Data Data-Driven Identification of Stable Non-Linear Systems Using Long Short-Term Memory

This work was a research seminar I completed during my degree program Computer & Systems Engineering (M.Sc.). It was supervised by Irene Schimperna. Responsible Professors were Prof. Dr.-Ing. Patrick Mäder (Data-intensive Systems and Visualization Group) and Prof. Dr.-Ing. Karl Worthmann (Optimization-based Control Group).

In the seminar, I investigated how to formally check an LSTM network for stability (more precisely: input-to-state stability) and how to enforce this property during the training process.

Figure 1: Schematic of an LSTM network. Adapted with changes from https://colah.github.io/

The results enable for guarantees regarding stability and boundedness and can be applied to e.g. controller design.

For further details please refer to the seminar report or the presentation slides can be checked.