The Monte Carlo technique is used to evaluate the performance of four techniques for making decisions in the presence of ambiguity. A modified probability approach (both weighted and unweighted) and weighted and unweighted Dempster-Shafer are applied to compare the reliability of these methods in producing a correct single decision based on a priori knowledge perturbed by expert or sensor inaccuracy. These methods are tested across multiple conditions which differ in condition mass values and the relative accuracy of the expert or sensor. Probability and weighted probability are demonstrated to work suitably, as expected, in cases where the bulk of the input (expert belief or sensor) data can be assigned directly to a condition or in scenarios where the ambiguity is somewhat evenly distributed across conditions. The Dempster-Shafer approach would outperform standard probability when significant likelihood is assigned to a particular subset of conditions. Weighted Dempster-Shafer would also be expected to outperform standard and weighted probability marginally when significant likelihood is assigned to a particular subset of conditions and input accuracy varies significantly. However, it is demonstrated that by making minor changes to the probability algorithm, results similar to those produced by Dempster-Shafer can be obtained. These results are considered in light of the computational costs of Dempster-Shafer versus probability.

Date Published: 13 January 2012
PDF: 5 pages
Proc. SPIE 8350, Fourth International Conference on Machine Vision (ICMV 2011): Computer Vision and Image Analysis; Pattern Recognition and Basic Technologies, 83502F (13 January 2012); doi: 10.1117/12.920195

Published in SPIE Proceedings Vol. 8350:
Fourth International Conference on Machine Vision (ICMV 2011): Computer Vision and Image Analysis; Pattern Recognition and Basic Technologies
Safaa S. Mahmoud; Zhu Zeng; Yuting Li, Editor(s)