Biological Metaphors In Designing Modular Artificial Neural Networks [chapter]

Egbert J. W. Boers, Herman Kuiper, Bart L. M. Happel, Ida G. Sprinkhuizen-Kuyper
1993 ICANN '93  
Preface This thesis is the result of a research done at the departments of Computer Science and Experimental and Theoretical Psychology at Leiden University, the Netherlands. It is a part of the project 'Architecture and Function of Modular Neural Networks', done at the department of Experimental and Theoretical Psychology by Bart L.M. Happel 1 . The research comprised four parts: first, extensive reading was done to get familiar with the areas of neural networks, genetic algorithms and formal
more » ... rammars. Secondly, new methods were developed to design modular structures, which were then transformed into software with which a number of experiments were done. Finally, this thesis was written, together with manuals for the software developed. We wish to thank Bart Happel for his stimulating input and excellent suggestions whenever we got stuck. Also, we wish to thank Ida Sprinkhuizen-Kuyper for her guidance throughout the project. Finally, we thank the numerous proofreaders who helped to write (hopefully) a faultless thesis and all the Sun users at the Computer Science department who had to put up with our simulations. Leiden Abstract In this thesis, a method is proposed with which good modular artificial neural network structures can be found automatically using a computer program. A number of biological metaphors are incorporated in the method. It will be argued that modular artificial neural networks have a better performance than their non-modular counterparts. The human brain can also be seen as a modular neural network, and the proposed search method is based on the natural process that resulted in the brain: Genetic algorithms are used to imitate evolution, and L-systems are used to model the kind of recipes nature uses in biological growth. A small number of experiments have been done to investigate the possibilities of the method. Preliminary results show that the method does find modular networks, and that those networks outperform 'standard' solutions. The method looks very promising, although the experiments done were too limited to draw any general conclusions. One drawback is the large amount of computing time needed to evaluate the quality of a population member, and therefore in chapter 9 a number of possible improvements are given on how to increase the speed of the method, as well as a number of suggestions on how to continue from here. vi Contents Preface iii Abstract v
doi:10.1007/978-1-4471-2063-6_219 fatcat:hwr5mgg6x5edxpaktfocup5hlq