Th. Wetter K.-D. Althoff J. Boose
B.R.Gaines M.Linster F. Schmalhofer (Eds.)
Current Developments in Knowledge Acquisition - E K A W '92
6th European Knowledge Acquisition Workshop Heidelberg and Kaiserslautern, Germany,
May 18-22,1992 Proceedings
3
Springer-Verlag
Berlin Heidelberg New York London Paris Tokyo Hong K o n g Barcelona Budapest
Series Editor Jorg Siekmann University of Saarland
German Research Center for Artificial Intelligence (DFKI) Stuhlsatzenhausweg 3, W-6600 Saarbriicken 11, F R G
Volume Editors Thomas Wetter
I B M Germany, Scientific Center
WilckensstraBe l a , W-6900 Heidelberg, F R G Klaus-Dieter Althoff
University of Kaiserslautern, Dept. of Computer Science R O . Box 30 49, W-6750 Kaiserslautern, F R G
John Boose
Computer Science Organization, Boeing Computer Services R O . B o x 24346, 7L-64, Seattle, W A 98124, U S A
Brian R. Gaines
Knowledge Science Institute, University of Calgary
2500 University Dr. N W , Calgary, Alberta T 2 N 1 N4, Canada Marc Linster
German National Research Center for Computer Science ( G M D ) P. O . Box 13 16, W-5205 Sankt Augustin 1, F R G
Franz Schmalhofer
German Research Center for Artificial Intelligence P. O . Box 20 80, W-6750 Kaiserslautern, F R G
Untv.-BiblMhak! 4 Regewtare J
C R Subject Classification (1991): 1.2.4-6,1.2.8 / r~ T r /\
I S B N 3-540-55546-3 Springer-Verlag Berlin Heidelberg New York I S B N 0-387-55546-3 Springer-Verlag New York Berlin Heidelberg
This work is subject to copyright. A l l rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting, reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer-Verlag.
Violations are liable for prosecution under the German Copyright Law.
© Springer-Verlag Berlin Heidelberg 1992 Printed in Germany
Typesetting: Camera ready by author/editor
Printing and binding: Druckhaus Beltz, Hemsbach/Bergstr.
45/3140-543210 - Printed on acid-free paper
Preface
M e t h o d o l o g i c a l knowledge acquisition and knowledge engineering have achieved increasing attention over the last years due to both active research projects a n d successful practical applications. B o t h aspects have over the years been reflected in the structure of the European Knowledge Acquisition Workshops ( E K A W ) , where a users' forum has always been combined with a scientific w o r k s h o p .
K n o w l e d g e acquisition workshops also take place a n n u a l l y i n N o r t h A m e r i c a (the
" B a n f f " workshops) and A s i a or A u s t r a l i a . Intense interaction between these communities, reflected in international conference attendance, shared authorship from different continents, and international p r o g r a m committees guarantees fast exchange and critical review of results, whereas the p a r t i c i p a t i o n o f practitioners in the scientific exchange and o f scientists in practical projects enhances technol- ogy transfer.
A l l these elements can be found i n this volume. Therefore it seems w o r t h w h i l e not merely to distribute it as selected collection of isolated papers but to provide at least a rough and partly subjective m a p of the field as it can be presented i n M a r c h 1992 on the basis of the texts included.
First of a l l we find a clear segmentation into extended abstracts of the invited speakers of the users' f o r u m a n d into full papers to be presented at the scientific w o r k s h o p . T h i s distinction on the one h a n d reflects the "research notes" character of the Lecture Notes in A r t i f i c i a l Intelligence: the m a i n purpose of fast c o m m u n i - cation of original research is captured by these full papers. O n the other h a n d the strongly application oriented character of the field of knowledge acquisition q u a l - ifies short analyses about the need a n d impact of knowledge acquisition ( K A ) in high-tech industries ( A l l a r d1) , project management for K A projects ( K i l l i n ) , the European marketplace for methodological K A (Georges), and an assessment of the industrial use of machine learning ( M o r i k ) as highly valuable s u p p o r t i n g a n d directing evidence to be published together w i t h front end research results.
A s far as research contributions are concerned the E u r o p e a n m a p is increasingly dotted w i t h general modelling approaches w h i c h make up the second section of the book and to a considerable extent are also present i n contributions i n other sections (Jonker, N c u b e r t , D i e n g , A l l c m a n g , Porter). T h e i r c o m m o n u n d e r l y i n g
1 In the preface, contributions arc indicated only by first author's name for the sake of readability.
principle is to be based on an explicit abstract h u m a n l y perceivable model of the expertise to be captured for a knowledge based system. T h e most influential modelling approach presently is K A D S , w h i c h also plays a role i n several of the papers included.
Eriksson generalizes M u s c n ' s w o r k of generating domain-specific knowledge ac- quisition tools. G r c b o v a l presents a method of c o m p i l i n g efficient code from K A D S conceptual models. L i n s t c r complements the K A D S p a r a d i g m o f starting analysis w i t h the problem solving behavior by using tools that equally support the static relations of a d o m a i n as starting point for modelling. G a p p a provides an in-depth comparison of two strategically different approaches to modelling that haye both been pursued over the years: strong modelling approaches (such as her o w n expansions of Puppe's work) require close correspondence between a m o d e l of the d o m a i n a n d of the problem solving behavior a n d allow fast transition to operational representations whereas weak modelling (such as K A D S ) assumes the applicability of problem solving models to deliberate d o m a i n structures. In the latter case the type of connection has to be specified as p a r t of the modelling process. T h e contributions of Schreiber and v a n Heijst specify two aspects w i t h i n the K A D S research p r o g r a m : Schreiber provides a detailed analysis of two s i m i l a r diagnostic problem solving methods a n d how f o r m a l representation allows one to clearly identify distinctions. V a n Heijst describes details a n d theory based tool support for the process of further specifying roughly specified models of p r o b l e m solving ("interpretation models")- F i n a l l y , G c c l e n uses f o r m a l models as a n objec- tive basis for deriving problem solving models from expert protocols.
T h e section on knowledge formalization and automated methods starts w i t h two of the three full papers i n this book about machine learning. T s u j i n o has enhanced the mechanical induction of decision trees by methods of q u a l i t y assessment of resulting trees. Nedcllec has closely coupled m a n u a l acquisition w i t h automated learning in such a w a y that the v a l i d a t i o n a n d maintenance activities based o n new cases become a genuine part of the architecture. Schwcigcr presents a tool based on a logical theory of configuration w h i c h allows automated generation o f knowledge based systems for the respective subclass of applications. J o n k e r a n d N e u b e r t treat special aspects of K A D S based modelling. Jonker's f o r m a l language for K A D S conceptual models emphasizes the aspect of d o m a i n signatures that correspond to the models of problem solving and hence comes close to p r o v i d i n g a bridge between the above weak and strong modelling approaches. N e u b e r t p r o - vides a detailed specification of the activities required to achieve K A D S concep- tual models.
Elicitation and diagnosis of human knowledge ranges f r o m foundations i n theory of science to practical guidelines a n d tools for knowledge acquisition activities.
N w a n a provides a possible rationale for the stepwise justified transition f r o m manifestations of expertise towards models in a wider sense t h a n discussed above.
