Seminar: Safe Multi-Agent Learning in Dynamic and Competitive Environments

SEMINAR 

DEPARTMENT OF INDUSTRIAL ENGINEERING 

Safe Multi-Agent Learning in Dynamic and Competitive Environments

Ceyhun Eksin,

Texas A&M University

Abstract:

Multi-agent systems consist of multiple autonomous decision-makers that seek to achieve their objectives in the absence of a central coordinator, often while operating under safety, resource, or operational constraints. Examples include teams of robots, smart meters in electricity grids, and networked sensors and actuators in manufacturing systems. In such settings, agents typically lack precise information about the environment and about other agents’ objectives, making strong coordination assumptions unrealistic. As a result, agents must rely on learning mechanisms that adapt strategies over time while respecting safety constraints. In this talk, we adopt constrained Markov games as a modeling framework for safe multi-agent decision-making and study decentralized learning algorithms with performance and feasibility guarantees. In the first part, we introduce a best-response learning algorithm for constrained Markov games, where agents’ actions and state dynamics are coupled through shared constraints. Each agent computes approximate best responses using a probably approximately correct solver for constrained Markov decision processes, treating other agents’ policies as fixed. We establish finite-time convergence to approximate Nash equilibria with high probability in games that exhibit limited violations of the potential game property. In the second part, we present an augmented Lagrangian game approach for solving constrained Markov games. The method constructs and solves a sequence of relaxed Markov games over a finite horizon. We show that the resulting sequence of policies constitutes an approximate Nash equilibrium of the original constrained game. Together, these results advance the design of decentralized, safety aware learning algorithms that are applicable to complex real-world multi-agent systems.

Keywords: Multi-agent systems, game theory, Markov decision processes

Short Bio: 

Texas A&M University. He also holds a courtesy appointment with the Electrical and Computer Engineering Department. He received his Ph.D. in Electrical and Systems Engineering from the University of Pennsylvania in 2015 and was later a Postdoctoral Fellow at the Georgia Institute of Technology, affiliated with both the School of Electrical & Computer Engineering and the School of Biological Sciences. He also holds a M.S. degree in Industrial Engineering from Boğaziçi University, Istanbul, Turkey in 2008. His B.S. degree is in Control Engineering from Istanbul Technical University, Istanbul, Turkey, which he received in 2005. He is a recipient of the NSF CAREER award in 2023. His research interests are in the areas of distributed optimization, network science, game theory and control theory. His current research focuses on game theoretic modeling and learning in multi-agent systems with applications in autonomy, epidemics and energy systems.

All interested are cordially invited.  

DATE:  April 17, 2026

TIME:  Friday, 15:00

PLATFORM: https://us06web.zoom.us/j/83416706686?pwd=AyRllMHnBPKdmSq8IYp7P40kuRM35c...

Meeting ID: 834 1670 6686