Evolutionary Design of Artificial Neural Networks

Abstract

WE saw in Chapter 2 that artificial neural networks are biologically-inspired computational models that have the capability of somehow “learning” or “self-organizing” to accomplish a given task. They are particularly efficient when the nature of the task is ill-defined and the input/output mapping largely unknown. However, many aspects may affect the performance of an ANN on a given problem. Among them, the most important is the structure of the neuron connections i.e., the topology of the net, the connection weights, the details of the learning rules and of the neural activation function, and the data sets to be used for learning. There are guidelines for picking or finding reasonable values for all of these network parameters but most are rules of thumb with little theoretical background and without any relationship with each other.