P o r l m a n demonstrates productive use of the metaphor of thinking for getting ac- cess to those facets of knowledge that are hard to elicit by methods that emphasize a question-response rather than a resolution of conflict view. L a r i c h c v differen- tiates a m o n g several settings in the process of eliciting expert classification k n o w - ledge. T h e next two approaches involve tools for elicitation activities. D i c n g suggests a n architecture where the so far neglected aspects of dealing w i t h multiple experts and of laying the g r o u n d for explanation at the beginning of b u i l d i n g a system arc taken into account. C h a r l c t introduces the a d d i t i o n a l guidance that c a n be made use of when a d o m a i n is k n o w n to be determined by causal relations.
Practice and experiences of knowledge acquisition starts w i t h the subjects o f knowledge base maintenance and consistency checking w h i c h f o r m i m p o r t a n t re- quirements to be met by systems in practice. A c c o r d i n g l y , M a u r e r describes an extension of A l t h o f f s M O L T K E w o r k b e n c h of w h i c h several aspects have already been introduced d u r i n g the previous E K A W s . A l l c m a n g reports on an c v a l u a t o r y study about how one of the early model based approaches - C h a n d r a s c k a r a n ' s generic tasks - is applied by practicing knowledge engineers. Porter reports practical experiences in a p p l y i n g K A D S elements in similar large scale financial applications. T h r o u g h his large-scale experiences w i t h machine learning projects, M a n a g o has arrived at the reported enhanced description of cases i n order to i m - prove the efficiency o f inductive learning as w e l l as to overcome some of its defi- ciencies by case based reasoning. Schmalhofcr's h y p e r m e d i a based support system unifies the two practical needs of p r o v i d i n g easy access to existing industrial case bases a n d of using them i n the development of knowledge based systems. F i n a l l y , B r a d s h a w contributes a large in-house application of modelling business processes for the purpose of better capture of the processes themselves a n d for computerized support of selected functions.
A total of 65 persons from a r o u n d the w o r l d have done a great job in serving as the program committee. T h e i r recommendations and p a r t l y very detailed c o m - ments have helped both w o r k s h o p organizers a n d i n d i v i d u a l authors a great deal to achieve the q u a l i t y that we hope the reader w i l l notice. T h e i r fast a n d reliable responses have allowed us to hold to the planned schedule in almost every detail.
Therefore all the organizers of the conference w o u l d like to express their great gratitude to the colleagues listed below.
Tom Addis Dean AHemang Klaus-Dieter Althoff Nathalie Aussenac John Boose Guy Boy Jeff Bradshaw
U n i v e r s i t y o f R e a d i n g
1st. D a l l e M o l l e d i S t u d i sull'Intell. A r t i f i c , L u g a n o U n i v e r s i t y o f K a i s c r l a u t c r n
A R A M I I H S , Toulouse
Boeing C o m p u t e r Services, Seattle, W A O N E R A - C E R T , Toulouse
Boeing C o m p u t e r Services, Seattle, W A
Ivan Bratko Joost Breuker Clifford A . Brunk Steve M . Easterbrook Henrik Eriksson Rose Dieng Brian R. Gaines Jean Gabriel Ganascia Ute Gappa
Heiner Gertzen Catherine Greboval Thomas G ruber Andreas G (inter Frank van Hamiclen Koichi Hori
Willem Jonker Jonathan Killin Georg Klinker Yves Kodratoff Jean-Paul Krivine Kevin Lano Hubert Lelunann Marc Linster Frank Maurer Riichiro Mizoguchi Katharina Morik Hiroshi Motoda Bernard Moulin Shogo Nishida Susanne Neubert Hyacinth S. Nwana Ray Paton
Karsten Pocck Bruce W . Porter Angeliki Poulymenakou Frank Puppe
Thomas E . Rothenfluh Beate Schlenker Franz Schmalhofer Gabricle Sclimidt Guus Schreiber Johann Schweiger Joacliim Selbig Nigel R. Shadbolt
T h e T u r i n g Institute, G l a s g o w U n i v e r s i t y of A m s t e r d a m
U n i v e r s i t y of C a l i f o r n i a , Irvine, C A U n i v e r s i t y of Sussex, B r i g h t o n
Stanford U n i v e r s i t y School of M e d i c i n e I N R I A - C E R M 1 C S , V a l b o n n e
U n i v e r s i t y of C a l g a r y
Institute Blaise Pascal, U n i v c r s i t c Paris V I K a r l s r u h e U n i v e r s i t y
H o c h s t A G , F r a n k f u r t
U n i v c r s i t c de Technologic de C o m p i c g n e Stanford U n i v e r s i t y , Palo A l t o , C A U n i v e r s i t y of H a m b u r g
U n i v e r s i t y of A m s t e r d a m U n i v e r s i t y of T o k y o P T T Research, G r o n i n g e n T o u c h e Ross, K B S C , L o n d o n
D i g i t a l E q u i p m e n t C o r p . , M a r l b o r o , M A L R I , U n i v e r s i t y of P a r i s - S u d a n d C N R S , O r s a y E . D . F . D i r e c t i o n des Etudes et Rccherches, C l a m a r t L l o y d ' s Register of S h i p p i n g , C r o y d o n
I B M G e r m a n y Scientific Center, Heidelberg
G e r m . Res. Inst. f. M a t h , a n d D a t a P r o c , St. A u g u s t i n U n i v e r s i t y of Kaiserslautern
Institute of Scientific a n d I n d . Res., O s a k a U n i v e r s i t y U n i v e r s i t y of D o r t m u n d
H i t a c h i L t d . , H a t o y a m a S a i t a m a Universite L a v a l , S t c - F o y , Quebec M i t s u b i s h i Electric C o r p o r a t i o n , H y o g o U n i v e r s i t y of K a r l s r u h e
U n i v e r s i t y of Kecle U n i v e r s i t y of L i v e r p o o l U n i v e r s i t y of K a r l s r u h e U n i v e r s i t y of Texas at A u s t i n
T h e L o n d o n School of Economics a n d P o l i t i c a l Science U n i v e r s i t y of K a r l s r u h e
O h i o State U n i v e r s i t y , C o l u m b u s , O H U n i v e r s i t y of F r e i b u r g
U n i v e r s i t y of Kaiserslautern U n i v e r s i t y of Kaiserslautern U n i v e r s i t y of A m s t e r d a m
T e c h n i c a l U n i v e r s i t y of M i i n c h c n
G e r m . Res. Inst. f. M a t h , a n d D a t a P r o c , St. A u g u s t i n U n i v e r s i t y of N o t t i n g h a m
Mildred L . G . Shaw Ingeborg Solvberg Maarten W . van Sonieren Marcus Spicss
Susan Spirgi Werner Stephan Hirokazu Taki Wolfgang Tank Bernd Welz Dieter Wenger Bob J . Wielinga J . Brian Woodward Nora Yue
Manuel Zacklad
U n i v e r s i t y of C a l g a r y S i n t c f D E L A B , T r o n d h e i m U n i v e r s i t y of A m s t e r d a m
I B M G e r m a n y Scientific Center, Heidelberg Swiss B a n k C o r p o r a t i o n , Basel
U n i v e r s i t y of K a r l s r u h e
M i t s u b i s h i Electric C o r p o r a t i o n , K a n a g a w a T e c h n i c a l U n i v e r s i t y of Berlin
U n i v e r s i t y of K a r l s r u h e Swiss B a n k C o r p o r a t i o n , Basel U n i v e r s i t y of A m s t e r d a m U n i v e r s i t y of C a l g a r y U n i v e r s i t y of L o n d o n
O N E R A , G r o u p c d'Intelligence A r t i f i c i c l l c , C h a t i l l o n T o o m a n y more persons a n d institutions have helped to make E K A W 9 2 a success for a l l of them to be mentioned here. Nevertheless I cannot close w i t h o u t t h a n k i n g I B M G e r m a n y and m y manager Peter G r c i s s l for p r o v i d i n g me the liberty to p u t E K A W 9 2 on top of m y agenda whenever necessary. F u r t h e r m o r e I h a d highly efficient and continuous s u p p o r t from Christine Sperling a n d M a r k Beers of I B M a n d from m y students Simone B u r s n c r , R a l f N i i s c , and W o l f r a m S c h m i d t , w h o never hesitated to assist me w h e n things h a d to be ready yesterday.
