Content uploaded by Roar Arne Fjellheim
Author content
All content in this area was uploaded by Roar Arne Fjellheim on Nov 15, 2020
Content may be subject to copyright.
c)
rt
f..-
I
$
o
cY)
o
z
a
U)
a quarterly review
ON PAGE: 1984
VOL,l2
A AN EXPERT SYSTEM
Z FoR SEsAM-69
PROGRAM SELECTION
1 ' SULPHIDE,AND HzS
r + srREss coRRos/oN
rt, MIROS-AMICROWAVE
Z+ REMoTE sENsoR
FOR OCEAN SURFACE
/rrr vtBRATtON D/AGNOS/S
dV DATA BANK FAR
MACHINERY MAINTENANCE
fia FTBRE OPTTCS tN
O { SUBSEA eoNTRol.sys TEMS
r \I
l \l
(
.-4...
fft
tN ,Nr
t(t
Ut
An expert system for Sesam-69 program selection.
by
R. Fiellheim
AS Computas, Havik, Norway
Abstract: The technology of expert systems in a prornising
new approach to the probletn of preserving anddisserninating
hurnan expertise. Such systetns are on outgrowth of artificial
intelligence research, and rely on forrnal rnethods foi repre-
senting specialized knowledge of facts, relationships and
rules.
This paper describes SesCon, a prototype expert systern
for advising a user of the Sesarn-69 structural analysis
package on which method to use for a particular analysis.
The decision is based on properties ofthe structure andthe
loads applied to it, as well as on analysis goals. SesCon was
built on top ofErnycin, an expert systetn tool frorn Stanford
University.
An overall assessrnent ofthe SesCon developrnent is positive.
A working, albeit lirnited, decision aid was constructed,
which due to explanation facilites also can be used for teaching.
More irnportant, the project provided practical experience
with, and insight into, the construction of knowledge'based
expert systerns. It clearly dernonstrates the potential value
of this technology to the Sesarn-69 dornain in particular, and
to other related dornains in general.
Keywords: Expert slstems, structural analtsis, knowledge engineering,
Sesam-69.
I. INTRODUCTION
This paper describes an experimental application
of the emergingexpert slstemtechnology [1] to atask
of practical interest to AS Computas and its owner
institution, Det norske Veritas. The task is to advise
inexperienced users of the structural analysis pro-
gram package Sesam-69 [ll] on how to make
proper use of that package.
Expert systems represent a practical application
of artificial intelligence research [2]. The basic idea
is to formalize an experts' special knowledge of a
particular domain, then make that formalized body
of knowledge available in computerized form such
that non-expert users can benefit from it without
the expert's presence.
The development of our expert system (SesCon -
Sesam-69 Consultant) was part of an internal quality
assurance project in Veritas. We wanted to deter-
mine whether expert systems may contribute to
the quality of work performed in the organization'
In particular, the aim was to test the idea that an
expert system for a particular field may be an efficient
vehicle for transferring expertise to newcomers to
that field.
R. Fjellheim
The particular domain of Sesam-69 consultation
was selected for two reasons. First, it is a domain
where a real need for an advise system is felt. The
package is large and complex.l/2-l yearisnormally
required before a new user becomes proficient.
Second, the Sacon [3] expert system from Stanford
University was available to us. Sacon is a consultation
system for the Marc structural analysis package,
and served as a source ofideas in the early phase oi
our project.
Using Sacon as a model was more than a coin-
cidence. Our approach to expert systems is based
on work done at Stanford. In particular, SesCon
is built using the Emycin [4] expert system package
developed by the Heuristic Programming Project
at Stanford. We installed this system on a DEC-20
computer at the University of Linkciping (Sweden)
and used it over a dial-up-line.
In order to have a manageable task to work with,
we had to find a suitable subdomain of Sesam-69
consultation. The selection of an appropriate pro-
gram for analyzing a particular structure was
chosen. This is fairly straightforward decision for
.lontegian,\7 arilime Ret earch
..\0.3i1984
2
experienced users, as evidenced by the small Emycin
implementation (approximately 20 parameters
and 35 rules). Still, it is definitely not trivial, and
it has served us well as a testing ground for gaining
experience with expert systems.
The next section of this paper provides back-
ground information on Sesam-69 and expert systems.
This prepares the ground for a presentation of the
SesCon knowledge base in section three, followed
by an example of usage in section four. Section five
summarizes our experience, while the final section
contains a few concluding remarks.
The work described below was done jointly by Roar .1-ellheim,/
A.S Computas (..knowledge engineeringrr, Emycin) and Per
Slwersen,/Veritec (Sesam-69 expertise). The paper is an abridged
version of Computas report No. 83-6010.
Acknowledgements are due to James Bennett of Stanlord
University (now Teknowledge, Inc.) for providing us with
Emycin and Sacon and lor helping us over a few spots in using
Emycin, and to Mats Andersson, Jan Axing and Sture Hiiggund
ol the Inlormatics Laboratory, Linkoping University, for their
assistance and helpful attitude in bringing up Emycin on their
DEC-20.
2. BACKGROUND INFORMATION
2.1 SESAM.69
Sesam-69 is the collective name of a suite of com-
puter programs for doing structural analysis Il l].
The system was developed by Det norske Veritas,/
Computas, and has been used in a large variety of
applications, especially in the shipping and oflshore
industries. A new generation of the product, Sesam-
80, is currently under development.
Sesam-69 is based on the finite element method
(FEM), and is particularly well adapted for use in
the so-called multilevel or substructure technique.
The system covers static structural analysis within
the linear theory of elasticity as well as forced or
free vibration of linear elastic structures. FEM is a
method for finding the approximate solution to
structural analysis problems, i.e. the determination
of stresses, strains, displacements and reaction
forces in a structure exposed to external loads under
a given set of boundary conditions.
