Fraunhofer Institute for Experimental Software Engineering IESEAI model with dependable uncertainty estimates
In this bilateral technology transfer project, the “Uncertainty Wrapper” of Fraunhofer IESE was applied to an existing AI component together with IAV GmbH.
Last modified:
Reference Project: IAV GmbH
The challenge: How to deal with unavoidable residual uncertainty in AI models
To cope more effectively with control engineering tasks, IAV product solutions from the field of automotive engineering increasingly include components based on Machine Learning methods and Artificial Intelligence (AI). Such methods are used, in particular, when interrelationships cannot be fully specified in advance, but must be learned on the basis of data. Therefore, even after intensive testing of such components, some residual uncertainty remains regarding the occurrence of faulty results in certain situations.
The support: Use of the “Uncertainty Wrapper” of Fraunhofer IESE for dependable uncertainty estimates
The “Uncertainty Wrapper” architecture and analysis methodology developed at Fraunhofer IESE allows estimating this situation-specific degree of uncertainty reliably and und thus lays a dependable foundation for decisions made at development time as well as at runtime.
For example, certain types of situations with increased uncertainty can already be identified during development and can be mitigated by concrete design decisions. At runtime, dependable and at the same time situation-specific uncertainty information can be used, for example, as part of dynamic risk management. In this way, higher performance or availability of the normal function can be achieved, as it is not always necessary to rely on worst-case estimates to provide safety assurance for it.
Here, the “Uncertainty Wrapper” architecture addresses all three types of uncertainty sources in the corresponding shell model, i.e., uncertainty factors related to the model, the input, and the application context.
This is what Dr. Christian Kruschel, Team Leader Data Science at IAV GmbH, has to say:
“The integration of data-based processes into safety-critical systems creates many opportunities for new impulses in digitalization. An essential part of this is the safety assurance of these processes, which is indispensable for licensing.
By collaborating with Mr. Kläs and his colleagues, we have managed to apply the Uncertainty Wrapper architecture to our existing AI components as part of a bilateral project.”
The result: Transfer of the know-how into industrial practice
As part of a bilateral transfer project, the IESE methodology was applied to an existing AI component together with IAV GmbH. In a series of workshops, the corresponding know-how was successfully transferred into industrial practice based on customer-specific questions and many hands-on training opportunities for the participants.