A n o t h e r special " t h a n k y o u " goes to C h r i s t i n e H a r m s , whose experience in o r g a n - izing conferences took considerable load from m y shoulders a n d prevented me from overseeing i m p o r t a n t details.
Heidelberg M a r c h 1992
T h o m a s Wetter on behalf of the editors
Contents
Technology Transfer
A U s e r ' s V i e w o f C u r r e n t P r a c t i c e a n d Possibilities
F. Allard /
T h e M a n a g e m e n t a n d M a i n t e n a n c e o f a n O p e r a t i o n a l K A D S S y s t e m D e v e l o p m e n t
J. Killin 6
K n o w l e d g e E n g i n e e r i n g T r e n d s i n E u r o p e
M. Georges 7
A p p l i c a t i o n s o f M a c h i n e L e a r n i n g
K. Morik 9
General Modelling Approaches
C o n c e p t u a l M o d e l s for A u t o m a t i c G e n e r a t i o n o f K n o w l e d g e - A c q u i s i t i o n T o o l s
H. Eriksson, M. A. Musen 14
A n A p p r o a c h to O p e r a t i o n a l i z e C o n c e p t u a l M o d e l s : T h e S h e l l A i d e
C . Greboval, G. Kassel 37 L i n k i n g M o d e l i n g to M a k e Sense a n d M o d e l i n g to
I m p l e m e n t Systems i n a n O p e r a t i o n a l M o d e l i n g E n v i r o n m e n t
M. Linster 55
C o m m o n G r o u n d a n d Differences o f the K A D S a n d S t r o n g - P r o b l e m - S o l v i n g - S h e l l A p p r o a c h
U. Gappa, K. Poeck 75
D i f f e r e n t i a t i n g P r o b l e m S o l v i n g M e t h o d s
G. Schreiber, B. Wielinga, H. Akkermans . 95
U s i n g G e n e r a l i s e d D i r e c t i v e M o d e l s i n K n o w l e d g e A c q u i s i t i o n
G. van Heijst, P. Terpstra, B. Wielinga, N. Shadbolt 112
T o w a r d s a F o r m a l F r a m e w o r k to C o m p a r e P r o t o c o l Interpretations a n d T a s k S p e c i f i c a t i o n s
P. Geelen, Z. Ruttkay, J. Treur 133
Knowledge Formalization and Automated Methods
K n o w l e d g e A c q u i s i t i o n D r i v e n b y C o n s t r u c t i v e a n d Interactive I n d u c t i o n
K. Tsujino, V. G. Dabija, S. Nishida 153
K n o w l e d g e R e f i n e m e n t U s i n g K n o w l e d g e A c q u i s i t i o n a n d M a c h i n e L e a r n i n g M e t h o d s
C. Nedellec, K. Causse 171
G e n e r a t i n g C o n f i g u r a t i o n E x p e r t S y s t e m s f r o m C o n c e p t u a l S p e c i f i c a t i o n s o f the E x p e r t K n o w l e d g e
/. Schweiger 191
Y e t A n o t h e r F o r m a l i s a t i o n o f K A D S C o n c e p t u a l M o d e l s
W. Jonker,J. W. Spee 211
T h e K E E P M o d e l , a K n o w l e d g e E n g i n e e r i n g Process M o d e l
S. Neubert, R. Studer 230
Elicitation and Diagnosis of Human Knowledge
D o m a i n - D r i v e n K n o w l e d g e M o d e l l i n g : M e d i a t i n g a n d I n t e r m e d i a t e R e p r e s e n t a t i o n s for K n o w l e d g e A c q u i s i t i o n
H. 5 . Nwana, R. C . Paton, M. /. R. Shave, T. J. M. Bench-Capon 250
P M I : K n o w l e d g e E l i c i t a t i o n a n d D e B o n o ' s T h i n k i n g T o o l s
A / . - M . Portman, S. M. Easterbrook 264 A N e w A p p r o a c h to the S o l u t i o n o f E x p e r t C l a s s i f i c a t i o n
P r o b l e m s
O. 1. Larichev 283
K n o w l e d g e A c q u i s i t i o n for E x p l a i n a b l e , M u l t i - E x p e r t , K n o w l e d g e - B a s e d D e s i g n S y s t e m s
R. Dieng, A, Giboin, P.-A. Tour tier, O. Corby 298
C a u s a l M o d e l - B a s e d K n o w l e d g e A c q u i s i t i o n T o o l s : D i s c u s s i o n o f E x p e r i m e n t s
/. Charlet; J.-P. Krivine, C. Reynaud 318
Practice and Experiences of Knowledge Acquisition
K n o w l e d g e Base M a i n t e n a n c e a n d C o n s i s t e n c y C h e c k i n g in M O L T K E / H y D i
F. Maurer 337
A c q u i r i n g K n o w l e d g e o f K n o w l e d g e A c q u i s i t i o n : A S e l f - S t u d y o f G e n e r i c T a s k s
D. Allemang, T. E. Rothenfluh 353
R e u s a b l e A n a l y s i s a n d D e s i g n C o m p o n e n t s for K n o w l e d g e - B a s e d S y s t e m D e v e l o p m e n t
D. Porter 373
A c q u i r i n g D e s c r i p t i v e K n o w l e d g e for C l a s s i f i c a t i o n a n d I d e n t i f i c a t i o n
M. Manago, N. Conruyt, J. le Renard 392
Intelligent D o c u m e n t a t i o n as a C a t a l y s t for D e v e l o p i n g C o o p e r a t i v e K n o w l e d g e - B a s e d S y s t e m s
F. Schmalhofer, T. Reinartz, B. Tschaitschian 406
c Q u a l i t y : A n A p p l i c a t i o n o f D D u c k s to Process M a n a g e m e n t
/ . M . Bradshaw, P. Holm, O. Kipersztok, T. Nguyen 425
A User's view of current practice and possibilities
Francois A i l a r d European Space Agcncy(ESA)
Kcplerlaan 1 220 A G Noordwijk, N L
1.0 Introduction
E S A is funding currently a n u m b e r of knowledge-based system ( K B S ) projects.
A l l of them are confronted to the knowledge acquisition ( K A ) p r o b l e m . T h e i n - tention of this paper is to present current practice and possibilities as can be found in various projects a n d to highlight issues and unfulfilled needs. T h i s report is broken d o w n into four m a i n parts:
• the practice of K A across the K B S L i f e - C y c l e illustrated b y various projects
• T h e particular role of technical documentation and its relation w i t h knowledge acquisition
• T h e emergence of reusable conceptual models
• Issues and Needs 1.1 K B S Life-Cycle and K A
K n o w l e d g e A c q u i s i t i o n across the L i f e - C y c l e will be decomposed i n a set of tasks and viewed in the perspective of a p a r t i c u l a r project. These tasks arc: knowledge identification, knowledge elicitation, knowledge editing, computer-supported knowledge acquisition.
