SCCS Colloquium - Sep 18, 2019
|Date:||September 18, 2019|
|Time:||15:00 - 16:00|
Julia Konrad: Multifidelity Monte Carlo Sampling in Plasma Microturbulence Analysis
Multifidelity Monte Carlo sampling aims to increase the accuracy of standard Monte Carlo estimators for statistical moments of model output based on uncertain input parameters. Standard Monte Carlo sampling estimates statistics by evaluating the underlying model for a number of samples and computing estimators from the results. In order to obtain acceptable precision, a prohibitive amount of samples is often required and simulations become computationally infeasible. To overcome this shortcoming, multifidelity Monte Carlo sampling distributes the model evaluations between the model under consideration and one or more computationally less expensive surrogate models. In this thesis, multifidelity Monte Carlo sampling was applied to test cases from plasma microturbulence analysis and the reduction of the achieved estimator‘s mean squared error was studied.
Keywords: Multifidelity methods, Monte Carlo sampling, plasma microturbulence analysis
Vincent Bautista Anguiano: Visualization of High Dimensional Models within the SG++ Data Mining Pipeline
This is a Guided Research project talk. Vincent is advised by Kilian Röhner.
The Spatially Adaptive Sparse Grid Toolbox SG++ is an under development open source library mainly used to tackle a variety of high dimensional problems using the sparse grid numerical technique. This open source library is comprised of several modules, each one developed specifically to solve a specific kind of problem like Partial Differential Equation Solvers, Function Interpolation, Uncertainty Quantification etc.
From all of these modules, a specific focus will be brought to the ”datadriven” module, specially to one of its components, the ”SG++ Datamining Pipeline”. This is the one in charge of generating high dimensional machine learning based models like predictive models, regression models, density estimation models etc. Unfortunately, it has until now no way of visualizing these high dimensional models in a user friendly manner. It is due to this, that the focus of this Guided Research will be to develop a new component within the SG++ Datamining Pipeline, which enables the end user to visualize in an intuitive and friendly manner the results of the high dimensional model obtained by the pipeline within a greatly reduced dimensional space.
In order to achieve that, research will be conducted on dimensionality reduction techniques applied to high dimensional data and data visualization. Based upon this research, the module will be developed and integrated into the pipeline. This visualization module must be implemented in such a way that it fulfills the following requirements:
• The output of the module must be so that it can serve as an input for a variety of graphic library packages, like mathplotlib.
• The output must be saved in a configurable file, which enables the user to customize its visualization.
• The output of the module must be independent of the model generated by any other component of the pipeline.
• The module must do its processing efficiently without blocking the other processes of the SG++ Datamining pipeline.
• The module must be able to handle a significantly large amount of data points within a high dimensional space.
Keywords: Visualization, High Dimensional Models, Sparse Grids