Anda di halaman 1dari 29

See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/220231728

Learning Design Knowledge as Generalization.

Article in Cybernetics and Systems · March 1998


DOI: 10.1080/019697298125786 · Source: DBLP

CITATION READS

1 4

2 authors, including:

Heng Li
The Hong Kong Polytechnic University
351 PUBLICATIONS 8,328 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Intelligent solutions for the sustainable built environment View project

Megaproject management View project

All content following this page was uploaded by Heng Li on 10 July 2017.

The user has requested enhancement of the downloaded file.


Cybernetics and Systems

ISSN: 0196-9722 (Print) 1087-6553 (Online) Journal homepage: http://www.tandfonline.com/loi/ucbs20

LEARNING DESIGN KNOWLEDGE AS


GENERALIZATION

HENG LI XINDONG WU

To cite this article: HENG LI XINDONG WU (1998) LEARNING DESIGN KNOWLEDGE AS


GENERALIZATION, Cybernetics and Systems, 29:2, 181-207, DOI: 10.1080/019697298125786

To link to this article: http://dx.doi.org/10.1080/019697298125786

Published online: 15 Dec 2010.

Submit your article to this journal

Article views: 13

View related articles

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=ucbs20

Download by: [Hong Kong Polytechnic University] Date: 10 July 2017, At: 04:13
LEARNING DESIGN KNOWLEDGE AS
GENERALIZATION

HENG LI
Department of Building and Real Estate
Hong Kong Polytechnic University

XINDONG WU
Department of Software Development, Monash
University, Melbourne, Australia

Ge ne ralizing de sign knowle dge is a proce ss of extracting knowle dge from


de sign-provid e d data. Although it is one of the most com mon ways of
``le arning’ ’ de sign knowle dge , gene ralization has a fundame ntal we akne ss:
e xce pt for spe cial occasions, the re sults of ge ne ralization can ne ver be
validate d. Inquirie s into gene ralization have the re fore de alt with que stions
of what are the be st crite ria for guiding the gene ralization. This pape r
argue s that ge ne ralizing de sign knowle dge involve s thre e esse ntial tasks;
knowle dge re pre se ntation, a de scription language for de sign exam ple s, and
gene ralization ope rators. An application of ge ne ralizing e mpirical ne tworks
from de sign exam ple s is de scribe d to illustrate the se thre e tasks in the
de ve lopm ent of a ge ne ralization syste m.

De sign is one of the most difficult real-world proble ms s He rrmann &


B e ckmann, 1994 .. The ability of de signe rs to make gene ralizations from
facts and data of de sign is we ll pe rce ive d s E astman, 1969 .. The ne e d for

Addre ss corre sponde nce to He ng Li, De partme nt of Building and Real Estate , Hong
Kong Polyte chnic Unive rsity, Kowloon, Hung Hom , H ong Kong. E-m ail: BSH engLi@
Polyu.e du.hlc
The authors are inde bte d to the two anonymous re viewe rs for the ir constructive
comm ents on the first version of this pape r.

Cybernetics and Systems: An International Journal, 29:181 ] 20 7, 1 998


Copyright Q 19 98 Taylor & Francis
019 6-97 22 ¤9 8 $1 2.0 0 + .0 0 181
182 H. LI AND X. WU

this ability in knowle dge -base d de sign systems motivates the re search on
ge ne ralizing de sign knowle dge from de sign e xample s. Ge ne ralization
can be de scribed as a proce ss of taking into account a large num ber of
de sign e xample s, the n e xtracting and re taining prope rtie s and value s
that characte rize the se e xample s. The e xtracted prope rties and value s
are re pre sented in forms that can assist future de sign proje cts.
Ge ne ralizing conce ptual de scriptions s conce pts . from de sign e xam-
ple s is an inte rsectional re se arch area be twe e n machine le arning and
de sign re search. Its e mphasis on ge ne ralizing de sign-supporting, rathe r
than ge ne ric, de scriptions distinguishe s itse lf from othe r ge ne ralization
proble ms. The re are a num ber of applications that de m onstrated the
strength of using ge ne ralization to acquire de sign knowle dge s Arcisze w-
ski e t al., 1987; He rrmann & Be ckmann, 1994; Kamal e t al., 1989; Li,
1993, 1994; Mahe r & Li, 1994; Lu & Tche ng, 1991; Re ich, 1993;
W hite hall e t al., 1990 .. Although the ge ne ralization technique s and
re sults are we ll pre sented in the se applications, the the ore tical princi-
ple s upon which the y are built are rare ly e xplaine d. Lack of a the ore ti-
cal foundation make s re search in ge ne ralizing de sign knowle dge an
ill-de fine d task, conducte d in a ``blind-searching’’ manne r.
This pape r attem pts to ide ntify e sse ntial tasks for ge ne ralization
proble ms in de sign. It focuse s on the differe nces betwe e n ge ne ralization
proble ms in de sign and ge ne ralization proble ms in othe r proble m-
solving dom ains. The role s of the repre sentation and the de scription
language for de sign e xamples are ide ntified. Se ve ral ge ne ralization
strategie s are de scribed through an illustration of ge ne ralizing de sign
conce pts from de sign e xample s.

