Past Event: Center for Autonomy Seminar

Safe Learning and Alignment in Multiagent Systems

Soham Das, Ph.D. Candidate, Department of Industrial and Systems Engineering, Texas A&M University

Abstract

The rapid advancement of artificial intelligence has ushered in a new era of decentralized decision making. From autonomous driving and collaborative robotics to online marketplaces, intelligent agents are increasingly driving efficiency and innovation.  While the opportunities are endless, significant challenges remain. Self-interested agents may overly prioritize individual outcomes at the expense of the larger system. Ensuring that multiagent learning aligns with broader system-level goals requires a principled approach to intervention and safety.

This talk presents two complementary approaches to ensuring safety and alignment in learning in games. The first part will focus on alignment through interventions, specifically studying how an external system designer can steer best-responding agents in an anti-coordination network game toward socially optimal outcomes.  The second part of the talk will center itself on safe multiagent reinforcement learning, using the constrained Markov game (CMG) formalization. We provide a decentralized algorithm that has provable guarantees for learning constrained Nash Equilibria (NE) in CMGs. Additionally, we introduce the concept of a Lagrangian game, and demonstrate that solving a trajectory of Lagrangian games yields a constrained nonstationary NE. By integrating external interventions with intrinsic safety mechanisms, we advance a framework that ensures alignment between individual incentives and collective well-being.

Biography

Soham Das is a Ph.D. Candidate in the Networked Multiagent Systems Lab within the Department of Industrial and Systems Engineering at Texas A&M University. He holds a Bachelor’s degree in Mechanical Engineering from the National Institute of Technology, Durgapur, India. His research focuses on safety and alignment in complex multiagent interactions, integrating theoretical insights with practical applications in dynamic, interconnected systems. His work spans intervention design in network games, safe learning in Markov games, dynamical processes on graphs, with applications in areas such as social network analysis, epidemic modeling, additive manufacturing and safe multiagent reinforcement learning. In solving these problems, Soham draws on methodologies from game theory, optimization and applied probability. He is a student member of IEEE, INFORMS and SIAM.