If we look at the evaluation of the results given by the expert, we see that the systems perform equally well in Series 5 and 8, KATE is better in Series 6, and CBRDM is better in Series 7 and 9. We here try to explain why these results occurred for the series that did not get an equal evaluation for the two systems. In the discussion we decide to ignore the feature value DC_CODE because it is always equal for the systems.
Series 6 is a query for failure description (FD_NO) with the value MISCFAIL, where the failure remark is GASLEAK and the failure mode (FM_CODE) varies. The reason that the expert viewed the results of KATE better than CBRDM was that KATE had ESD/PSD for the failure remark feature (in two queries), where the CBRDM returned ``NA''. For the two other queries, M_DOWNTIME was MODERATE for KATE and LOW for CBRDM. From the Bayesian Network in Fig 5.3, we see that the failure remark feature is dependent on FM_CODE, and M_DOWNTIME is dependent on MAINT_EVENT_SU_CODE. The latter is not stated in the query, so CBRDM have no way of using this value to get better results for M_DOWNTIME. The first feature is given, so this indicates that the experts knowledge differs from the network used. If the expert was allowed to modify the network given in appendix D, these two results could have been the same for CBRDM.
Series 7 is a query for failure remark GASLEAK where FM_CODE is varied. The results are here thought to be better for CBRDM because it returns GASLEAK in three queries where KATE returns NA or ESD_PSD. The features given are connected in the Bayesian network.
Series 9 is a collection of various queries where the four last are similar in that they all have many feature values in the query. The first query only has PRODUCTION in the FC_CODE feature.
In general, it seems that CBRDM produces better results when more feature values are given in the query, and when the feature values that are given are a sub-node in the Bayesian network.