Robustification of Self-Optimising Systems via Explicit Treatment of Uncertain Information

Abstract

Uncertainty treatment in self-optimising systems touches two design-issues. Firstly, a valid estimation of uncertainties within the system is impossible beforehand as the uncertainties as well as the systems behaviour changes during run-time due to self-optimisation. Secondly, the design of a selfoptimising system needs to mediate between the often conflicting goals of optimality and robustness. Here we present the concept for a lightweight algorithmic add-on for self-optimising function approximators that enables to reflect uncertainties related to the current state and to flexibly combine optimality and robustness in one design. Illustrating examples of TS-fuzzy systems highlight the properties of our approach.

Gaia-AgStream: An Explainable AI Platform for Mining Complex Data Streams in Agriculture

“…One key aspect of Agri-Gaia with respect to providing data for AI at scale is the generation of synthetic data for training AI models in order to leverage existing high-quality real data. Here, we want to complement this line of work by developing a uniform uncertainty management embedded in the proposed AI and data streaming platform to support the generation of 'AI ready data' and find an application-agnostic data quality indication, cf., [13,14,28] for according related work about uncertainty estimation, architecture concepts and robust self-optimization. The importance of handling uncertain data is inherent to AI applications, but the current balance between the effort for handling data and tuning the model puts a disproportional weight on data handling.…”