Title: Fast inference in expensive computational models
Authors: Umberto Noe - University of Glasgow (United Kingdom) [presenting]
Dirk Husmeier - Biomathematics and Statistics Scotland, Edinburgh (UK)
Maurizio Filippone - University of Glasgow (United Kingdom)
Nicholas Hill - University of Glasgow (United Kingdom)
Weiwei Chen - University of Glasgow (United Kingdom)
Abstract: Inference in expensive computational models, involving the numerical solution of a system of Partial Differential Equations (PDEs), is discussed. These models are typically referred to as ``black-box'' functions and could arise from Biomechanical, Engineering or Financial problems. They are not suitable for MCMC or standard likelihood based inference due to the high computational resources needed for a single output and the non-availability of gradient information. We present an extension of the Efficient Global Optimization (EGO) algorithm that allows for hidden constraints. The latter often arise in practice when a simulation returns no output at a point that violates the model assumptions.We illustrate the improved algorithm on a PDE model of the double sided Human Pulmonary Circulation. The reduction in the computational cost, compared to standard likelihood-based methods, is by two orders of magnitude.