Optimum binary codification for genetic design of artificial neural networks
D. Barrios, Daniel Manrique, Juan Felipe Porras, José Ríos · 2002
This paper describes a new scheme of binary codification of artificial neural networks designed to be used for generating automatically neural networks using genetic algorithms. Instead of using direct mapping of chromosomes in network connectivities, this particular codification abstracts genetic encoding so that it does not reference the artificial indexing of network nodes; thus this codification employs shorter chromosome length while avoids illegal individuals but does not exclude any legal neural network. With this purpose, a particular internal operation, called superimposition, has been designed in the set of artificial neural networks that allows building complex neural nets from minimum useful structures while it preserves the important feature that similar neural networks only differ in one bit, which is very desirable when using genetic algorithms. Experimental results are reported showing that this encoding scheme exhibits scaling properties when encoding large networks while the decoding process is very simple.