Comparison and Analysis of Mutation-based Evolutionary Algorithms for ANN Parameters Optimization

Mutation-based Evolutionary Algorithms, also known as Evolutionary Programming (EP) are commonly applied to Artificial Neural Networks (ANN) parameters optimization. This paper presents a comparative analysis of mutation approaches, based on the different distributions for the purpose to examine their performance for ANNs with the predefined architectures, evaluate the average improvement of chromosomes and investigate their ability to find solutions with the high precision. The experiments have been provided for training feed-forward ANNs using the XOR as a case problem. Besides established EP techniques, i.e. the classical EP (CEP), based on a standard normal distribution, the Fast EP (FEP), based on the Cauchy distribution, the improved FEP (IFEP), based on both Cauchy and Gaussian distributions, a novel selfadaptive dynamic mutation strategy, based on a uniform distribution is considered. The proposed approach consists of two components; the first component describes the ANN "...

• Type: Conference
• Year: 2006
• Where: DMIN
• Authors: Kristina Davoian, Alexander Reichel, Wolfram-Manfred Lippe
• Keywords: DMIN 2006, DMIN 2007, Hidden Layers, Mutation, Mutation-based Evolutionary Algorithms