Many expert systems for diagnosis, prediction, and analysis in complex dynamic scenarios use Bayesian networks for reasoning under uncertainty. These networks often benefit from adaptations to their specific conditions by machine learning on operational data. The knowledge encoded in these adapted networks yields insights as to typical modes of operations, c configurations, types of usage, etc. To utilize this knowledge for the revision and refinement of existing and future expert systems, we developed a context-sensitive machine learning process that uses a multitude of Bayesian networks as input for concept discovery. Our algorithms allow the identification of typical network fragments, their relation, and the context in which they are valid. With these results, we are able to substitute parts of existing networks that are not yet optimally adapted to their tasks and initiate a knowledge engineering process aiming at a precise network generation for future expert systems which accounts for previously unknown characteristics.
展开▼