1.2 E X A C T and Knowledge Identification:
E X A C T is a model-based diagnostic system meant to support the task of satellite diagnostic a n d p l a n n i n g . A long part of the project, w h i c h is still ongoing has been used to actually identify the knowledge required and adopt the correct represen- tation format. It went from rule-based to model-based diagnostic through several steps. Once this was defined, the knowledge identification task c o u l d be performed m u c h more easily. A l t h o u g h this is w h a t should stay an atypical case, m a n y les- sons c a n be derived:
• starting w i t h a pre-conceived idea of w h a t the knowledge should look like creates m a n y unnecessary problems.
• at the same time, predefined conceptual models do help i n the identification task. A c t u a l l y , one m u s t remember that i n the K A task one is looking for
knowledge that is useful, i.e. operating knowledge. There experience w i t h m a n y conceptual models helps considerably.
• no formal means of knowledge identification beyond test cases is used and could be useful.
1.3 The Battery Management System and knowledge elicitation
T h i s system is one of the first developed at E S A . It concerns the s u p p o r t to the problem of managing batteries of in-orbit satellites, basically a n a l y z i n g historical trends to predict battery output. T h e system was successfully developed to a level of prototype. K n o w l e d g e elicitation was used w i t h interview techniques as they are in m a n y of our projects. T h e questions raised were:
• how can the output from such sessions be m a x i m i z e d . N o clear answers are available today. A complete interviewing methodology is still to be defined.
• how to qualify (or train) experts themselves for such tasks. T h e y arc not a l l adequate.
1.4 M A R S and Knowledge editing
M A R S is a scheduling tool. Its m a i n K A module is composed of a sophisticated object editor. T h e M A R S case is interesting because object editors are becoming standard features of K B S s . U s i n g the tool, w i t h certainly one of the best object editor available at E S A , has shown several aspects:
1. It is possible to standardize the characteristics of an object editor (context sensitive help, local verification, etc.) a n d that s t a n d a r d i z a t i o n should be p u r - sued.
2. O n the other h a n d Object editors impose an editing environment that is heavy.
Users actually often bypass the object editor a n d write directly in K B files.
N o r speed, nor lack of functionality can be here questioned. In this system for reason of coherence of attribute values, the knowledge editor is actually re- q u i r e d . W e have found that it is the format of the frame editor w h i c h seems to be undesired. W h a t users wanted is actually that the K B page provides them w i t h the same functionalities of the object editor while being able to scroll a n d have constantly at their disposal the detailed view of other objects.
T h i s indeed requires very sophisticated functionalities. It could be argued that users should adapt a n d they do. B u t the lesson to us is that K A functions s h o u l d always allow a very easy j u m p between the local view a n d the global v i e w of a knowledge base, in a w a y m i m i c k i n g the easiness w i t h w h i c h persons actually d o that mentally.
1,5 P R E V I S E and the development of knowledge acquisition module
P R E V I S E is a system under development meant to support the edition a n d checking of crew procedures and meant to support the evolution of the procedure concept a n d of the operations knowledge. S y n t a x , vocabulary, edition rules, o p - erational knowledge, all is there represented as objects. It can be said that P R E - V I S E knowledge acquisition features arc the first example, at E S A , of full b l o w n knowledge acquisition modules, except for machine learning w h i c h is not used at E S A for now.
1. T h e K A module is actually part of the functions of the system a n d is intended for s u p p o r t i n g the evolution of concept a n d knowledge.
2. Test cases defined include evolution test cases.
3. T h e K A module does not rely on the shell's o w n ( P r o K a p p a ) w h i c h is viewed as insufficient.
4. T h e procedure syntax is variable a n d since the K B on each procedure is b u i l t by " c o m p i l i n g " a procedure's text, it means that:
a. T e x t (a very specialized one) is used here as a knowledge acquisition fea- ture
b. W h e n the syntax is changed the compiler has to change w h i c h implies the user is given a means to change it ( Y A C C ) .
5. there is a syntax editor, a v o c a b u l a r y editor, a procedure editor as well as a general knowledge editor.
6. the H C I of the K A acquisition module was built using O p e n Interface, a H C I p r o t o t y p i n g tool. It allowed to defined desirable features such as object filters, objects trees, template based editors, dictionaries, etc. W e can say that this p r o t o t y p i n g approach is here very beneficial because it allows experts a n d us- ers to understand the K A process.
2.0 Technical Documentation
F o r a n organization such as E S A , documentation is a w a y of life, not to say the m a i n product. Tremendous quantities of knowledge arc stored in documents i n a very passive form. In order to exploit better this knowledge E S A has begun two parallel efforts:
1. the evaluation of text analysis tools such as K - S t a t i o n , based o n the K O D a p p r o a c h . K - S t a t i o n is a very interesting tool. T h e first one really c o m m e r - c i a l l y available based on a solid methodology. O u r evaluation concluded that:
• it does support the task of text analysis whether it is used for interview transcripts or for existing d o c u m e n t a t i o n ;
• the concepts manipulated (actions, objects, etc.) are coherent a n d w e l l - defined
• it allows to slowly accumulate knowledge w h i c h can easily transformed into a knowledge base
• it has a few technical d r a w b a c k s like the concept of m u l t i - v a l u e d attributes is not supported and the notion of instances can o n l y be a w k w a r d l y i n - troduced; it docs not support the analysis o f graphics.
• its user interface suffers from the lack of hypertext capabilities
• the aggregation of different analysis versions (improvements) of the a n a l - ysis of the same text is not really supported.
• F i n a l l y and mostly the process is heavy. It is felt that the cognitive process by w h i c h text is analyzed is only partially matched by K - S t a t i o n . T h e r e is a very i m p o r t a n t w o r k equivalent to understanding the structure o f the document a n d the key concepts (building a semantic network) w h i c h is not supported by K - S t a t i o n . T h u s we t h i n k K - S t a t i o n is a limited tool but w h i c h in some cases can really be of use. A consultancy s t u d y w i l l later this year assess its use for requirements analysis purposes.
2. the development of a n u m b e r of tools to support the life-cycle of systems a n d the support the slow gathering of formalized knowledge; these are:
• P R E V I S E already mentioned, i n w h i c h it is foreseen to slowly acquire the operational knowledge used in procedure w r i t i n g and verification.
• T h e R A M S ( R e l i a b i l i t y , A v a i l a b i l i t y , M a i n t a i n a b i l i t y a n d Safety) i n i t i - ative where in particular cases o f problems a n d failures arc being a n d w i l l be stored for use in hazard analysis,
• F S D K O where the transferred of satellite knowledge from development to operations is p a r t i c u l a r l y considered a n d supported v i a a set o f
Hypertext-based tools.
These arc very i m p o r t a n t developments because they are the first explicit at- tempts to develop tools before the actual knowledge on new space systems (under development) is available a n d written i n documents. K n o w l e d g e A c - quisition is seen here as a w a y to support knowledge formalization (and reuse).
These efforts still remain s m a l l a n d can only be considered as experimental.
3.0 Conceptual Models
T h e experience gathered at E S A allows to safely say that one of the very real w a y out of the K A bottleneck is the use of pre-defined conceptual models. T h e term is used here in a loose definition of a pre-defined model of task and d o m a i n de- scription, possibly using a predefined knowledge representation.
