Prediction of peak ground acceleration by genetic expression programming and regression: A comparison using likelihood-based measure

Abstract Peak ground acceleration (PGA) has still been considered one of the important factors that plays significant role on the earthquake-induced damage of structures. Thus, prediction of the PGA and selection of appropriate ground motion models often have become a valuable topic for seismic hazard assessments. This paper presents an application in order to predict the PGA by a relatively new prediction tool, genetic expression programming (GEP) and the conventional method of regression, and then aims to select the appropriate ground motion models derived by the GEP and regression methods. The selection was carried out by comparing the prediction performances through a ranking process using a new goodness-of-fit measure, likelihood estimation (LH), recently proposed by Scherbaum et al. (2004, on the use of response spectral reference data for the selection and ranking of ground-motion models for seismic hazard analysis in regions of moderate seismicity: the case of rock motion. Bulletin of the Seismological Society of America 94, 1–22). The Turkish earthquake data properly organized were used for deriving the candidate ground motion models of PGA attenuation equations. The validations of the LH method as well as the GEP and regression models were performed by the key attenuation characteristics, the attenuation equations for Turkey and the case records of Turkish strong ground motion data. The results indicate that majority of the PGA candidate models (GEP and regressions) that are ranked as good qualifications (class A and B) by the LH method generally pass the model validations, but the ones ranked as lower levels (class C) fail. This finding indicates that the LH method can be beneficial for the comparisons of the candidate ground motion models, and the selection of appropriate models for attenuation studies. Even though the GEP-based models that were found to be appropriate by the LH method need some improvements to obtain simple functional forms of PGA predictive relationships, their PGA predictions were reasonably consistent with the model validations. Thus, this study suggests that the GEP approach can be employed for predicting the PGAs in seismic hazard studies at least as a supplement to conventionally derived predictive relationships from the regressions.