Trivial Geography in Genetic Programming

Abstract

Geographical distribution is widely held to be a majordeterminant of evolutionary dynamics. Correspondingly,genetic programming theorists and practitioners havelong developed, used, and studied systems in whichpopulations are structured in quasi-geographical ways.Here we show that a remarkably simple version of thisidea produces surprisingly dramatic improvements inproblem-solving performance on a suite of testproblems. The scheme is trivial to implement, in somecases involving little more than the addition of amodulus operation in the population access function,and yet it provides significant benefits on all of ourtest problems (ten symbolic regression problems and aquantum computing problem). We recommend the broaderadoption of this form of 'trivial geography' in geneticprogramming systems.