Perceptrons play repeated games with imperfect monitoring

This paper studies two-person repeated games with imperfect monitoring without discounting through perceptrons, which are feedforward artificial neural networks. Under a fairly standard informational condition, we establish the folk theorem through perceptrons with at most three linear classifiers. The maximum number of linear classifiers is independent of the number of actions in the component game or the target payoff vector. In particular, the perceptron dictates that each player monitor the opponent's action by computing the ordinary least-square estimator of the opponent's expected payoff. (C) 1996 Academic Press, Inc.