Requirements for machine learning in expert systems (abstract)

Abstract

Listen

Translate article

As Dietterich pointed out at the 1989 Machine Learning Workshop, the field needs analyses of the kinds of tasks that require machine learning (ML), and of the suitability of existing ML techniques for solving real problems. An area often mentioned as one that would benefit from ML is that of expert systems. Learning from experience could alleviate the knowledge acquisition “bottleneck”. Although there have been some impressive successes in learning practical knowledge from examples, existing ML techniques lack the power to learn to conduct different phases of an expert “consultation”. There are usually many dozens of variables to consider in an expert task, and their relevance changes as the consultation evolves. ML systems such as ID3, AQ11, and neural networks bog down quickly as extraneous variables are introduced. It is not practical to submit all data derived from an example consultation to such programs, expecting them to produce monolithic knowledge structures. Based on our work in developing a diagnostic expert system, we offer suggestions for future work in ML. One is that the system be given knowledge that is easily available from a human expert. The ML program should be given a “proto-rule” for a specific decision, indicating the most relevant variables to be used by the rule and its “roughcut” logic. Such knowledge can be gotten from an expert quickly. ML routines can then edit the logic and consider other variables as examples are considered. Search is reduced drastically. Other suggestions include ways to combine ML techniques.