Evaluating complex fusion systems based on causal probabilistic models

The paper evaluates a class of fusion systems that support interpretation of complex patterns consisting of large numbers of heterogeneous data obtained from distributed sources at different points in time. The fusion solutions in such domains must be able to process large quantities of heterogeneous information of different quality and adapt at runtime to accommodate for new data sources. This requires models consisting of many variables representing different types of correlated phenomena. In addition, the models are typically severe abstractions associated with significant uncertainties. By using probabilistic causal models we can efficiently build reliable fusion systems that can cope with the above mentioned challenges in a robust manner in a relevant class of applications. The resulting solutions simultaneously satisfy a range of criteria associated with the correctness, scalability, performance, knowledge handling, evidence handling, etc. In addition, causal models also facilitate tractable evaluation of the overall accuracy of complex fusion systems. In particular, we show that the locality of causal relations supports sound decomposition of the system into smaller components, each of which can be evaluated separately using subsets of the data, which reduces the overall evaluation effort. The challenges and the concepts are illustrated with the help of a running example, a system that supports localization of chemical leaks. The paper is concluded with experimental results.