Evolutionary Design of Neural Networks

Master of Science thesis by Marko Grönroos. Computer Science, Department of Mathematical Sciences, University of Turku, Finland. 1998.

Evolutionary Design of Neural Networks (PDF, 2.8MB) .ps.gz 1018 kB)

Bibliography entry (.bib)

Abstract
This thesis compares some methods for finding neural network architectures suitable for learning particular problems. We use an evolutionary algorithm with four different genetic encoding methods to search for the suitable architectures. We train the neural network weights with a separate neural learning algorithm. We use eight different learning problems for benchmarking the encoding methods. Four of the problems are artificial (XOR, Encoder and two function approximation problems), three are real-world classification problems from the Proben1 benchmarking problem set, and one is a bankruptcy classification problem studied earlier in one of our projects. Our evaluation criteria are classification accuracy and efficiency for using only the relevant variables. The classification results are compared also to those for network architectures found by a systematic search.
Keywords: neural networks, evolutionary algorithms, genetic algorithms, encoding methods


Annalee

Annalee is a C++ library for the evolutionary engineering of artificial neural network architectures. It is the software developed for conducting the experiments require by the thesis.

Sources On-line - Annalee v.0.4 (not always online!) These will be updated irregularly.

A version of the library will be publicly available at some time, although no support for the installation can be given at the moment. It currently compiles under Solaris 2.5 with EGCS v1.0.3 C++ compiler, but works in Linux with some minor changes.

And then, the deep heart's core of this work (yeah, right)


Pääsivu Takaisin Last modified: Thu Oct 2 22:24:03 EEST 2003