Causal Belief Inference in Multiply Connected Networks

The belief function theory is an efficient tool to represent causal knowledge under uncertainty. Therefore, causal belief inference process is important to evaluate the impact of an observation or an intervention on the system. However, existing algorithms only deal with the propagation of observational data in belief networks. This paper addresses propagation algorithms of causal knowledge in multiply connected causal belief networks. To handle this propagation, we have first to transform the initial network into a tree structure. Therefore, we propose some modifications to construct a new structure by exploiting independence relations in the initial network. This structure is called hybrid binary join tree composed of conditional distributions and non conditional ones. Then, we develop a causal belief propagation algorithm using the belief graph mutilation or the graph augmentation methods.