Sparse Gaussian Process

Prof. Dr. Bernard Haasdonk, Holger Ulmer

M.Sc. Roman Föll

01/04/2016

01/04/2019

The focus in this research project is on the topic of improving existing methods of modelling dynamical systems from pure data, i.e. system identification. The problem to solve is that we just have raw data from sensors or other measurements which are noisy because of the difficulty to measure or just because of some missing knowledge. Gaussian processes are a well-known tool to this modelling task because they not only provide a good approximation, which has proven not to suffer from overfitting, but also provide noise prediction, Rasmussen (2006). They inherently make use of kernels, which are well known functions in machine learning. The challenge in real world application is often the huge amount of data, noisy data, as mentioned before and huge input dimensions for the modelling task,
which makes it difficult to approximate and which lead to a huge computational cost for the approximation, especially the optimization of the model parameters, and the prediction.

We want to investigate three main topics, 1. Sparse Gaussian Processes, 2. Dynamic modeling and 3. Data assimilation.

In practice there is a frequenting need for realizing the modelling task and the predictions very fast, especially in online applications for online simulations/predictions, e.g. virtual sensors. Therefore we want to focus in the project improving existing and inventing new Sparse Gaussian Process methods (O(nm2 ) , m <<n, operational cost instead of O (n3) ) for dynamical modelling and trying to combine it with model reduction to tackle these issues.

The driving application in this research project is engine control and dynamical modeling of an engine control unit in cooperation with ETAS GmbH.

External P.h.D funded by ETAS GmbH