Roland's Master Thesis Page (Angewandte Informatik)

"Pruning, Growing and Evolutionary Design of Artificial Neural Networks"

Abstract:

In our work we describe three approaches how to automatically determine the architecture of artificial neural networks (ANN) for a given problem. The first approach, network growing , starts with a basic number of neurons and adds neurons (grows) until the overall architecture fulfills the problem to a certain extent. The second approach, network pruning , works the other way round. From the network known to solve the problem neurons are removed (pruned) as long as the architecture preserves its performance. The third approach, evolutionary development of neural architectures , mimicks nature and evolves generations of neural networks. At the end of the evolutionary cycle the network which is best adopted to the problem, according to given criteria, is the result of the design process. The netGEN system which belongs to the class of systems simulating the evolutionary approach and which was co-designed by the author will be introduced in detail at the end of our thesis. Furthermore we give examples to illustrate the approaches and algorithms in detail.

Literature on Neural Network Growing, Pruning and Evolutionary Design

• Articles Bibliography by Roland Schwaiger (.bib)
• Neural Network FAQ by Warren S. Sarle