Machine Learning Ph.D. Student

Home School: Aerospace Engineering

Georgia Institute of Technology

Committee

1 Dr. John-Paul Clarke (Advisor, School of Industrial and Systems Engineering, School of Aerospace Engineering, Georgia Institute of Technology)

2 Dr. Siva Theja Maguluri (Co-advisor, School of Industrial and Systems Engineering, Georgia Institute of Technology)

3 Dr. Justin Romberg (School of Electrical and Computer Engineering, Georgia Institute of Technology)

4 Dr. Benjamin Van Roy, Department of Electrical Engineering, Department of Management Science & Engineering, Stanford University) (external)

Abstract

Reinforcement Learning (RL) captures an important facet of machine learning going beyond prediction and regression: sequential decision making, and has had a great impact on various problems of practical interest. The goal of this proposed thesis is to provide theoretical performance guarantees of RL algorithms. Specifically, we develop a universal approach for establishing finite-sample convergence bounds of RL algorithms when using tabular representation and when using function approximation. To achieve that, we consider general stochastic approximation algorithms and study their convergence bounds using a novel Lyapunov approach. The results enable us to gain insight into the behavior of RL algorithms.