cmiller@SRC.Honeywell.COM (Chris Miller) (01/30/91)
Okay, here is my vague understanding: At various points during the development of the MYCIN medical diagnostic expert system, attempts were made to provide MYCIN with the ability to explain/justify its decisions by providing an account of the reasoning which led to the decision. This account was based on a trace of the rules which fired during the construction of the recommendation. This explanatory ability either was the product of, or was augmented by the ancillary system TEIRESIAS. The general consensus was that this approach to making "why"-type explanations was less than satisfactory. Here are my questions: 1. What's right or wrong about the above paragraph? 2. Assuming that the last sentence is accurate, what was wrong with the approach? 3. What's been done since?? Reading recommendations are thoroughly welcome. Another vague source told me that there was a full issue of an AI journal devoted to this topic a while back, but couldn't remember what journal, how far back, or much else-- does this ring any bells for anyone?
geb@dsl.pitt.edu (Gordon E. Banks) (01/30/91)
In article <1991Jan29.210221.7984@src.honeywell.com> cmiller@SRC.Honeywell.COM (Chris Miller) writes: > >Okay, here is my vague understanding: > >At various points during the development of the MYCIN medical diagnostic >expert system, attempts were made to provide MYCIN with the ability to >explain/justify its decisions by providing an account of the reasoning >which led to the decision. This account was based on a trace of the rules >which fired during the construction of the recommendation. This >explanatory ability either was the product of, or was augmented by the >ancillary system TEIRESIAS. The general consensus was that this approach >to making "why"-type explanations was less than satisfactory. > >Here are my questions: > >2. Assuming that the last sentence is accurate, what was wrong with the >approach? > A recapitulation of the rules that have fired in a backward chaining system doesn't always provide a satisfactory explanation to the clinician who is using the system. It's much better than nothing, but it is a very brittle explanatory system, and it is not robustly interactive. >3. What's been done since?? > A lot. Check Joanna Moore's work (her thesis was done under Bill Swartout at USC). She has an explanation system for Lisp tutoring.
vansoest@cs.utwente.nl (Dick van Soest) (01/31/91)
A researcher in our group did her PhD thesis on explanation, in which, among other topics about explanation, the topic of the last sentence of your first paragraph is discussed. The reference is: P.M. Wognum, 1990 Explanation of automated reasoning: How and why? PhD thesis University of Twente, Enschede, The Netherlands Her email address is wognum@cs.utwente.nl Abstract: Automated-reasoning systems need large amounts of knowledge to solve complex problems. Knowledge engineering focuses on techniques for acquiring, structuring, and representing knowledge, and on techniques for reasoning with the knowledge. Users may wish a computer to explain how the reasoning has been performed and why it made certain statements. This thesis adresses the topic of explanation of the reasoning performed by an automated reasoning system. We describe how the reasoning performed by a computer may serve as the basis for explanation. First, we show that a resolution proof which is not very transparent can be transformed into a natural-deduction proof that is more suitable for explanation. Through such a transformation, a resulution-based automated theorem prover can combine the efficiency of resolution with the transparency of natural-deduction. Second, we describe our model of reasoning which defines the architecture a knowledge-based system must have to reason in an understandable way. We show that this architecture is suitable to produce reasoning traces which can be used to generate a wide range of explanations. In the literature, the importance of explanation in knowledge-based systems has frequently been emphasized but has hardly been assessed in practice. This thesis contains the results of investigations to determine the importance of explanation. We describe the results of a study of the use made of explanation in a number of knowledge-based systems which are actually used in the Netherlands. Second, we describe a study of the impact of explanation on users' decisions in a complex domain. Third, we describe how we used our model of reasoning to analyze explanations that can be found in the medical literature. This analysis has yielded criteria for a knowledge-based system with explanation that is acceptable to physicians. The results presented in this thesis offer a knowledge engineer useful guidelines for acquiring and structuring knowledge for knowledge-based systems which are transparent to their users. -- Dick van Soest University of Twente Computer Science Department Internet: vansoest@cs.utwente.nl P.O. Box 217 Bitnet: vansoest@utwente.nl 7500 AE Enschede SURF-net: UTRCV1::VANSOEST The Netherlands Tel. +31 53 893736/893690 FAX: +31 53 339605