Chapter 9 Neural networks for encoding and adapting in dynamic economies

This chapter describes feedforward neural networks as approximators and relates them to statistical discriminant functions, and explains the ways in which neural nets of varying complexity can represent equilibria in two repeated games and one dynamic economic model. Because linear strategies are simple to implement, a large class of equilibrium outcomes can be represented with simple perceptions. However, in the presence of imperfect monitoring, the linear proxies can overestimate the hidden variables. The chapter presents bounded rationality as a constraint on computational capabilities. Thus, the main exercise is to construct equilibria sustained by simple neural networks, and to study the ways in which the restriction on strategies reduces the set of equilibrium outcomes. The approach of Rubinstein (1986) presumes that players have access to finite automata that encode transition and outcome functions. It is believed that neural networks are useful tools to study the strategic implications of complexity cost, especially when, the feasible computing machines may have infinite states, and the computing machine can adapt itself as the game progresses.