Skip to content

Professor Simon WoodBSc(Manch.), PhD(Strath.), RSS Grad. Dipl.

Professor of Statistical Science

Simon Wood

Professor Simon WoodBSc(Manch.), PhD(Strath.), RSS Grad. Dipl.

Professor of Statistical Science

Member of

Research interests

I am interested in modern regression modelling, especially using smooth functions and random effects, and in statistics applied in ecology, especially to ecological dynamics. I am author of the R recommended package 'mgcv' for generalized additive models. Particular current interests are in spatio-temporal modelling, sparse methods, scalable statistical computing for big models and data, and model selection issues. Recent applications have been in air-pollution modelling and electricity demand prediction.

PhD projects

* I am interested in advising PhD students in a range of projects fitting in with the above areas. Some  examples are given below. 

* General smooth models for spatial ecological capture recapture data (with university of St Andrews). Recent advances in methods for camera trap networks and other capture re-capture networks have opened the possibility for spatial modelling from capture recapture data, in order to better understand the spatial distribution of hard to observe animals. This project will use recent advances in smooth modelling methods with general regular likelihoods in conjunction with the new capture recapture modelling framework.

* Faster parallel computation of large smooth regression models. Consider a smooth regression model with 10000 coefficients fitted to 10 million data. For a subset of such models new methods based on discretization and parallelization can estimate such models in less than an hour on a desktop workstation. Can the methods be applied to a wider class of models, and what about 100,000 coefficients and 1 billion data?

* Sparse parallel methods for smooth regression. Many large smooth regression problems can best be addressed by combining random fields represented using sparse matrix methods, with smooth functions represented using reduced rank approaximations. To make these methods scalable to very large models and datasets requires that they be extended to exploit parallel processing. This should be possible by combining recent developments in sparse parallel matrix decompositions with the development of new parallel sparse codes for tasks specific to these statsitcal regression models.

 

 

View research connections

Postal address:
University Walk
Clifton
Bristol
United Kingdom