IFIP 2013 - Special Session on Information Fusion with Imprecise Probabilties and Related Approaches

In classical Bayesian estimation, a typical assumption is that the models for prior and likelihood are completely known. However, in many cases, due to a lack of information about the system to be modeled, it may not be possible to characterize prior and/or likelihood precisely. For example, sensor noise is usually characterized by a Gaussian distribution, but a measurement device is often also affected by systematic errors. Systematic errors can generally be narrowed down to intervals rather than to precise values. Hence, one may only know that the mean of the sensor noise lies in an interval or, in more general cases, one may only be able to state that the distribution of the prior/likelihood belongs to some set. In this context, a more systematic approach to estimation can be obtained by means of imprecise probabilities. The basic idea is to solve the estimation problem by dealing with all elements of the set of distributions characterizing the prior and the likelihood. The inferences will therefore be less dependent on the model assumptions and, thus, be intrinsically more robust and reliable. The purpose of this Special Session is to share and discuss ideas on these topics and to develop, evaluate and apply novel robust/reliable methods for information fusion based on imprecise probabilities and related approaches, such as random sets, robust Kalman filters, set-membership estimation