Embedding astronomical computer models into principled statistical analyses Seminar
- Date:
- 2 May 2012
- Venue:
- Building 54 Lecture theatre 4A
For more information regarding this seminar, please email Mrs Jane Revell at j.revell@southampton.ac.uk .
Event details
Special seminar
Abstract
Sophisticated and computational expensive computer models are routinely used to describe complex processes in the physical, engineering, and social sciences. In Astronomy, for example, computer models are used to describe such processes as the evolution of stars and galaxies, the formation of the Universe, and the workings and calibration of sophisticated space-based telescopes. Like a statistical likelihood, these models typically predict observed quantities as a function of a number of unknown parameters. Including them as components of a multi-level statistical model, however, leads to significant modelling, inferential, and computational challenges. In this talk, I describe how we tackle these challenges in the context of two examples: (i) a principled statistical analysis of stellar evolution and (ii) the calibration of space-based X-ray telescopes. In our stellar evolution model we must link together a number of separate computer models for various phases in the life of a star. These separate models are combined via parametric models the relate the output of one model with the inputs of the next and are themselves of direct scientific interest. In our calibration example, we employ computational simulators of the subassembly components of the detector to emulate the prior distribution of its response function. In both cases we embed the computer models into likelihood-based statistical models that allow for principled inference but present substantial computational challenges. We use a combination of sophisticated MCMC techniques and simple emulators of the computer models to tackle these challenges. This strategy allows us to apply the full force of powerful statistical tools to build, fit, check, and improve the statistical models and their computer model components. In contrast to traditional methods, we can estimate the uncertainty in our fit using principled statistical techniques, and we can ensure that the components of our overall statistical models are internally coherent.
Speaker information
David A. van Dyk , Imperial College London. Statistics Section