Zeroth-Order Learning in Continuous Games via Residual Pseudogradient Estimates

A variety of practical problems can be modeled by the decision-making process in multi-player games where a group of self-interested players aim at optimizing their own local objectives, while the objectives depend on the actions taken by others. The local gradient information of each player, essential in implementing algorithms for finding game solutions, is all too often unavailable. In this paper, we focus on designing solution algorithms for multi-player games using bandit feedback, i.e., the only available feedback at each player's disposal is the realized objective values. To tackle the issue of large variances in the existing bandit learning algorithms with a single oracle call, we propose two algorithms by integrating the residual feedback scheme into single-call extra-gradient methods. Subsequently, we show that the actual sequences of play can converge almost surely to a critical point if the game is pseudo-monotone plus and characterize the convergence rate to the critical point when the game is strongly pseudo-monotone. The ergodic convergence rates of the generated sequences in monotone games are also investigated as a supplement. Finally, the validity of the proposed algorithms is further verified via numerical examples.