Genetic Algorithm with Species for Regularization Network Metalearning

Authors: Petra Vidnerová, Roman Neruda
Publication year: 2010
Publication type: Konferenční příspěvek (mezinárodní konf.)
Periodical: Informačné Technológie - Aplikácie a Teória
Projects: Neural Networks Learning Algorithms Based on Regularization Theory
Bibliographic citation: VIDNEROVÁ, Petra, NERUDA, Roman. Genetic Algorithm with Species for Regularization Network Metalearning. In: PARDUBSKÁ, D., ed. Informačné Technológie - Aplikácie a Teória. Seňa: Pont, 2010, S. 111-116. ISBN 978-80-970179-3-4.
ID: 348394

Abstract

Regularization networks are one of the important methods for supervised learning. They benefit from very good theoretical background, although the presence of metaparameters is their drawback. The metaparameters are typically supposed to be given in advance and come ready as an input of the algorithm. Typically, they are set based on the task context by an experienced user. In this paper, we develop a method for finding optimal values of metaparameters, namely the type of kernel function, kernel parameters and regularization parameter. The method is based on co-evolutionary genetic algorithms with different species for different kind