Self-Organizing and Adaptive Systems
In self-organizing systems, components apply local rules to adapt their interactions in response to changing conditions and cooperatively realize adaptation. (-- Danny Weyns)
In our research, the dynamic nature of distributed systems is addressed by concepts and mechanisms of self-organizing systems. We work on architectures for the specification of configurations and algorithms for the automatic selection of suitable configurations at runtime.
Self-adaptive systems autonomously modify their behavior and/or structure in response to changes in their context and environment. If multiple such self-adaptive systems interact to form a self-organizing system, new forms of behavior can emerge.
In our research, we investigate transition mechanisms that enable proactive self-optimization within self-adaptive communication systems and their applications. We explore methodologies for model-based design of a system’s reconfiguration behavior. Furthermore, we are developing a runtime environment (middleware) to coordinate the triggering and execution of cross-application and cross-layer transition mechanisms.
Contact: Michael Matthé
With the ubiquity of computing devices, the call for new interaction paradigms has become louder. While one form of interaction may be useful in one situation, it can be rather impractical in another. For example, text input is inconvenient when the hands are needed for another task, whereas most users would feel uncomfortable sending and receiving audio messages in public spaces. Multimodal interfaces allow users to dynamically select the interaction modality that best suits their current context, but typically give output in a default modality as specified in the application settings. Adaptive multimodal interfaces therefore monitor and dynamically respond to the user’s changing situation and state. The goal is to improve the user experience by dynamically adapting either the content or its modality.
An indispensable prerequisite for responding to changing situations of the user is to correctly interpret them in the first place. One way to learn about the user in a non-invasive way is to use biosignals which have become readily available through ubiquitous devices like fitness trackers or smartphones. With the help of AI, biosignals can be used to decode different cognitive states. This enables applications not only for marketing, but also for healthcare or even driving assistants.
In our research, we explore strategies for presenting information in the best possible way given a users’ situation and cognitive state. We focus on the subtle cues from eye movements, facial expressions, and speech to decode cognitive states (e.g., preferences, emotions, and comprehension).
Contact: Melanie Heck
Software-Defined Deterministic Networking
Software-Defined Deterministic Networking aims to increase flexibility, efficiency, and quality of service with respect to deterministic real-time guarantees of communication networks.
The objective is to support and enable modern networked systems that, on the one hand, rely on deterministic real-time guarantees, including safety critical Cyber-Physical Systems from the Industrial Internet of Things (IIoT), smart grids, or automotive domain. On the other hand, we optimize the utilization of precious network resources to increase the scalability and efficiency of networked systems that need to connect a large number of devices (sensors, actuators, controllers, and appliances).
We utilize state-of-the art networking technologies and build on current industry standards, in particular, Software-Defined Networking (SDN), IEEE Time-Sensitive Networking (TSN), and IETF Deterministic Networking (DetNet). These technologies introduce and support concepts such as logically centralized network control which facilitates optimized network control, or scheduling mechanisms for deterministic real-time guarantees which we use as our basis for designing novel concepts. For instance, we investigate algorithms to calculate time-driven real-time schedules for TSN (IEEE 802.1Qbv) and novel scheduling mechanisms to support networked control systems based on a holistic view onto the control application and the network.
Our research covers a broad spectrum of networked systems along different dimensions:
- Wired systems, such as IEEE 802.3 TSN (Ethernet) networks, and wireless systems, such as 6G mobile networks or IEEE 802.11 (WiFi) networks.
- Switched local area networks, such as factory networks, and larger routed networks, such as networks of smart energy grids.
On the one hand, wireless networks have fundamentally different properties with respect to reliability and network latency compared to fixed networks. On the other hand, many future systems, including in particular safety critical mobile systems, require deterministic latency guarantees for wireless communication.
In order to tackle this challenge and enable deterministic communication over upcoming 6G mobile network infrastructures, novel robust scheduling concepts are required that deal with the specific properties of wireless communication links and also enable the integration with existing wired networking technologies such as IEEE TSN for deterministic end-to-end guarantees. Such novel end-to-end deterministic scheduling concepts together with the tools to validate these concepts are in the focus of our research.
Time-triggered network streams are critical for real-time machine-to-machine communication, e.g. in Industrial IoT applications. Thereby, the time-triggered streams have tight bounds in terms of deadline and jitter. Furthermore, the frame release times are usually known, allowing for precise per-frame scheduling.
In our research, we investigate fast and scalable scheduling heuristics to organize time-triggered real-time communication in large-scale deterministic Ethernet networks. In that, we are not necessarily limited to IEEE Time-Sensitive Networks, but also consider other deterministic Ethernet techniques.
