Can Learning-Agent Simulations be used for Computer Assisted Design in Economics?

Abstract

I follow Nelson and Winter (1982, page 372) in the judgement that “the ability of a theory to illuminate policy issues ought to be a principle criterion by which to judge its merit. ” Wanting to illuminate policy leads me to seek modelling tools that lead to better predictions about social systems (even negative predictions about the difficulty of making firm recommendations are useful). The notion of “Computer Assisted Economic Design” (eco-CAD) is meant to represent a common idea of the sort of model that will perform this task; it is a very detailed facsimile of the social process, and agents are represented by artificially intelligent processes (or “learning-agents”).¹ I describe some modelling experiments I undertook exploring the possibility of eco-CAD for the specific policy question of how to redesign electricity markets. But I argue that eco-CAD is, today at least, a chimera. Learning agent simulations that go in this direction cannot today “illuminate policy issues”, because we know too little about the applicability of the building blocks of these models. This suggests a modelling style for learning agent simulations different from the ideal of eco-CAD: the models should be as simple as possible, and be used to study the components of learning in controlled environments. These foundations do need to be in place before the models can be used in an “economic-engineering” mode.

“CAD” is the standard acronym for Computer Assisted Design. CAD progress has led to very much faster design times, to greater experimentation, and falling costs with increasing quality in most disciplines of engineering.