Zap Meets Momentum: New Stochastic Approximation Algorithms and Applications to Reinforcement Learning
Wed, Dec 12, 2018 @ 12:00 PM - 01:00 PM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Adithya Devaraj, University of Florida
Talk Title: Zap Meets Momentum: New Stochastic Approximation Algorithms and Applications to Reinforcement Learning
Series: Center for Cyber-Physical Systems and Internet of Things
Abstract: Stochastic approximation algorithms are used to approximate solutions to fixed point equations that involve expectations of functions with respect to possibly unknown distributions. Among many algorithms in machine learning, reinforcement learning algorithms such as TD- and Q-learning are two of its most famous applications.
This talk will provide an overview of stochastic approximation, with focus on optimizing the rate of convergence. Based on this general theory, the well known slow convergence of Q-learning is explained: the variance of the algorithm is typically infinite.
Three new Q-learning algorithms are introduced to dramatically improve performance:
(i) The Zap Q-learning algorithm that has provably optimal asymptotic variance, and resembles the Newton-Raphson method in a deterministic setting
(ii) The PolSA algorithm that is based on Polyak's momentum technique, but with a specialized matrix momentum, and
(iii) The NeSA algorithm based on Nesterov's acceleration technique
Analysis of (ii) and (iii) require entirely new analytic techniques. One approach is via coupling: conditions are established under which the parameter estimates obtained using the PolSA algorithm couple with those obtained using the Newton-Raphson based algorithm. Numerical examples confirm this behavior and the remarkable performance of these algorithms.
Biography: Adithya Devaraj is a Ph.D. student at the University of Florida where he works with Prof. Sean Meyn. The focus of his research has been variance reduction in stochastic approximation algorithms with application to reinforcement learning. He has held visiting/research positions at the Indian Institute of Science, Bangalore, Inria, Paris, and the Simons Institute for the Theory of Computing at UC Berkeley.
Host: Rahul Jain
Audiences: Everyone Is Invited
Posted By: Talyia White