Sesam-69 is an integrated package of nearly two
dozens individual programs. Some are pre- and
postprocessors for facilitating the data-handling
and presentation problems involved in building a
Iarge model. In this study, we are only concerned
with analysis programs. Each program covers a
class of analysis problems, for example thin-shell
analysis, frame analysis, etc. Conducting an analysis
involves the major steps outlined in figure l.
Becoming an expert modeller is a longwinded
process. The idea behind the current work is to
shorten the necessary time by capturing the expertise
of an experienced Sesam-69 user in a computerized
knowledge base, which may subsequently be used
Norruegian M aritime Research
No.3/1s84 3
Determine overall
characteristics of pro blem
I
^ Sel^e^ctappro.priate
Sesam-69 analysis program
t
Design detailed
. finite element model
I
- Prepare program input
for the constructed mbdel
I
Run the pro
t
Evaluate final
gram
results
Fig. l. Major steps of a Sesarn-69 structural
analysis
for consulting and teaching. This paper describes
an expert system for doing the second step shown
above, i.e. the selection ofan analysis program based
on problem characteristics.
2.2 EXPERT STSTEMS
Expert systems represent a practical application of
artificial intelligence research. From an initial
emphasig on general mechanisms for intelligence
(e.g. search algorithms), this field has in recent years
concentrated more on representation and use of
specialized knouledge as a basis for intelligent be-
haviour [2].
Expert systems are a direct outgrowth of this
work. The basic idea behind such systems is quite
simple, and is illu,strated by figure 2. An expert (or
several) formalizcd his specialized knowledge of a
particular domain in a form that can be stored in a
computer. The resulting knowledge base may then
be used by an interpreter mechanism to provide
consulting and teaching to non-expert users.
The crucial component of an expert system is its
knowledge base. A number of formal methods for
representing knowledge in a computer have been
devised. The currently best understood and most
widely used representation for expert system work
is the so-called rule representation In a rule-based
system, the knowledge is represented as a number
of fairly small and self-contained rules of the form
if condition then conclusion
Expert
Forrnalization
of expertise
Knowledge-based
expert system
Consulting,
teaching
tJser
Fig. 2" Basic principle of expert systerns
{lonciitions and conciusions rel-er to ihcts r:on-
iaincci in a separate ritttaha.;e. fhe rulebase represents
general knou,iedge oi.the dcimain. r'r,hile the Cata-
base holds lacts specific to a particr-riar case r-tnder
investigation.
Tire rulebase interpreter maY \vork in one of tvu'o
basicallr. diiferent r,r'a..,s (or possiblv a combination
of the tlr.,o):
r l'he interpretation may l:e datadrii.ten or.f oru'ard-
ihainine. The interpreter first finds a rule u'hose
condition matches a fhct. Its conclusion is added
to the database, a new matching rule is found. etc.
o Alternativelv, a goaldrit'en or backti'ard-chaining
interpretation mav be employ,ed. The goal
mar. be to establish (or eiiminate) a certain
fact. The interpreter starts br' finding a rule
n'hose conclusion ma,v help establish the goal.
II' tl-re condition of that rule is satisfied. the
soal has been reached; otherr,r'ise the conditic,n
is set up as a subeoal. etc.
The choise o{'interpretation method dcpends on
the nature o1'the domain as r.r,ell as on the specific
needs the s\.stem is meant to serve.
Expert svstems r'r,ere first developed for domains
'uvhere knorvledge is uncertain and not representable
in algorithnric lbrm. Examples are medicai diagnosis
and oil u'ell log interpretation. Decisions in such
domains must olien be based on incomplete, tln-
certain and conflicting information. In order to
handle such situations, most expert systems have
rt'avs of'representins and handling uncertain know-
ledge. For example. to a simple nrle of the kind
shor.r'n earlier mal' be attached a numerical certaint-1'
.factr-tr, a number that measures the expert's con-
fidence in the rule.
A major requirement for expert systerns rs account-
abilitl. Considering the importance of decisions
in for example the medical domain, it is clear that a
user will not accept advice from a computerized
consultant uniess he is convinced that the advice is
sound. Therefore, the ability to explain advice is
built into most expert systems. Explanations are
based on the same knowledge as is used for generating
advice. This is so important that it may be taken
as a defining characteristic of expert systems.
The current state of the art of expert systems does
not normaliy allow an expert to enter his knowledge
directly into the computer. An intermediary knour
ledge engineer is required, whose task is to elicit in-
formation from the expert and express it in a language
acceptable to the computer. The effort required to
build a formal representation of even a small part
of an expert's knowledge is substantial in all but the
simplest of cases. The knowledge acquisition problem
is the major hinderance to widespread use of expert
systems, and is therefore the subject of much current
research.
Theory and applications of expert systems are
still in their infancy. After a number of years in the
academic environment, the technology has recently
caught the interest of industry. Some successful
applications and some useful support tools are com-
mercially available. Most existing systems can be
classified as fault finding, diagnosis or data inter-
pretation systems. Some examples are:
o Diagnosis and therapy for meningitis (Mycin,
i5l).
o Interpretation of mineral prospecting data
(Prospector, [6]).
o Configuration of DEC VAX computer systems
(Rl, [7]).
o Advice on using a large program package
(Sacon, [3]).
o Fault finding in computer systems (Dart, [B]).
Much remains to be done, but there seems to be
widespread agreement that knowledge-based systems
will play an important role in sophisticated and
user-friendly computer systems in the future.
2.3 EMTCI.N
Stanlord University has been particularly active
in developing methods and tools for expert systems.
Mycin, one of the earliest and best known medical
systems, was developed at Stanford [5], and has
later served as a model for a number ofother systems.
The success of Mycin prompted the development
of Emycin (Essential or Empty Mycin), a domain-
independent package for constructing rule-based
'Mycin-like' consultation systems. We have used
Emycin for our Sesam-69 application.