Contact: Heiko Geppert
Cyber-physical systems (CPS) monitor the state of physical objects through sensors and control their physical state through actuators. Physical objects include machines in an Industry 4.0 factory, a fleet of cars, smart home devices, or energy consumers and producers in the smart grid. Such systems are highly sensitive to the delay and loss of communicated sensor data. Therefore, operating them over the Internet, which was not designed to support such time-sensitive applications, is a great challenge. To this end, we investigate networking concepts optimizing the quality of control of networked control systems operating on top of an unreliable IP-based networking infrastructure like the Internet.
Contact: Robin Laidig
Mobile & Edge Computing
IoT applications on the one hand produce large and fluctuating volumes of data and on the other hand require fast, highly available and resource efficient data processing to ensure low response times. For example, semi-autonomous car applications assist drivers in reaching their destination safely by analyzing traffic situations and driving patterns. Likewise, surveillance applications can assist police officers in identifying persons with suspicious activities in the nearby vicinity by processing camera streams.
Aggregating and processing data at a centralized cloud platform is not adequate to meet the requirements of these applications. Consequently, there is a paradigm shift towards distributing data and computation between remote cloud data centers and resources at the edge of the network (i.e., close to the devices) including routers and end user devices.
In our research, we investigate methods to support latency and availability requirements of IoT applications while ensuring efficient utilization of heterogeneous compute and storage resources at the different levels of the network hierarchy.
Today every smartphone, every application and every company, (i.e., every connected entity) gathers and analyzes data about the environment. For example, a smartphone can learn the best time to leave home depending on the weather conditions and travel time to work. Clearly, greater benefit can arise if the knowledge is not only used locally but also accessible to others. We therefore research efficient mechanisms for indexing, updating and retrieving large and changing quantities of data in a large-scale distributed system.
Contact: Lukas Epple
Graph neural networks (GNN) have proven their usefulness for learning on graph data in a variety of domains including recommender systems, traffic, finance, chemistry, etc. Unlike images, graph data is often not publicly available. Instead, the data of institutions and companies is often maintained in isolated data silos to protect privacy and intellectual property. This puts companies in the predicament of not having sufficient data on their own to train a model, while still wanting to be able to use centralized GNNs. An example are pharmaceutical research institutions that want to use GNN, but have limited and often confidential data.
Federated Learning is a machine learning setting where many clients collectively train a model without making the data itself accessible to collaborators. This enables institutions with smaller datasets to gain insights that they could not observe using their proprietary data alone. Due to its privacy preserving approach, Federated Learning is also a viable method for applications that rely on highly sensitive data, for example in the medical domain.
In our research, we investigate Federated Learning approaches where the original data stays with each participant, and the participants only share information about their local model. Information sharing can be handled by a central server or in a decentralized way. Currently, our main focus is on Knowledge Distillation approaches.
Contact: Michael Schramm
In order to support resource-restricted mobile devices with services based on simulation, we investigate appropriated simualtion models. For the execution, these have to be placed according to the available resources. Edge Computing provides means for latency-aware placement of computing tasks at or near the mobile access network.
The Tasklet project provides a serverless Edge Computing platform based on a minimal virtual machine and the C-- programming language, which is a subset of C. Based on this platform, a we investigate a variety of scheduling strategies, including fault-tolerant task scheduling, decentralized task placement, and energy-aware offloading.
Contact: Johannes Kässinger
Sensor data from IoT devices can feed simulations, enabling not only novel gaming, but also serious engineering applications. For instance, a structural engineer could simulate the current state of a building. Planned changes to its architecture can be evaluated immediately on-site by visualizing the results on a mobile device such as Microsoft’s HoloLens augmented reality headset as an overlay of the physical world.
Since simulations are typically compute-intensive, their calculation on mobile devices takes a long time and consumes lots of limited battery resources. Distributed mobile simulations therefore offload parts of the simulation to powerful remote cloud or a nearby edge computing infrastructures to bring even complex simulations to mobile devices. In this context, we investigate optimized simulation methods and manage the trade-off between communication overhead, computation overhead, and required quality of simulation results.
Contact: Johannes Kässinger
To meet the increasing energy demand and mitigate environmental impact, energy informatics seeks to apply digital technology and information management that facilitate the global transition towards sustainable and resilient energy systems.
After phasing out nuclear energy, energy generated from fossil fuels will also gradually be reduced. Energy suppliers respond to the ensuing electrification of energy demand by capitalizing on renewable energy sources, some of which - sun and wind in particular - are highly volatile and non-deterministic. As a result, flexibility is required both on the supply side and in terms of temporally (and in some cases geographically) shifting demand.
Additionally, heat is one of the major contributors to CO2 emissions. Novel approaches that aim to reduce emissions by, for example, integrating industrial waste heat supply, also require demand to be adaptive. And finally, in order to fully make use of the available energy in a system, cross-vector flexibility options need to be taken into account.
In our research, we tackle these issues both from an energy informatics and from a multi-disciplinary point of view, with the objective to create integrated solutions.
Contact: Sonja Klingert