Identifying irrelevant input variables in chaotic time series problems: Using the genetic algorithm for training neural networks

Many researchers consider a neural network to be a "black box" that maps the unknown relationships of inputs to corresponding outputs. By viewing neural networks in this manner, researchers often include many more input variables than are necessary for finding good solutions. This causes unneeded computation as well as impeding the search process by increasing the complexity of the network. The main reason for this rational is the dependence upon gradient techniques, typically a variation of backpropagation, by the vast majority of neural network researchers for network optimization. Since gradient techniques are incapable of identifying unneeded weights in a solution, researchers have not been able to determine contributing inputs from those that are irrelevant. By using a global search technique, the genetic algorithm, for neural network optimization, it is possible to identify unneeded weights in the network model, which allows for identification of irrelevant input variables. This paper demonstrates through an intensive Monte Carlo study, that the genetic algorithm can automatically reduce the dimensionality of neural network models during network optimization. The genetic algorithm is also directly compared with backpropagation networks to show effectiveness for finding global versus local solutions.