IMAGe Brown Bag Seminar - Sequential Estimation for Space-time Modelling

03/03/2017 - 12:00pm to 1:00pm
ML - Chapman Room

Matt Edwards
Newcastle University

Large space-time datasets often require models with many parameters to capture dependencies at different scales. Consequently, the memory and computational costs required to perform inference are very demanding. The numerical optimization algorithms used to perform maximum likelihood estimation (MLE) require multiple iterations that each involve evaluations of the likelihood. Most efforts to overcome the costs of inference have focused on reducing the costs of evaluation. With the exception of a few special cases, no non-exponential complexity numerical optimization algorithms exist for non-convex likelihood functions. As a consequence, performing MLE is infeasible for large parameter spaces, regardless of the costs of evaluation. This work proposes Sequential Likelihood Estimation (SLE); an approximate inference method that estimates subsets of the model parameters in multiple steps in order to reduce the parameter space over which numerical optimization is performed. This multi-step inference method requires a sequence of likelihoods, each likelihood employed to estimate a different subset of the model parameters. Independence likelihoods are ideally suited to SLE as they can facilitate various levels of parallelization and a consistency theorem provides conditions under which bias and variance propagation are controlled. Furthermore, this work introduces a flexible and general class of space-time models specifically designed for the SLE method. To demonstrate the effectiveness of the SLE method this model is applied to over 400,000 space-time points from the North American Regional Climate Change Assessment Program dataset.

(Bring your lunch)