Emycin has as its core a backward-chaining rule
interpreter identical to the one used in Mycin,
.\'oriL t g ian,lTari I ime Re t eart h
\0. 3'19814
SYSTEM DESIGNER
experti se debugging feedback
+
Knowledge Base
Construction Aids +
Domain
Knowledge
Base
EMYCIN
Consultation
Driver
+
t
case data advice
CLI ENT
Fig. 3. Structure of the Ernycin systern
surrounded by a number of tools for constructing,
debugging and using the rulebase. Emycin in
implemented in the Interlisp dialect of Lisp, which
in its most widely available version runs on DEC-10
and DEC-20 computers. Interlisp is also available
on the DEC VAX machine, as well as on several
Xerox professional workstations (the Scientific
Processor I 100 series).
A general description of Emycin can be found in
the Emycin Manual [4]:
Emycin is used to construct and run a consultation
program, a program that offers advice on problems
within its domain of expe rtise. The consultation
program elicits information about a particular
problem ("case") by asking questions to a user.
It then applies its knowledge to the specific facts
of the case and informs the user of its conclusions.
The user is free to ask the program questions
about its reasoning in order to better understand
or validate the advice given.
The overall structure of Emycin is shown on
figure 3, also taken from the Manual. Emycin
Norwegian M aritime Researc h
No.3/1984
accomodates both the designer (the "knowledge
engineer") of an expert system and its user(s). The
former has a knowledge acquisition interface to the
system, through which knowledge in the form of
rules is entered, debugged and maintained. The
user accesses the resulting knowledge base through
a consultation driaer.
As was descriTbed earlier, the knowledge base of
a rulebased expert system has two components: a
case-independent set of rules, and a dynamic case-
dependent set of facts. In Emycin, the latter infor-
mation is stored in a so-called context tree, a tree
structure where the nodes are contexts, or data objects.
Each context has a number of parameters, i.e. data
attributes. Each parameter, in turn, is described by
a numper of properties. A sample Emycin parameter
definition follows:
ERROR ISTRUCTURE-PARMS]
EXPECT: POSNUMB
UNITS: PERCENT
TRANS: (the adrnissible analysis error)
PROMPT: (Enter the admissible analysis error
in percent)
ASKFIRST: T
REPROMPT: (The error parameter is the maxi-
mum acceptable deviation between calcu-
lated stress, deflection, etc. and the real
values. Since analysis cost goes up with pre-
cision requirements, one should not specify
a more accurate analysis than is actually
necessary. )
UStrD-BY: (Rules 28 IB B 7 12ll)
CONTAINED-IN: (Rules 28 lB)
This describes the error parameter associated
with a structural analysis in our Sesam-69 system.
The error must be a positive number (EXPECT)
and is expected to be supplied in terms of percent
(UNITS). The TRANS, PROMPT and RE-
PROMPT properties are used for generating an
English-like dialog, while ASKFIRST : T (for
True) indicates that the value of this parameter
cannot be deduced by the system, but must be asked
directly from the user. The UStrD-BY and CON-
TAINED-IN are mainly for internal system use.
The rule Dasa consists of a numer of if ' then rules.
Each rule has a unique number, and is represented
internally as a Lisp data structure. The rule designer
may enter rules directiy in this format, or through
the more user-friendly ARL (Abbreviated Rule
Language) format. On output, a rule may be dis-
played in Lisp, ARL or in an automatically generated
English form. As an example, we show the same
rule in the three formats.
Lisp
PREMISE: ($AND(SAME CNTXT CON-
STRUCTION CONTINUUM)
(SAME CNTXT DIMENSION 2)
(SAME, CNTXT BENDING)
(SAME CNTXT SHE,AR)
ACTION: (CONCLUDE, CNTXT CLASS
THICK-SHELL TALLY I000)
ARL
If l) Construction : continuum
2) Dimension:2
3) Bending, and
4) Shear
Then: Class : thick-shell (1.0)
English
fThis rule applies to loads, and is tried in order to
find out about the analysis class(es) of the structure]
If t) The construction of the structure is
continuum,
2) The stress dimensionality due to the
Ioads is 2,
3) Bending is caused by the load, and
4) Significant transversal shear is caused
by the load
Then: It is definite (1.0) that thick-shell is one of
the analysis class(es) for the structure.
The conclusion of a rule may be more or less
certain. This is specified by giving a so-called certaintlt
factor, a number between I (absolute beliel) and -l
(absolute disbelief). The Emycin interpreter handles
certainty factors according to principles described
in [5]. For some domains, such as Sesam-69 consul-
tation, uncertainty is not an inherent property; all
our rules therefore use the values I og -1.
The rule interpreter of Emycin is of the goal-
directed, or backward-chaining type. Given a para-
meter whose value is to be determined, it will look
for rules that conclude about that parameter. The
premises of those rules are inspected. The values of
ASKFIRST parameters are requested from the
user, the remainder are set up as subgoals. The
result is a focused, meaningful dialog. A forward-
chaining interpreter might have caused questions
in a seemingly random order. The search is exhaustiue,
i.e. all possibly relevanf factors will be considered
during a consultation.
3. SESCON KNOWLEDGE BASE
3.1 INFERENCE STRUCTURE
An expert system is a logical model of the reasoning
process employed by an expert. A natural approach
to the construction of a rule-based system is first to
determine the major steps of the reasoning process,
then refine each step in terms of specific rules. A
graph describing the overall reasoning process is
called the inference structure.
An Emycin inference structure is a graph where
the main goal parameter is shown at the top. Below
it, all parameters that may contribute to the goal
by being mentioned in conditions of rules of the
form "if condition then goal", are listed. The graph
only depicts the existence of such a relationship,
while details are delegated to the rules. The next
level down shows how second-level parameters are
derived, etc. The structure need not be a pure tree,
i.e. link may cross several levels or leap to another
area of the graph.
