Ming Shi Learning · Optimization · Reliable Intelligence
G
Learning theory · Optimization · Networked systems

Ming ShiReliable intelligence.
Real-world constraints.

I develop mathematical foundations and algorithms for intelligent systems that must learn safely, adapt to change, and make decisions with limited information and resources.

Assistant Professor of Electrical and Computer Engineering
University at Buffalo, The State University of New York

Explore my contributions Research program
Google Scholar → UB faculty profile → Email me
Selected recognition ICLR 2023 SpotlightAdversarial learning with switching costs → WiOpt 2026 Best Paper Award Runner-UpStudent-led provisioning and scheduling → NeurIPS 2026 · AcceptedSafe cooperative multi-agent learning →
Scientific contributions

Four advances toward
reliable learning and decisions.

My research connects provable guarantees to the constraints that shape how a system can actually operate.

Complete publication record →

Learning safely at every step

Many learning algorithms limit average risk. My work studies instantaneous hard constraints, where an unsafe action or transition cannot be offset by safer behavior later.

We develop near-optimal safe reinforcement learning and extend certification to non-convex decision spaces and coupled agents, with safety guarantees that hold with high probability under the models’ stated assumptions.

Key resultsSafe exploration with mathematical guarantees

ICML 2023 & 2025 · NeurIPS 2026 accepted

Near-optimal safe RL · ICML 2023 →Non-convex safety · ICML 2025 →Coupled-agent safety · NeurIPS 2026 →

Adapting when change has a cost

Changing a model, route, or resource allocation consumes time and resources. I develop online optimization and learning methods that account for switching costs, ramp limits, and uncertain forecasts.

In our adversarial bandit model with switching costs, choosing two arms at a time changes regret scaling from Θ(T2/3) to O(√T), without full feedback—a precise gain from a small increase in choice.

Key resultsThe value of prediction, feedback, and flexibility

IEEE/ACM ToN · INFOCOM · ICLR · WiOpt

Power-of-2-Arms · ToN 2025 →Regularization with lookahead · ToN 2024 →

Understanding what information makes learning possible

Intelligent systems rarely observe the full state of the world. My work asks when partial online state information changes what a learning algorithm can achieve.

We establish worst-case limits and identify structured classes of partially observable problems that admit efficient learning, clarifying when additional observations are worth acquiring.

Key resultRegret bounds and limits under partial observability

IEEE Transactions on Information Theory, 2026

Partial online state information · IEEE TIT →Information and preference learning →

Keeping edge intelligence useful, not just up to date

A newer model is not always worth downloading. Our Age of Intelligence framework measures how model staleness affects a device’s decisions, then balances that loss against refresh costs.

Across 20 validation traces in the paper’s numerical study, parameter-aware synchronization reduced mean joint cost by 6.15% relative to the Instant-Param baseline.

Key resultDecision-aware model synchronization

AIoT 2026 · Accepted · Hui Wan, first author

Fresh Enough to Decide · AIoT 2026 →Distributed intelligence agenda →
Research vision

Intelligence that stays reliable
as the world changes.

My long-term goal is a theory of networked intelligence that treats safety, information, and adaptation as one decision problem. The central question: how can systems learn and coordinate when observations are incomplete, actions interact, and every update has a cost?

Explore the full research agenda →

Certify before acting

Expand what a system can safely do while learning unknown dynamics and coordinating with other agents.

Acquire information with purpose

Choose which states to observe, experiences to share, and preferences to query according to their decision value.

Adapt across the network

Co-design learning, communication, and model updates across cloud, edge, and devices under drift and resource limits.

National and societal relevance

Better decisions for systems
people depend on.

Reliable infrastructure, safe autonomy, and efficient AI are the broader goals motivating this work. The contributions above provide theoretical foundations and numerical evidence toward these goals.

Safety & autonomy

Learn without treating safety as an afterthought.

Hard-constraint learning addresses a central challenge for autonomous and multi-agent systems: improving decisions while respecting operational limits throughout learning.

Safety research →
Digital infrastructure

Make adaptation feasible for real networks.

Switching-aware optimization models the costs of changing routes, provisioning services, and reallocating resources in data centers and communication networks.

Network optimization research →
Resource-efficient AI

Spend bandwidth and compute where they matter.

Decision-aware synchronization and selective information sharing aim to keep edge AI useful under limited communication, computation, and changing conditions.

Edge intelligence research →
Education & community impact

Build a research community.
Open a path into engineering.

I mentor students to connect proofs with systems and communicate their work beyond a specialist audience. Our educational activities bring these habits into undergraduate research, K–12 teacher engagement, and robotics.

Teaching, mentoring & outreach →
Student-led research · 2026

Early research ownership at UB

Jialei Liu: first-author WiOpt Best Paper Award Runner-Up. Hui Wan: first-author AIoT acceptance.

Teacher & student engagement · 2026

Make mathematical ideas tangible

Student-led demonstrations for K–12 teachers connected modeling with smart grids and autonomous systems; EAS 202 Faculty Mentor Week introduced students to research pathways.

Ohio State · 2023–2024

Undergraduate research & youth robotics

Supported NSF AI-EDGE REU experiences in AI and networking, and engaged young learners through LEGO FIRST League in 2023.

Selected publications

Start with these papers.

Representative work in safe learning, information limits, and resource-aware decisions.

Search all publications →
Recent updates

From the group.

Professional service →
Work with me

Questions worth proving.
Systems worth improving.

I welcome prospective students and collaborators interested in learning theory, optimization, and reliable intelligent systems.

mshi24@buffalo.edu Prospective students →Leadership & service →