Activity Number:
|
172
- Machine Learning and Algorithms
|
Type:
|
Contributed
|
Date/Time:
|
Monday, July 31, 2017 : 10:30 AM to 12:20 PM
|
Sponsor:
|
Section on Statistical Computing
|
Abstract #323227
|
View Presentation
|
Title:
|
Matrix-Free Computation of Spatial-Temporal Gaussian Autoregressions and Related Stat-Space Models
|
Author(s):
|
Chunxiao Wang* and Debashis Mondal
|
Companies:
|
Oregon State University and Oregon State University
|
Keywords:
|
H-likelihood ;
Kalman filter ;
Lanczos algorithm ;
Matrix-free computation ;
State space model
|
Abstract:
|
We combines Besag's spatial-temporal Gaussian autoregressions within a state-space model and presents a novel matrix-free h-likelihood method for statistical inference. The inference proposed here is shown to have significant computational advantages compared with traditional Kalman filter method. It includes a novel matrix-free scalable Lanczos algorithm that the best linear unbiased predictors by solving Henderson's mixed equations and a novel matrix-free trust region method that finds the solution of the REML score equations. Furthermore, we pays careful attention to computations in small time steps and indicates how the h-likelihood method can be adapted for statistical inference of other space-time models, such as stochastic injection-diffusions, blurred generated models and other spatial-temporal processes based on stochastic partial differential equations. The method is illustrated through a simulation study and soil moisture data application.
|
Authors who are presenting talks have a * after their name.