ABSTRACT
In ensemble methods, as presented in ML, data drive both the learning of the individual models and their combination into an ensemble. This is not entirely the aim in our case; rather, we are provided with several physically plausible predictive models
INTRODUCTION
An ensemble is a collection of models where each model has been set-up or trained to solve the same problem. One of the main motivations for using an ensemble, rather than a single model, is the superior performance in terms of generalisation error. In the paper we wish to highlight certain ideas behind ensemble methods and motivate why such methods are relevant in the context of ground motion model (GMM) aggregation. GMMs are an important component of PSHA. GMM uncertainty is an important factor controlling the exceedence frequency of e.g. peak ground acceleration (PGA) (Bommer & Abrahamson 2006). Several GMMs have been proposed, and these differ considerably in the way they have been derived, resulting in different functional forms and in the number and kind of predictor variables used to characterize the earthquake source, path and site effects.
