Most computer networks are currently designed to maximize throughput and minimize latency, whereas energy consumption is ignored. As the energy consumption is nearly independent of the current demands a network has to serve, networks consume much more energy than needed when the demands are low.

To reduce the overall energy consumption of a network, routers that are idle or operate under low load could be disabled. This will increase the load on the other routers and will result in increased latency. A good energy-conserving algorithm will disable a set of routers so that large amounts of energy are saved and the latency only increases slightly.

Also, as the demands change over time, the algorithm has to dynamically enable and disable routers.

The problem can be approached with both centralized and distributed algorithms. As each of them has their own advantages and disadvantages, it is unclear which is better suited to solve the problem. To summarize, I develop and compare algorithms to conserve energy in computer networks and analyze how they affect the latency and how they behave when the demands change.

Automatic model generation for monitoring process plants

Dynamic evolution of today's production systems has brought a big attention to their reliability, efficiency and safety. An era of mechanical and electromechanical control systems is coming to an end, as they are slowly being replaced by so-called embedded or mechatronic systems that contain a high number of software-based components. In addition, another important trend is the growing adoption of distributed architectures that further increase the production systems' complexity. Systems such as big process plants are usually accompanied by nonlinearities, different types of noises, and external disturbances. Ensuring their proper functioning has led to the development of various monitoring, anomaly detection and diagnosis techniques. Model-based approaches have established themselves among the most successful ones in the field. However, they require a behavior model of a system, which most often needs to be derived manually. Manual modeling of the systems that exhibit state-based, continuous, temporal, and probabilistic behaviors (hybrid systems) is a very hard task that requires a lot of efforts and resources. In my research, which comes from the intersection of computer science and engineering, I tackle this modeling bottleneck by finding an alternative in statistical learning theory. The guiding research question is: Given a system's structure and its recorded observations (logs of control signals and process measurements), how can behavior models for its components be automatically learned?