Structuring Cooperative Teams for Multi-Agent Reinforcement Learning

Thursday, August 22, 2024 10 a.m. to 11 a.m.

Speaker: Dr. Qi Zhang

From: University of South Carolina

Abstract

Cooperative artificial intelligence equips a team of autonomous agents with the capability of planning and learning to maximize their joint utility, which finds a wide range of applications. Current solutions to cooperative AI, instantiated as cooperative multi-agent reinforcement learning frameworks, are still more computation-, data-, and communication-hungry than what practice can often afford. This talk presents a series of theories and methods for cooperative multi-agent teams to achieve efficiency in computation, data, and communication by discovering, engineering, and exploiting structures inherent in the problem. This talk presents a series of algorithmic innovations, focusing on the team structures exhibited in the form of 1) grounded communication where cooperative agents communicate their actions through context-aware Bayesian networks, 2) agent-wise homogeneity where policy parameter sharing is theoretically justified for the first time, and 3) Euclidean symmetries such as 3D rotations and translations. Each of the algorithmic innovations enjoys provable guarantees (e.g., first-order convergence to near-optimal joint policies) and achieves state-of-the-art empirical performance on simulation-based benchmarks. Then, the talk presents a future research vision that extends the theories and methods to electrical and computer engineering applications and with human teammates. 

For more info, please follow this link.

Read More

Locations:

HEC: 101A: 101A

Contact:


Calendar:

College of Engineering and Computer Science

Category:

Academic

Tags:

UCFCRCV