Learning Stationary Correlated Equilibria in Constrained General-Sum Stochastic Games

We study constrained general-sum stochastic games with unknown Markovian dynamics. A distributed constrained no-regret Q-learning scheme (CNRQ) is presented to guarantee convergence to the set of stationary correlated equilibria of the game. Prior art addresses the unconstrained case only, is structured with nested control loops, and has no convergence result. CNRQ is cast as a single-loop three-timescale asynchronous stochastic approximation algorithm with set-valued update increments. A rigorous convergence analysis with differential inclusion arguments is given which draws on recent extensions of the theory of stochastic approximation to the case of asynchronous recursive inclusions with set-valued mean fields. Numerical results are given for the exemplary application of CNRQ to decentralized resource control in heterogeneous wireless networks (HetNets).