|
Abstract:
|
The emergence in 1976 of Doug Lenat's mathematical discovery program AM [Len76] [Len82a] was met with suprise and controversy; AM's performance seemed to bring the dream of super-intelligent machines to our doorstep, with amazingly simple methods to boot. However, the seeming promise of AM was not borne out: no generation of automated super-mathematicians appeared. Lenat's subsequent attempts (with his work on the Eurisko program) to explain and alleviate AM's problems were something of a novelty in Artificial Intelligence research; AI projects usually 'let lie' after a brief moment in the limelight with a handful of examples. Lenat's work on Eurisko revealed certain constraints on the design of discovery programs; in particular, Lenat discovered that a close coupling of representation syntax and semantics is neccessary for a discovery program to prosper in a given domain. After Eurisko, my own work on the discovery program Cyrano has revealed more constraints on discovery processes in general in particular, work on Cyrano has revealed a requirement of 'closure' in concept formation. The concepts generated by a discovery program's concept formation component must be usable as inputs to that same concept formation component. Beginning with a theoretical analysis of AM's actual performance, this program presents a theory of discovery and goes on to present the implementation of an experiment — the CYRANO program — based on this theory. (This article is a preliminary version of an invited paper fro the First International Symposium on Artificial Intelligence and Expert Systems, to be held in Berlin on May 18-22 1987.) |
