The week April 22–25, 2014 will be a statistics week. Over these days there will be four statistics events.

All the talks will be in room V-157 in VR-II on the UI campus.

- April 22 at 14:00.
**Emtiyaz Khan** from École Polytechnique Fédérale de Lausanne: *Machine Learning*
- April 23 at 12:00.
**R-Iceland**: *Opening meeting of R-Iceland, a community of R-users in Iceland*
- April 23 at 13:00.
**Andrew Gelman** from Columbia University: *The statistics package Stan*
- April 25 at 12:00.
**Håvard Rue** from Norwegian University of Science and Technology: *Penalising model component complexity: A principled practical approach to constructing priors*

Continue reading 'Statistics Week April 22-25'»

Statistics Colloquium

### Speaker: Philippe Crochet, Research Specialist at the Icelandic Meteorological Office

Title: Estimating the flood frequency distribution at ungauged catchments using a regional flood frequency analysis

Location: V-157, VRII

Time: Thursday, April 10th, 12:00-13:00.

### Abstract:

Extreme events such as floods may have serious societal and environmental consequences. It is therefore necessary to develop adequate models for the estimation of such events. Often, this information is required at sites where measured streamflow series are either too short to allow a robust estimation of extreme flood quantiles, or where no data is available at all. We explore the possibility of conducting a regional flood frequency analysis (RFFA) to reliably predict flood quantiles at both gauged and ungauged sites. The principle of the method is to transfer information from gauged sites to a target site (gauged or ungauged) within a homogeneous region. The principle of the method is first presented and then tested in Northern Iceland, using available streamflow observations and then streamflow simulations from a distributed hydrological model.

Statistics colloquium talk

### Speaker: Guðmundur Einarsson, M.S. student of Statistics, University of Iceland

### Title: Discussion on evolutionary algorithms in optimization

Location: Room V-157 in building VR-II on the UI campus

Time: Thursday, April 3rd 2014, at 12:00 to 13:00.

#### Abstract:

Two papers will be discussed. First, a paper entitled “Metaheuristics-the metaphor exposed” by K. Sörensen. Summary of paper:

In recent years, the field of combinatorial optimization has witnessed a true tsunami of “novel” metaheuristic methods, most of them based on a metaphor of some natural or man-made process. The behavior of virtually any species of insects, the flow of water, musicians playing together — it seems that no idea is too far-fetched to serve as inspiration to launch yet another metaheuristic. In this paper we will argue that this line of research is threatening to lead the area of metaheuristics away from scientific rigour. We will examine the historical context that gave rise to the increasing use of metaphors as inspiration and justification for the development of new methods, discuss the reasons for the vulnerability of the metaheuristics field to this line of research, and point out its fallacies. At the same time, truly innovative research of high quality is being performed as well. We conclude the paper by discussing some of the properties of this research and by pointing out some of the most promising research avenues for the field of metaheuristics.

Secondly, a paper entitled “On the Hunt: Competitive Coevolution as a Metaheuristic” by Guðmundur Einarsson, Tómas Philip Rúnarsson and Gunnar Stefánsson. Summary of paper:

A predator prey model for competitive coevolution is used as a metaheuristic. Two scenarios are created. In each scenario the predator genetic setup consists of parameters for optimization procedures. The prey genetic setup consists of starting points for the optimization of a specific function. A sampling method is explored for the relative fitness assessment where the number of fitness evaluations can be controlled similar to K-fold relative fitness assessment. The historical evolution of the prey is explored as a diagnostics tool for multimodality.

Málstofa í stærðfræði

### Speaker: Óli Páll Geirsson

### Title: Weak solution to stochastic partial differential equations and applications in spatial statistics

Location: Room **V-147** in building VR-II on the UI campus

Time: **Friday** March 21, 2014, at **12:00 to 13:00**.

#### Abstract:

The literature on spatial statistics is rich and recognized, where Gaussian fields play a dominant role in statistical modeling. Gaussian fields (GFs) are both practical and readily interpretable due to the flexible structure of the parameterization of the GF. One of the most commonly used covariance model in the Gaussian field class is the Matérn covariance model which has gained recognition within statistical climatology in recent years.

Although Gaussian fields suit well from both analytic and practical point of view, they become computationally demanding as data sets get larger as the covariance matrices tend to be fully populated. Gaussian fields can be approximated with Gaussian Markov random fields (GMRF), which increases the speed of computation significantly. Even though GMRFs have very good computational properties, using them for involved spatial models has not been feasible as there has been no good way to parametrize the precision matrix of a GMRF to achieve a predefined behaviour in terms of correlation between two sites and to control marginal variances.

In recent work it has been shown that using an approximate stochastic weak solution to (linear) stochastic partial differential equations (SPDEs), it is possible to provide an explicit link, for any triangulation over a spatial domain of interest, between GFs and GMRFs. The consequence is that we can take the best from the both worlds and do the modelling using GFs but do the computations using GMRFs. The approximate is then used to construct a GMRF representation of the desired GF on the triangulated mesh. This allows for continuous spatial predictions by choosing appropriate basis function for the approximation.

#### Note that the time, date and location is different from the usual one.

Málstofa í tölfræði

### Speaker: Óli Páll Geirsson, Doctoral student, Department of Mathematics, Faculty of Physical Sciences, University of Iceland.

### Title: Modelling annual maximum 24 hour precipitation in Iceland using the SPDE approach and split sampling MCMC strategy

Location: Room V-157 in building VR-II on the UI campus

Time: Thursday, March 20th 2014, at 12:00 to 13:00.

Abstract: To obtain predictions of distributional properties of extreme precipitation on a high resolution grid both observations and outputs from climate models are relied upon. A method for quantile predictions of extreme precipitation which combines these two sources is developed. The method is based on observations of annual maximum 24 hour precipitation from 86 observation sites in Iceland and outputs from a climate model on a 1 km by 1 km regular grid.

A covariate based on the climate model which represents the intensity of precipitation can be computed at each grid point. A latent Gaussian model is built for the observations with the intensity measure as a covariate which in turn allows for spatial predictions on the high resolution grid. The observations are assumed to follow the generalized extreme value distribution. In order to make fast continuous spatial predictions, SPDE spatial models are implemented at the latent level of the model for both the location and scale parameters of the proposed data distribution.

A novel MCMC strategy called split sampling, is implemented for model inference. The strategy greatly improves the mixing and convergence properties of the MCMC sampler. Inferring the parameters of the model yields posterior estimates for the distribution of maximum annual 24 hour precipitation in every grid point. In particular, a quantile estimate can be given at every grid point.

Málstofa í tölfræði

### Speaker: Daniel Simpson, Researcher, Department of Mathematical Sciences, Norwegian University of Science and Technology

### Title: Towards computationally efficient spatial statistics

Location: Room V-157 in building VR-II on the UI campus

Time: Thursday January 23rd 2014, at 12:00 to 13:00.

Abstract: Spatial statistics is hard! The combination of large datasets and complicated models leads to a large number of computational and modelling challenges. In this talk, I will outline a framework for reducing this complexity through a combination of approximate models and approximate inference.