JR Centre for Symbolic Regression

In this JR Centre, new algorithms for modelling technical systems and methods for monitoring model quality are being developed. The performance of the algorithms is demonstrated using drive components.

Symbolic regression is a special type of regression analysis. The functional relationship between dependent variables is determined and a prediction model is developed from this. In this specific case, models are generated as mathematical symbolic formulae. The model structure and the model parameters are simultaneously adapted to the given data. In the context of machine learning, the models are trained purely empirically. However, prior knowledge about the context to be modelled can also be taken into account in order to create semi-empirical models.

The ability to find relatively simple mathematical formulas for an unknown context and to incorporate prior knowledge of physics is particularly interesting for modelling technical systems. At the JR Centre, the modelling of drive components and friction systems is considered as an example. The aim is to further develop and adapt methods for symbolic regression in such a way that they can be used as a standard tool for modelling in industrial applications.

Algorithms currently used for symbolic regression are stochastic, which means that their results cannot necessarily be reproduced when the same process is repeated. They are also significantly slower than standard methods for regression analysis and expert knowledge of the special solution methods is required. This JR Centre will therefore develop efficient deterministic algorithms for symbolic regression that generate robust and accurate models, enabling users to generate and validate symbolic regression models within a few hours.

For industrial applications, it is necessary to adapt models to constantly changing conditions and systems and their prediction accuracy decreases over time (this is called "concept drift"). At present, models are usually generated from scratch when the prediction accuracy decreases. In contrast, the JR Centre will develop a methodological and technical framework for monitoring model quality, detecting concept drift and managing data and models. To this end, adaptable solution methods for symbolic regression will be developed that reuse existing models in order to improve the efficiency of the training phase.

The algorithms and frameworks developed will be used for modelling powertrains and friction systems to demonstrate the performance of the algorithms and support the development of powertrains and friction systems.