GENERALIZATION PROBLEMS IN DESIGN


``Ge ne ralization’’ is a wide ly use d term cove ring a broad range of
activities. It refe rs to a proce ss of conde nsing a num ber of e xamples
re prese nte d in some language , which we shall call the de scription
language , into term s or forms de fine d in a knowle dge structure that is
normally calle d a knowle dge repre sentation. Give n the knowle dge re p-
re sentation, the de scription language , and the e xample s, the gene raliza-
tion problem is to sele ct or de velop an ope rator to e xtract information
from de sign e xample s to fit into the knowle dge re prese ntation. A
ge ne ralization proble m in de sign can the refore be formalize d as follows.
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 183

Give n a se t of de sign e xample s


De termine

1. A knowle dge re pre se ntation in which ge ne ralized knowle dge can be


accom modate d in form s use ful in ne w de sign situations.
2. A de scription language in which de sign e xample s are de scribed in a
way in which important de sign information is refle cted.
3. An ope rator that e xtracts de sign inform ation from the de sign e xam-
ples to fit into the knowle dge re prese ntation.

Thre e tasks constitute the proce ss of ge ne ralizing de sign knowl-


e dge . The first task is to de te rmine a knowle dge repre sentation to me e t
the nature of de sign and spe cific ne e ds in various de sign proce sses. This
distinguishe s a ge ne ralization problem in de sign from a ge ne ralization
proble m in othe r dom ains. The latter has be e n de scribe d by Mitche ll
s 1982 . as a se arch proce ss in which the represe ntation is given, but how
to se le ct a repre sentation according to the nature of a problem domain
is not spe cified. Be cause de sign involve s various com plex tasks, the
comple xity of ge ne ralizing de sign knowle dge is too high s H e rrmann &
B e ckmann, 1994 ..
The second task is to de te rmine a de scription language for de sign
e xample s. To ide ntify and de scribe re le vant information as the input
data for the ge ne ralization is important, particularly in domains such as
de sign, whe re inform ation involved in e ven a sm all task is e normous. A
de scription language cove ring overall de tails of information of a de sign
e xample m ay conce al important prope rtie s and make it difficult for a
ge ne ralization program to gle an any use ful ge ne ralizations. De scribing
de sign e xample s in an oversimplified style may result in significant loss
of useful de sign information. To de cide an appropriate le vel of abstrac -
tion for de sign information is the major conce rn in de te rmining a
de scription language .
A knowle dge repre sentation can be imple me nted in diffe re nt me dia
s e .g., programming language s ., and a me dium s such as Prolog. can also
be use d to imple me nt diffe rent knowle dge re prese ntations. The refore ,
in som e systems, the knowle dge re pre se ntation and the de scription
language have the same or similar data structures. The LE FT system
s He rrmann & B e ckm ann, 1994 ., for e xample , use s we ighte d pre dicate s
to de scribe de sign e xample s and pre dicate logic as the knowle dge
re prese ntation.
184 H. LI AND X. WU

The third task is to de term ine or de vise a gene ralization ope rator
that will match the de sign e xample s to the knowle dge re pre sentation.
This require s colle cting e xisting gene ralization s le arning . me thods and
choosing suitable one s to adopt, or cre ating ne w me thods based on an
unde rstanding of the limitations of e xisting one s.
The following subsections furthe r address the se tasks in turn.
Spe cifically, the y ide ntify the role of knowle dge re pre sentation in ge ne r-
alization, discuss the importance of a de scription language for de sign
e xample s and the ir ge ne ralizations, and outline ge ne ral principles for
se le cting and de ve loping ge ne ralization te chnique s to ge ne rate the
targe t knowle dge re prese ntation from de sign e xample s. In the following
se ction, we will de monstrate the utility of the se three tasks through an
application of ge ne ralizing de sign concepts.

Knowledge Representation in Generalizing Design Knowledge


Knowle dge repre sentation is always the first task in machine le arning
re search s W u, 1997 .. Knowle dge is the targe t of le arning, and knowl-
e dge represe ntation plays an e sse ntial role in the de sign of a le arning
system. It is pointle ss trying to build a le arning system until one knows
what kind of knowle dge re prese ntation is e xpe cte d for use . A le arning
system is a compute r program; one has to have the input and output
structure s in mind be fore de signing the system. O n the othe r hand, e ven
if one has de ve lope d very good knowle dge repre sentations, the y are
use le ss if the y cannot be le arne d by the le arning system . The refore, we
have to pursue both line s of re search in parallel.
Ge ne ralizing de sign knowle dge is de scribe d as a task-oriented proc-
e ss aime d at producing a de sign-supporting conce ptual de scription. The
conce ptual de scription is e xpresse d in te rms and forms that are de fine d
in a knowle dge repre sentation. The conce ptual de scription is re quired
to refle ct the nature and comple xities of de sign and is e asy to use for
cre ating computer mode ls of de sign. Thus the unde rlying re quisite in
any applications of ge ne ralizing de sign knowle dge is to me asure the
re le vance of the knowle dge re pre sentation to the nature of de sign.
De termining a knowle dge represe ntation is a particularly crucial
part in the task of ge ne ralizing de sign knowle dge . If the re pre sentation
capture s the e ssential prope rtie s of de sign e xample s, ge ne ralization
be come s primarily a proce ss to e xtract the prope rties from the de sign
e xample and arrange the m into the represe ntation. If the te rms and
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 185

