In order to handle self-optimization and uncertainty treatment, we focus on the a priori controllability at design-time for dealing with uncertainties and on-line learning as both introduce specific kinds of dynamically changing uncertainties. We address the challenge of fast and robust adaptation on sparse data as well as shift and drift issues with an emphasis on the interpretability and engineerability of the learning system. To this end, we combine our learning techniques with trust management to Trusted On-line Learning and build a toolset of extensions for the guidance of on-line learning algorithms in combination with coarse- as well as fine-grained trust management.
It is important for us to bring our methods and architectures into practice as early as possible. Such concrete practical applications range from data analysis to real-time control in order to proof the concepts and to demonstrate the benefits on a broad spectrum of real-world problems under real conditions.
For systematic scientific investigations we built the UOSlib-framework. It collects different state-of-the-art learning algorithms and offers tools for their easy assessment and comparison. It features the definition of the investigation scenarios by individual ´footprints´ which uniquely define the learning algorithm and its parameterization as well as the data to be processed. In this way, a particular scenario can be reproduced for world-wide comparisons just by typing in the footprint into the UOSLib-framework. It can be downloaded here.