Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.
Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.
Author ORCID Identifier
N/A
AccessType
Open Access Dissertation
Document Type
dissertation
Degree Name
Doctor of Philosophy (PhD)
Degree Program
Computer Science
Year Degree Awarded
2016
Month Degree Awarded
February
First Advisor
Sridhar Mahadevan
Second Advisor
Andrew G. Barto
Third Advisor
Shlomo Zilberstein
Fourth Advisor
Weibo Gong
Subject Categories
Artificial Intelligence and Robotics
Abstract
This thesis presents a general framework for first-order temporal difference learning algorithms with an in-depth theoretical analysis. The main contribution of the thesis is the development and design of a family of first-order regularized temporal-difference (TD) algorithms using stochastic approximation and stochastic optimization. To scale up TD algorithms to large-scale problems, we use first-order optimization to explore regularized TD methods using linear value function approximation. Previous regularized TD methods often use matrix inversion, which requires cubic time and quadratic memory complexity. We propose two algorithms, sparse-Q and RO-TD, for on-policy and off-policy learning, respectively. These two algorithms exhibit linear computational complexity per-step, and their asymptotic convergence guarantee and error bound analysis are given using stochastic optimization and stochastic approximation. The second major contribution of the thesis is the establishment of a unified general framework for stochastic-gradient-based temporal-difference learning algorithms that use proximal gradient methods. The primal-dual saddle-point formulation is introduced, and state-of-the-art stochastic gradient solvers, such as mirror descent and extragradient are used to design several novel RL algorithms. Theoretical analysis is given, including regularization, acceleration analysis and finite-sample analysis, along with detailed empirical experiments to demonstrate the effectiveness of the proposed algorithms.
DOI
https://doi.org/10.7275/7935541.0
Recommended Citation
Liu, Bo, "Algorithms for First-order Sparse Reinforcement Learning" (2016). Doctoral Dissertations. 588.
https://doi.org/10.7275/7935541.0
https://scholarworks.umass.edu/dissertations_2/588