F o r example, this is true of the C o l u m b u s F D I R ( F a i l u r e Detection, Isolation a n d Recovery) w h i c h uses a pre-defined diagnostic conceptual m o d e l , simple a n d r o - bust enough to be brought o n - b o a r d . K A is then reduced to its simplest ex- pression. T h e experience has been repeated successfully with C O M P A S S , a system for the diagnostic of payloads. C u r r e n t l y an effort is made to standardize the approach to knowledge-based scheduling tools in the same spirit of p r o v i d i n g a reusable conceptual models
4.0 Conclusion: Issues and Needs
T h e experience so far of knowledge acquisition has shown that there are still m a n y issues and needs not really solved or addressed. M a n y are mentioned above. T h e following is a complementary list w h i c h concludes this report:
1. T h i s concerns the distance between the expert language a n d the knowledge engineering language. A l t h o u g h , conceptually, experts clearly manipulate classes, objects, attributes, instances, etc. they do not usually manipulate such abstract notions. W e have found that abstraction, although necessary for K B S development, is not useful as a tool. T h i s m a y seem a t r i v i a l problem but we t h i n k it is not. First it is a real one, and although it can be helped with some training of the expert or user, this training has to be done every time. B u t even more i m p o r t a n t is that knowledge representation ( K R ) s h o u l d respect expert terminology a n d methods of representation. K A modules provide seldom to- d a y support for basic conceptual tools, used by experts such as decision trees, network diagrams, matrices a n d tables a n d general graphics not to mention text analysis support.
2. Experts and documentation use formalization as a tool not as an end, like i n A I . T h i s means that the flexibility w i t h w h i c h they use conceptual tools such as types, classes, tables, graphs and text is m u c h greater than the one offered by K A tools and K R in general. M o r e o v e r , the K A process increases formalization of knowledge beyond w h a t experts a n d users arc f a m i l i a r w i t h , find comfortable, useful, to the extent that they might oppose it. T h i s was identified as one of the m a i n problems a n d it is still not clear how it can be solved.
3. In order to greatly facilitate the knowledge acquisition process emphasis should be given to text analysis support, character recognition systems, the development of specific recognition tools for tables, drawings, etc. and the de- finition of a nomenclature in graphics. These m a y seem m u n d a n e problems but the goal is here to hide the internal representation layer that the machine uses. In the same spirit considerable efforts must be made so that K A modules arc defined, labelled and coded in terms of the used knowledge.
4. Increasing the variety and a v a i l a b i l i t y of conceptual models at various levels is one w a y to solve the K A bottleneck: we have s h o w n these are used for that purpose, sometimes inappropriately. T h u s w h a t is needed is the following:
• define the criteria for m a t c h i n g conceptual models w i t h problems
• define a p r o g r a m of training on various conceptual models.
• define a structure for m a n a g i n g these repositories.
There might be considerable k n o w - h o w i n conceptual models a n d competition is never far away. T h e proposition is here to allow them to be public (for a fee). E S A is currently w o r k i n g on a p l a n to define a knowledge repository of knowledge on E u r o p e a n satellite payloads. T h i s p r o g r a m of w o r k can only succeed if some of the issues addressed above arc solved.
The management and maintenance of an operational K A D S system development
Jonathan Killin
The Knowledge-Based Systems Centre, Touche Ross Management Consultants London, U . K .
A b s t r a c t . The Barclaycard Fraudwatch knowledge-based system identifies potentially fraudulent transactions on Visa and Mastercard credit cards. It makes use of two generic tasks, Select and Assess. Select reduces the number of transactions to be considered from about one million to about fifty thousand.
Assess further processes these fifty thousand to arrive at a set of about four hundred accounts, which are output to fraud operations staff for further action. O f these four hundred, about fifteen will eventually turn out to be fraudulent.
Fraudwatch was developed by Touche Ross and Barclays, using the K A D S method for knowledge-based system development. It is fully embedded within the Barclaycard suite of business and cardholder programs running on I B M mainframes. Select is largely written in C O B O L , Assess is written in KnowledgeTool™, and the peripherals and communications are written in J C L . Fraudwatch runs daily in batch mode. It saves about thirty percent of pre-status fraud losses on the products to which it is applied.
A n initial implementation went live in November 1990. The system has undergone two cycles of knowledge maintenance, in order to adapt performance to organisational changes and changes in business objectives, and to enhance overall performance. Further systems and enhancements are planned during 1992 and 1993, which are intended to deploy Fraudwatch at earlier stages of transaction processing and to look at devolving processing.
Knowledge Engineering Trends in Europe
M a r i Georges Cap Gemini Innovation
118, rue de Tocqueville 75017 Paris
e-mail: mari@crp.capsogcti.fr
T h e industrial w o r l d is attracted by knowledge-based techniques. A certain n u m - ber of increasingly pertinent a n d c o n v i n c i n g experiments have been carried out.
B u t for some reason, K B S technology just doesn't seem to take i n c o m m e r c i a l culture. W h a t ' s wrong?
O n the one h a n d , research has attacked a n d succeeded in controlling one of the p r i n c i p a l technical issues: the so-called knowledge-acquisition bottleneck. O n the other h a n d , commercial software product developers have p u t a m y r i a d of expert system generators at the market's disposal. B u t one can say that both approaches are incomplete. S c h e m a t i z i n g :
• T h e former doesn't go far enough practically: the K A methods arc not c o m - prehensive as (i) they d o n ' t cover the full application development life cycle, (ii) they don't sufficiently address quality aspects, (iii) there is not computer-based support i n the form of integrated tools.
• T h e other is superficial: due to encouraging ad hoc, specificity-driven develop- ment, it does not go far enough i n the generalization of basic methods a n d techniques permitting scale-up to p r o d u c t i o n q u a l i t y elements: it passes u p opportunities both on reuse aspects and on exploiting the basic principles o f the A I a p p r o a c h .
T h e latter a p p r o a c h m a y not appeal to the purists, but it certainly pleased (some of) the commercial w o r l d (for awhile). It offered fast, visual results. B u t taken i n the long t e r m , it broke d o w n due both to its lack of generality a n d reusability, a n d to its unsatisfactory q u a l i t y as a modern development a p p r o a c h . A s a result, the E u r o p e a n c o m m e r c i a l w o r l d (and it appears, increasingly in the U S ) is shifting its interest towards methodologically-sound, comprehensive approaches - something that makes the junction between the KA research results and the expert system shell.
T h e E u r o p e a n m a r k e t is l o o k i n g for a method, supported by tools. T h e method must respond to all of the concerns of conventional software engineering, but even more rigourously due to the perceived complexity a n d nebulousness of K B S - a n d to their perturbation potential i n the organizational fabric - a n d due to the i n -
creased expectations of a more informatics-mature market. It must be comprehen- sive, addressing (i) V&V: b u i l d i n g the right p r o d u c t {adequacy and appropriateness) a n d building it right {reliability), (ii) evolution: b u i l d i n g for change, m a i n t a i n a b i l i t y , (iii) interoperability: w i t h conventional software for hy- brid system development, (iv) life-cycle vision and control: process m o d e l l i n g , project management a n d facilities for cooperative (team)work. It must offer computer- based tool support for h a n d l i n g knowledge complexity and v o l u m e , a n d for leading into i m p l e m e n t a t i o n .
T h e offer a r o u n d the method a n d s u p p o r t i n g tools m u s t also be complete, r a n g i n g from awareness ' c a m p a i g n s ' through seminars a n d training courses to technical consultancy.
