Modelling Bounded Rationality In Agent-Based Simulations using the Evolution of Mental Models

1 Introduction

There are many possible ways of modelling economic agents. These traditionally fall into one of two camps, dating from Simon's distinction between substantive and procedural rationality: this is often characterised as those with bounded rationality and those with no such bounds (although this is not strictly correct [24]). Although the latter type is more analytically tractable we are interested in the former type.

The purpose of this paper is to report a particular approach to simulating economic agents with procedural and bounded rational agents such that their behaviour matches some of the broad qualitative characteristics of people. In particular it aims to capture a context-dependent, resource-bounded, open-ended, satisficing cognition such as might be feasible in a real agent*1. The approach taken is to introduce these characteristics using evolutionary techniques.

By using an approach to modelling learning that is close to that used in genetic programming (GP) [17], we open up a new range of possibilities in the credible modelling of such agents. Here an agent in the simulation has a population of candidate beliefs (or models) of its environment which evolve. This evolution is its learning mechanism. As well as differing from traditional economic models of agency, this also contrasts with agent modelling approaches that use "crisp" logic-like beliefs, and those approaches that only involve some inductive learning. In particular multiple and frequently inconsistent beliefs are held as a resource for future model development. In this way I simultaneously embrace Simon's emphasis on the importance of the learning process and reject the sequential symbol processing picture of cognition he adopted.

After describing the basic model, I will then describe two examples that use these techniques: the first is of agents attempting to learn the form of a function in the presence of structural change and the second is an extension of a model of Brian Arthur's (the `El Farol Bar' model) where evolutionary learning and communication has been added.

Modelling Bounded Rationality In Agent-Based Simulations using the Evolution of Mental Models - 17 MAR 98

[Next] [Previous] [Top] [Contents]

Generated with CERN WebMaker