The DFG is funding the project "Plausible Inference and Plausible Knowledge Revision in AI along Two Dimensions: Syntax Splitting and Kinematics" for three years.
The goal of this project is to extend plausible reasoning and knowledge revision in symbolic artificial intelligence to include two basic probabilistic techniques that have contributed significantly to success there: Syntax Splitting and Kinematics. While syntax splitting subdivides the exponential space of models by means of subsignatures, kinematics allows further subdivision by considering exclusive cases on the basis of which conditionalization is performed. Thus, the overall result is the possibility of reducing semantic approaches to plausible inference and knowledge revision along two dimensions, opening up entirely new perspectives for improved implementations. In particular, the successful locality principle of reasoning in probabilistic networks is made useful for the symbolic/qualitative domain. This also involves solving another challenge: how to coherently fuse local solutions into global solutions.
The results of this project will have far-reaching implications both theoretically and practically, underpinned by a repository of benchmark problems addressing the two splitting dimensions of plausible reasoning and iterated knowledge revision, and by evaluation with a demonstrator system. These methodological extensions will lead to novel axioms and techniques that essentially advance the current state of the science, and eventually result in more efficient algorithms. A methodologically broad foundation based on knowledge states and conditionals throughout expands the narrow space of classical logics from the outset, resulting in a coherent and unified framework for jointly addressing plausible inference and knowledge revision. To represent knowledge states, we use two widely used semantic approaches: total preorders (TPOs) and Spohn's ordinal conditional functions (OCFs), between which there are intuitive relations. We will systematically elaborate these relations formally in order to make the stronger structures of OCFs useful for the domain of TPOs. The main tools for this are the so-called c-representations and c-revisions for OCFs, which have their roots in probabilistic methods and can be controlled axiomatically and coherently by strategies. These strategies make interactions between conditional knowledge explicit and allow deeper general insights into the problem of plausible inference and knowledge revision, respectively, which will be useful for other approaches as well. Furthermore, the use of c-representations/revisions for the fusion of local to global solutions represents a completely novel methodological perspective of this project.
Project leader: Prof. Dr. Gabriele Kern-Isberner (retired professor, TU Dortmund)
Project partner: Prof. Dr. Christoph Beierle (FernUni Hagen)
Collaborators: n. n.