The domain of our SesCon expert system is to
Program
selection
,il
I
lnternal phenomena :
Stress, bending, etc.
Analysis
objectives External properties
of structure and load
Fig. 4. Overview of SesCon inference structure
.\'aru e gi.an M arilirne Researc h
.\'0. 3i1984
Program
t
CIass
,/ /x
Load-direction Slenderness High-slenr
,/\
Error
Typical-span
/\
Scope Load-ty pe
I
Reduction Bendi ng Shea r
Fig.5. SesCon inference structure
select a Sesam-69 program for a particular analysis
job, based on properties of the mechanical structure
and the loads applied to it, as well.as on the objectives
of the analysis. A first approximation to the inference
structure of this system is shown in figure 4.
Parameters describing analysis objectives and
external properties of the structure and load cause
certain internal phenomena in the structure to be
focused upon. Sample phenomena are the dimen-
sionality of stresses (2 or 3 dimensions) and presence
of bending forces. Each particular Sesam-69 pro-
gram has been designed for a particular set ofvalues
of internal phenomena (an anaQsis class), and, may
therefore be selected as soon as su{ficient data on
these phenomena have been collected or deduced.
This overall model is straightforward and captures
the expert's reasoning in this domain. The next step
. is to refine the inference structure. Through several
major revisions we have arrived at the detailed in-
ference structure in figure 5. Each identifier in this
figure is the name of a parameter in the SesCon
knowledge base. A directed arc between two para-
meters indicate that at least one rule mentions the
parameter pointed to in its conclusion, and the
other (lower) parameters in its condition.
.\lrtL'e p ian ;\I ar it ime R e s e ar c h
.\'0. 3 ' 1981
Thickness
Load-span
Support-span
Study-span
3.2 PARAMETER DATABASE
As explained in section 2.3, an Emycin database is
tree-shaped. The SesCon context tree is a very simple
two-level, two-context stnlcture:
o The root context is the STRUCTURE, which
has parameters for describing the physical
structure under consideration, as well as para-
meters of the consultation.
o The only se,:ond-level context is the LOAD, of
which there may be any number belonging to
the same structure. Parameters describe a single
load and the effects it causes in the structure.
Figure 6 summarizes the context tree structure,
including the names of all parameters belonging to
two context types.
3.3 RULE BASE
The <.expertise>> of SesCon resides in the rule base,
which contains approximately 35 rules. The following
few rules illustrate the contents of the rule base:
RULEOI5
[This rule applies to loads, and is tried in order
to find out whether bending is caused by the
loadl
Dimension
H ig h-slenderness
Constructi on
STRUCTURE
Name
Construction
Curved
Thickness
Support-span
Study-span
Advice
Program
Class
Scope
Error
Load-type
Load-direction
Load-span
Typical-span
Dimension
Bending
Shear
Slenderness
Low-slenderness
High-slenderness
Reduction
Fig. 6. SesCon context tree structrrre
If I ) The dire ction(s) of the load are one of:
length-width width-length, and
2) A typical distance associated with the
Ioad is greater than 0
Then: It is'definite (I.0) that bending is caused
by the Ioad
Justification: Bending forces are present in the
structure if the load has a component normal to
the Iength-width dimension, and this component
does not bear directly on a support.
RULEOOS
fThis rule applies to loads, and is tried in order
to find out about the analysis class(es) for the
structure]
If: I ) The construction of the structure is
continuum,
2) The stress dimensionality due to the
load is 2,
3) Bending is caused by the load,
4) Significant transversal shear
caused by the load,
5) The shape of the structure is curved, and
6) The admissible analysis error is less
thanor equal to l5 percent
It is definite (1.0) that curved-thin-shell is
one of the analysis class(es) for the structure.
Justification: This rule describes the more com-
plex thin-shell analysis case: curved thin-shell.
As for flat thin-shells. it is relevant Ibr continuous
structures u,ith tr,r,o-dimensional stress variation.
bending, and absence of sisnificant shear forces.
In addition, the structure has a curved shape.
and a fairly accurate analysis is required.
RULEOO3
['fhis rule applies to structures. and is tried in
order to find out about the appropriate Sesam-
69 program to anaivze the structurel
l) The analvsis class(es) ibr the stmcture
are one of: curved-thin-shell thick-
shell. and
2) Solid is not one of the analvsis class(es)
for the structure
It is definite (1.0) that the appropriate
Sesam-69 program to anal,vze the struc. is
n\'.).1z.
Justification: If the worst-case analysis class due
to any load is curved thin-shell or thick-shell,
the appropriate analysis program is NV 332.
RULEOI5 defines when Bending may be assumed
to be present in the structure (a load is applied some
distance away from a support point). RULE0OB
classifies an analysis case, using, among other factors,
the presence of Bending. Finally, RULE003 selects
a Sesam-69 program for a particular analysis class
that includes bending.
4. USING SESCON
4.1 SAMPLE DIALOG
A user interacts with SesCon in much the same way
as he would with a Sesam-69 expert (except, of
course, that one does not normally speak to an expert
via a terminal). The system poses questions, to which
the user provides information or asks for clarification.
Based on the information, the system offers its con-
clusions with respect to internal phenomena in the
structure, and advice on which Sesam-69 program
to select.
A typical interaction with SesCon requires around
I0 questions before a program selection can be
made. We now reproduce one such short terminal
session. The emphasis here is on the decision aspect,
i.e. get the right answer as fast as possible. In the
next section we shall see how the user may inter-
.\' oru:e gian l,Iarit ime Research
.\'0. 3 | l9B1
rogate the reasoning process and ask for clarification
of questions.
The questions posed by the system are numtrered,
while the lines starting with** contain user input
responses. The lines that start with neither a nurnber
nor** are conclusions,/advice generated by the
system. The example construction is a hollow
cylinder, cornpletely fixed at one end, and with a
single pointload acting radially two-thirds of the
cylinder length out from the supporting end. Figure
7 illustrates this situation.
