Faculty of Computer Science

Research Group Theoretical Computer Science

Oberseminar: Heterogene formale Methoden

Abstract:

This study explores the potentiality of simplifying the determination of machine control parameters by examining the use of systems of equations and neural networks in the context of calendering - a process for refining materials by rolling. One focus of the study is on the optimal compression of electrodes, a process that is controlled by a variety of parameters. Neural networks have proven powerful in various fields in recent years, outperforming existing methods and opening up entirely new application areas. For example, they have made significant progress in image processing and speech recognition, with developments such as GPT-3. However, their high complexity and large number of parameters, which make them difficult to track and interpret, is a significant drawback. As an alternative, this study considers genetic programming approaches, in particular the use of GPLearn, a technique that allows the solution of tasks such as regression by evolving systems of equations over many randomly controlled iterations. These approaches produce easily understood, though potentially large, systems of equations that allow clear interpretation of results. Despite their advantages, systems of equations have their limitations, especially when dealing with complex data. They tend to become imprecise and require significant computational resources, especially when different training parameters need to be tested multiple times. In the context of machine control, particularly in electrode calendering, the creation of comprehensible systems through the use of systems of equations is considered essential to better understand errors and effectively use simulations for future applications. The challenge is to select and measure the appropriate parameters to create realistic data sets that can form the basis for developing systems of equations. In this thesis, the focus is to explore approaches that use a system of equations as the output, with vectorized data and possibly a starting system of equations as the input, in order to navigate and potentially simplify the complexity of parameter settings in machine control systems. It is hypothesized that the comparative approach of neural networks and genetic programming can provide an optimized solution to machine control challenges.

Kommission:

Vorsitzender: Prof. Dr. Till Mossakowski
1. Gutachter: Prof. Dr. Till Mossakowski
2. Gutachter: Prof. Dr. Andreas Nürnberger
Beisitzer: Dr. Fabian Neuhaus