Research Topics

Reinforcement Learning

Reinforcement learning can learn powerful policies which enable autonomous systems to dynamically adapt to unknown situations and still perform well in maximizing expected rewards. In our group, we develop novel solutions for spatial mobility tasks such as resource collection and allocation in highly dynamic environments. We aim to make our agents as versatile to adapt to changed conditions and variations of the environment. We further investigate risk and constraints to enforce stable outcomes in financial settings such as portfolio allocation.

Running projects:

  • Routing and Resource Collection in dynamic Spatial Environemnts
  • Financial Portfolio Allocation
  • Robust Policy Learning and Meta Reinforcement Learning

Computer Vision and Deep Learning for Raster Data

Computer vision is one of the core areas driving the development of deep neural networks. In our group, we examine settings with noisy, low-quality data and weak supervision. Examples include super resolution for low resolution raster data and point annotation for object detection. We are particularly interested in hyperspectral raster data common in satellite measurements. As data in most application areas is highly volatile, we also investigate video data and time series of other raster data.

Running Projects:

  • Out-of-distribution detection in Multispectral Raster Data
  • Super Resolution in Remote Sensing Applications
  • Weak Supervision for Segmentation and Object Detection
  • Change Detection in Remote Sensing Imagery
  • Video Instance Segmentation

Spatial Artificial Intelligence

Sequential planning allows optimizing behaviour in sequential decision problems with known mechanics. We investigate how to acquire sensor data best and learn the underlying transition mechanics to formulate sequential planning problems in spatial environments. We develop novel algorithms for spatial optimization problems and compute effective heuristics to improve decision-making and autonomous behaviour.

Running Projects:

  • Sensor Placement
  • Routing in Dynamic Environments
  • Spatial Resource Allocation

Analyzing Relational Data

Relational data generally describes non-idd data, which can be modeled as graphs and networks. As with other data, relational data develop over time, and thus, analyzing the development of networks is crucial in understanding the dynamics between relations. In our group, we examined various types of homogenous and heterogenous networks like dynamic encounter graphs in video games, knowledge graphs, and heterophile networks. Though we currently do not pursue this direction, it is still present in other projects as dynamic and non-idd data can be found in the application area.