Robust decision-making for AI-powered cyber-physical systems: Learning and optimization perspectives
This event is in the past.
11 a.m. to noon
Speaker
Sihong He, University of Connecticut
Abstract
With the development of sensing, communication, computation technologies, and the promise of AI to the next generation of intelligent physical systems, AI-powered cyber-physical systems (CPS) are pivotal in both research and many real-world applications. However, AI-powered CPS face unique challenges in decision-making, such as uncertainty, complex dynamics, scalability, and coordination. We will focus on learning and optimization-based methodology design for robust and efficient decision-making of AI-powered CPS. First, we present a distributionally robust optimization (DRO)-based method with data-driven uncertainty quantification, for vehicle rebalancing problems in intelligent transportation systems. It minimizes the worst-case expected rebalancing cost considering uncertainties in both passenger mobility demand and electric vehicles supply. Then, we design a robust multi-agent reinforcement learning (MARL)-based method, to provide fair and robust mobility and charging decisions through adversarial uncertainty modeling. Finally, a general robust MARL framework is proposed to address state uncertainty in multi-agent systems. We will introduce the Markov game with state perturbation adversaries (MG-SPA), including theoretical analysis, algorithm designs with performance guarantees, and experimental results. Future work on uncertainty-aware robust decision-making, AI methodologies, and AI-powered CPS applications will also be briefly introduced.
Bio
Sihong He is a Ph.D. candidate in Computer Science and Engineering at the University of Connecticut. Sihong's research contributions and interests include Multi-Agent Reinforcement Learning (MARL), Robust Reinforcement Learning, Uncertainty Quantification, Distributionally Robust Optimization, and Cyber-Physical Systems (CPS). Her research goal is to lay the foundations for AI-powered CPS, including ensuring efficiency, robustness, safety, and security for CPS through learning and optimization methods. Her work has provided practical robust and efficient decision-making strategies for a variety of CPS including Intelligent Transportation, Connected Autonomous Vehicles, and Smart Cities. She published her work in several top conferences and journals including TMLR, IEEE TITS, IROS, ICRA, TCPS, and TMC. She has also been honored with Rising Stars in AI (KAUST AI Initiative), the NC State Building Future Faculty Program Fellow, the Synchrony Fellowship, the Predoctoral Prize for Research Excellence, the GE Advanced Manufacturing Fellowship, and the Cigna Graduate Scholarship. More information is available on her website:www.sihonghe.com.