A Comparison of Architectures for Exact Inference in Evidential Networks

This paper presents a comparison of two architectures for belief propagation in evidential networks, namely the binary join tree using joint belief functions [9] and the modified binary join tree using conditional belief functions [2]. This comparison is done from the perspective of graphical structure, messagepassing scheme, computational efficiency, storage efficiency, and complexity analysis. As a main result, we show that the implication of the conditional relations between variables in evidential networks reduces the computational complexity of the inference process.