T h e E u r o p e a n C o m i s s i o n ' s E S P R I T p r o g r a m m e has been a n d is f u n d i n g v a r i o u s projects that address these K E issues, the most renowned K A D S1. T h e i n d u s t r i a l element of such a projects' consortia propel tangible results onto the m a r k e t place:
"original K A D S " application have numbered over 30, a p p r o x i m a t e l y 20 o f w h i c h have been carried out under the auspices o f C a p G e m i n i I n n o v a t i o n . K A D S industrial-quality competition i n E u r o p e is at the methodological level a n d is p r a c - tically limited to S K E (Bolesian) a n d K O D ( C i s i Ingenieric), while expert system shells are seen solely as upstream experimentation aids.
Nevertheless, it is clear that while the K A D S methodology is the most a d v a n c e d to date, it requires enhancement towards
• comprehensiveness: this is the role o f the K A D S - I I project2.
• industrialization: this is being addressed by members of the K A D S - I I consor- t i u m w h o are m a k i n g available seminars a n d courses ( C a p G e m i n i I n n o v a t i o n , L l o y d ' s Register, T o u c h e Ross M a n a g e m e n t C o n s u l t a n t s , ...), methodological guides (e.g. C a p G e m i n i Sogeti's " P E R F O R M - K B A " ) , and tool s u p p o r t (e.g.
I L O G / C a p G e m i n i Innovation's " K A D S - T O O L " ) for the present, o p e r a t i o n a l version of K A D S , responding to strong market d e m a n d .
1 T h e partners of K A D S project ( E S P R I T P1098) were S T C Tecnology L t d . ( U K ) , C a p Scsa Innovation (F), N T E Neutcch G m b H (D), S D - S c i C o n L t d . ( U K ) , Touche Ross Management Consultants ( U K ) , University of Amsterdam ( N L ) .
2 T h e partners of the K A D S - I I project ( E S P R I T P5248) arc S T C Tecnology L t d . ( U K ) , C a p Sesa Innovation (F), N T E Neutech G m b H (D), S D - S c i C o n L t d . ( U K ) , T o u c h e Ross Management Consultants ( U K ) , University of Amsterdam ( N L ) .
Applications of Machine Learning
K a t h a r i n a M o r i k University Dortmund, Dept. Computer Science, LS VIII, P . O . B o x 500 500, 4600 Dortmund 50, e-mail: morik@kilo.informatik.uni-dortmund .de
A b s t r a c t . During the last 10 years, machine learning has been successfully applied. Most often, the applications are confidential. Therefore, only few publications about real world applications exist. In this paper, an overview of machine learning applications is given with their scenarios. Some typi- cal applications are described. T h e n , future directions of machine learning applications are proposed. It is argued that machine learning is now ma- ture enough to be incorporated into standard systems as well as algorithms.
T h e integration of learning modules into database and retrieval systems is one of the trends. Another trend is to automatically select an appropriate learning tool out of a toolbox. T h e third trend, which is even more challeng- ing, no longer requires a distinguished learning module, but offers methods of machine learning to be applied by programmers in their regular system development. Software engineers of the future can use inductive techniques as they now use message passing, for instance. T h e n , any program can be enhanced by some learning ability.
1 Experience with Machine Learning
In the past 10 years, machine learning ( M L ) h a d several applications of two types of a l g o r i t h m s , n a m e l y
— t o p - d o w n i n d u c t i o n of decision trees,
a f a m i l y of algorithms f r o m which I D 3 [ Q u i n l a n , 1983] is the most famous one.
— conceptual clustering,
a f a m i l y of algorithms f r o m which A Q [Michalski and Stepp, 1983] is the most f a m o u s one. T h e first break-through of a p p l y i n g machine learning was achieved by e x p l o i t i n g conceptual clustering for the b u i l d i n g of a rule base on soy bean diseases [Michalski and C h i l a u s k i , 1980].
B o t h a l g o r i t h m s learn f r o m examples w h i c h are represented by attribute values.
C u r r e n t research enhances the algorithms to deal w i t h relations [ Q u i n l a n , 1990]
a n d restricted first-order predicate logic [Michalski and Stepp, 1983]. Other, logic- oriented approaches have been developed, w h i c h use background knowledge for learn- i n g a n d even learn the background knowledge itself [ M o r i k , 1987], [ M o r i k , 1990], [Kietz a n d W r o b e l , 1991], [Bisson, 1991], [Muggleton and Feng, 1990]. T h e new, more powerfull algorithms are not yet products on the market. So, the following descrip- t i o n of applications refers to learning from attribute-value representations. A lot of knowledge processing can be performed using this representation [Morales, 1990].
T h e r e are two scenarios for a p p l y i n g M L :
- p a r t i a l l y b u i l d i n g up the rule base of an expert system:
a software house applies M L in order to solve customers' problems more effi- ciently , or
a company or p u b l i c institution uses an expert system shell w i t h an integrated M L tool;
- finding an o p t i m a l procedure:
using M L technology, on the grounds of experience (either i n the f o r m of case d a t a or i n the f o r m of interviews), a procedure for decision m a k i n g or a p l a n for a w o r k i n g routine is developed which is then used by experts.
Several applications are of the first type. For instance, D o n a l d M i c h i e reports two applications of this type [Michie, 1989]. W i t h the help of E x p e r t E a s e , a system w i t h an integrated learning module, 3 000 rules which m o d e l the design of a gas-oil separa- tor could be acquired for B P . Westinghouse, by using E x p e r t E a s e achieved increased t h r o u g h p u t i n an i m p o r t a n t factory to the extent of increasing business v o l u m e by more t h a n ten m i l l i o n dollars per a n n u m . In b o t h applications, the g a i n was achieved by a p p l y i n g a knowledge-based technique w h i c h , i n t u r n , was enabled a n d became appropriate w i t h respect to the cost-benefit relation because of a learning m o d u l e . B r a i n w a r e G m b H , a G e r m a n software house w i t h a particular expertise i n M L a n d neural networks, reports an application for Siemens [Brainware G m b H , 1990]. T h e expert system called B M T configures fire detection equipment. B M T d r a s t i c a l l y cuts a d m i n i s t r a t i v e efforts thus reducing the t i m e required for m a k i n g quotations a n d processing orders. T h e large number of such applications can be seen f r o m the fact that some companies make their l i v i n g b y M L technology. For instance, the c o m p a n y ISoft at Paris, funded i n 1988, now already has a turn-over of 1 m i l l i o n D o l l a r s .
However, M L cannot be reduced as a means to b u i l d - u p rule bases. In m a n y cases, the customers need a decision tree as a k i n d of a check list. W h e r e it is k n o w n how to carefully analyze a l l features of a situation, under some circumstances the expert does not have the t i m e to do so. A quick decision based on only a few features w h i c h are easy to determine, is necessary. A n example is the i m m e d i a t e help needed by n e w b o r n children w i t h a yellow skin colour. T h e u n k n o w n expertise is: w h i c h features correctly indicate a particular test to be necessary? I n w h i c h order are w h i c h tests necessary? M L analysis of cases can provide a check list to be used by the doctor 1. A n o t h e r example of this type is reported by D o n a l d M i c h i e [Michie, 1989]. A p i l o t of a space shuttle has to decide whether to use the autolander or not. In some u n o b v i o u s cases, m a n y factors have to be taken into account. A s the t i m e for decision m a k i n g is too short to do so, the few indicating factors have to be found out. These m a y then serve as a check list. If there are still too m a n y factors, a system can propose the decision.
1 O f course, the doctor keeps in charge and the check list does not prescribe any procedure!
However, doctors are often overloaded with work and cannot be specialists of all diseases.
A s the child dies if a particular diagnosis is missed, this diagnosis needs to be excluded immediately.
