Learning and adaptation in graphs, networks, and autonomous systems

Lubars, Joseph

Permalink

https://hdl.handle.net/2142/113817 Copy

Description

Title

Learning and adaptation in graphs, networks, and autonomous systems

Author(s)

Lubars, Joseph

Issue Date

2021-10-06

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

Srikant, Rayadurgam

Doctoral Committee Chair(s)

Srikant, Rayadurgam

Committee Member(s)

• Beck, Carolyn L
• Hu, Bin
• Varshney, Lav R

Department of Study

Electrical & Computer Eng

Discipline

Electrical & Computer Engr

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Date of Ingest

2022-04-29T21:34:07Z

Keyword(s)

• Reinforcement Learning
• Autonomous Driving
• Wireless Scheduling

Abstract

Reinforcement learning and adaptation are widely used in a variety of applications, whenever we want to control a dynamical system in some optimal manner. For example, in wireless networking, we can use these tools to learn to route packets more efficiently, leading to lower latency. In games, we can learn better strategies through self-play. In this dissertation, we focus on three problems of reinforcement learning and adaptation. We first consider a theoretical problem in reinforcement learning called optimistic policy iteration (OPI). We prove convergence of a variant of OPI whose convergence properties have been previously unknown. Next, we consider the problem of designing an algorithm to allow a car to autonomously merge onto a highway from an on-ramp. Two broad classes of techniques have been proposed to solve motion planning problems in autonomous driving: Model Predictive Control (MPC) and Reinforcement Learning (RL). In this dissertation, we present an algorithm which blends the model-free RL agent with the MPC solution and show that it provides better trade-offs between a number of relevant metrics: passenger comfort, efficiency, crash rate and robustness. Finally, we analyze a wireless scheduling problem with a limited probing constraint. We show how a throughput optimal solution can be computed, even without knowledge of the channel statistics. Interestingly, although this problem requires adaptability in the face of unknown conditions, we find that reinforcement learning is not needed for optimal performance and may even be harmful if applied without care.

Graduation Semester

2021-12

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/113817

Copyright and License Information

Copyright 2021 Joseph Lubars

Owning Collections