In the project KI Wissen, two deliverables on the topics of domain knowledge (automotive knowledge) and validation of the predictions of an AI were pending. The results of the work on the two deliverables have now been published in a joint white paper and present more than 50 methods concerning those topics.
Over the three-year duration of the project KI Wissen, a total of seven public deliverables are planned to document and share the project's progress with the public. A joint white paper now presents the results for two of these seven deliverables. Specifically, it concerns Deliverable 1 "Catalogue, characterisation and representation of relevant domain knowledge for L3 to L5 driving functions" and Deliverable 4 "Methods for data-independent validation of the predictions and decisions of an AI function". Based on the collaborative work in the white paper, a project-wide required and uniform basis of knowledge types, their formalization, and their representation methods is established and thus forms the foundation for a variety of methodological work in the project.
Deliverable 1 (D1) introduced initial methods and concepts for knowledge formalization, knowledge representation, and conformance checking, and addressed various knowledge sources on which they are based. These methods are used in the context of KI Wissen to integrate prior knowledge into the training of an AI component for automated vehicles with a degree of automation between L3 and L5.
Deliverable 4 (D4) on the other hand dealt with methods regarding the conformance checking of AI functions against integrated prior knowledge based on physical and mathematical knowledge, as well as world and expert knowledge.
After an intensive literature research, idea generation, and subsequent concept and method development, which characterized the first year of the project, all partners involved were able to record and document the first technical concepts. In D1, more than 50 different concepts, 16 of which are related to the conformity test discussed in D4, could be presented and a large number of methods could be discussed.
Due to the uniform understanding and the method basis, which was developed by the collaborative white paper in the project, the presented partner-specific methods can now benefit from each other and synergize. In the second iteration, which will be published in early 2023, first realizations and evaluations of the methodological concepts will be presented. Furthermore, a catalog will be presented, which shows the already used knowledge from the initial state, over the representation to the applicability.