ONLINE LEARNING FOR THE OPTIMIZATION OF WIRELESS COMMUNICATIONS
Open Access
- Author:
- Gan, Chao
- Graduate Program:
- Electrical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- October 09, 2020
- Committee Members:
- Jing Yang, Dissertation Advisor/Co-Advisor
Jing Yang, Committee Chair/Co-Chair
Viveck Ramesh Cadambe, Committee Member
Minghui Zhu, Committee Member
Uday V. Shanbhag, Outside Member
Kultegin Aydin, Program Head/Chair - Keywords:
- Online Learning
Multi-armed Bandits
Resource Allocation - Abstract:
- This dissertation focuses on online learning based approaches to the optimization of wireless communication systems operating in dynamic environment. In contrast to traditional network management policies that exploit perfectly known operating environment to deliver optimal performances, online learning based network management requires the system to explore the unknown environment and exploit learned knowledge to guide the optimal decision-making. In order to balance the the tension between exploration and exploitation in this process, this dissertation leverages multi-armed bandits, a classical tool in sequential decision-making and control, to obtain principled solutions with provable performance guarantees. The dissertation is organized as follows: Chapter 1 provides an introduction to the background and the problems studied. Chapter 2 investigates the dynamic spectrum access problem with unknown channel statistics. Motivated by the special observation model in dynamic spectrum access, Chapter 3 proposes a novel cost-aware cascading bandits model, which may have wide applications in cognitive radio, clinical trials, etc. Chapter 4 takes the diverse user requirements into consideration, and investigates the impact of heterogeneous user requirements on the corresponding learning performance. Chapter 5 studies optimal spectrum access where rateless codes are adopted for transmissions. Chapter 6 concludes the dissertation. This dissertation presents several adaptive and intelligent online learning based solutions to cope with random and non-stationary environment facing modern communication systems. Thanks to the unique characteristics of the considered engineering application, the resulting online learning algorithms also hold their own intellectual merit in machine learning. iii