A General Theory for the Fusion of Data

Abstract : The problem of data fusion is in a real sense the problem of how to model the real world with all of its great complexities. A miniaturized version of this is the multiple target tracking and data association problem. There, a number of pieces of information arrive, typically from disparate sources - such as from various sensing systems and from human sources in the form of narrative descriptions in natural language. A procedure has already been established for dealing with this type of situation, called succinctly the PACT algorithm. (PACT = Possibilistic Approach to Correlation and Tracking.) The technique is based upon the premise that all arriving information can be adequately treated through some appropriate choice of classical or multivalued logic such as Probability Logic, fuzzy Logic, Lukasiewicz - aleph1 Logic, or some (t-norm, t-conorm, negation function) general logic as discussed in a recent text of Goodman and Nguyen. Uncertainty Models for Knowledge-Based Systems. Moreover, it can be demonstrated that for a large class of logics chosen, a version of a partially specified Probability Logic may be used instead. Indeed, other approaches to uncertainty, such as the Dempster-Shafer approach, can also be strongly related to Probability Logic through the vehicle of random set modeling. In any case, the structure of the PACT algorithm is based upon a generalized chaining and disjunction relation, which in a classical probability setting reduces to the usual posterior probability description as a weighted sum of intermediate probabilities, an alternative form of Bayes' formulation.