Core Concepts
Deep reinforcement learning algorithm for population-dependent Nash equilibrium in Mean-Field Games.
Stats
MFGs provide a framework for large-population games.
Convergence of Banach-Picard fixed point iterations relies on strict contraction condition.
Fictitious Play method smoothens mean field updates by averaging historical distributions.
Online Mirror Descent method stabilizes learning process using past iterations.
Master policy enables attainment of Nash equilibrium from any initial distribution.
Quotes
"The resulting policy can be applied to various initial distributions."
"Algorithm is more efficient than FP in learning master policies."
"Numerical experiments demonstrate algorithm's superiority."