A Comparative Analysis of Methodologies for Automatic Design of Artificial Neural Networks from the Beginnings until Today

Despite being extensively used in many fields of science, there is still no general procedure to determine the most suitable neural architecture for a given task. Even today the design of an artificial neural network (ANN) is still very dependent on the designer's experience. Several methods for automatic design of ANNs have been proposed, but they are not biologically inspired, in this paper it's presented a state of the art about this research theme, from the earliest methodologies emerged in the 80's to the present day focusing on aspects such as biological plausibility, scalability and neural computability. Furthermore, we introduce a biologically plausible methodology, section III.C, that can automatically generate Artificial Neural Networks (ANNs) with optimum number of neurons and connections. The same can generate recurrents neural networks with an optimal number of neurons and connections, good generalization capacity, smaller error and large tolerance to noises. The methodology was tested in three well known simple problems where recurrent networks topologies must be evolved. A more complex problem, involving time series learning was also proposed for application. The experiments results show that our methodology is very promising and can lead to optimal design of ANNs. Finally, we make a comparative analysis of the various methodologies showing advantages and disadvantages for the different encoding methods from two aspects: genome evolution and neural network development.