A new preprint on arXiv examines decentralized bandit learning in discounted infinite-horizon, general-sum Markov games. The setting combines continuous action spaces with agents that observe only the public state and their own realized rewards, so each agent must learn from limited feedback without seeing opponents' payoffs.

The paper targets Markov perfect equilibria, a standard solution concept for stochastic games. According to the abstract, the proposed approach relies on simultaneous perturbation, a gradient-estimation technique suited to bandit settings.

Because this is a single source and the abstract is brief, the write-up does not include comparisons with other methods or experimental results.