Efficient reinforcement learning through variance reduction and trajectory synthesis

Reinforcement learning is a general and unified framework that has been proven promising for many important AI applications, such as robotics, self-driving vehicles. However, current reinforcement learning algorithms suffer from large variance and sampling inefficiency, which leads to slow convergent rate as well as unstable performance. In this thesis, we manage to alleviate these two relevant problems. For enormous variance, we combine variance reduced optimization with deep Q-learning. For inefficient sampling, we propose novel framework that integrates self-imitation learning and artificial synthesis procedure. Our approaches, which are flexible and could be extended to many tasks, prove their effectiveness through experiments on Atari and MuJoCo environment.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2019

Author dc:creator

• Zhao, Xiaoming

Contributors dc:contributor

• Peng, Jian

• Machine Learning, Reinforcement Learning, Optimization

Statement dc:rights

• Copyright 2019 Xiaoming Zhao

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/105067

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/105067