SCCS Colloquium - Mar 28, 2019
|Date:||Mar 28, 2019|
|Time:||15:00 - 16:30|
Frank Schraufstetter: Development of a prototype to quantify the uncertainty of the water balance model LARSIM
Hydrological models are prone to uncertainty due to their complexity. The water balance model LARSIM provides a high degree of control through configuration parameters, which have to be calibrated for each application. The uncertainty originating from parameter calibration was evaluated. A prototype was developed to facilitate the propagation of uncertainty through the LARSIM model by using stochastic collocation with pseudo-spectral approach. The snow module parameters were exemplarily used for uncertainty propagation. The output of interest was evaluated using global sensitivity analysis and statistical measurements to determine the impact of uncertain parameters on the simulated runoff. Uncertainty of snow module parameters could not be proven definitively, but the large deviation of simulated and measured runoff indicate uncertainty in the model.
Keywords: Uncertainty Quantification, LARSIM
Nikolaos Ioannis Bountos: Combining the Combigrid method with the SG++ data-mining pipeline
This is a Guided Research talk. Nikolaos is advised by Kilian Röhner.
In this guided research we try to extend the data mining pipeline of SG++ to support solving PDEs using the combination technique. The main idea is to solve a PDE in small independent components and then combine the results. We expect this to majorly improve the performance of the current probability density estimation technique in the SG++ project, as well as create further opportunities for efficiency boosting via parallelization.
Subhan-Jamal Sohail: Embedded SGDE image classification on mobile devices
This is a Bachelor's Thesis submission talk. Subhan-Jamal is advised by Kilian Röhner.
This thesis describes the integration of the geometry aware sparse grids into thedatamining pipeline of the SG++ framework. The datamining pipeline is used tomake SG++ easier available and fasten the process of using the framework. Geometryaware sparse grids have a great use in image classification and therefore this thesis isadditionally about creating a mobile application that classifies hand drawn numbersusing these kind of grids. For this to be done the trained data had to be exported fromSG++ and then imported again into the application. Also the evaluation method forsparse grids was implemented into the application. This new evaluation method isthen validated by comparing it to the evaluation method of SG++. The outcome of thevalidation resulted in a high deviation for random datasets and very low deviation forrelevant datasets. The high deviation for random datasets can be neglected, since it hasno use case in the user application. Finally the application was tested for numbers 0, 2and 6 with various relevant datasets to calculate the accuracy of the implementation.The tests have shown that by following certain rules while drawing the numbers intothe application an average accuracy of 90% could be observed.