Assessment of Relative Contribution of Input Variables to Pavement Performance Prediction by Artificial Neural Networks

Abstract

Neural networks can be quite efficient at classification, pattern recognition, and prediction. However, the knowledge embedded in the neural network is opaque; that is, it is difficult to understand how neural networks arrive at the results. Here three explanatory mechanisms (attempts to explain the knowledge) are reviewed from the literature. Neural networks are trained to predict the present serviceability rating (PSR) of pavements with structural number, age, and cumulative equivalent single-axle loads as input variables. To determine the relative contribution of each input variable to the prediction of PSR, the method of partitioning of connection weights is employed. Neural networks are also trained with a few connection weights pruned, and the relative contribution of the input variables to the prediction of PSR is assessed. Apart from the three input variables mentioned above, another input variable, random numbers, is introduced. Its contribution to the prediction of PSR is found to be minimal.