gps0@bunny.UUCP (05/22/87)
This is a promised follow-up on the subject of extracting knowledge from databases. I have received about a dozen replies and my thanks to all the respondents. I have also tried to reply individually, but not always succeeded (all blame is on the mailer). I have found two recent references in this area. The Spring, 1987 issue of IEEE Expert contains a good article by Michael Walker on "How Feasible is Automated Discovery?". This article also contains references to other relevant systems, including Meta-Dendral, AM, Bacon, RX, Prospector and others. There is also an article by Gio Wiederhold et al, on "KSYS: An Architecture for Integrating Databases and Knowledge Bases". It was submitted to IEEE transactions on Software Engineering and it can be obtained by writing to Prof. Wiederhold at Stanford. I have found that there is some work on extracting expert system rules from databases at GM (contact samy@gmr.com). There is also a company in Hawaii working on automatic analysis of medical databases and there is a small start-up in Boston area working on extracting data models from databases. However, none of the above have published anything. There are some commercial expert system tools that interface to databases: Intellicorp has KEE Connection to interface KEE to SQL databases Inference is working on a similar tool for ART Arity Prolog has an interface to SQL Guru from mdbs combines an ES and DBMS (and other stuff). Insight 2+ interfaces to dbase II, III VP Expert also has an interface to dbase II, III Mad Intelligent Systems from San Jose, CA has produced Relational Lisp - Lisp extended by relational operations Herman Rubin from Purdue expressed doubts that it is possible to come up with new theories in a mechanical way. He says >I do not trust anyone to come up with anything new by that device. Data >analysis is necessary, but it should only be done by geniuses, or at least >very bright people, who are constantly aware of the dangers of incorrect >analysis, or even accidently incorrect analysis. True - "extracting knowledge from data" will not come up with radically new theories. However this approach can and does come up with new relationships, the general form of which is known - look at Bacon, Meta-Dendral, RX, Prospector. >If Kepler had one more decimal place to work with, his laws would not fit; >The data analysis problem is to get theories >which are certainly incorrect, and which fit "more or less." An excellent observation. But who says that a computer cannot search for relations that hold more or less? In fact, accounting for approximate relationships is a must prerequisite in analyzing any real data, and it was done before. Comments are welcome.