forms in a repre sentation are irre le vant or trivial to de sign reasoning,


no gene ralization me thods will be able to construct conce ptual de scrip-
tions that will be use ful in de sign re asoning.
To de te rmine a repre sentation for de sign knowle dge is an intricate
re search topic of nume rous de bates s see Brachman & Le ve sque , 1985;
Coyne , 1992; Ge ro, 1990 .. It is be yond the scope of this pape r to give a
full discussion of how a de sign re prese ntation can be se le cted or
de ve lope d. Rathe r, the following discussions e xplicate m ajor characte r -
istics of de sign, so that the re le vance of a repre se ntation to the nature
of de sign may be me asure d against how the re pre se ntation addre sse s
the se characteristics.
A com mon de finition of de sign is that it is an ill-structure d prob-
le m-solving proce ss s Simon, 1973 .. De sign involves a continuous inte r-
play betwe e n unde rstanding de sign require me nts, setting up goals, and
finding plausible solutions to satisfy the de sired goals. The re are many
discussions of de sign nature and me thodologie s s Coyne e t al., 1990;
Dym e t al., 1988; Finge r & Rinde rle , 1990; Pahl & B e itz, 1984; Thom as,
1985 .. For the purpose of clarifying de sign characte ristics that affect the
de te rmination of a de sign repre se ntation. The following discussions
focus on the aspe cts of de sign that diffe r from ge ne ral proble m-solving
proce sse s, the aspe ct in which de sign appe ars to be ill structured, the
aspe ct in which de sign facilitate s a one -to-many mapping proce ss, and
the aspe ct in which de sign appe ars to be pragmatic.

The ill-structure d nature. The ill-structure d nature of de sign include s


thre e aspe cts. First, constraints of a de sign proble m are always
incomple te. A de sign task is ofte n pre se nte d as a se t of require -
me nts. The require me nts are insufficie nt in te rms of de ducing
solutions. Parts of the missing information are the constraints
im pose d by the social and physical world in which the artifact will
be posite d s Finge r & Rinde rle, 1990 .. Se cond, the de sign goals are
am biguously de fine d at the be ginning s Dasgupta, 1989; Ge ro, 1990 ..
To set a goal re quire s a de termination of what the potential
solution space is and what appropriate de sign m e thod it is worth-
while to follow. Third, de sign require me nts are irre gularly de fine d.
An inte rest in computer applications ge ne rally favors a formalized
set of require me nts for a problem ; de sign, on the othe r hand,
appe ars to be the re ve rse . Se ts of require me nts for de sign proble ms
186 H. LI AND X. WU

vary. A de sign variable can be give n a value as a require me nt in one


de sign, and it can also appe ar in anothe r de sign as an e ntirely
unknown factor that ne e ds to be de signate d.
The incomple tene ss of de sign constraints require s a knowle dge
repre sentation to capture associations and constraints among de -
sign variable s and provide accom modations for the se associations in
the structure of the repre sentation. The ambiguity of se arching
goals require s that the ove rall structure of the re pre sentation
manifests a possible solution space for a de sign task. The irregular-
ity of de sign re quireme nts de m ands that variable s in a repre senta-
tion are not fixe dly de fine d as input and output one s.
The one -to-many mapping. The alte rnatives for a de sign proble m are
multiple . Thus the ge ne ralize d knowle dge is re quired to be e mbe d-
de d into a ``fle xible ’’ repre sentation that can propagate many de sign
solutions from one set of de sign re quireme nts. In orde r to achie ve
the one -to-m any m apping, it is ne ce ssary that the re pre sentation
supports a wide spe ctrum of re asoning paradigms, particularly rea-
soning paradigm s de aling with unce rtainty, which often le ads to
multiple solutions.
The pragmatic nature . In most de sign proble ms the re are additional
variable s s attribute s . that cannot be optimized. Usually differe nt
de signs are ge ne rate d according to differe nt com promises m ade to
the trade -offs among de sign variable s. It must be note d that the re
could e xist conflicting obje ctives in a de sign problem. Conseque ntly,
ge ne ralized de sign knowle dge ne e ds to re pre sent the m ajor te nd-
e ncy by which most com promises are made . In othe r words, results
of ge ne ralization do not have to be corre ct for e ve ry case but only
most of the m .