Fig. 7. Hollow cylinder
SesClon helps vou decide rvhich Sesam-69 direct
anail'sis program (N\'331, N\'332, NV333, N\r
337, N\'339) to use for a particular structural
analy,sis. It is limited to a linear and static analr.sis
ol' a relativelv homogenous part of a possiblv
more complex total structure. and to a sinsle
load-condition composed of one or more loads.
---- STRUC]TURE.2 ----
l) \\'hat is the name of SII'RUC'I'URE-2?
d.,t THE CYLINDER
2) Have r,ou identilled one or more tvpical load
tvpes to include in the analysis?
,tr( YES
---- LOAD-I ----
3) Is LOAD-I a point load or distributed?
,rn PoINT
4) Are there more loads associated r,r,ith the
cvlinder?
,()Y \O
5) Describe the constr.uction o1'the cvlinder as
a continuum or a netu,ork.
i(,I CONTINUI]NT
6) Do you $,ant to studv local or globai pher.ro-
mena in the ct.linder?
,t{' GLOB,\L
7) \\Ihat is a tr,pical distance betrveen supports
in the cr,linder?
.\'a rte,qi an .\ I ar i.| ime Rue arc h
.\'0. 3 , 1981
,tJ. I
8) What is the thickness (shorrest dimension) of
the cylinder?
,f ,r 0.01
The slenderness ratio associated with LOAD-l is
as follows: 100.0.
9) Enter the admissible analysis error in percent.
,tn B
The stress dimensionality due to'LOAD-I is as
follows: 2.
l0) Enter the direction(s) of LOAD-1 relative to
length/width./thickness.
'I'f LENGTH-WIDTH
Bending is caused by LOAD-I.
Significant transversal shear is not caused by
LOAD-I.
I I ) Is the shape of the cylinder cur-ved?
,rn YES
The analysis class(es) for the cylinder are as fol-
lows: CURVED-THIN-SHEtL.
The appropriate Sesam-69 program to analyze
the cylinder is as follows: NV332.
4. 2 t) X P t.1.\'.1T I O.\t t',.|C; t Ll T I E,\
A motivation lbr this r,r,ork has bccn to test the
teaching capabilities of expert svsrems. Intuitir"'el1,.
one'"r,ould expect that a knon,ledge base dcveloped
for consultine also might be uscd fbr tutorins in tlte
domain of expertise. After all. u,e expecrt an expert
to be able to explain his knor,r'ledse to a novicc,
Ite:ide. gir irre advit r.
This idea is not original. C)f particular interest to
us is the r.vork done br,\\. Clancev and his collcagrrcs
at Stanlbrd f-iniversitr.. Thev built thc (]uiclon
svsterl [9] that turns anv Emr,cin knor,r,'ledge base
into a computer-aided instruction s\.stem bv addir-rg
tutoring ruLes to the ordinarv domain rules. As \r,c
understand it, this approach is not entireh,strcccssfirl;
sr-rbtle assumptions made in constructing a c:orr-
sultancv rule base make it less appropriate Ibr
teaching.
\\e did not have acccss to Guidon. and relv in-
stead on the ltasic explanation mechanisms iruilt
into Emvcin to provide an adrnittedlv cnrde turorins
capability. 'I'hei.e are for.rr such mechanisrns. :rll
available as options that the user mav enter as the
answer to a question posed b1, the sYstem. instcad
:f:"r*,..t"g the question directlr,. The lbur options
? .\ question-mark u.,iil cause Emvcin to
displav the list of. legal ans\\.ers to the
quesrion iusr posed. and in addition to
displal, a <<canned, explanatorv text. i{'
the krrolvledee engineer has provided
such text (the RIiPR()N,{PT paramcrer
propertl,).
RULE 'I-his ansn,er rvill cause thc rule that tire
Pointload
system is currently considering to be dis-
played, in its English translation.
WHY This is a request to show the reasoning
that caused the current rule to be con-
' sidered. The method used is to show that
the question will help establish the con-
dition of the current rule. If the user is
curious to know why the current rule is
considered, another WHY will take him
one level up and display the rule that
caused the current one to be invoked, etc.
HOW This last option may be used to interrogate
Emycin how it found (or will find) the
validity of a certain statement. The state -
ment is for example a part of the con-
dition of a rule, and is identified by a
number that refers to a previous listing
of that rule in the session (such a listing
may have been produced by a previous
WHY request).
Together, these options provide a surprisingly
rich facility for exploring the rule base in the setting
of a concrete case. There are in addition other
explanation facilities in the Emycin system (e.g.,
QA - Question Answering), but we have not explored
these to any great depth.
As an illustration of use of explanation facilities,
we now show parts of a rerun of the sample dialog
of section 4. l, this time with a rather more curious
(or inexperienced) user at the terminal.
2) f{ave you identified one or more typical load
types to include in the analysis?
*i( wHY
[i.e. WHY is it important to determine the loads
of STRUCTURE-5?l
[2.0] This will aid in determining the analysis
class(es) for STRUCTURE-5.
[2.1] This can be done bye using RULE020, but
first it is necessary to find out about a load.
lback to question 2 ...]
rf*r How 2.0
ii.e. HOW is the analysis class(es) for STRUC-
TURE-5 determined?]
[3.0] Currently RULE020 is being used.
Based on current data, the following may also
prove useful later in the consultation:
3.ll RULEOIg
3.21 RULE0l0
3.31 RULE009
3.41 RULE0oB
3.51 RULE007
3.61 RULEO12
3.71 RULE0U
back to question 2 .. ]
,k* How 3.I
[i.e. HOW will RULE0l9
If [4.1] the construction of the
continuum, and
14.2) the stress dimensionality
load is 2, and
cylinder is
due to the
[4.3] bending is not caused by the load
then it is definite (I.0) that membrane is one of
the analysis class(es) for the structure.
