Reinforcement learning in Markovian evolutionary games

摘要

A population of agents plays a stochastic dynamic game wherein there is an underlying state process with a Markovian dynamics that also affects their costs. A learning mechanism is proposed which takes into account intertemporal effects and incorporates an explicit process of expectation formation. The agents use this scheme to update their mixed strategies incrementally. The asymptotic behavior of this scheme is captured by an associated ordinary differential equation. Both the formulation and the analysis of the scheme draw upon the theory of reinforcement learning in artificial intelligence.