Dendral
Dendral
Dendral
233
Elsevier
ARTINT 983
DENDRAL and
Meta-DENDRAL
roots of knowledge systems and
expert system applications
Edward A. Feigenbaum
Knowledge Systems Laboratory, Department
University,
Bruce G. Buchanan
Computer Science Department, University
.
k
ofPittsburgh, Pittsburgh, PA
15260, USA
Correspondence
to:
of Computer Science,
234
Salient among the early applications of AI was the work of the DENDRAL
group, applying AI to problems of the analysis of the mass spectra of organic
molecules and the induction of new rules of mass spectral fragmentation.
The paper "Dendral and Meta-Dendral: their applications dimension" [3]
was solicited by the editor of the special issue. The article appeared in 1978,
thirteen years after the start of the DENDRAL Project. In those thirteen
years, many DENDRAL papers had been published in the literature of both
AI and chemistry. Some of the results of DENDRAL and Meta-DENDRAL
as applications to chemistry had been reported to chemists. The time was
ripe for reporting these results not merely as chemistry but as applied AI,
and the special issue provided us with the appropriate vehicle.
In this note, we will look both backward and forward from the 1978
publication date of the special issue. Though this note is necessarily short,
two other and longer works have done this job thoroughly [8,9], the more
recent paper having the advantages of perspective of time and experience
with the technology transfer of DENDRAL to an industrial setting.
.
i
post-docs.
*
.
235
several years. We made a decision (in retrospect correct) to achieve experimental results and gain system-building experience on a more concrete
problem first: the hypothesis formation problem of inferring from one set of
spectral data one (or a few) candidate molecular structure (s). Success with
DENDRAL led us back to the original problem of theory formation, which
now appeared in a quite specific and concrete form that was "meta" to
DENDRAL (hence the project's name, Meta-DENDRAL). If the knowledge
of mass spectrometry was crucial to the progress of DENDRAL, then we
must codify it, and we knew of only two ways. Either work in the painstaking one-on-one fashion of our interaction with Djerassi's chemists (which
has since become known as the knowledge acquisition part of knowledge
engineering). Or, infer the knowledge directly from electronic libraries of
mass spectral data and the known structures that gave rise to the data. The
latter was the task of the Meta-DENDRAL learning program.
.
.
geometry exam).
The DENDRAL group was strongly focused on the performance dimension of AI and played a key role in reinstating the view (the goal, the
dream) that AI programs can perform at the level of the most competent
humans performing the task (and perhaps beyond). It did this with enough
specificity that its results could be extended by the group itself, and by
others. For the DENDRAL group, the extension was to MYCIN, then to
EMYCIN (the software generalization), then to applications in medicine,
engineering, molecular biology, x-ray crystallography, submarine detection,
etc. This system-building experimental AI effort, sustained over a period of
236
DENDRAL was not the only agent that brought about the shift of
paradigm (at MIT, Moses and the Mathlab/Macsyma group were influential supporters of and early contributors to the expert systems viewpoint),
but it was one of the most significant agents.
By 1967, the DENDRAL project faced a crisis of knowledge representation. The amount of new knowledge (represented as LISP code) that was
pouring in via the knowledge acquisition interactions with the chemists
produced a complexity of the knowledge base that we could neither manage nor sustain. Inspired initially by the Newell-Simon use of productions
as an architecture for problem solving, we conceptualized productions as
modular situation-action "rules" in terms of which we could represent the
,f
r.
237
238
"
239
"
"
Noisy data: We could not assume that the empirical data given to the
program were complete and correct. The data were known to contain
spurious (noisy) data points and to omit data points that the theory
predicted should be present.
Multiple concepts: Meta-DENDRAL had to learn the preconditions
(LHSs) for more than one concept (mass spectral process), but did
not know how many concepts needed to be learned.
Unclassified data: Meta-DENDRAL was given sets ofx-y points without
havingthose points labeled as positive or negative instances of a concept.
Thus we first had to generate possible explanations of each x-y point
before we could consider positive and negative evidence associated with
each explanation.
References
/
[1] D.G. Bobrow and P.J. Hayes, Artificial Intelligencewhere are we? Artif. Intell. 25 (3)
(1985) 375-415.
[2] J. Brinkley, R. Altman, B. Duncan, B.G. Buchanan and O. Jardetzky, The heuristic
refinement method for the derivation of protein solution structures: validation of
Cytochrome-b562, /. Chem. Inf. Comput. Sci. 28 (4) (1988) 194-210.
[3] B.G. Buchanan and E.A. Feigenbaum, Dendral and Meta-Dendral: their applications
dimension, Artif. Intell. 11 (1978) 5-24.
[4] B.G. Buchanan, D.H. Smith, W.C. White, R. Gritter, E.A. Feigenbaum, J. Lederberg and
C. Djerassi, Applications of artificial intelligence for chemical inference, XXII: automatic
rule formation in mass spectrometry by means of the Meta-DENDRAL program, /. Am.
Chem. Soc. 98 (1976) 6168.
[5] E.A. Feigenbaum, P. McCorduck and H.P. Nii, The Rise of the Expert Company (Times
Books, 1988).
240
[6] P.D. Karp, Hypothesis formation and qualitative reasoning in molecular biology,
Knowledge Systems Laboratory Tech. Report 89-52, Stanford University, Stanford, CA
(1989).
[7] D.B. Lenat and E.A. Feigenbaum, On the thresholds of knowledge, in: Proceedings
IJCAI-87, Milan, Italy (1987) 1173-1182; also: Artif. Intell. 47 (1991) 185-250.
[8] R.K. Lindsay, B.G. Buchanan, E.A. Feigenbaum and J. Lederberg, Applications of
Artificial Intelligence for Organic Chemistry: The DENDRAL Project McGraw-Hill, New
York, 1980).
[9] R.K. Lindsay, B.G. Buchanan, E.A. Feigenbaum and J. Lederberg, DENDRAL: a case
study of the first expert system for scientific hypothesis formation, Artif. Intell. 60 (1993)
(to appear).
[10] T.M. Mitchell, Version spaces: an approach to concept learning, Doctoral dissertation,
Department of Electrical Engineering, Stanford University (1978).
[11] A. Newell, H.A. Simon and C. Shaw, Chess playing programs and the problem of
complexity, E.A. Feigenbaum and J. Feldman, eds., Computers and Thought (McGrawHill, New York, 1963).
[12] E. Rich and K. Knight, Artificial Intelligence (McGraw-Hill, New York, 2nd cd., 1991 ).
"