Pseudo Credal Networks for Inference With Probability Intervals

Abstract The computation of the inference corresponds to an NP-hard problem even for a single connected credal network. The novel concept of pseudo networks is proposed as an alternative to reduce the computational cost of probabilistic inference in credal networks and overcome the computational cost of existing methods. The method allows identifying the combination of intervals that optimizes the probability values of each state of the queried variable from the credal network. In the case of no evidence, the exact probability bounds of the query variable are calculated. When new evidence is inserted into the network, the outer and inner approximations of the query variable are computed by means of the marginalization of the joint probability distributions of the pseudo networks. The applicability of the proposed methodology is shown by solving numerical case studies.