2 Future Directions of Machine Learning
2 . 1 I n t e g r a t i n g M a c h i n e L e a r n i n g i n t o S t a n d a r d E n v i r o n m e n t s
F o r several years, M L programs suffered from their dependency of A I computer e n v i - r o n m e n t s . M L tools were either stand-alone programs or integrated into expert sys- t e m s w i t h o u t access to the companies' conventional programs. T h i s was an obstacle p r e v e n t i n g companies from a p p l y i n g M L . C u r r e n t M L programs overcome this obsta- cle. F o r instance, the expert system shell T W A I C E 2 has access to database systems.
A l s o K E T 3 has access to databases. T h e system analyzes databases and detects reg- ularities w h i c h can be displayed as descision trees or If-Then-rules. T h i s way of ob- t a i n i n g knowledge on the basis of given databases is also the success of R U L E A R N 4. T h e concentration of acids could be determined successfully by this system inspect- i n g an already existing database. G i v e n the s i t u a t i o n that m a n y databases exist, w h i c h do no longer correspond to their documented d a t a base schema, or which are not well documented at a l l , the analysis of databases becomes i m p o r t a n t . T h e t e r m 'database m i n i n g ' illustrates the s i t u a t i o n . M L techniques are a suitable means to p e r f o r m database m i n i n g . I n particular, the understandability of the learning results helps database managers to determine what to do about the databases. T h e y m a y even t u r n ' d a t a m i n e s ' into effective knowledge bases. M L techniques can become the m i s s i n g l i n k between the conventional database technology a n d knowledge-based systems.
2.2 M u l t i s t r a t e g y L e a r n i n g
T h e use of just one learning m o d u l e cannot cover a l l the applications. Instead of t r y i n g to find the one universal learning mechanism, the current trend is to develop specialized learning a l g o r i t h m s . T h e n , the user or even a system selects the appro- priate a l g o r i t h m for a p a r t i c u l a r problem. Founded by the E u r o p e a n C o m m u n i t y ( E S P R I T P 2 1 5 4 ) , the project " M a c h i n e Learning T o o l b o x " is currently developing such a s y s t e m . I n the U n i t e d States of A m e r i c a , the first conference on m u l t i s t r a t - egy learning indicates a s i m i l a r trend [Michalski, 1991]. T h i s trend can already be observed i n industry. For instance, a combination of neural network learning a n d i n - ductive learning was applied by Brainware G m b H to classify 22 000 complex signals, each c o n t a i n i n g 8 192 n u m e r i c a l features. A p p l y i n g just one a l g o r i t h m to this huge a m o u n t of d a t a is not feasible. However, seperating a preprocessing step w h i c h m a y e x p l o i t one strategy a n d then a p p l y i n g different learning strategies to the resulting compressed d a t a allowed for cross-validating the results. T h e results were 20-25 per- cent better t h a n those achieved by statistical techniques [Brainware G m b H , 1991].
2 . 3 I n d u c t i v e P r o g r a m m i n g
Since l e a r n i n g i n (restricted) predicate logic has been better understood, the pos- s i b i l i t y of i n d u c t i v e logic p r o g r a m m i n g is now given. T h e first approach into t h a t
2 registered trademark of Nixdorf Computer A G , now SNI
3 K E T is a product of Brainware G m b H , running on P C and compatibles.
4 R U L E A R N is a product of K r u p p Technologie-Transfer G m b H at Duisburg, Germany.
d i r e c t i o n was S h a p i r o ' s debugging m e t h o d for P r o l o g p r o g r a m s [Shapiro, 1983]. A p r o g r a m m i n g e n v i r o n m e n t can use i n d u c t i v e m e t h o d s i n order to s u p p o r t the P r o l o g p r o g r a m m e r . Q u i c k e r p r o g r a m development and easier debugging becomes possible u s i n g i n d u c t i v e l e a r n i n g techniques.
However, the real challenge w h i c h is not yet realized is to teach software engineers such that they can e x p l o i t basic a l g o r i t h m s in whatever they p r o g r a m . In fact, any s y s t e m i n c o r p o r a t i n g conventional a l g o r i t h m s can be enhanced by i n t r o d u c i n g l e a r n i n g capabilities into these a l g o r i t h m s . A learning text e d i t o r , a l e a r n i n g database m a n a g e m e n t s y s t e m , a l e a r n i n g h u m a n - c o m p u t e r interface, a l e a r n i n g knowledge a c q u i s i t i o n s y s t e m , a l e a r n i n g scheduling p r o g r a m s h o u l d be superior to any other s y s t e m of the same type. T h i s is the goal of i n t r o d u c i n g M L into everyday life of software industries. It is a far reaching goal, but it m a y guide our t h i n k i n g a b o u t M L a n d current research.
References
[Bisson, 1991] Bisson, G . (1991). Learning of rule systems in a first order representation.
Rapport de Recherche 628, L R I , University de Paris-Sud.
[Brainware G m b H , 1990] Brainware G m b H (1990). News from Companies - Germany.
page 10.
[Brainware G m b H , 1991] Brainware G m b H (1991). Machine Induction on Complex Sig- nals. Pragmatica, page 14.
[Kietz and Wrobel, 1991] Kietz, J . - U . and Wrobel, S. (1991). Controlling the Complexity of Learning in Logic through Syntactic and Task-Oriented Models. In Muggleton, S., editor, Proc. of Int. Workshop on Inductive Logic Programming, pages 107 - 126, V i a n a de Castelo, Portugal. Also available as Arbeitspapiere der G M D No. 503.
[Michalski, 1991] Michalski, R. (1991). Inferential Learning Theory as a Basis for M u l t i - strategy Task-Adaptive Learning. In Michalski and Tecuci, editors, Multistrategy Learn- ing. George Mason University, U S A .
[Michalski and Chilauski, 1980] Michalski, R. and Chilauski, R. (1980). Learning by Being T o l d and Learning from Examples: A n Experimental Comparison of the T w o Methods of Knowledge Acquisition in the Context of Developing an Expert System for Soybean Disease Diagnosis. Int. Journal of Policy Analysis and Information Systems, 4(2): 125 - 161.
[Michalski and Stepp, 1983] Michalski, R. S. and Stepp, R. E . (1983). Learning from O b - servation: Conceptual Clustering. In Michalski, R., Carbonell, J . , and Mitchell, T . , edi- tors, Machine Learning, volume I, pages 331 - 363. Tioga* Palo Alto, C A .
[Michie, 1989] Michie, D . (1989). New Commercial Oppertunities Using Information Tech- nology. In Brauer, F . , editor, Wissensbasierte Systeme, number 227 in Informatik Fach- berichte, pages 64-71, Berlin, Heidelberg, new York, Tokio. Springer.
[Morales, 1990] Morales, E . (1990). T h e Machine Learning Toolkit Database. Deliverable T I - M L T - 5 . 5 , T h e Turing Institute, Glasgow, U K .
[Morik, 1987] Morik, K . (1987). Acquiring Domain Models. Intern. Journal of Man Ma- chine Studies, 26:93-104. also appeared in Knowledge Acquisition Tools for Expert Sys- tems, volume 2, J . Boose, B . Gaines, eds., Academic Press, 1988.
[Morik, 1990] Morik, K . (1990). Integrating manual and automatic knowledge aquisition - B L I P . In M c G r a w , K . L . and Westpha!, C . R., editors, Readings in Knowledge Acqui- sition - Current Practices and Trends, chapter 14, pages 213 - 232. Ellis Horwood, New York.
