Nash versus Reinforcement Learning on a Search Market Some Similarities and Differences between Individual and Social Learning
In this paper, we consider a simple search market extended from H. Varian's (Amer. Econ. Rev., 1980) classical model and ask whether less informed sellers are able to learn such sophisticated price strategies in this framework. We therefore compare the choices made by adaptive sellers (reinforcement learning) to those made by Nash sellers. We confront the results of two types of learning models: individual learning (where sellers only observe their own price performance) and social learning (where sellers observe the pricing experiments of other sellers). In the case of individual learning, we show that although sellers are not able to learn the Nash price distribution, they are able to qualitatively mimic Nash predictions when buyers' search information varies. In the case of social learning, first results suggest that the process is highly path dependent. Again, the choices made by adaptive sellers do not converge to the Nash equilibrium. In addition, some (but not all) qualitative properties are no longer preserved.
Eric DARMON, Roger WALDECK
Imperfect Information, Nash Equilibrium, Mixed Strategies, Reinforcement Learning, Individual and Social Learning