1 EFFECTS OF ZERO-OVERTOPPING DATA IN ARTIFICIAL NEURAL NETWORK PREDICTIONS Hajime Mase 1 , Maria T. Reis 2 , Shunji Nagahashi 3 , Takehisa Saitoh 4 and Terry S. Hedges 5 This study examined the applicability of artificial neural networks (ANNs) to the estimation of wave overtopping over sloping seawalls, especially with regard to the best structure for an ANN. Correlation coefficients between measurements and predictions were best when 6 input units and 12 hidden layer units were employed. Bayesian Regularization, recommended in this study, does not require a validation data set. It was found that the ANNs could not recognize when wave overtopping failed to occur if data on zero overtopping were omitted. INTRODUCTION Coastal flood disasters caused by wave overtopping are expected to increase due to sea level rise and increased storminess. Thus, it is becoming more important to accurately predict wave overtopping for given conditions of structural configuration, nearshore bathymetry and wave climate. An artificial neural network (ANN) is an effective method for estimating wave overtopping. ANNs have been used in various fields, including coastal engineering; for example, in the stability analysis of rubble-mound breakwaters (Mase, 1994; Mase et al., 1995), in the prediction of the occurrence of wave impacts (Mase and Kitano, 1999), in the prediction of tides (Deo and Chaudhar, 1998) and in wave forecasting (Deo and Sridhar, 1999). This paper reports the use of ANNs for the prediction of the mean wave overtopping discharge over sloping seawalls, especially with regard to the best structure for an ANN (Mase et al., 2005). Engineers are often interested in the very small values of overtopping discharge which are allowable for vehicles and pedestrians in close proximity to seawalls. These small values can be over-predicted if zero data points are excluded when calibrating overtopping models. Over-prediction leads to over-design, with all of the associated costs and the implications for adverse impacts on the natural environment caused by unnecessarily high sea defences. Here, the authors examine the effects of zero-overtopping data when applying ANNs to estimate the mean wave overtopping discharge. 1 Disaster Prevention Research Institute, Kyoto University, Gokasho, Uji, Kyoto 611-0011, Japan 2 National Civil Engineering Laboratory (LNEC), Avenida do Brasil 101, 1700-066 Lisbon, Portugal 3 Construction Bureau, Osaka City, Minato-ku, Osaka 552-0012, Japan 4 Department of Civil Engineering, Kanazawa University, Kanazawa-shi, Ishikawa 9201192, Japan 5 Department of Engineering, University of Liverpool, Liverpool L69 3GQ, United Kingdom