A biologically inspired methodology for neural networks design

The aim of this paper is to introduce a biologically plausible methodology that can automatically generate artificial neural networks (ANNs) with an optimum number of neurons and connections, good generalization capacity, smaller error and larger tolerance to noises. In order to do this, three biological metaphors were used: genetic algorithms (GA), Lindenmayer systems (L-systems) and ANNs. At the end of the paper some experiments are presented in order to investigate the possibilities of the method, especially in problems where a recurrent neural network should be evolved. The proposed problems are the parity generator and the recognizers for some regular languages proposed by Tomita. Some of the advantages of the proposed methodology is that it increases the level of implicit parallelism of genetic algorithm and seems to be capable to generate an economical satisfactory neural architectures that solve specifics tasks, reducing the project costs and increasing the performance of the obtained neural network.