The comple xitie s of de sign imply that m any e xisting re pre sentation
sche me s do not address the nature of de sign. Re pre sentations in many
of the ge ne ralization programs are used to pre dict or re cognize an
e xample that has not be e n e ncounte re d. In de sign, the repre sentation is
use d to ge ne rate ne w de signs, not to re cognize or cate gorize e xamples
that have not be e n see n be fore. Many re pre sentation paradigms for
de sign knowle dge have bee n propose d. E xample s include de sign proto-
type s Ge ro, 1990 ., de pe nde nce ne tworks s Duffy & MacCallum, 1989 .,
the object-orie nte d approach s Sriram e t al., 1986, 1991 ., and e mpirical
ne tworks s Li, 1993 .. E ach re pre sentation paradigm offers differe nt
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 187

approache s to working around complying with the nature of de sign, and


e ach paradigm confronts difficulties that may not be common to othe r
paradigms. The use of a re pre sentation de pe nds on the nature of the
de sign dom ain and the proce ss that ope rate s the re prese ntation, which
may make one re prese ntation more use ful than anothe r in a spe cific
task.
Pre vale nt ge ne ralization programs tend to be ambiguous in e xplain-
ing whe re the re pre se ntation come s from and reluctant to conce de that
the re pre se ntation is actually the pe rce ption of an intellige nt obse rver
s most like ly the de ve lope r of the ge ne ralization system .. Although the re
is a notable e ffort s Riddle , 1989 . to cre ate com puter system s that will
automatically de fine a repre sentation during the ge ne ralization proc-
e sses, the se systems re ly on a few he uristics to choose am ong a numbe r
of pre de fine d re prese ntation sche me s. The he uristics are ``crude ’ ’ as-
sumptions of how to se le ct a re pre se ntation according to the type of
training data, while the utility of the re prese ntation in proble m domains
is rare ly conside re d.
In orde r to de ve lop a com puter system that will pe rform ge ne raliza-
tion in a full se nse of this te rm, furthe r re se arch is re quire d to e xplore
the possibility of allowing the program to re de fine the re pre sentation
and repair the de ficie ncie s in the re pre sentation during the gene raliza-
tion proce sse s.* For a ge ne ralization system in de sign, the system ne e ds
to be inte grate d within a compute r-based de sign system, so that the
re le vance of the re pre sentation to the nature of de sign can be re gularly
me asure d. At pre sent, the re is no me thod by which a ge ne ralization
system would automatically adjust the re pre sentation to minimize the
irrele vance . Any bre akthrough in this would re prese nt a significant
advance in this fie ld.

The Design Examples and Description Language


De sign take s place in a rich conte xt that involves many de cision-m aking
proce sse s and ge ne rate s various kinds of information and facts. Som e of
the m are pre sented in the final solutions in forms of drawings or te xts;

* The re is inde ed a re se arch branch in m achine learning taking care of knowle dge
re fine m ent. Som e of the most re ce nt re sults can be found in Ginsbe rg e t al. s 1988 . and
Craw & Sle e man s 1995 .. H owe ver, knowle dge re fine m ent improve s knowle dge , such as
dropping an unne ce ssary condition from a rule or incre asing the stre ngth of particula r
rule s, but doe s not change the knowle dge re pre se ntation sche me of a given syste m .
188 H. LI AND X. WU

othe rs are late nt or lost in de sign proce sses, such as the infe re ntial
proce dure s and he uristics used by the de signe r in making de cisions. To
de te rmine the ir e sse ntial prope rties and a language to de scribe the m
be come s anothe r challe nging issue . Traditionally, the training e xamples
in de sign have be e n handcrafted in an e ffort to strike a balance be twe e n
the following two e xtre me s:

Too much de tail. As stated above , information in e ven a small de sign


proble m is e normous. If a de sign e xample is only an accumulation
of de sign information, it would be very difficult for a le arning
program to gle an me aningful ge ne ralizations.
O versimplification. O n the othe r hand, oversimplification m ay result in
loss of knowle dge . The gre ate r the sim plification, the more de tail is
conce aled.

The fact that the succe ss of ge ne ralization proble ms de pe nds in


part on finding appropriate de scription language s is indisputable . If a
feature A is important to a conceptual de scription, the task of ge ne ral-
ization is simple r whe n training e xamples e xplicitly contain A . If A is
implicit in the e xample s, pe rhaps as the conjunct of two attribute s, the
ge ne ralization program must pe rform additional work to ide ntify the
conce pt.
The choice of a de scription language has a major influe nce on the
capabilities and usefulne ss of the ge ne ralization results. Howe ver, few
e xisting ge ne ralization systems e xamine the fundame ntal que stion of
what is an appropriate way to de scribe e xamples. At prese nt, e xamples
use d in a ge ne ralization system are crafte d by the system de ve lope r
based on the de ve lope r’s inte rpretation of what is the rele vant informa-
tion and how the information may be de scribe d. This is diffe rent from
most othe r ge ne ralization problems, such as attribute-based induction
s W u, 1993 ., whe re attributes of e xample s are taken for grante d. Thus
the de scription language is he avily de pe nde nt on the ``correctne ss’ ’ of
the de velope r’s interpretation. In orde r to re duce the bias e ffect, the re
is ce rtainly a ne e d to e xplore a formalism by which de sign information
can be e xtracted and de scribe d with ``rightful’’ fe atures that e nable a
ge ne ralization program to find a de sire d targe t conce pt.
Anothe r difficulty that a de scription language has to de al with is
that fe atures of de sign e xamples usually posse ss grade d structures.
Inste ad of be ing e quivalent, fe ature s m ay be characte rize d by a de gre e
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 189

of typicality, and typical fe atures may outpe rform atypical feature s in


re prese nting a de sign concept. H ow to re flect the typicality in de sign
feature s is anothe r challe nging issue in de te rmining de scription lan-
guage s.

