- Data Portals
- User Support
- About Us
University College London
Approximate Bayesian Computation (ABC) is a family of algorithms used to perform Bayesian inference. ABC, being a likelihood-free approach, is used to infer parameters when it is difficult or very expensive to evaluate the likelihood. Formally, the posterior distribution is approximated by simulating data sets for a given set of parameters. The parameter values are accepted if data simulated (from these parameters) are close to the observed data. If the data are high-dimensional, summary statistics are often used to make this evaluation.
LatticeKrig is a computationally efficient methodology for spatial inference and prediction. Its multi-resolution feature captures spatial dependence at various levels (local and global) and the use of sparse approximations to the precision matrix ensures computational efficiency. For parameter inference, it uses the profile maximum likelihood estimation (MLE) method.
In this talk, I’ll discuss the use of ABC as an alternative inference methodology for LatticeKrig. I will show parameter inference results comparing ABC and profile MLE using simulated data where the true covariance parameters are known.
(Bring your lunch)