next up previous
Next: Introduction

Genetic Programming Bloat with Dynamic Fitness

W. B. Langdon and R. Poli

School of Computer Science, University of Birmingham, Birmingham B15 2TT, UK
{W.B.Langdon,R.Poli}@cs.bham.ac.uk http://www.cs.bham.ac.uk/wbl, 126 rmp
Tel: +44 (0) 121 414 4791, Fax: +44 (0) 121 414 4281

Technical Report: CSRP-97-29, 3 December 1997

Abstract:

In artificial evolution individuals which perform as their parents are usually rewarded identically to their parents. We note that Nature is more dynamic and there may be a penalty to pay for doing the same thing as your parents. We report two sets of experiments where static fitness functions are firstly augmented by a penalty for unchanged offspring and secondly the static fitness case is replaced by randomly generated dynamic test cases. We conclude genetic programming, when evolving artificial ant control programs, is surprisingly little effected by large penalties and program growth is observed in all our experiments.





William B Langdon
Mon Dec 8 19:56:15 GMT 1997