The Intelligent Sensing and Control Group is a research group at the Institute for Information Processing (tnt) at the Leibniz University Hannover, Germany. Our mission is to solve complex problems found in control theory, sensing and other related research fields with novel and innovative machine learning methods.
With the rise in popularity of deep learning based approaches, many research areas have been divided into two camps, with one side continuing to build on classical methods and the other side committing to fully machine learning based approaches. Currently, while there is some interdisciplinary collaboration, many methods exist in parallel, with each research area offering its own solutions.
We aim to use the best of both worlds in our research by combining classical concepts with deep learning approaches.
Reinforcement Learning: While methods from classical control theory (e.g., PID controllers, model predictive control, etc.) are well established in many applications, the use of reinforcement learning (RL) in industrial applications has been limited due to its perceived lack of safety and convergence guarantees. However, a reinforcement learning controller offers many advantages over classical controllers: For example, RL does not necessarily require an explicit system model, as it is able to learn a complex model representing the state transition probabilities by itself, removing the burden and potential inaccuracy of a manually defined model from the process. Furthermore, the deep learning architecture of most RL methods is capable of learning the dynamics of a nonlinear system, which may prove difficult for a classical approach based on classical numerical optimization. Recent research has also shown that so-called "safe RL" approaches show promising results in incorporating safety constraints into the learning process.
Thus, our group focuses on the development of safe and explainable reinforcement learning methods for single and multi-agent systems that exploit concepts from classical control theory. This enables these intelligent agents to make traceable autonomous decisions in challenging and complex real-world environments, such as those commonly found in robotics and other applications.
Another part of our research deals with finding mathematical proofs by means of reinforcement learning.
State-Space Models: State-Space Models (SSM) represent a well-established area of research in the fields of control theory and mechanical engineering. They are used to express complex relationships and transitions between different states of a system. Recently the general structure of state-space models has attracted renewed interest in the machine learning community. Architectures such as Structured State Space (S4) and Mamba learn state transition matrices using neural networks to process sequential data. Since State-Space Models do not rely on the attention mechanism, their memory requirements do not scale quadratically with the sequence length. This represents a significant advantage over the traditionally used Transformer architectures.
We develop novel methods using state-space models in seq2seq tasks like Natural Language Processing (NLP), time-series forecasting (TSF) and information extraction from visual data. Furthermore, we try to address problems regarding the matrix initialization of SSMs by optimizing for suitable initial values using non-linear physics. Moreover, we aim to build physics-informed systems whose internal structure builds upon the mathematical foundations used to describe the properties of the system. State-Space models are particularly promising for this, as their structure is already grounded in physics.
Overall, our interdisciplinary research focus allows our research to contribute to a wide range of fields relevant in many real-world applications, including:
If you are interested in cooperation in any of the research areas mentioned above, feel free to contact Dr. Melanie Schaller.
Furthermore, if you are a student looking for a bachelor/master’s thesis or practical experience as a student assistant to complement your theoretical studies, we are always looking for motivated students. Simply send an informal application to Tristan Gottwald.
