MARL: Population Training with Exact PSRO

Last meeting we revisited the population training algorithm introduced in Chapter 9 of Multi-Agent Reinforcement Learning: Foundations and Modern Approaches and considered only tabular problems with exact solution techniques. That version of the algorithm has some limitations regarding the ability to learn stochastic equilibrium solutions which is often required in stochastic games with simultaneous actions.

In this meeting we will explain the origin of the deficiency and how the stochastic game scenario differs from the problem used in the original paper describing the double oracle algorithm. Then we will derive the method needed to apply Kuhn's theorem and build a tabular MDP from a population distribution of opponent behavior. Finally, we can then create an exact tabular version of the PSRO algorithm and demonstrate the equivalence of its equilibrium policies with that of a value iteration solution. For additional reading, see the paper introducing PSRO

Meetup Links:
Recordings of Previous RL Meetings
Recordings of Previous MARL Meetings
Short RL Tutorials
My exercise solutions and chapter notes for Sutton-Barto
My MARL repository
Kickoff Slides which contain other links
MARL Kickoff Slides

MARL Links:
Multi-Agent Reinforcement Learning: Foundations and Modern Approaches
MARL Summer Course Videos
MARL Slides

Sutton and Barto Links:
Reinforcement Learning: An Introduction by Richard S. Sutton and Andrew G. Barto
Video lectures from a similar course