Effect of Noisy Data on Pavement Performance Prediction by Artificial Neural Networks

Abstract

Artificial neural networks are increasingly employed in prediction modeling and are particularly advantageous when the relationship between the response and the predictor variables is complex. For the purposes of prediction, neural networks are to be trained with data that are accurately compiled. Frequently, the data collected either from field or laboratory observations are noisy in nature. The effect of noisy data on the predictive capability of neural networks has been studied. Present serviceability rating (PSR) of pavements is the attribute to be predicted. Six noisy databases are created and are employed to train the neural networks to predict PSR. Regression equations are developed with the same noisy databases, and the predictions from neural networks are compared with those of regression. The results show that the neural networks predict PSR as accurately as regression models with a given noisy data. In addition, neural networks are trained with data containing no noise. If no noise is present in the data, neural networks predict PSR accurately while properly capturing the effect of each explanatory variable on the response variable.