Learning & Optimization
for Autonomous Intelligent
Networked Systems.面向自主智能网络系统的
学习与优化
I develop mathematical foundations and algorithms with provable guarantees for practical intelligent networked systems that learn safely, adapt efficiently, and coordinate under uncertainty.
Assistant Professor
Department of Electrical and Computer Engineering
University at Buffalo,
The State University of New York
Modeling · Learning Theory · Optimization · Intelligent Networked Systems
Safe cooperative learning with coupled instantaneous hard constraints.
Read PDF → AcceptedAge of Intelligence (AoInt): decision-aware model freshness and cost-aware updates.
Read PDF → SpotlightNear-optimal adversarial reinforcement learning with switching costs.
Read PDF → Best Paper Award Runner-UpJoint online provisioning and queue-aware scheduling with switching costs and cross-level constraints.
Read PDF →In the news · UB Engineering: preparing networks for 6G →
Modeling, learning, and optimization for decisions with provable performance and safety guarantees.
Three connected questions shape my research on autonomous intelligent networked systems.
Learning under hard constraints on each action, including when agents interact.
Near-optimal safe reinforcement learning, non-convex safety certification, and coupled-agent learning. Our NeurIPS work recombines local safety certificates and uses optimistic Bellman planning within certified safe policies. Its safety and regret guarantees depend on the paper’s stated assumptions.
Online decisions that weigh performance against switching, reconfiguration, and resource costs.
Competitive online optimization, switching-aware learning, and resource provisioning. In the studied adversarial bandit model with switching costs, allowing two arms to be selected per round improves the achievable dependence of regret on the time horizon T from Θ(T2/3) to O(√T), without full-information feedback.
Understanding the value of partial observations, preference feedback, and fresh models.
Regret bounds and information limits for partially observable learning, alongside decision-aware synchronization that balances model staleness against refresh costs.
From mathematical foundations to dependable autonomy and digital infrastructure.
My long-term goal is to unite safety, information, and adaptation in one mathematical framework for networked intelligence.
The linked papers establish mathematical guarantees and evaluate algorithms numerically, providing foundations for these application goals. Current questions include safe coordination with limited communication, selective observation, and decision-aware model updates across cloud platforms, edge servers, and end devices.
In the NeurIPS model, a known safe fallback supports safety during exploration, but sublinear regret also requires the stated safety-margin, information-coverage, and planning assumptions. Communication restrictions can limit the best achievable safe value.
Research →I mentor student-led research and teach Convex Optimization (EE 441/541) and Principles of Networking (EE 434/534) at UB, connecting mathematical foundations with engineering judgment.
Through National Grid K–12 teacher engagement and high-school research mentoring, I connect mathematical modeling, AI, and engineering with learning beyond the university.
My service includes co-chairing the SEAS AI Cluster, serving on the ECE Tenure and Promotion Guideline Committee, conference leadership, and peer review across learning, optimization, and networking.
Outside research, I enjoy photography and spending time with my cat, Maomi. A few quiet moments, through my lens.
Select a photograph for a larger view.
Original photographs credited to Ming Shi are © Ming Shi. All rights reserved. For reuse permissions, please contact me. Copyright & Image Reuse →