Run Transferable Libraries — Learning Functional Bias in Problem Domains

Maarten Keijzer¹, Conor Ryan², and Mike Cattolico³

¹Prognosys, Utrecht, The Netherlands
mkeijzer@xs4all.nl

²University of Limerick, Ireland
conor.ryan@ul.ie

³Tiger Mountain Scientific Inc., Kirkland, WA USA
mike@tigerscience.com

Abstract. This paper introduces the notion of Run Transferable Libraries, a mechanism to pass knowledge acquired in one GP run to another. We demonstrate that a system using these libraries can solve a selection of standard benchmarks considerably more quickly than GP with ADFs by building knowledge about a problem. Further, we demonstrate that a GP system with these libraries can scale much better than a standard ADF GP system when trained initially on simpler versions of difficult problems.

LNCS 3103, p. 531 ff.

Full article in PDF

lncs@springer.de
© Springer-Verlag Berlin Heidelberg 2004