Prototypical Knowledge for Expert Systems: A Retrospective Analysis

The article, "Prototypical knowledge for expert systems" [1 ], published in 1983, was a summarization of my dissertation research at Stanford University in the late 1970s. This was the era of the first rule-based expert systems--MYCIN, Meta-DENDRAL, PUFF, and others from Stanford, and the early frame systems from MIT, such as PIP and NUDGE. I remember being intrigued by the simplicity of the rule-based architectures and the ease with which one could express logical heuristics that did indeed seem to be part of what an expert knew about his area of expertise. I also remember the frustration that was felt by experts and application developers alike over a lack of control in the dialog between application end users and the expert consultation system. Backward-chaining of rules was certainly efficient, but unpredictable, and resulted in dialogs that were not consistent with the way a human expert would have acquired information to solve the problem. I realized that there was not only domain-specific knowledge that was necessary to infer new information for applications, but also control knowledge that was specific to each application, and that would direct the acquisition and application of domain knowledge. Many of the frame systems represented control knowledge in the frame structure itself, directing the flow of the consultation for each context. The knowledge representation that I devised took the best ideas from both sets of research and created what we would call today one of the first "hybrid" architectures.