| Abstract: | In real classification problems intrinsically vague information often coexist with conditions of “lack of specificity” originating from evidence not strong enough to induce knowledge, but only degrees of belief or credibility regarding class assignments. The problem has been addressed here by proposing a fuzzy Dempster–Shafer model (FDS) for multisource classification purposes. The salient aspect of the work is the definition of an empirical learning strategy for the automatic generation of fuzzy Dempster–Shafer classification rules from a set of exemplified training data. Dempster–Shafer measures of uncertainty are semantically related to conditions of ambiguity among the data and then automatically set during the learning process. Partial reduced beliefs in class assignments are then induced and explicitly represented when generating classification rules. The fuzzy deductive apparatus has been modified and extended to integrate the Dempster–Shafer propagation of evidence. The strategy has been applied to a standard classification problem in order to develop a sensitivity analysis in an easily controlled domain. A second experimental test has been conducted in the field of natural risk assessment, where vagueness and lack of specificity conditions are prevalent. These empirical tests show that classification benefits from the combination of the fuzzy and Dempster–Shafer models especially when conditions of lack of specifity among data are prevalent. ©1999 John Wiley & Sons, Inc. |
