Inductive learning for expert systems in manufacturing

The authors evaluate several inductive learning techniques using semiconductor wafer failure data gathered during its manufacturing process and where there is currently an expert system in use with rules derived from experts. The learning systems include symbolic (ID3, GID3, CN2), connectionist (Quickprop) and a hybrid model (SC-net). A year´s worth of data and expert system diagnoses were available for training these systems. The learning systems were evaluated according to three criteria: the accuracy of the induced rules, the quality of the induced rules as judged by two domain experts and by direct comparison with the existing expert system rules, and the flexibility of the systems in learning to diagnose multiple failures on a wafer. Based on these evaluations, the use of machine learning tools for automatic knowledge acquisition in a real manufacturing domain is discussed