Why Do Machine Learning Based Techniques Fail to Accelerate the Evolution of Neural Networks? Is the Long Bitlength or the Nature of Neural Net Chromosomes to Blame?

Abstract

The first author's primary research interest is in building artificial brains, defined to be interconnected assemblages of 10,000s of electronically evolved neural network circuit modules, whose neural signaling is run in real time in an ordinary PC, to control autonomous robots, etc. An alternative to the electronic evolution of the NN modules is to use software based machine learning techniques in a PC, for example Michalski's LEM algorithm [Michalski 2000]. LEM can be very successful at accelerating the evolutionary optimization of multivariable mathematical functions, but appears to fail to accelerate the evolution of neural networks [Aleti & de Garis 2004]. This paper reports on experiments comparing the evolution times of mathematical optimization problems with those of neural network evolution problems, where the bit string chromosomes of both sets of problems are the same, to see if there is something special about neural network chromosomes that make them unsuitable to have their evolution accelerated. © Springer-Verlag 2004.