Propagation of Multiple Observations in QPNs Revisited

The sign-propagation algorithm for inference with a qualitative probabilistic network has been designed to handle a single observation at a time. Multiple observations can in essence be dealt with by entering them consecutively and combining the results of the successive propagations, or by entering them for a newly added dummy node. We demonstrate that both approaches can yield weaker results than necessary. We identify the causes underlying this unnecessary weakness and adapt the propagation algorithm so as to provide for the strongest possible results upon inference.