Evolving Sum and Composite Kernel Functions for Regularization Networks
In this paper we propose a novel evolutionary algorithm for regularization networks. The main drawback of regularization networks in practical applications is the presence of meta-parameters, including the type and parameters of kernel functions Our learning algorithm provides a solution to this problem by searching through a space of different kernel functions, including sum and composite kernels. Thus, an optimal combination of kernel functions with parameters is evolved for given task specified by training data. Comparisons of composite kernels, single kernels, and traditional Gaussians are provided in several experiments.
Keywordsregularization networks kernel functions genetic algorithms
Unable to display preview. Download preview PDF.
- 2.Kůrková, V.: Learning from data as an inverse problem. In: Antoch, J. (ed.) Computational Statistics, pp. 1377–1384. Physica Verlag, Heidelberg (2004)Google Scholar
- 3.Poggio, T., Girosi, F.: A theory of networks for approximation and learning. Technical report, Cambridge, MA, USA (1989); A. I. Memo No. 1140, C.B.I.P. Paper No. 31Google Scholar
- 5.Neruda, R., Vidnerová, P.: Genetic algorithm with species for regularization network metalearning. In: Papasratorn, B., Lavangnananda, K., Chutimaskul, W., Vanijja, V. (eds.) Advances in Information Technology. Communications in Computer and Information Science, vol. 114, pp. 192–201. Springer, Heidelberg (2010)CrossRefGoogle Scholar
- 9.Prechelt, L.: PROBEN1 – a set of benchmarks and benchmarking rules for neural network training algorithms. Technical Report 21/94, Universitaet Karlsruhe (9 (1994)Google Scholar
- 10.LAPACK: Linear algebra package, http://www.netlib.org/lapack/