Combined Biological Metaphors
Egbert J. W. Boers, Ida G. Sprinkhuizen-Kuyper · The MIT Press eBooks · 2001
As computer hardaware becomes more powerfll, increasingly large artificial neural networks will becomme feasible. There will be no limit to the size of the networks, in principle, and it will become increasingly difficult to design them and understand their internal operation. That is why it is very important to find design methods that are scalable to large network sizes. In this chapter we describe our explorations into a new scalable method for the construction of good neural network architectures for a given task. Reverse-engineering the only example available, the evolution of our own brains, we have combined several ideas from evolution, embryology, neurology, and theoretical psychology. The method we describe in this chapter allows the coding of large modular artificial neural network architectures while restricting the search space of the genetic algorithm that is used for the optimization. This is done by using a scalable, recipe-like coding (as opposed to a blueprint-like coding). Currently, it is still very time-consuming to train artificial neural networks, which has to be done many times in our method to establish their evolutionary fitness. Therefore, we are working on all kinds of improvements to accelerate the process. In this chapter we will describe some new techniques that can make architectural improvements to a given artificial neural network during their fitness evaluation. These techniques can speed-up and improve the evolutionary search for good network arhectures, an effect that has become known as the 'Baldwin effect'.