Model-free learning with imitation

Optimizing sample efficiency, or the experience needed in an environment to gain satisfactory performance, is a core challenge for developing reinforcement learning agents. While imitation learning resolves this issue, it is constrained by expert performance. On the other hand, model-based strategies, which learn a world model of the environment, typically fail to approach the asymptotic performance of model-free approaches. In this thesis, we focus on combining imitation learning with model-free reinforcement learning to maximize sample efficiency and achieve higher asymptotic performance. We propose an intuitive approach to leveraging the strengths of each paradigm to produce higher rewards over a fixed number of frames when observing learned experts. We further investigate our method’s applicability to knowledge distillation for reduced-complexity agents. These studies and results lay the foundation for further study which will benefit model-free reinforcement learning as a whole.

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

2023

Author dc:creator

• Walia, Nikash

Contributors dc:contributor

• Lazebnik, Svetlana

• Reinforcement Learning
• Knowledge Distillation

Statement dc:rights

• Copyright 2023 Nikash Walia

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/120276

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F120276