Evaluation of heterogeneous uncertain information fusion

Abstract

In this paper our motivation is to provide evaluation approaches for information fusion where both possibilistic uncertainty and probabilistic uncertainty occurs. For effective utilization of such diverse data, fusion is used to assist decision-making. However it is necessary to evaluate the results in order to assess their value. Our innovation is the use of information and specificity measures to provide assessments of the fusion results. In particular we investigate transformation-based approaches to the problem of combining possibility and probability distributions. We consider a total heterogeneous fusion process as having three phases in general. Phase 1 is a transformation phase used to produce homogenous data representations. Specifically we explore two transformations—probability to possibility and vice versa. Phase 2 consists of specific aggregation functions operating on the homogenous formatted data. For aggregation functions we representatively cover a range of possible functions by using min, max and average. Phase 3 consists of the applicable assessment measures on the fusion results. Two examples of the complete approach for representative probability and possibility distributions are worked out in full detail and evaluation techniques are used to compare the various results. Finally, general evaluative comparisons of the approaches are given based on extreme bounding cases of completely certain and uncertain probability and possibility distributions. Our contribution then has been to provide approaches to understand which aggregation functions from the min, avg, max spectrum and which transformations would be most useful for the fusion result.