10:00 〜 10:15
[SCG52-05] 地震動予測式のサイト汎化性能：単調ニューラル・ネットワークの適用
Ground motion models (GMM) provide empirical relationships between earthquake parameters and ground-motion intensities. In addition to regression models developed over the past 50 years, machine learning (ML) methods recently came into use. ML models can obtain input-output relationships without assumptions, but are in danger of overfitting owing to their flexibility. To evaluate the predictive power (generalization performance) of ML GMMs, observational records are customary divided into training and test datasets based on records or earthquakes.
In this presentation, however, we demonstrate that the above division of dataset does not work for monitoring overfitting to recorded site when a model includes site-condition proxies (SCP) as input variables; complex deep neural network (DNN) models with many SCPs apparently exhibit good predictive power at trained sites but show serious overfitting at new sites. Therefore, it is crucial to divide dataset based on recorded site for correctly evaluating site generalization performance. As a possible solution to maintain generalization performance in DNN models, we propose to impose monotonic dependence on input variables. An experimental application supports the effectiveness of this simple approach.
In this presentation, however, we demonstrate that the above division of dataset does not work for monitoring overfitting to recorded site when a model includes site-condition proxies (SCP) as input variables; complex deep neural network (DNN) models with many SCPs apparently exhibit good predictive power at trained sites but show serious overfitting at new sites. Therefore, it is crucial to divide dataset based on recorded site for correctly evaluating site generalization performance. As a possible solution to maintain generalization performance in DNN models, we propose to impose monotonic dependence on input variables. An experimental application supports the effectiveness of this simple approach.