Learning by examples with uncertainty for the adaptive diagnostic system

A customized AQ/STAR series learning-by-examples technique was used to learn uncertain diagnostic knowledge for a real-world domain. Built-in-test (BIT) patterns in the form of if-then rules were created which will deduce faulty subcomponents, given a history of BIT results. Software in which the diagnostic knowledge was created from a portion of the data was developed. Three different portions of the data (90%, 80%, and 70%) were utilized to generate different versions of diagnostic knowledge. Each version was tested in a simple production rule-based expert system against that portion of the data that was not used to generate the rules. This process was repeated several times. The results of the tests are presented. The learning algorithm handles identical BIT results from different diagnostic sessions in which different subcomponents were found to be faulty. A lexical evaluation function which generated a certainty factor for each of the rules learned is presented