Maximal Consistency, Theory of Evidence, and Bayesian Conditioning in the Investigative Domain

1 Introduction During the upstream phase of an inquiry, much of the detectives' and magistrates' everyday task consists of • acquiring information from investigations on the spot and from witnesses' depositions • introducing hypothetical rules to link the various hypotheses into evidential networks • finding contradictions and incompatibilities inside and across the various depositions • judging the credibility of the information items • judging the reliability of the witnesses In Artificial Intelligence, finding contradictions (or incompatibilities) and rearranging the knowledge base in order to remove them is often referred to as " belief revision ". Somewhat departing from the literature on this subject (see [37,38] for a survey), in the late eighties Aldo Dragoni conceived a model for belief revision that suited the multi-source characteristic of the inquiry domain [43]. That ATMS-based approach [20] laid down the architecture of an Inquiry Support System (hereafter called ISS) that could help a detective to perform the activities 3, 4 and 5 (ISS was published only five years later [16]). That ISS's ultimate tasks were those of: 1. finding the maximal consistent subsets of the beliefs about the case under consideration 2. ordering them w.r.t. their global degree of believability. However, the credibility of the information items, as well as the reliability of the witnesses, were estimated in some strange and rather naive ways. One of ISS' worst characteristic was that such values were dependent (at least theoretically) on the particular chronological sequence of the depositions received from the various witnesses. Recently the author realized that the " belief-function formalism " adopted to aggregate audit evidence [13,18,19] could work as well here, and, especially, that Dempster's Rule of Combination is a powerful