Selecting and Developing Generalization Techniques


O nce the knowle dge represe ntation and de scription language are de ter-
mine d, the proble m of ge ne ralization may be vie we d as an induction
proble m guide d by the e xpe ctation impose d by the knowle dge re prese n-
tation. Thre e possible case s m ay arise in the se arch for ge ne ralization
technique s. O ne is that e xisting m achine le arning te chnique s are suffi-
cie nt to provide tools for ge ne ralizing de sign knowle dge in the e xpe cte d
forms. This is de sirable but not always re alistic. The se cond case is that
the re are e xisting le arning te chnique s that partially satisfy the de ter-
mine d purposes; e xte nsions of the se technique s to handle de sign prop-
e rties are the n re quired. The third and pe rhaps the most difficult case is
that, if no le arning technique s are available that can le arn de sign
knowle dge and frame it into the pre de termine d knowle dge repre senta-
tion, the n ne w gene ralization technique s have to be de ve lope d.
W e have so far ide ntifie d the crucial role of a knowle dge re prese n-
tation in directing the de ve lopme nt of a gene ralization system in de sign.
The de scription language for de sign e xample s is de te rmine d based on
the unde rstanding of the e ssence of de sign and important prope rtie s in
a de sign e xample. The de ve lopme nt of ge ne ralization te chnique s re -
quire s se arching for e xisting te chnique s and choosing suitable one s to
adopt or de ve loping ne w te chnique s based on the unde rstanding of the
limitations of e xisting one s. In orde r to de scribe furthe r the usability of
the se three e ssential tasks for ge ne ralization proble ms in de sign, the
ne xt se ction shows how a knowle dge represe ntation, a de scription
language , and some ge ne ralization te chnique s are sele cte d or de ter-
mine d in an application of ge ne ralizing de sign conce pts.

LEARNING EMPIRICAL NETWORKS BY GENERALIZATION


De sign is an intellige nt activity that has differe nt goals at differe nt
de sign stage s and draws on many information source s. It is pe rform e d in
a wide varie ty of domains and usually involves a num ber of proce sse s
de signate d to achie ve differe nt goals. Le arning systems must not only
190 H. LI AND X. WU

be able to unde rstand and ge ne ralize knowle dge from de sign proce sse s
but also be able to fully utilize the information source s to ge ne ralize
re sults that are use ful in assisting future de sign proble m s. The varie ty of
information sources provides many pote ntial application areas for ap-
plying gene ralization strate gies. The re fore, in orde r to de ve lop a le arn-
ing system for de sign domains, an inte gration of se ve ral ge ne ralization
strategie s may be use d.
The aim of this section is not to e nume rate all ge ne ralization
strategie s e xhaustively but rathe r to give an illustration of how gene ral-
ization strate gies can be applie d to le arn a pre de term ine d de sign
knowle dge repre sentation. Spe cifically, the following subsections pre -
se nt how a de sign re prese ntation is de ve lope d and le arne d using an
inte gration of ge ne ralization strategie s.
Figure 1 shows the overall proce ss of ge ne ralizing e mpirical ne t-
works from de sign e xamples. E mpirical ne tworks are de ve lope d to
re prese nt de sign conce pts that support conce ptual stages of de sign. The
ge ne ralization system has three major le arning m odule s: CO NCE P-
TO R, E FD 1 , and DPF. E ach module use s a spe cific le arning strategy to
de al with part of the de sign information and produce s part of the
e mpirical knowle dge . The combination of the se module s e nables the
formation of e mpirical ne tworks. Spe cifically, CO NCE PTO R is a le arn-
ing module that groups similar de sign e xamples into conce ptual space s
that provide the basis for ge ne rating e mpirical ne tworks. This module
also me asure s de fault value s and value range s for de sign variable s.

F igu r e 1 . A ge ne ralization mode l for learning empirical ne tworks.


LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 191

E FD 1 use s guide d se arching and re gressional tools to discove r pie ce -


wise line ar e quations to form e mpirical formulas. DPF applie s e mpirical
pattern re cognition te chnique s to e xtract useful de sign patterns from
qualitative de sign information. The colle ction of the se le arning m odule s
forms a multistrategy le arning approach for ge ne rating e m pirical ne t-
works. Major compone nts of this ge ne ralization mode l are de scribed
be low.
The subse que nt prese ntation is rathe r brie f, as it is only to de mon-
strate the ide ntifie d three tasks in ge ne ralizing de sign knowle dge. A
more de taile d de scription of the le arning process is give n in Mahe r and
Li s 1994 .. The e mphasis is placed on de scribing E FD 1 , which is an
e volved version of the E FD module used previously s Mahe r & Li, 1994 ..

