Multiagent Reinforcement Learning: Rollout and Policy Iteration

Explore multiagent reinforcement learning through a 37-minute lecture by Dimitri Bertsekas from ASU & MIT, focusing on rollout and policy iteration techniques. Delve into finite-state infinite horizon problems, the Policy Iteration (PI) algorithm, and the underlying theory of trading off control and state complexity. Compare standard and multiagent approaches to rollout and policy iteration, and examine approximate policy iteration with agent-by-agent policy improvement. Gain insights into this well-researched field dating back to the 1960s, presented as part of the Simons Institute's series on reinforcement learning from batch data and simulation.

Sources
Multiagent Problems - A Very Old (19608) and Well-Researched Field
For this Talk we Focus on Finite-State Intinite Horizon Problems
Policy Iteration (PI) Algorithm
Outline of Our Approach for Multiagent Problems
Underlying Theory: Trading off Control and State Complexity (NDP book, 1996)
Comparing Standard with Multiagent Rollout/Policy Iteration
Approximate Policy Iteration with Agent-by-Agent Policy Improvement
Concluding Remarks