|
Abstract:
|
The availability of georeferenced data sets on large scales has generated interest on sophisticated models that can deal with different sources of information, account for space and time dynamics, include expert opinion, and incorporate the uncertainty induced by observational errors. Bayesian hierarchical models are ideally suited to deal with these goals, as they provide a coherent paradigm to propagate the different sources of uncertainty in a probabilistic framework. For this purpose it is necessary to develop methods that allow for the exploration of highly multivariate and complex probability distributions. In this talk we present an approach to perform Markov chain Monte Carlo simulations that leverages the computing power of multi-processors computing platforms. The method is focused on problems involving long time series of spatial data. It is based on slicing the time domain among the different computing nodes of a parallel machine, and developing efficient schemes of information exchange between nodes, in order to reduce the idle times of each of the processors. The method is illustrated with a model developed to obtain very high resolution predictions of sea surface tem
|
