An evolutionary approach to learning in a changing environment

Abstract

The evolutionary fitness of the ability to learn a best response (at a cost) is explored in an environment where the underlying game changes stochastically over time. We provide a sufficient condition for asymptotic dominance by the learners. It does not require that the learners are the most fit type in any single state of the environment. We identify a necessary and sufficient condition that ensures the survival of the learners when learning costs are small. Last, we describe a medium-run phenomenon where the learners “spike”—their share temporarily rises and gets arbitrarily close to one following a regime shift.