet and magnet-free baselines -- but is the better hyperparameter default; we report it both ways: over the full grid collapse rates are statistically indistinguishable, yet at conventional parameterizations a matched-mean-lag setting collapses in 0/40 vs 10/40 seeds (a snapshot matches it only by knowing to shorten $K$). The boundaries: an anticipatory (negative-weight) reference does better still on the stale side, and the advantage appears only where naive self-play cycles (five deep self-play loops). All experiments are pure JAX and reproducible.