Thesis supervisor: Balázs Renczes
Location of studies (in Hungarian): Department of Measurement and Information Systems Abbreviation of location of studies: MIT
Description of the research topic:
"Frequency domain system identification is an efficient tool to gain insight into the functionality of a dynamic system. If the system is excited with an appropriate signal, by measuring its output, the modeling of its behavior becomes possible.
The problem is widely researched and published in the case when the dynamic system is linear. In this case, the only question is whether the linear system identification framework is applied appropriately to obtain system parameters. Contrarily, if the system is non-linear, the problem becomes much more complex, involving several unknown challenges in the previous case.
In non-linear system identification, the models can be divided into categories from white-box models to black-box models. In the former one, detailed physical knowledge is needed for the system, which is usually costly to gain. Black-box models do not require prior knowledge of the system, which is a great advantage. The drawback of this modeling is that the resulting model will contain a large number of parameters, and therefore the result will be hard to interpret.
The proposed research belongs to improving this interpretability. One possible way is to apply a tensor decomposition. By this means, involved multivariate polynomials are described as a linear combination of univariate polynomials. This functionality is similar to the behavior of neural networks with one hidden layer. Besides, it is also essential to investigate whether all braches (a univariate polynomial belongs to a branch) are needed to describe the non-linear behavior. If not and the number of branches can be decreased significantly, the explanatory potential of the obtained system can be much higher than that of the original black-box model.
Another exciting and not expansively researched area is the implementation of deep learning methods and increasing the robustness of non-linear system models by reinforcement learning methods. Along with conventional parameter estimation techniques, the application of neural networks may lead to the improvement of currently existing processes.
"
Required language skills: English Number of students who can be accepted: 1
Deadline for application: 2021-09-01
2024. IV. 17. ODT ülés Az ODT következő ülésére 2024. június 14-én, pénteken 10.00 órakor kerül sor a Semmelweis Egyetem Szenátusi termében (Bp. Üllői út 26. I. emelet).
Login