[Muggleton and Feng, 1990] Muggleton, S. and Feng, C . (1990). Efficient induction of logic programs. In Proceedings of the 1th conference on Algorithmic Learning Theorie.
[Quinlan, 1990] Quinlan, J . (1990). Learning Logical Definitions from Relations. Machine Learning, 5(3):239 - 266.
[Quinlan, 1983] Quinlan, J . R. (1983). Learning Efficient Classification Procedures and Their Application to Chess E n d Games. In Michalski, R., Carbonell, J . , and Mitchell, T . , editors, Machine Learning - An Artificial Intelligence Approach, pages 463 - 482.
Tioga, Palo Alto, C A .
[Shapiro, 1983] Shapiro, E . Y . (1983). Algorithmic Program Debugging. A C M Distin- guished Doctoral Dissertations. T h e M I T Press, Cambridge, Mass.
This article was processed using the M j p C macro package with L L N C S style
Conceptual Models for Automatic Generation of Knowledge-Acquisition Tools
Henrik Eriksson* M a r k A . M u s e n Medical Computer Science Group
Knowledge Systems Laboratory Stanford University School of Medicine
Stanford, C A 94305-5479, U.S.A.
A b s t r a c t . Interactive knowledge-acquisition ( K A ) programs allow users to en- ter relevant domain knowledge according to a model predefined by the tool de- velopers. K A tools are designed to provide conceptual models of the knowledge to their users. Many different classes of models are possible, resulting in different categories of tools. Whenever it is possible to describe K A tools according to explicit conceptual models, it is also possible to edit the models and to instan- tiate new K A tools automatically for specialized purposes. Several meta-tools that address this task have been implemented. Meta-tools provide developers of domain-specific K A tools with generic design models, or meta-views, of the emerging K A tools. The same K A tool can be specified according to several alternative meta-views.
1 Introduction
Numerous knowledge-acquisition ( K A ) tools have been implemented i n research l a b - oratories. F r o m a research point of view, implementations of tools provide the means to test K A models and methods i n realistic situations. In many ways, the creation of these K A tools is still a research issue, and sometimes is an art. Nevertheless, it is desirable to classify and understand more clearly the principles behind various K A tools so that we can, for instance, outline new generations of tools.
One way of classifying K A tools is to group them according to the conceptual model that they present to their users [25]. A conceptual model, i n this context, is the metaphor for the user interaction; for instance, it is the way i n which knowledge is entered, edited, and presented. (We distinguish such conceptual models for K A tools from conceptual domain models, which describe the experts' view of the d o m a i n and the relevant d o m a i n knowledge.) Examples of conceptual models include symbol- level, method-based, and task-based conceptual models. K A tools supporting s y m b o l - level conceptual models are concerned w i t h rules, objects, and other symbol-level entities [28]. Tools adopting method-based conceptual models present a m o d e l of a particular problem-solving method—for example, methods for classification, p l a n n i n g ,
*On leave from the Department of Computer and Information Science, Linkoping University, S- 581 83 Linkoping, Sweden
Knowledge
engineer Meta-tool
D o m a i n
expert K A tool
E n d user A p p l i c a t i o n
F i g u r e 1: T h e fundamental basis of meta-tools. T h e knowledge engineer uses a m e t a - t o o l to develop a specialized K A tool that is used by a domain expert to enter know- ledge for an application system.
or synthesis. Tools providing task-based conceptual models are tailored to a specific task i n a particular domain.
T h e tradeoff between general tools that can be used i n a broad variety of situations a n d highly supportive, specific tools is a classical software-engineering problem that applies to K A tools as well as to conventional computer programs. (In fact, this tradeoff is a general engineering problem, too.) Meta-tools are designed to provide the means to escape from this d i l e m m a by making it easy to produce and custom-tailor new K A tools. A s depicted i n Figure 1, knowledge engineers can use meta-tools to create new K A tools suited to their particular needs; the K A tools i n t u r n are used by experts to create knowledge bases. In particular, meta-tools are useful for development of domain-oriented K A tools that incorporate task-based conceptual models, since the range of domain-specific models is large.
A s we shall see, the use of meta-tools uncovers a different set of problems at the metalevel, which is the tool-design and K A - m o d e l level, because the task of a K A tool quite different from the task of the application system. There are many ways in which a target K A tool can be described, and there also are m a n y ways i n w h i c h K A tools can be specified (i.e., described i n such detail that they can be implemented automatically). T h i s article discusses a number of such meta-views (or conceptual models for meta-tools), and describes the strengths and weaknesses of each. Different implementations of meta-tools are presented according to the meta-views that the meta-tools support, and the meta-tools are categorized according to their meta-views.
T h i s article is structured as follows. Section 2 provides the background i n terms of conceptual models of software. Section 3 describes several meta-views. Section 4 compares these meta-views and discusses approaches to combine and improve t h e m . F i n a l l y , a summary and conclusions are given i n Section 5.
2 Background
T h e purpose of this section is to provide the general ideas behind automated generation of K A tools and to summarize different conceptual models for K A tools.
2.1 A u t o m a t i c Generation of K A Tools
T h e major impediment for designers of specialized K A tools, including those adopting task-based conceptual models, is the problem of implementing such tools for a broad variety of domains at a reasonable cost, In principal, the problem can be approached in several ways. One approach is to try to find the "right" level of generality for K A tools (i.e., to balance the cost of implementation and the level of support supplied by the tool). Note that it is not clear that such an o p t i m a l level exists. E v e n if such a level was proposed, it might be difficult for tool users—experts and knowledge engineers—to agree on i t .
A second approach is to develop generic K A tools that can be configured for specific needs—for instance, for each domain. Examples of such configurations are specifica- tion of various tool properties i n resource files, and wiring of sub tools i n the overall K A tool. These techniques form a knowledge-acquisition workbench. S i m p l e config- urations of generic K A tools can be performed by the domain expert (e.g., changing user preferences), whereas more sophisticated configurations can be performed by only the knowledge engineer (e.g., editing of resource databases).
A third approach, which is discussed i n this article, is to introduce an a d d i t i o n a l layer of tool support that is used to build specialized K A tools. In this approach, knowledge engineers use supportive environments, or meta-tools, to develop the K A tools that axe to be used by the domain experts. Meta-tools have the advantage of providing more generality than can generic K A tools, since knowledge engineers specify target K A tools, rather than merely parameterize generic tools. For instance, m a n y conventional programs allow their users to custom-tailor certain predefined aspects of the program behavior. Resource editors can be used to custom-tailor forms, menus, accelerator keys, colors, and so on. Such resource editors, however, cannot m a k e more radical changes to programs (e.g., they cannot turn a document editor into a spreadsheet program, or vice versa). Hence, the range of options for custom-tailoring is l i m i t e d for conventional programs, as well as for K A tools. Sometimes, there is no clear distinction between generic K A tools and meta-tool systems. Indeed, a sophisticated generic K A tool could provide the same flexibility and support as do the K A tools produced by a meta-tool.
2.2 Conceptual Models of K A Tools
Interactive K A tools must adopt a conceptual model that allows their users to c o m m u - nicate w i t h the tools and to provide the relevant domain knowledge. T h e conceptual model forms the language i n which the tool and its users communicate. F o l l o w i n g M u s e n [25], we identify three major conceptual model types for K A tools:
1. Symbol-level conceptual models: Symbol-level conceptual models comprise i n - dividual knowledge-base entities, such as rules, frames, and parameters. T h u s ,
