Evolutionary Algorithm vs. Other Methods for Constructive Optimisation of RBF Network Kernels
Three methods for optimising Radial Basis Function (RBF) neural network receptive field are compared in the paper, namely: gradient descent, simulation annealing, and evolutionary algorithm. An incremental RBF network training scheme is considered, i.e., in which RBF kernels are added one at the time and individually optimised. Algorithmic implementations of the tested optimisation methods for configuring the RBF receptive field are shown and their computation costs are compared. The considered optimisation methods yield excellent results for the classification benchmark of Iris flowers. For the genetic optimisation scheme three RBF Gaussian kernels are sufficient to achieve average classification accuracy of the Irises at the level of 98%.
KeywordsRadial Basis Function Radial Basis Function Neural Network Radial Basis Function Kernel Network Error Average Classification Accuracy
Unable to display preview. Download preview PDF.
- 5.Michalewicz Z. (1996) Genetic algorithms + data structures = evolution programs (in Polish), WNT, WarszawaGoogle Scholar
- 6.Koszlaga J., Strumillo P. (2000) Rule extraction from trained Radial Basis Function neural networks, Colloquia in Artificial Intelligence, Poland, October 2000, 97–106Google Scholar