Adaptive learning from demonstration in heterogeneous agents: Concurrent minimization and maximization of surprise in sparse reward environments

Clark, Emma

Permalink

https://hdl.handle.net/2142/122173 Copy

Description

Title

Adaptive learning from demonstration in heterogeneous agents: Concurrent minimization and maximization of surprise in sparse reward environments

Author(s)

Clark, Emma

Issue Date

2023-12-05

Director of Research (if dissertation) or Advisor (if thesis)

Mehr, Negar

Department of Study

Aerospace Engineering

Discipline

Aerospace Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• reinforcement learning
• learning from demonstration, curriculum learning

Abstract

Learning from Demonstration (LfD) is a reinforcement learning method where an agent learns a policy by imitation demonstrations from an expert. This expert can be another agent, already trained to have an optimal policy or a predefined control system. Or, the expert can be a human. LfD is useful for learning very complex tasks or in settings with strict behavior guidelines or restrictions. One of the major limitations of LfD is an inability to learn when there are differences in dynamics between the student and teacher agents. This limits LfD methods to homogenous agents; however, real-world scenarios may often have differences in dynamics or environmental constraints between the student and teacher. Such as, a robot learning from human demonstration, or two different models of robot with variations in maximum joint angles or actuator power. Even analogous systems may have small variations in robot capabilities, due to noise or under-performance from technological limitations. To address this challenge, we propose a Student-Teacher framework, where the Teacher agent uses the Student’s surprise with, respect to demonstration trajectories, to infer differences in dynamics between itself and the Student. The teacher is then able to adapt its demonstration trajectories to consider the dynamics or constraints of the Student. In contrast to most common LfD methods, we assume the Teacher is not already an expert, but instead is learning in parallel to the Student.

Graduation Semester

2023-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/122173

Copyright and License Information

Copyright 2023 Emma Clark

Owning Collections