Abstract
This study examines the applicability of artificial neural networks (ANN) to theestimation of wave overtopping over sloping seawalls, especially for searching the best structure of ANN. The linear activation function was found to be a good choice for output units. Correlation coefficients between measurements and predictions were best when 6 input units and 12 hidden layer units were employed. The Levenberg-Marquardt method with Early Stopping (LM) and with Bayesian Regulation (BAYESIAN) both gave reasonable predictions. The LM requires a validation data set to prevent over-fitting and to judge the convergence and generalization. The BAYESIAN, recommended in this study, does not require a validation data set, butrequired more iterations of learning. If there are no data for non-overtopping conditions, the ANN cannot recognize when wave overtopping fails to occur. It is concluded that the ANN proposed here gives reasonable predictions for mean wave overtopping rates.
