Fixed-point error bounds for mean-payoff Markov decision processes

Information Systems and Operations Management & Economics and Decision Sciences

Intervenant: Roberto Cominetti (UAI)

Salle S118

Abstract

We discuss the use of optimal transport techniques to derive finite-time error bounds for reinforcement learning in mean-payoff Markov decision processes. The results are obtained as a special case of stochastic Krasnoselski—Mann fixed point iterations for nonexpansive maps. We present sufficient conditions on the stochastic noise and stepsizes that guarantee almost sure convergence of the iterates towards a fixed point, as well as non-asymptotic error bounds and convergence rates. Our main results concern the case of a martingale difference noise with variances that can possibly grow unbounded. We also analyze the case of uniformly bounded variances, and how they apply for Stochastic Gradient Descent in convex optimization.