AI 2012 - LNAI 7691

Local Search in Parallel Linear Genetic Programming for Multiclass Classification

Aaron Scoble¹, Mark Johnston², and Mengjie Zhang¹

¹School of Engineering and Computer Science, Victoria University of Wellington, P.O. Box 600, Wellington, New Zealand
aaron.scoble@ecs.vuw.ac.nz
mengjie.zhang@ecs.vuw.ac.nz

²School of Mathematics, Statistics and Operations Research, Victoria University of Wellington, P.O. Box 600, Wellington, New Zealand
mark.johnston@msor.vuw.ac.nz

Abstract. Parallel Linear Genetic Programming (PLGP) is an architecture that addresses instruction dependencies in Linear Genetic Programming (LGP). The Co-operative Coevolution (CC) methodology has previously been applied to PLGP but implementations have not been able to improve performance over vanilla PLGP. In this paper we present Hill Climbing Parallel Linear Genetic Programming (HC-PLGP) which uses a local search to discover effective combinations (blueprints) of partial solutions that are evolved in subpopulations. By introducing a new caching technique we can efficiently search over the subpopulations, and our improved fitness function combined with normalisation and blueprint elitism address some of the weaknesses of the previous approaches. Hill Climbing Parallel Linear Genetic Programming (HC-PLGP) is compared to three PLGP architectures over six datasets, and significantly outperforms them on two datasets, is comparable on three, and is slightly worse on one dataset.

LNAI 7691, p. 373 ff.

Full article in PDF | BibTeX