ECE Seminar by Yasin Yazıcıoğlu: Graph Theoretic Methods for Autonomy of Multi-Agent Systems

Armour College of Engineering’s Department of Electrical and Computer Engineering will welcome Yasin Yazıcıoğlu, research assistant professor in the Department of Electrical and Computer Engineering at the University of Minnesota, to present a lecture, “Graph Theoretic Methods for Autonomy of Multi-Agent Systems.”

Abstract: Multi-agent systems appear in many applications of robotics and cyber-physical systems such as warehouse logistics, precision agriculture, environmental monitoring, and smart infrastructure/city. Autonomous operation of these complex systems underperformance guarantees requires scalable and provably correct control, learning, and optimization algorithms/architectures. In this talk, I will present some graph theoretic methods for distributed control and reinforcement learning. In the first part of the talk, I will focus on multi-agent networks, where each agent (node) has a state that evolves under some rules based on the states of its neighbors on the underlying graph (e.g., rendezvous or flocking by multiple robots). I will show that the underlying graph structure of such networks has a significant impact on how much the overall system is influenced by external disturbances and control inputs. More specifically, I will present fundamental relationships (expressed as tight bounds) between some simple graph properties (node degrees and the distances between the nodes) and complex system properties (controllability and robustness). The proposed bounds can be computed efficiently and used for the analysis and design of multi-agent networks. In the second part, I will talk about a constrained reinforcement learning (RL) problem where the goal is to learn an optimal control policy in a Markov Decision Process while ensuring a probabilistic satisfaction of a complex spatio-temporal constraint on the resulting trajectories. In such constrained scenarios, standard RL algorithms can eventually learn a feasible policy by penalizing constraint violations. Accordingly, these methods do not yield any guarantees on the constraint satisfaction in the early episodes of learning. Here, I will present a novel graph-theoretic approach, which can be integrated into standard RL algorithms to ensure constraint satisfaction with a desired probability throughout learning (even in the first episode). I will conclude the talk with some ongoing work and future directions.

Biography: Yasin Yazıcıoğlu is a research assistant professor in the Department of Electrical and Computer Engineering at the University of Minnesota. Prior to joining the University of Minnesota, he was a postdoctoral researcher in the Laboratory for Information and Decision Systems (LIDS) at MIT. He received a Ph.D. degree in Electrical and Computer Engineering from the Georgia Institute of Technology, and B.S. and M.S. degrees in Mechatronics Engineering from Sabancı University, Turkey. His research is primarily focused on distributed control, learning, and optimization with applications to robotics, cyber-physical systems, and networks.