When: Thursday 15th February 2018, 13:00-14:00.
Included in meeting:
The fortnightly BRAG meeting will be held tomorrow 15/02/18 at 1pm in room Y801. This week we will have presentations by Alan and Brodie.
Alan’s Talk: Exploiting the natural structure of spatial datasets to cut down on processing time: an example from the Great Barrier Reef
Abstract: The mantra of “working smarter, not harder” is reminiscent of the bad advice you might have received from your high school careers counsellor. However, when working with large spatial datasets, it pays to take the saying to heart. In this talk, I present the example of a geoprocessing problem I faced with data on the Great Barrier Reef. I had to delineate reef passages between all pairs of neighbouring reefs on the edge of the continental shelf, and find the average depth and slope of each of these passages. I present my methodology for exploiting the natural structure of the reef and the passages to drastically reduce the time required for each operation in R. Though the example is highly specialised, I hope it will help you to develop your own ideas for exploiting the natural spatial structure of otherwise prohibitively large datasets that you may be faced with.
Brodie’s Talk: Treed Gaussian Processes: Supervised and Unsupervised Approaches
Abstract: Gaussian processes are a powerful regression and emulation technique, but can be strongly limited by their underlying assumption of smoothness. Because discontinuities in the data typically arise due to sudden shifts in model behaviour, one approach is to use separate Gaussian process surfaces for each distinct behavioural regime. In my previous talk I introduced the idea of using a priori identification of distinct model behaviours to allow for a supervised approach to this problem. In this talk, I will share my experiences with how this approach performed, including the performance of different machine learning techniques to the supervised learning problem. Then I will compare those results to the unsupervised Bayesian approach that uses no a priori information, instead combining regression trees with reverse-jump MCMC to learn where behavioural shifts appear to occur. The latter is implemented by R’s package ‘tgp’, which I will discuss.
We look forward to seeing you all there!
Farzana and Trish