Justification: In the absence of bending forces, a
continuous structure may be analyzed as a
membrane if the internal stress field is (one-) or
two-dimensional and completely contained in
the length-wisth plane.
IRULE019]
fback to question 2 .. ]
,*r. YES
6) Do you want to study local or global pheno-
mena in the cylinder?
iin wHY
[i.e. WHY is it important to determine the scope
of the analysis?]
[9.0] This will aid in determining a typical
distance associated with LOAD-1.
If [9.1] the scope of the analysis is global
then it is definite (1.0) that a typical distance be-
tween supports in the cylinder is a typical
distance associated with LOAD-1.
Justification: For an analysis with a global scope,
a typical distance between structure supports
can be used as the typical planar span in establish-
ing the slenderness ratio.
IRULE027]
fback to question 6 .. .]
,(n wHY
[i.e. WHY is it important to determine a typical
distance associated with LOAD-l?]
p0.0] ... in order to determine the slenderness
,ratio associated with LOAD-1.
If p0.I] a typical distance associated with
LOAD-l is greater than 0, and
[10.2] the thickness ofthe cylinder is greater
than 0
then it is definite (1.0) that the slenderness ratio
associated with LOAD-I is a typical distance
associated with LOAD-l divided by the
thickness of the cylinder.
Justification: The way in which an external load
causes internal stress fields in a structure depends
to a larger extent on the slenderness ofthe struc-
ture with respect to that load. As a rough measure
of the slenderness, this system uses the ratio be-
tween a typical length in the length-width plane,
and the average thickness of the structure. The
appropriate measure to use as a typical length
depends on the scope ofthe analysis and the type
of the load, and is derived by other rules.
.\ :o ru' e ! i an,\.1 ari tim e Re s ear c lt
.\0. 3,t 1984
be used?]
t0
IRULE0r7]
fback to question 6 .. .]
*,I GLOBAL
I I ) Is the shape of the cylinder curved?
*if ?
The shape of the structure is considered curved
if part or all of its surface is curved. This covers
both regular (e.g. exisymmetric) and irregular
curvatures.
Expected responses are: YES or NO.
Enter HELP for list of user options.
{.i* WHY
[i.e. WHY is it important to determine whether
the shape of STRUCTURE-S is curved?]
[29.0] This will aid in determining the analysis
class(es) for STRUCTURE-5.
It has already been established that
[29.1] the construction of the cylinder is
continuum, and
129.2) the admissible analysis error is less
than or equal to 15 percent
Therefore,
If 129.3] the stress dimensionality due to the
load is 2, and
[29.4] bending is caused by the load, and
[29.5] significant transversal shear is not
caused by the load, and
[29.6] the shape of the cylinder is curved
then it is definite (1.0) that curved-thin-shell is
one of the analysis class(es) for the structure.
Justification: This rule describes the more com-
plex thin-shell analysis case: curved thin-shell.
As for flat thin-shells, it is relevant for continuous
structures with two-dimensional stress variation,
bending, and absence ofsignificant shear forces.
In addition, the structure has a curved shape,
and a fairly accurate analysis is required.
[RULE00B]
fback to question ll ...]
jrn YES
4.3 CA,SE LIBRART
A very useful feature of Emycin is the possibility of
building libraries of cases that may be stored and
rerun. The knowledge engineer may use these to
check out the correctness of modifications by re-
running standard test cases with known results.
Another use of this facility is as an aid to using an
Emycin expert system for teaching purposes. The
tutoring capabilities of such systems are case-
oriented, i.e. ..lessons>> ore always cast in terms of
specific cases. A teacher may build a library of
representative cases that demonstrate various aspects
of the knowledge domain. The student may later
run the cases and thereby learn how experts handle
these aspects.
The library produced for SesCon include the fol-
lowing constructions: Transverse shipframe, Engine
Norrue qian Maritime Research
ttro.3/1984 ll
bracket, Hollow Cylinder, Cantilever beam, Cubic
brick, Thin square plate, and Simple ring.
5. SUMMARY OF EXPERIENCE
5.1 EVALUATION OF SESCOAT
Our project is probably quite typical of experimental
work : we used most of our time constructing the
system, and very little actually using it. Therefore,
we do not have extensive or systematic experience
on which to base an evaluation of usefulness.
Nevertheless, we have su{ficient experience to form
an opinion on the possible benefit of SesCon, and
more importantly, on promising directions for
further work.
SesCon as described in this report is nearly adequate
for its limited domain. We estimate that a 50 7 in-
crease in the size of the knowledge base would be
required to cover the more obscure cases, but this
would still be a rather small system. The resulting
system would be useful mainly for training of novices,
since the program selection skill is one that is rapidly
acquired by Sesam-69 users.
As for the specific teaching capabilities of the
system, we believe that the built-in explanation
mechanisms (essentially, WHY and HOW) , may
be sufficient if the student is well motivated and
has a basic understanding of the domain and of the
knowledge base (in our case, this means the inference
structure). Under these conditions, we feel that the
unenhanced explanation mechanisms may be suf-
ficient as a basic training facility.
However, for a complete novice, this approach
would almost certainly be confusing. An augmented
system for tutoring could still be based on the same
knowledge base, but would in addition contain
mechanisms specifically for teaching purposes.
Examples are sequencing of material presentations,
scoring of student responses, monitoring of student
progess, etc. With such extensions, an instruction
system based on a knowledge base should prove
superior to an ordinary text-based CAl-system,
which after all only has a ..surface understanding>>
of the domain.
Another objection to using the present system for
teaching purposes is that its behaviour is entirely
case-oriented. A more complete tutoring system
should in addition to case examples be able to teach
general knowledge from the domain. To a iimited
extent Emycin-based experts systems exhibit such
capability, in that the domain rules (representing
general knowledge) are presented as part of case-
specific explanations.
