Murali Haran
Penn State University
Gaussian processes for inference with implicit likelihoods
Complex deterministic and stochastic models are often used to describe
dynamic systems in disciplines like climate science and infectious
disease modeling. Inferring unknown parameters of these models is of
interest, both for understanding the underlying scientific processes as
well as for making valid predictions. Some of the challenges typically
involved in inference for these models are: (i) intractable or implicit
likelihood functions; (ii) computationally expensive model simulations;
(iii) high-dimensional, multivariate model output and observations. I
will describe an approach that uses a Gaussian process to fit
simulations of the complex model at several different parameter
settings, while also accounting for measurement error and a data-model
discrepancy term. This method provides a computationally tractable
approach for performing Bayesian inference for the parameters of
interest. I will outline the application of this approach to climate
models; this work is useful in making projections of the North Atlantic
Meridional Overturning Circulation (AMOC). AMOC projections are of
considerable interest since changes in the AMOC may result in major
impacts on global climate. Time permitting, I will also briefly discuss
a different Gaussian process-based approach used in the context of
infectious disease modeling.