Microsoft research has developed two prototype systems for fault diagnosis which are codenamed Aladdin. The systems uses three-layer Bayesian networks to store information about faults in Microsoft Word and Microsoft Windows NT. The layers describe one or more cause - a contributing factor for an error, an issue which occurs if all the causes occur, and one or more symptoms which is caused by the issue. The Bayesian network is constructed by an expert, and is updated each time it is used. The system is not very dynamic, as Microsoft has stopped using it because the knowledge base is too small. The approach is described in [10].
Two Bayesian Networks are used in a system developed at the University of Salford and at ITESM, Mexico. In an exemplar-based CBR system, one network is used for indexing of categories - collections of cases with some degree of similarity, and one network is used for identifying exemplars - or single cases within a category. The work is described in [28].
D-SIDE is a software package developed at the University of Helsinki [30]. They focus on data-intensive domains where cases are viewed as vectors, and they assume that the casebase has some incorrect cases. During the reuse step in the CBR process, they use Bayesian model selection to do adaption on the case, for instance, if a retrieved case has missing features which can solve a problem, the most probable feature value is predicted and used.
At the Navy Center for Applied research in AI a system for action selection based on Bayesian networks and CBR has been outlined. The networks are used as the environmental context, and agents act according to action that fulfills a plan with the highest probability, given a current state. CBR is used for implementing the selected action. The system is called INBANCA after Integrating Bayes Networks with Case-Based Reasoning for Planning, and is described in [5]. The paper also gives a survey of existing integrated methods.
The existing integrated systems use various approaches. The two first systems use Data Mining to extract domain knowledge, the INBANCA approach uses CBR to implement an action plan, the Aladdin system uses CBR to solve an error which is found by using a Bayesian network. The system developed at Salford uses networks for management of cases, and finally D-SIDE uses cases as the base for classification. In Tab 4.1, we list the Data Mining techniques used when describing the systems using the CBR cycle. Most techniques can be described as finding domain knowledge using DM techniques for usage in the CBR process.
The aim here is to show how integration can be done at a lower level by implementing a prototype which uses a Bayesian network to compute similarity metrics during retrieval of cases in a CBR system. First we give an example of possible use of such a system, and then we give a description of the system.
| System | Aladdin | INBANCA | D-SIDE | Salford | Rostock | Case-Method |
| Retrieve | Use BN to | Identify case | Use DM method | |||
| select case | case in BN | |||||
| Reuse | Adapt Solution | |||||
| using BN model | ||||||
| Revise | ||||||
| Retain | Update BN | Update BN, | Index cases | Learn by extracting | ||
| add case | add case | using BN | rules from casebase |