A strategy for Bayesian inference for computationally expensive models with application to the estimation of stem cell properties Seminar

Abstract
Bayesian inference is considered for statistical models that depend on the evaluation of a computationally expensive computer model. For such situations, the number of evaluations of the likelihood function, and hence of the unnormalised posterior probability density function, is determined by the available computational resources and may be extremely limited. A new example of such a computer model that describes the properties of human embryonic stem cells using data from optical trapping experiments is presented. This application is used to motivate a strategy for Bayesian inference which exploits a Gaussian process approximation of the computer model and allows computationally efficient MCMC inference. The advantages of this strategy over previous methodology are that it is less reliant on the determination of tuning parameters and allows the application of model diagnostic procedures that require no additional evaluations of the computer model.