SC²S Colloquium - September 11, 2014

From Sccswiki
Jump to navigation Jump to search
Date: September 11, 2014
Room: 02.07.023
Time: 1 pm, s.t.

Carsten Uphoff: Fitting generalized additive models with BiCGStab

Additive models are a well established tool in statistics. They extend standard linear regression to include non-parametric models, yet avoid the "curse of dimensionality". In 1989 Buja et al. introduced the Backfitting algorithm to fit such models. Although it converges fast, it is not suited for a parallel implementation. Hence, Khakhutskyy and Hegland proposed using the BiCGStab method. It converges fast and is well suited for a parallel implementation. In this talk, we derive additive models, show a method to stabilize the iterations and present the (failed) attempt to build a preconditioner.