Explore the intricacies of decentralized policies in multiagent systems through this comprehensive lecture by Na Li from Harvard University. Delve into the Scalable Actor Critic (SAC) framework, which leverages network structures to find local, decentralized policies approximating global objectives. Examine the performance of stationary points in scenarios where states are shared among agents but actions follow decentralized policies. Investigate the use of stochastic game frameworks to characterize policy gradient performance in multiagent Markov Decision Process systems. Learn about opportunities and challenges in decision-making, data-driven approaches, and control of networked Markov Decision Processes. Discover numerical results in multi-access wireless communication and explore various multiagent learning settings, including decentralized control, optimality guarantees, and optimization landscapes. Gain insights into gradient play for identical interest cases and general stochastic games, concluding with a discussion on convergence and a comprehensive summary of the presented concepts.
Learning Decentralized Policies in Multiagent Systems - How to Learn Efficiently
Overview
Syllabus
Intro
Opportunities and Challenges Decision-making
Learning (Data-driven decision-making) is a promis
Control of Networked Markov Decision Process
Examples of Systems with the local interact
Scalable RL for Network Systems
Review: Policy Gradient in the Full Information C
RL in the Network Setting
The Exponential Decay Property
Truncation of Q-function
Numerical results: Multi-Access Wireless Communic
Other (Multiagent) Learning Settings Decentralized Control
Optimality Guarantee
Optimization Landscape
Gradient play for identical interest case
General Stochastic Games
Convergence of gradient play?
Summary
Taught by
Simons Institute