Empirical Networks as the Representation for Design Concepts


As pre viously de scribe d, the first step in ge ne ralizing de sign knowle dge
is to de te rmine a repre sentation for de sign knowle dge . This re quires
e xamining the nature of de sign and me asuring the rele vance of a
re prese ntation to the comple xitie s of de sign. The re pre sentation re le -
vance is an important consideration in ge ne ralizing de sign knowle dge .
Com putationally, de sign can be conside re d as a proce ss of ide ntifying
re le vant variable s and assigning value s to the variable s. De sign variable s
include qualitative and quantitative variable s. Associations m ap into
what we call inte ractions among de sign variable s. Q uantitative variable s
are presumably associate d by e m pirical formulas; qualitative variable s
are pre sumably linked by e mpirical patterns. From this vie wpoint,
e mpirical ne tworks have be e n propose d as a knowle dge re pre sentation
s Li, 1993 .. An e mpirical ne twork has two compone nts, node s and
inte rrelationships, as shown in Figure 2.
A node comprises a de sign variable , its de fault value , and value
range. The variable s e ncom pass both the require me nt and de scription

F igu r e 2 . Compone nts in an em pirical ne twork.


192 H. LI AND X. WU

attributes, although a single variable is not ide ntifie d as one or the


othe r. Inte rrelationships re fle ct e mpirical associations among de sign
variable s. Two kinds of associations are include d: e mpirical formulas
capturing associations among quantitative variable s and de sign patte rns
capturing associations among qualitative variable s. An e xample of an
e mpirical ne twork for a de sign conce pt is illustrate d in Figure 3. This
e mpirical ne twork is actually ge ne rated by the ge ne ralization mode l
from a num ber of de sign e xample s that will be de scribed later. The

F igu r e 3 . An e mpirical ne twork for continuou s girde r bridge de sign.


LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 193

node s are variable s for a continuous girde r bridge de sign. Re lationships


among variable s are e mpirical formulas or de sign patterns re le vant to a
continuous girde r bridge de sign.
For e xample , two quantitative variable s, de ck ] width s DW . and
girde r ] width s GW ., are associated by an e m pirical formula v DW y
10.32GW s 0 4. As anothe r e xam ple , two qualitative variable s,
girde r ] mate rial and supe rstructure ] type, are linke d by a de sign patte rn
v pattern 1 4 indicating that the se two variable s and the ir value s as
spe cifie d in the pattern fre que ntly occur toge the r in de sign de cisions.
Abbre viations used in Figure 3 are e xplaine d in Table 1.

Ta b le 1 . Attribute s and value s for a continuou s bridge de sign e xample

Abbre viation Attribute s V alue s

BN Bridge nam e Mr-124


LC Location Sutton-in te rchange
TT Traffic type V ehicle
CY Constructi on ye ar 1984
CS Cost Low
PH Pie r he ight 1.3 m
TL Traffic lane s 3
CR Cross a rive r r road Rive r
RB River be d condition Me dium
PL Supe rstructure de sign load NA ARSA -1976
BL Substructu re de sign load NA ARSA -1976
LN Bridge length 95.4 m
DW De ck width 9.3 m
DC Se ction de pth at the cente r 1.42 m
DS Se ction de pth at the support 2.03 m
SN Span numbe r 3
MS Main spain 93.4 m
ST Supe rstructure type Continuou s-girde r
GH Girde r he ight 1.04 m
GW Girde r width 3.24 m
GS Girde r shape Trape zoid
NG Numbe r of girde rs in a se ction 5
NTC Numbe r of te ndons at the cente r 124
NTS Numbe r of te ndons at the support 156
FT Foundation type E xcavative
BT Bearing type Pot-type
DJ De ck joint Claw-and-gland
DA De ck are a 1206 cm 2
TA Tendon are a 49.3 cm 2
194 H. LI AND X. WU

Compared with othe r re pre sentation sche m as s Li, 1994 ., such as


rule s, object-orie nted approache s, and Baye sian and Markov ne tworks,
an e mpirical ne twork has quite a few advantage s. First, an e mpiri-
cal ne twork can e xpress complicate d de sign de pe nde nce s. In an e m-
pirical ne twork a node indirectly influe nce s its ne ighbors through an
e mpirical formula or de sign patte rn, whe re as node s in the obje ct-ori-
e nte d approach and Bayesian and Markov ne tworks are associate d with
dire ct links only. Se cond, an e mpirical ne twork can distinctive ly re pre -
se nt quantitative and qualitative variable s, whe reas othe r ne tworks are
unable to spe cify the diffe re nce s be twe e n variable s of differe nt type s.
This se cond ability is important in de sign knowle dge re pre sentation
be cause de sign activities must involve both quantitative and qualitative
variable s s He rrmann & B e ckmann, 1994 .. Q uantitative and qualitative
variable s be have dissimilarly in de sign reasoning and thus must be
re prese nte d in diffe rent ways. Third, as we will de monstrate later,
e mpirical ne tworks can be autom atically ge ne rated by a se t of le arning
me thods.

