REPAIR LEARNING FOR A WORKING KNOWLEDGE-BASED SYSTEM

This paper addresses the issues of using a learning mechanism for clos ed-loop updating of the repair knowledge base of a working knowledge-b ased system (KBS). Issues addressed are stability under noisy data and errors arising from learning from cases in which several repairs are attempted. Simulated data are used to investigate the effects of the l atter feature. It is shown that the learning method can cause a signif icant systematic error in learnt knowledge. A knowledge-based method, which aims to intelligently compensate for the systematic error using diagnostic domain knowledge, is investigated It is shown that the meth od greatly reduces the systematic error in learnt repair knowledge.