Скачать или смотреть Shayegan Omidshafiei: Multiagent Behavioral Analysis

Abstract: Each year, expert-level performance is attained in increasingly-complex multiagent domains. This rapid progression is accompanied by a commensurate need to better understand how such agents attain this performance, to enable their safe deployment, identify limitations, and reveal potential means of improving them. This talk takes a step back from performance-focused multiagent learning, and instead focuses on behavioral analysis of multiagent systems. I will provide an overview of a model-agnostic method for discovery of behavior clusters in multiagent domains, using variational inference to learn a hierarchy of behaviors at the joint and local agent levels. I'll then demonstrate applications of our method to enable coupled understanding of behaviors at the joint and local agent level, detection of behavior changepoints throughout training, and disentanglement of previously-trained policies in multiagent MuJoCo and OpenAI's hide-and-seek domain.

Bio: Shayegan Omidshafiei is a research scientist in the People + AI Research (PAIR) team in Google Research, where he works on interpretable and human-in-the-loop reinforcement learning (RL). Before joining PAIR, Shayegan was a research scientist in DeepMind, where he worked in the Game Theory team and co-led their Sports Analytics effort. Shayegan's research interests include multiagent RL, game theory, interpretable RL, robotics, and control systems. He previously received his Ph.D. at the Laboratory for Information and Decision Systems (LIDS) and Aerospace Controls Laboratory (ACL) at MIT, specializing in Autonomous Systems. He also received his S.M. degree in Aeronautics and Astronautics from MIT in 2015, and his B.A.Sc. degree in Engineering Science from the University of Toronto in 2012. He is co-inventor of five patents filed with the United States Patent Office and co-author of the book "AI for Sports".