A Formalism for Describing Design Examples


The formalism use d to de scribe de sign information is based on the
assumption that the e sse nce of de sign e xamples can be de scribed by
feature s from both de sign re quire me nts and structural de scriptions.
De sign re quireme nts state an e nvironme nt within which the structural
de scription is ge ne rated. The structural de scription, on the othe r hand,
de picts a plausible solution to the de sign proble m. Toge the r, de sign
re quireme nts and structural de scription display a logical continuity of
how the de sign require me nts can be transform e d to a structural de scrip-
tion. Ne e dle ss to say, fe atures of a de sign e xample have to be drawn
from de sign re quire me nts and structural de scriptions. O the rwise , it will
be difficult for a ge ne ralization program to ``unde rstand’’ the de sign
e xample , not to m e ntion to ge ne ralize any me aningful re sults. A contin-
uous bridge de sign e xample with attribute-value pairs de scribing de sign
re quireme nts and structural de scriptions is illustrated in Table 1.
Attribute s s feature s . in a de sign e xample are colle cte d from de sign
re quireme nts and structural de scriptions. V alue s of attributes can be
qualitative or quantitative. Attributes that have nume rical value s are
re garde d as quantitative attributes; attributes that do not have nume ri-
cal value s are qualitative attribute s. For instance , traffic type has
qualitative value s of ve hicle and pe de strian; pie r he ight has quantitative
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 195

value s of 1.2 and 39.7 m. Some attribute s, e ve n though the y can have
quantitative value s, such as cost, cannot be pre cise ly calculate d be cause
the crite ria used to me asure the ir values are changing. The value s of
the se attribute s are base d on a ve rbal asse ssme nt of the quantitative
e le me nts and the refore are re garde d as qualitative. For e xample , the
attribute cost is value d in te rms of low, m e dium , and high.
The fore going formalism for de scribing de sign e xample s is consis-
tent with the way de sign solutions are prese nte d and docume nted. In a
de sign spe cification, de sign re quire me nts are usually re corde d as part of
the docume nt, and structures are mainly de scribed as drawings. The
information on de sign history is usually not include d in the final
docume nts. Howe ve r, the formalism provides little aid in how to trans-
late de sign re quire me nts and structural de scriptions into fe atures and
the ir related value s. This re late s to the e le m e ntary-term proble m
s Saxena, 1989 ., which de als with finding e le m e ntary units of de scription,
or te rms, that are commonly acce pted and ``hardwire d’’ in a spe cific
domain. Conside rable re se arch e ffort is re quired to clarify whe the r
the re are e le me ntary term s in a spe cific de sign domain so that de sign
e xample s can be ``legitimate ly’’ de scribed by the se terms.

Generalizing Empirical Networks


O nce the e mpirical ne tworks are de term ine d as the targe t repre senta-
tion for de sign knowle dge and de sign e xample s are de scribed by fea-
tures from de sign require me nts and de sign de scriptions, it be come s
possible to use ge ne ralization strategie s s including e xisting we ll-known
me thods. to produce e mpirical ne tworks from de sign e xample s. An
e mpirical ne twork, illustrate d in Figure 3, consists of node s, e mpirical
formulas, and de sign patterns. The task of ge ne ralizing e mpirical ne t-
works m ay the refore be subdivide d into ge ne ralizing de sign variable s,
the ir de fault value s, and value range s; ge ne ralizing e mpirical formulas;
and ge ne ralizing de sign patterns. For the sake of clarity, the se subtasks
are de scribe d in turn.

Generalizing Design Variables, Their Default Values, and Value


Ranges as Nodes for Empirical Networks. Ge ne ralizing de sign vari-
able s from de sign e xample s involve s assigning an e xample to a spe cific
conce pt space . If the e xample doe s not fit into an e xisting conce pt
space , a ne w conce pt space is forme d. E ach concept space is characte r -
196 H. LI AND X. WU

ized by a se t of attributes that are mutually share d by the e xamples in


the space . The se attribute s contribute to the variable s of a de sign
conce pt. The algorithm for ge ne ralizing de sign variable s is as follows.

NoO fConce pts s 0


DO I s 1 TO NoO fE xample s
IF I s 1 TH E N
NoO fConce pts s 1
assign E xample s 1 . into Conce pt s 1 .
E LSE
DO j s 1 to NoO fConcepts
Sim ilaritys j. s No of matche d attributes with
Conce pts j. r No of attribute s
LO O P
MaxSim ilarity s MAX v Similaritys j., j s 1 TO NoO fConce pts4
MaxNo s value of j which give s the Maximum Similarity
IF MaxSimilarity is gre ate r than threshold TH E N
assign the e xample into Conce pts MaxNo.
E LSE
NoO fConce pts s NoO fConcepts q 1
assign E xample s i. into Conce pts NoO fConce pts.
E ND IF
E ND IF
LO O P

The algorithm proce sses all available e xample s, rathe r than charac-
teristic sample s only, and puts the m into differe nt conce pt space s. For
e ach concept space, the algorithm ide ntifies the attribute s associate d
with all the e xample s in the conce pt space .
The intere sting aspe ct of this algorithm is that it gene ralize s multi-
ple conce pt space s; this e nable s the system to circumve nt a lim itation
confronted by many othe r gene ralization systems, as m any curre nt
conce pt formation systems can ge ne ralize only one conce pt from e xam-
ple s. The ability to le arn multiple conce pts is important in de sign,
be cause de signs are ofte n distinguishe d by the diffe rence in the way
the y are de scribe d. This differe nce occurs on at le ast two le vels: the
name s of the attributes use d to de scribe the de sign and the value s that
the attribute s assum e for a give n de sign. Grouping de sign e xamples
according to the similarity of the ir attribute s provide s a technique for
ge ne ralizing differe nt de sign conce pts rathe r than being lim ited to
ge ne ralizing a single conce pt.
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 197

