Final theses

The efficient searching and identification of links in specialized document databases is essential in order to make relevant information quickly accessible to users. This thesis compares different strategies for the development of a RAG (retrieval-augmented generation) agent system which should make this possible. Among other things, the exact composition of the system of generalist and specialist agents and possibilities to effectively generate keywords are to be examined and evaluated with regard to their usefulness. A further aim is to use the system to link relationships between different specialist concepts across different documents in order to provide specialist users with targeted support for their queries.

In chemical batch processes, a chemical product is produced by continuously combining starting materials. The chemical reactions are dependent on unknown environmental influences that cannot be controlled in the laboratory. Despite these uncertainties in the manufacturing process, the aim is to achieve an optimum result, e.g. low number of by-products, avoidance of overheating, minimum time required, etc., in order to ensure the economic efficiency of production.

The aim of this master’s thesis is to use modern machine learning methods to make predictions during the running process about different metrics of the further course of the process (e.g. quality, time requirement, temperature, etc.). The resulting prediction models then form the basis of an agent-based control system that allows the production process to be optimally controlled.

In many industrial processes, time-delayed and variational dependencies, such as those caused by different flow rates or intermediate storage of materials along the production line, lead to challenges in data analysis and predictive modeling. Although theoretical studies indicate that such variations can negatively influence the prediction performance, the question remains whether the prediction performance is sufficient under real conditions. This thesis aims to close this knowledge gap by validating the theoretical findings through practical investigations using real sensor data along the production line. Suitable ML algorithms are to be sought, implemented and their performance evaluated in order to effectively meet the challenges of time-delayed dependencies.

Synthetic data are artificially generated imitations of sensitive data and allow users to train AI and ML models without accessing real, in-house data. In recent years, research into generating this data has become very popular. However, it is unclear to what extent companies in Germany are utilizing the potential of synthetic data and what the challenges are that stand in the way of a broad application of this technology. With this thesis we want to contribute to closing this gap between research and industry by discussing the requirements that companies have for synthetic data.

Tender documents contain key requirements for products and services, but are usually available in unstructured text form and require a great deal of manual analysis. The aim of this work is to develop and evaluate an automated approach for extracting requirements from such documents. For this purpose, methods of natural language processing (NLP) with a focus on large language models are to be examined, implemented and compared with each other. The quality of the extracted requirements will be evaluated using suitable metrics and in comparison to manual reference analyses.

Predicting bed requirements is a key challenge in the healthcare sector, as it enables efficient resource planning. In addition to classical machine learning methods, quantum machine learning approaches are gaining increasing attention. The aim of this thesis is to implement a quantum ML approach for predicting bed requirements and to systematically compare its performance with classical ML methods. Prediction accuracy as well as computational effort and scalability will be considered.

Data warehouse/data warehouse platforms form the backbone of modern data-driven organizations. In addition to established closed-source solutions, there are increasingly powerful open-source alternatives. The aim of this thesis is a systematic comparison of selected closed-source and open-source data warehouse platforms in terms of performance, scalability, costs, maintainability and range of functions. The evaluation is based on defined realistic benchmarks and realistic application scenarios, including the implementation of such a new structure.

Data synthesis plays a major role in maintaining data privacy and generating realistic test data. While many methods are based on single tables, the synthesis of relational multi-table data poses particular challenges (foreign keys, n:m relationships, other constraints). The aim of this work is a comprehensive research of current state-of-the-art approaches for multi-table data synthesis and their systematic benchmarking. The methods are to be evaluated in terms of data quality, consistency between tables and practical applicability.

In a preliminary work, ball and player positions were extracted rule-based (or with blob detection) from a video stream during table soccer. Learning-based approaches are now mostly used in image processing methods. The aim of this work is to compare classic rule-based methods with modern machine learning and deep learning methods for real-time recognition. Recognition accuracy, robustness and real-time capability under real conditions (in particular changing light and camera conditions) are to be investigated.

Automated sports commentary is a growing field of application for natural language generation. In a preliminary project, simple commentaries were generated in real time for a table football game. In addition to content accuracy, creativity and linguistic variance play a decisive role in user acceptance. The aim of this work is to investigate different approaches to automatic sports commentary in the context of table football with regard to their creativity and variance. The evaluation is based on both quantitative metrics and qualitative analyses.

Background knowledge, such as game strategies, player history or game situations, can significantly improve the quality of automated sports commentary. In a preliminary work, simple commentaries were generated in real time for a table soccer game. The aim of this work is to investigate how such background knowledge can be integrated into systems for automated sports commentary. Using foosball as an example, different knowledge representations and integration strategies are to be developed and evaluated in order to analyze their influence on the comprehensibility, information content and naturalness of the comments.

The reconstruction of three-dimensional radiation clouds from exposure data along paths is an exciting problem, which was investigated in cooperation with the Federal Office for Radiation Protection. In existing preliminary work, an approach based on cuboid splatting was used for this purpose, which uses the sum of cuboids (boxes, so to speak) to approximate the radiation distribution. The aim of this master’s thesis is to further develop this approach and to replace or supplement it with density-based methods, in particular models based on Gaussian density functions. To this end, suitable density-based reconstruction methods are to be researched, implemented and compared with the existing cuboid-splatting approach in terms of reconstruction quality, robustness and computational effort.