A discipline of evolutionary programming

https://doi.org/10.1016/S0304-3975(99)00263-7Get rights and content
Under an Elsevier user license
open archive

Abstract

Genetic fitness optimization using small populations or small population updates across generations generally suffers from randomly diverging evolutions. We propose a notion of highly probable fitness optimization through feasible evolutionary computing runs on small-size populations. Based on rapidly mixing Markov chains, the approach pertains to most types of evolutionary genetic algorithms, genetic programming and the like. We establish that for systems having associated rapidly mixing Markov chains and appropriate stationary distributions the new method finds optimal programs (individuals) with probability almost 1. To make the method useful would require a structured design methodology where the development of the program and the guarantee of the rapidly mixing property go hand in hand. We analyze a simple example to show that the method is implementable. More significant examples require theoretical advances, for example with respect to the Metropolis filter.

Cited by (0)

Preliminary version published in: Proc. 7th Int'nl Workshop on Algorithmic Learning Theory, Lecture Notes in Artificial Intelligence, Vol. 1160, Springer-Verlag, Heidelberg, 1996, 67–82.

2

URL: http:/www.cwi.nl/oaulu

1

Partially supported by the European Union through NeuroCOLT II Working group, and by NWO through NFI Project ALADDIN under Contract number NF 62-376. Author's affilliations are CWI and the University of Amsterdam.