Overview
Explore the challenges and lessons learned in partner modeling for decentralized multi-agent coordination in this 55-minute lecture by Dorsa Sadigh from Stanford University. Delve into the role of representation learning in developing effective conventions and latent partner strategies, and discover how to leverage these conventions within reinforcement learning loops for coordination, collaboration, and influencing. Examine strategies for stabilizing latent partner representations to reduce non-stationarity and achieve more desirable learning outcomes. Investigate the formalization of decentralized multi-agent coordination as a collaborative multi-armed bandit with partial observability, and learn how partner modeling strategies can achieve logarithmic regret. Gain insights into autonomous driving, coordination tasks, and collaborative robotics while exploring topics such as intrinsic motivation and the superiority of human intelligence in complex scenarios.
Syllabus
Introduction
Autonomous Driving
Changing Lanes
Coordination Tasks
Understanding the Action Space
Nonstationary Agents
Nonstationarity
Reducing Nonstationarity
Representation Learning
Merrill
Stable
Lilly
Summary
Modular Architecture
Collaborative Multiarm Bandits
Robotics Example
Thank you
Humans are smarter than robots
Collaborative tasks
Intrinsic motivation
Taught by
Simons Institute