Research Branch: Artificial Intelligence

Machine learning involves algorithmic learning and the generalization of existing experience, based on data or through interaction with the environment. It includes supervised learning, unsupervised learning, and reinforcement learning. While supervised learning involves the learning of connections, for example for a regression analysis, from labeled data, unsupervised learning involves the learning of associations from data without any knowlegde of labels. The field of reinforcement learning, deals with an artificial agent that can acquire a complex goal-directed action strategy by interacting with its environment. Concrete research topics at the department are the learning of features out of labelt images by means of artificial neural networks which are used for image classification for example. The objective here is not only to achieve an outstanding classification rate, but also to improve the interpretability of what the neural network learns. With the increasing number of systems from the area of machine learning in our everydays life, the question of trustworthiness and secure use of such systems also arises. This is especially important in security-relevant application areas and is also being researched at the department.

As a concrete method of machine learning, the research of the department also deals with artificial neural networks. Inspired by biological neurons, artificial neural networks are able to independently learn correlations or features from data by optimizing the internal parameters and to include them in classification tasks. They are increasingly used in everyday life and in industry, for example in camera systems of automobiles, speech recognition of smartphones or in modern logistics warehouses with autonomous robots. In some cases, the learned capabilities of neural networks now already surpass those of humans. Open research areas are investigating, for example, the generalization capabilities of the networks. Although neural networks are often able to achieve very good results on data that correspond to the distribution of the optimization data, their performance is still often significantly reduced on strongly deviating data distributions. The practical applicability of artificial neural networks is also being researched at the department. In the automotive field, for example, the classification of audio recordings from road traffic and a fast evaluation of planned trajectories play a major role.

Game theory is used to model and solve optimization problems in multi-agent systems in which the agents' objective functions are coupled via their decision variables.

The applications of the game theoretical approach can be found, for example, in electricity markets, smart grids, robot swarms, and communication networks. The so-called Nash equilibria in games are the solutions for the corresponding optimization problems.

In order to efficiently solve game theoretic problems in modern complex multi-agent systems, distributed algorithms are developed to ensure the convergence of collective behavior to a solution. Attention must be paid to the fact that the rules of such algorithms are based only on the local information about the system available to each agent.

Optimization is a branch of applied mathematics. The goal of an optimization process is to find a configuration of variables so that the value of a certain objective function, which depends on the variables to be found, is minimized or maximized. Often, constraints are formulated for an optimization problem that limit the set of allowed solutions. To be solved optimization problems can be found in a variety of applications in engineering and are of great importance for industrial processes as well as in research.

Lectures in the department teach, among other things, the design of optimal controllers in aspect of time or control deviations and show how strategies for solving multimodal optimization problems can be derived from evolutionary processes. Optimization also plays a special role in the research projects of the rmr. Different optimization methods are used in the fields of robotics (optimal controllers, optimal trajectories), energy systems (model predictive control of microgrids, distributed solution of game theoretic problems in smart grids) and autonomous driving.