Analysing uncertain data in decision support systems

Abstract

Decision support systems use two basic strategies: the pursuit of a small set of hypotheses and the sequential partitioning of hierarchical hypothesis structures. We present an alternative method based on the maximization of information gain. In each step, we evaluate the difference between the actual and potential future evidence distributions. The data "promising" the maximum information gain are then inquired by the system. In simple situations, the new method behaves like traditional strategies but in divergent and inconsistent evidence situations, it avoids the drawbacks induced by the predetermined standard strategies by adapting itself continuously to the actual data configuration. Our method can be extended to layered hierarchical data structures, where its behavior is reminiscent of the cognitive phenomenon of "restructuring".