Chengshuai Shi has earned a Bloomberg Data Science Ph.D. Fellowship to support his research in wireless communication combined with machine learning.
Shi is pursuing his Ph.D. in electrical engineering at the University of Virginia, advised by Cong Shen, assistant professor of electrical and computer engineering.
“The Fellowship will help me explore the landscape and new possibilities in wireless communications research. I am looking forward to meeting and collaborating with leaders in the data science field who can help me expand my tool kit and systematically apply their techniques to wireless communications,” Shi said.
Shi met Shen as an undergraduate at the University of Science and Technology of China, where Shen was a professor of electrical engineering.
“Professor Shen has worked in wireless communications for many years; he has a very clear motivation to connect wireless communications with modern techniques like machine learning, with a focus on applications,” Shi said.
As a third-year undergraduate, Shi had expressed his interest in doing a research project with Shen. When Shen joined the University of Virginia in the Fall of 2019, just as Shi was applying to graduate schools, he recruited Shi to join his Laboratory for Intelligent Communications and Networking as his first student.
Shi is working on a complex problem in wireless communications, how to schedule, re-allocate or re-use scarce spectrum resources between many users while maintaining high network performance.
“When a device transmits a signal, it occupies some part of the spectrum. Maybe someone controls that resource, but they haven’t used it for a long time, or they use it only a few hours a day. We are developing machine learning tools that can re-allocate or re-use that resource for better efficiencies,” Shi said.
Shi specializes in a machine learning tool called multi-armed bandits, proven to help devices find open spectrum. The arms correspond to channels and players represent the distributed devices trying to communicate over the channels. This tool enables wireless devices to learn from the success or failure of past communications to estimate the value of a particular channel at a particular time.
“The key challenge is to reach a globally optimal solution without coordination among the players in an unpredictable, real-world environment,” Shen said.
The Bloomberg Data Science Ph.D. fellowship recognizes Shi’s development of machine learning applications and his publications in academic journals and conference proceedings. Shi first-authored On No-Sensing Adversarial Multi-player Multi-armed Bandits with Collision Communications, published in IEEE Journal on Selected Areas in Information Theory with Shen as co-author.
Their paper fills a gap in state-of-the-art no-sensing adversarial multi-player multi-armed bandits research. Shi and Shen introduce a new dimension of hardness called attackability to categorize all possible adversaries from either a local or global view and introduce a family of algorithms with forced-collision communication among players that can handle known or unknown attackabilities. Their proposed algorithms can operate with a much better robustness when an adversarial attacks the multi-armed bandits game, which often arises in spectrum sharing where the adversaries are intentional or unintentional interferers.
Shi also presented Federated Multi-armed Bandits with Personalization at the 24th International Conference on Artificial Intelligence and Statistics, AISTATS 2021, one of the top 50 of 1,500 contributors invited to give an oral presentation. Shi co-authored this research paper with Shen and Jing Yang, associate professor of electrical engineering and computer science at Pennsylvania State University.
Federated learning trains a machine learning algorithm across multiple decentralized devices without uploading any data to the server, deploying machine learning models to devices at the edge of the network over a certain period of time. This technique helps the algorithm become more savvy while satisfying privacy and security constraints. Shi described the first successful attempt to connect multi-armed bandits to federated learning, with a fundamental twist between generalization and personalization. This new framework allows the spectrum sharing system to select a channel that is not only good for the overall system but also tailors to any particular device.