In summaryr we conclude that Emycin-based
systems like SesCon (but with much larger know-
Iedge bases) probably are more appropriate for the
advise,/consultation task than for the teaching task.
5.2 POSSIBLE EXTE],{SI ONS
It is not dilficult to see how the present work could
be continued in other parts of the Sesam-69 (or
Sesam-80) domain. Refering to figure l, u,e consider
the steps .<Design detailed finite element model,
(step 3) and <.Er.aluate final resultsr, (step 6) to be
ol.particular interest. An effective adr.ice svstem in
these tr,vo domains u,ould have much sreater impact
than SesClon, since these tasks are more difllcult.
For exarnple) e\.en \rer\r experienced Sesam-69 users
admit to have onl\, r,ague notions on hor.v to check
propcrlv the final resuits of an analvsis (as opposecl
to sirnplv ar:cept that the results are meaningiili).
Tvpical tasks involved in the modellins step in-
clude [2]:
r Cllioose elemcnt tr pes lbr the finite clement
model. based trpon geometrical and mechanical
properties oi' the problem.
r lJetermine the llneness oi tirc elernent mesir.
'l'he llneness ma\' \,ar\. across the stnlctrrre.
and depends on seometn, and anair.'sis objectives
i'e.e. admissible error in the final results).
o Divide ihe total strllcture into a nurnlter ol
substructures isr"rperelemcnts').
f'here r,r,ould be one knou,ledge base for each of
tlie Scsam-69 analvsis prosrams (N\'331, N\'332,
ctc.). since ther,. have dill'erent propcrties. The
crnphasis on seometricai considerations in these
clomains almost necessitates a terminal i.r,ith graphical
capalrilities. The simple tele-tr,.pe inter{'ace oI
Emr,'cin r,r,ould be vert. ar,r,ku,ard.
Ir.'pii:ai tasks in the result evaluation step include
[12]:
o Check that the ecluation s\.stent generated on
basis o{'the finite element model is reasonable
(c.g. stillness matrix half-bandr,r.idth is optimal;.
o Chcck the numerical stabilitv of the solution
c.q...inqularirics.
o Clircck that simplil'vinu assumptions made
'uvhen formulating the modei are satisfied ie.s.
that the masnitude of stresses does nor inr.alidate
a linearitr, assumption).
o N1ake sure that the calculated reaction fbrces
and rnoments balance out the applied lbrces
and moments.
r If some o1'the chccks eir,'e negative results^ trr-
to track dor,r.'n the modelling mistake that most
likell' caused the error.
N{ost of these tasks r,r,ould be common to all
analr.sis program (possiblf u,ith the exception of
the last diasnostic task), since all programs use
esser-rtiallv the same eqrration solution procedure.
I)trrins the project u.e sketched a first version of a
knor'r'ledge base lbr the result evaluation task.
So Iar, u,e have considered each step of an analy'sis
in isolation. The Iogical next step is to combine
subsvstems into an integrated consultation system
for Sesam-69. This system should ideall,v be coupled
to other Sesam-69 tools lbr data-generation and
-presentation, and utilize an graphics terminal.
Such an intellisent interface would be extremelv
vaiuable to both novice and experienced users of
Sesam-69. It i,r,ould contribute to the uniformitv"
quality, and efllciencv of structural analvsis pro-jecrs.
and r,r,ould definitelr.enhance the market potential
ol the package.
5 . 3 h' \' O tL' L E D C; E 8,1.\ E ( ; O.\,t "t- R t.: C T I O.\"
\!hen constn-rction the knon,ledse base. u,e found
that the natirral procedure is to start r,r,ith the in-
lerence strllcture. forrnr.rlate rules. check those rules
against actual case data. and finailv revise on basis
of the observeci results. Figure B illustrates tiris
gcncla[ flc,ir oi evcnl:.
Discuss and
design inference structu re
I
Y
Construct rules
I
v
Try out rules on
specific case data
I
v
Evaluate
Revise
Fig. 8. Knowledge base construction
We found the inference structure to be the key to
knowledge base construction. A recommendation
to others attempting to build a rule-based system
is to postpone the formulation of rules until at Ieast
an initial version of the inference structure exists.
This should not come as a surprise. After all, soft-
ware engineers have for a long time advocated
design before coding. It rapidly became clear to us
that knowledge engineering is not different in this
respect. Rather than independent ..little chunks of
knowledge>>, rules may be regarded as small <<pro-
cedures> that must fit into an overall design.
In fact, our experience is that knowledge
engineering has many similarities to ordinary soft-
ware engineering. The basic tasks (interview client,
set up requirement specifications, design and code)
are the same. Some of the methods and tools are
different, and allow rapid construction of expert
systems with the attractive properties we already
have described extensively. In other words, there
Yorttegian,\Iaritinc Reseorch
\a. 3 1981
system responses
{ok
r2
is a difference, but it is one of degree rather than
of basic nature.
An often cited effect of having to explicitly define
decision rules for a domain, is an increased under-
standing of the domain itself. We also discovered
this, despite the simplicity of the SesCon domain.
While most analysis cases fall neatly in one particular
analysis class, there are less clear-cut borderline
cases, especially between the thin-shell, thick-shell
and solid classes (this is not very consequential,
since the corresponding Sesam-69 programs are
sufficiently versatile to cover all cases, but at dif-
ferent costs and <<naturalness>r). One result of our
work is to have established precise criteria for dif-
ferentiating between such cases.
5.4 SOFTWARE TOOLS
Using Emycin as the implementation base for
SesCon certainly influenced and constrained the
way the system could be built. Generally, we found
Emycin well adapted to the task. This may in part
be caused by our adaption to the tool, or *odifi-
cation of the task to fit, or lack of alternatives to
compare with. Nevertheless, we have been able
to express the domain knowledge in a fairly natural
and straightforward manner, and the system per-
forms as it should.
