Artificial Intelligence is one of the most active current fields of research. By combining approaches from modeling, simulation, optimization, and statistical learning, machines are enabled to learn complex relationships from data by finding underlying structures and to derive decisions for the process under investigation. There exist various fields of applications, e.g., connected to manufacturing, logistics, finance, robotics, and business process management.
Within the field of Artificial Intelligence, we develop, analyze, and investigate algorithms for solutions to data-driven problems. These tasks require expertise in the areas of modeling, simulation, and optimization. Although simulation and optimization are similar and share a lot of computational techniques and algorithms, they are used for different scenarios. Simulations are considered to be more exploratory and are helpful to identify interactions in complex systems. Optimization methods support both tactical and strategic planning decisions because they provide a single “best” answer to a given problem.
One focus area within the research center is the combination of Evolutionary Algorithms and Machine Learning. Such combinations allow us to optimally design Machine Learning algorithms and to optimize systems and processes that are observed by data collection procedures. On the application side, problems in the area of Predictive Maintenance and Finance, text analysis with Natural Language Processing, and handling of missing or erroneous data are investigated. In the future, we intend to strengthen our expertise in the area of Neural Networks, especially how the decisions derived can be better explained for human decision-makers. Within the research, we apply concepts from High-Performance Computing to scale the solutions and to achieve efficient run times. Next to cloud and on-premise server resources, self-developed systems like the Distributed Execution Framework for parallel and distributed computation are in use.
For a computer algorithm to understand a real-life problem, a model needs to be created. A model of a real-world system approximates the underlying relations and interactions between the system input, the system output, and the system's inner processes.
For many complex systems, data on the system's input and output are used to infer the respective relations and the model building process is performed by machines. Accurate models are the basis for simulations, optimizations, and provide essential support for managerial or technical decision making.
Machine Learning is the process of learning a model of a complex system under investigation. The learning process relies on the observed data and aims at predicting certain outcomes or identifying internal structures within the collected data.
One particular application area is Predictive Maintenance via estimation of the Remaining Useful Life (RUL) for a machine. RUL refers to the remaining time or number of cycles a machine can operate before it requires repair or replacement. By taking RUL into account, engineers can schedule maintenance, optimize operating efficiency, and avoid unplanned downtime. Estimating RUL is a top priority in predictive maintenance programs. In joint work with our project partners, we are developing RUL solutions based on state-of-the-art Machine Learning algorithms that are especially aiming at problems coming with sparse data.
Data generated from conversations, declarations, or process descriptions are examples of unstructured data. Such data can be messy and hard to process. Natural Language Processing (NLP) is a field of Artificial Intelligence that provides machines the ability to read, understand, and derive meaning from human languages. In particular, we investigate NLP-related strategies to validate and classify textual data concerning the origin and the level of product risks.
Through simulations, analysts can take a model of a complex system as a basis for analyzing the system's behavior. Simulations are state-of-the-art in many different disciplines like meteorology, sociology, biology, physics, engineering, and business.
Different analysis use cases are possible, from observing changes in the system output by varying the respective input to using simulations as a basis for optimization approaches.
Another application of a computer's capability of solving problems associated with complex systems is optimization. Next to the model for the complex system, an objective function that models the question of interest is required.
The aim is to maximize or minimize the objective function, typically involving some further constraints on the solution. Depending on the structures of the model and the objective function algorithms ranging from classical linear optimization to specialized procedures from the field of Computational Intelligence are applied.
Evolutionary Algorithms are a special class of optimization procedures that can be applied to problems where no analytical description of the model exists. By using concepts from the natural evolution and behavior of animals, iterative procedures are designed that aim at finding improved model outputs.
Evolutionary Algorithms are scalable and can be applied to a large variety of problems due to their flexible representations. We have strong expertise in Evolution Strategies, an optimization algorithm for real-valued problems. Our focus is on the analysis and the development of such strategies. Besides basic research in this field, our group is working on adopting Evolutionary Algorithms to real-world problems. Examples are modeling and prediction of economic risks, the integration of optimization strategies into particular automation processes, as well as the evolutionary algorithm-supported improvement of Machine Learning techniques. In several projects, we worked on theoretical analyses, developed numerical benchmarks, and performed extensive empirical investigations for constrained problems and noisy problems.