Computable Learning, Neural Networks and Institutions

Francesco Luna
Department of Economics, Università di Venezia
FLuna@unive.it

Abstract

We propose a tractable simplification of Gold's (1965, 1967) inductive inference model based on neural networks. In this way, we can introduce explicitly institutions in a learning model. In particular, the hidden layer of neural network is shown to perform certain functions-data preprocessing and uncertainty and complexity reduction-that are typically attributed to institutions.

In an evolutionary context based on selection and successive generations, we study under what circumstances individuals employing particular institutions-that improve their learning capabilities-are successful. Simultaneously, we are interested in how this process determines the adoption of different institutions in the population. From this perspective, this paper can be considered a contribution to the literature dealing with the emergence of a dominant design. We record lock-in phenomena as well as ``adoption externalities".

However, our main goal is that of showing that in the context of a major change in the environment, the more rigid and strictly specialized an institution is, the longer and more complex the learning process will be of any economic actor subject to the by-now obsolete institution.

The metaphor we suggest is for the firm intended as organization and productive process. A major change in the environment is, in this case, the introduction of an innovation or of compulsory standards. By describing an economic actor in structural terms-via neural networks-, we automatically impose bounded rationality. This is so because each agent will behave according to well defined procedures, but also because of the obvious cognitive and computational limitations.