Events Calendar

CS Colloquium: Rakshit Trivedi (MIT) - Foundations for Learning in Multi-agent Ecosystems: Modeling, Imitation, and Equilibria

Speaker: Rakshit Trivedi, MIT

Talk Title: Foundations for Learning in Multi-agent Ecosystems: Modeling, Imitation, and Equilibria

Series: CS Colloquium

Abstract: The growing presence of AI in critical domains such as information communication, service, financial markets and agriculture requires designing AI systems capable of seamlessly interacting with other AI, with humans and as part of complex systems in a manner that is beneficial to humans. For an AI to be effective in such settings, a key open challenge is for it to have the ability to effectively collaborate across a broad group of interdependent agents (AI or human) in a variety of one or few-shot interactions. A crucial step towards addressing this is to enable rapid development and safe evaluation of AI agents and frameworks that can incorporate the richness and diversity observed in human behaviors and account for various social and economic factors that drives interactions in the multi-agent ecosystems. In this talk, I will set forth the research agenda of real-world in silico design for such AI systems and discuss methodological advancements in this direction. First, I will focus on automated design of central mechanisms tasked to shape the behavior of self-interested agents and drive them towards improving social welfare. I will introduce a novel multi-agent reinforcement learning technique to solve the resulting bi-level optimization problem and present its effectiveness in a simulated market economy. Next, I will discuss the setting where the self-interested agents interact with each other in a strategic manner to form networks and present our approach on discovering the underlying mechanisms that drives these interactions. This approach considers a game-theoretic formalism, and leverages recent advances in inverse reinforcement learning, thereby serving as a preliminary step towards learning models of optimizing mechanisms directly from observed data. Finally, I will focus on the use of AI agents as surrogate for human actors that can provide simulations of real-world complexity and discuss challenges and opportunities on designing AI that is capable of handling the diversity, richness, and noise that is inherent to human behaviors. I will conclude my talk with an outline of my forward-looking vision on this agenda.

This lecture satisfies requirements for CSCI 591: Research Colloquium

Biography: Rakshit Trivedi is a Postdoctoral Associate in the Computational Science and Artificial Intelligence Laboratory (CSAIL) at MIT and a Researcher in EconCS at Harvard School of Engineering and Applied Sciences (SEAS). His research focuses on the development of AI that is capable of learning from human experiences, quickly adapt to evolving human needs and achieve alignment with human values. He is further interested in studying the effectiveness of such an AI in the presence of various socio-economic mechanisms. Towards this goal, he is currently leading a set of efforts on developing and evaluating design strategies for building helpful and prosocial artificial agents in mixed-motive settings, in collaboration with Deepmind and Cooperative AI Foundation. Previously, Rakshit completed his Ph.D. at Georgia Institute of Technology, where he focused on learning in networked and multi-agent systems to improve predictive and generative capabilities of downstream applications, by accounting for the structure and dynamics of interactions in such systems.

Host: Bistra Dilkina

Location: Olin Hall of Engineering (OHE) - 132

Audiences: Everyone Is Invited

Contact: Assistant to CS chair