Approximate Bayesian Computation by Subset Simulation for model selection in dynamical systems

Abstract

Approximate Bayesian Computation (ABC) methods are originally conceived to expand the horizon of Bayesian inference methods to the range of models for which only forward simulation is available. However, there are well-known limitations of the ABC approach to the Bayesian model selection problem, mainly due to lack of a sufficient summary statistics that work across models. In this paper, we show that formulating the standard ABC posterior distribution as the exact posterior PDF for a hierarchical state-space model class allows us to independently estimate the evidence for each alternative candidate model. We also show that the model evidence is a simple by-product of the ABC-SubSim algorithm. The validity of the proposed approach to ABC model selection is illustrated using simulated data from a three-story shear building with Masing hysteresis.