Features of Emycin we specially appreciated in-
clude:
o Flexible rule format, in particular the abbre-
viated rule language and the automatic trans-
lation to English on output.
o Adequate knowledge engineering interface
for debugging and updating the knowledge
base (rule tracing, editors, case libraries, etc.).
o Built-in user interface mechanisms for auto-
matic generation of prompts and explanations
in near-English.
We were less satiffied with some other characteristics
of Emycin, including the absence of modularity in
the rule base, semantic networks for more complete
data descriptions, and a more natural mixed-initiative
human interface. The overriding concern when
considering a possible extension of SesCon, is how-
ever the fact that Emycin is a non-commercial,
non-supported system. After the conclusion ol the
project, several commercial expert system tools
have become available, some of which are appropriate
for SesCon-like applications.
On the whole it is fair ro say that using Emycin
may have removed some degrees of freedom, com-
pared to constructing SesCon <<from scratch>>, but
has on the other hand allowed us to concentrate on
the domain-specific problems and let us progress
much f,aster towards a working system. Therefore we
conclude that a tool like Emycin is useful for con-
structing a limited class of rule-based expert svstems.
\'lru e qian,\I aritine Re surch
\n. 3,'1981
However, \,\re can easiiy envisage domains n,here
another approach may be called for.
An alternative to Lisp/Emycin is Prolog II0],
the best known lotic prograrnming Languaga. Many ol'
the features required for rule-based expert s),stems
are alread,v built-in primitives of the Prolog lang-uage.
Examples include rule-like basic statements,
associative database of facts. and a backr,r,,ard-
chaining rule interpreter. Other mechanisms. such
as a user interl'ace, explanations, and English
translation of rules, would have to be added.
The development and use of a knowledge base
are ver). dissimilar activities. and different rools
should be provided lbr the tr,r,o. Hopefully, a much
smaller environment r,r.ould then be needed for
runnins consultations. so that downloading ol'the
knor,r,ledge base to a small computer would be
feasible. \\'e leel that this is a necessarv condition
for industrial acceptance of expert svstems.
6. CONCLUSIONS
()ur overall conclusion is that the project rvas
successlirl. mainll. in givine us insisht into and
practical experiencc r,r,ith expert s),stem technolos\..
It points to a number of interesting and use{irl
extensions. Such extensions must be motir.'atcd ltv
real needs and the perceir.ed benellts of har.ins an
automated adr.ice component of'thc Sesarn package.
This paper has described one parricular appli-
cation of cxpert systems in lfet norske \/eritas. ltut
it seems clear that a numller of other dornains coulcl
be ol. interest as r,r,.ell. \'eritas. Iike several similar
organizations, is a .<knor,r,ledee-intensir,'c> instittrtion.
based on the expertise o1'a larqe numtrer o{'rvcll-
qualilied persons rt,hose time becomes more and
more expensive relative to other costs. 'I'hus. it is
likelv that increased investment in computcrizecl
tools lbr preservation and distribution of sur:h kno',r.,-
ledqe is ( o.l-e['lecrive.
REFERENCES
l. ^\I. Stellk ct al.. "TIrc ()rsanization ol l.)xpclr Sr'_stt'nr:. .\'.['Lrror ialo.
.lrtilicial lnttllistntt. \'ol. lB ilgfl2r, I35-173.
2. ,\. Barr anci fl..\. Feisenbilrm 'ecl". ..The Ilancllrook o1 .\rri1i ial
Intclligencc,,, ll'i I liarn hurliuattn. lnt.. l rr,\1tos. Cali1.. I 9Ul.
3..J. Bennett et al.. .A Knotleclec-Based (lonsLrlt:rnt li)r Stru(.rLlral
.\nalvsis". .\itanfotd ('ni.t er.ri1j., Sepl. l97ti.
4. \\'. \'an \lellc t:t al.. "'I'he lilrnvr.in \lanrral..,\ttnli,td [-niittritl,
Rep. S'I''\N-fiS-81-8t]5. ()<.t. l98l.
5. I-.H. Slxrrtlilli. <<flompurer-B:rsed \lt:ril<.al (irnsLrllatirin: \I\'(tl\".
lilrriirr. Neu York. l!176.
{i. R.C). Drrda ct al.. <A (iornpLrler-Based (lonsultanr lbr \lineral
Lxploration..,\lll lntttnati,ntal, Rcp. ttl72. .\pril lgt)0.
7..J. h1<:l)errnort. "Rl: -,\ Rrrle-Bascrl (lorrllsurer o1 (irmputcr Srslrrns,,.
(.lat ne sie -.\l t l lon l -n i r tr.r i t l, Rcp. Cl \1 I.--OS-80- l 1 9,,\pril l 9ti0.
8. \LR. Cirncscreth. .I)iaqnosis II-.ing Hit:rarchical l)e-*iqn \fudel...
Prot. tf .1.1-11-82. 2i8-283..\ug. 19U2.
9. \\'..J. C)lanct:r', "Trarrsfi'ro1'Rtrle-Basecl l-)xprtisc l'hrorrclr zrTLrtori:rl
Dialogui:,'. \'tanlhd ['ttirtrtitt. Rep. ST.{\-(]S-76g. .\uc. l97l).
10. \\'. 1". ( llocksin ancl Cl. S. \Iellish. .Prograrnmirrs in Proloq,. \2r rrrla
I rlaq. BcrLn. I 98 I .
ll.I{.F. Klcm et al.. "Sesam-6Q Cient:ral f)csr:ription,. _1.\'(.orrrltutat.
Rep. 78-9ttl, .Juh l9t30.
12. P. Svversen. ..Sesarrr-69 \Iodellirrg \lanual,. .l\' (.otnlrttat. l\tp.
82-tj126. Stpt. 1982.
t3
