Approximate inference in dynamic possibilistic networks

This paper describes an approximate algorithm for inference in dynamic possibilistic networks (DPNs). DPNs provide a succinct and expressive graphical language for representing sequential data and factoring joint possibility distributions and they are powerful models using only the concepts of random variables and conditional possibilities. The proposed algorithm, to perform inference in such networks, is an approximate one and it is based mainly on the standard Boyen-Koller (BK) algorithm well defined for dynamic probabilistic networks. The new possibilistic framework, proposed in this paper, is notable because it gives a counterpart of traditional probability framework, generally used to represent uncertainty in sequential data. The possibilistic BK algorithm is based on the junction tree technique where inference is done via an interface clusters that decrease the size of the dynamic network structured and amenable to a very simple form of inference. We present this algorithm in terms of two possibilistic conditioning; the product based and the min-based one.