A simple method for selection of inputs and structure of feedforward neural networks

When feedforward neural networks of multi-layer perceptron (MLP) type are used as black-box models of complex processes, a common problem is how to select relevant inputs from a large set of potential variables that affect the outputs to be modeled. If, furthermore, the observations of the input-output tuples are scarce, the degrees of freedom may not allow for the use of a fully connected layer between the inputs and the hidden nodes. This paper presents a systematic method for selection of both input variables and a constrained connectivity of the lower-layer weights in MLPs. The method, which can also be used as a means to provide initial guesses for the weights prior to the final training phase of the MLPs, is illustrated on a class of test problems.