Speaker: Nicolas Gast, Inria
Title: Approximating the bias of stochastic processes (with applications to stochastic approximation and mean field limits)
Time and place: May 14, 2024 at 10:00 in Salle Lagrange Gris, Inria, Sophia Antipolis
Abstract: Stochastic approximation algorithms are quite popular in reinforcement learning notably because they are powerful tools to study the convergence of algorithms based on stochastic gradient descent (like Q-learning of policy gradient). Mean field limit are quite popular in performance evaluation because they simplify the analysis of system of interacting queues. In this talk, I will talk about the link between mean field limits and constant step-size stochastic approximation. I will present methods to characterize the asymptotic bias of such methods. The analysis is based on tools similar to Stein’s method.
