Building New Spatial Interaction Models Using Genetic Programming

Building computer models of human geographic systems is di cult because of the poverty of applicable theory that can be used to specify well formed and soundly based mathematical models. As a result there has been little progress in this area since the late 1960's when mathematical model building procedures based on entropy-maximisingmethods were rst applied to spatial interaction data. In the mid{1980's, attention was focused on the use of genetic algorithms to explore the universe of di erent models that could be built from the available symbolic pieces; that is there is a very very large number of permutations of the model building blocks; viz. a mix of unary and binary operators, unknown parameters and observed variables, and rules of syntax for well formed equations. The so-called Automated Modeling System (AMS) worked but never ful lled its expectations. The paper revisits the problem but re-casts the AMS approach into a genetic programming framework. Some preliminary results based on Cray-YMP runs are reported.