UB researchers recognized for paper on preparing networks for 6G and AI-enabled edge computing
By Peter Murphy
Published June 18, 2026
Jialei Liu, a first-year PhD student, and her advisor, Ming Shi, assistant professor in the Department of Electrical Engineering, won the Best Paper Runner-Up Award at the Institute of Electrical and Electronics Engineers (IEEE) WiOpt 2026, the International Symposium on Modeling and Optimization in Mobile, Ad hoc, and Wireless Networks.
Modeling and Optimization in Mobile, Ad hoc, and Wireless Networks.
“Our goal is to develop AI-ready network control that is both efficient and safe,” said Shi. “In practice, you want the network to adapt to changing traffic while avoiding actions that violate online resource budget decisions or destabilize performance. This paper provides a principled way to do both simultaneously.”
The paper, “Bi-Level Online Provisioning and Scheduling with Switching Costs and Cross-Level Constraints,” authored by Liu, Shi and co-authored by C. Emre Koksal, professor of electrical and computer engineering at The Ohio State University, details how Shi and other researchers developed a bi-level optimization and learning framework that matches the way many networks are operated in practice, based on different needs. The implications for delay-sensitive services are paramount — addressing this problem can significantly improve the experience using large language models, autonomous vehicles and robotics — according to Shi.
Ming Shi, assistant professor
“As networks move toward 6G and AI-enabled edge computing, they are increasingly asked to support applications that are both interactive and resource intensive. AI-powered traffic such as agentic AI systems, augmented reality and virtual reality streaming, real-time perception for robotics and autonomous systems, and conversational assistants powered by large language models can create bursty demand and tight latency requirements,” Shi said. “At the same time, frequently reconfiguring network slices or shifting edge resources is not free. It can introduce signaling overhead, hardware retuning, and transient instability that degrade performance. For delay-sensitive services, reliability is not optional. Autonomous driving and connected mobility systems cannot wait for long delays when reacting to the physical world, and AI-assisted medical systems and remote clinical workflows may depend on timely delivery of data and decisions.”
The researchers’ goal is to develop an AI-ready network control that is efficient, safe and adaptable. They designed the control with budget decisions and performance in mind. Shi said one of the biggest reasons their work matters for real deployment — and a significant reason the control is safe — is because of the budget-adaptive safe reinforcement learning approach the researchers developed.
“Reinforcement learning can discover effective scheduling policies, but naïve exploration can violate resource limits, overload queues and cause cascading failures that users experience as severe delay or service interruptions,” Shi said. “Our approach enforces the dynamic episode-wise budget constraint during learning, so the scheduler can improve performance without taking unsafe actions.”
This work sits at the intersection of future-generation networking, optimization and safe machine learning.