Two proce sses are use d to ge ne ralize de fault value s for de sign
variable s. The de fault value s for quantitative variable s are de te rmine d
by the we ighte d ave rage of value s of the variable s in a conce pt space .
The de fault values for qualitative variable s are de term ine d using the
freque ncie s in which ce rtain nominal value s occur in the de sign e xam-
ple s of a conce pt space , the m ost fre que nt one s becoming the de faults.
V alue range s for quantitative variable s are range s de fine d by the ir
maximum and minimum limits of the nume rical distribution. V alue
ranges for qualitative variable s are colle ctions of all values that appe ar
in the e xample s of a conce pt space .

Generalizing Empirical Formulas. The E mpirical Formula Discove ry


ve rsion 1 s E FD .1 system is a le arning m odule that e mploys a guide d
se arching m e thod to ge ne rate pie ce wise line ar e quations by pie ce wise
re gre ssion to approximate com plex re lationships betwe e n two quantita-
tive de sign variable s. It ope rate s in e ach conce pt space within which
similar de sign e xample s are groupe d. This le arning module e m ploys a
ge ne rate-and-te st approach to ge ne ralizing e mpirical formulas. The
ge ne rate-and-te st approach has a data-inde pe nde nt proce dure for se -
le cting or gene rating mode ls for e mpirical formulas. E ach mode l is the n
tested against the e ntire se t of available training data and ide ntifie d as
e ithe r an acce ptable mode l of ge ne ralization or an unsuitable mode l for
the give n data set. Param e ters of an acce pte d m ode l are e valuate d by
re gre ssion tools so that a usable e mpirical formula is produce d. Be cause
the ge ne rate-and-test strategy judge s the pe rformance of a mode l ove r
many e xample s, rathe r than making de cisions based on individual
e xample s, it can accomm odate quite ``noisy’’ e le me nts in the training
data. Moreover, since the strategy doe s not e mploy domain knowle dge
in the ge ne ralization, it can be use d in any de sign dom ain.
E FD 1 e m ploys a line ar formula template w k s x r y . q b s 0 x in ge n-
e ralizing pie ce wise e mpirical formulas; x and y in the te mplate are
de sign variable s, and k and b are constants. Two criteria are used to
control the searching of pie cewise e mpirical formulas. MR is the
majority ratio de noting the minim um fraction of instance s that must be
cove red by an e quation within 1 " « ; « is the e rror bound. The
majority ratio also e xhibits the accre dibility of the e quation.
Figure 4 illustrate s how nonline arity distributed data se ts can be
mode le d by a se t of line ar e quations. Give n two de sign variable s X and
Y and a set of coordinate s w s x j , yj ., j s 1, . . . , n , whe re n is the total
numbe r of e xample s x that are e xtracted from the de sign e xample s, the
198 H. LI AND X. WU

F igu r e 4 . Ge ne rating pie cewise line ar equations to approxima te a nonline ar distributio n.

me chanism for ge ne rating pie ce wise line ar e quations is outline d by the


pse udocode displaye d in Figure 5.
The le arning proce ss can be be st de scribe d as a se arching process
guide d by the control criteria s MR and « .. It first sorts input data se ts
in se que nce according to the distance from s x j , yj . to the origin s 0, 0 ..
The n the le arning module incre me ntally searche s through the data
space in orde r to re trie ve the data se gme nts that fit line ar e quations. A
data segme nt is an annular shape de fine d by two radii d 1 and d 2 , as
shown in Figure 4. For e ach data se gme nt, E FD 1 use s the line ar
e quation k s X r Y . q b s 0 to mode l the data e nclosed in the se gme nt.
After computing constants k , b and form ulating the e quation, the
le arning module me asure s the pe rce ntage of the e xample s in a data
se gme nt that are cove red by the e quation in the 1 " « bound. All
e quations that cove r at le ast MR pe rce nt of the e xample s in a data
se gme nt are e xtracted to sample the nume rical associations in the data
se gme nt. The results of this le arning proce ss are a se t of pie ce wise
line ar e quations and the ir applicable boundarie s.
E FD 1 is applie d separately to conce ptual space s produce d by CO N-
CE PTO R. Q uantitative variable s in a conce ptual space , as we ll as the ir
corre sponding value s, are e xtracted from the e xample data. Candidate
e quations are forme d by re trie ving all possible pairs of variable s. E FD 1
the n ope rates on e ach pair of variable s and the ir value s to ge ne rate
pie ce wise line ar e quations. The re sultant e quations ge ne rate d from
space 3 are listed in Table 2. For comparison purpose s, re sults from
E FD are also listed in Table 2.
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 199

1
F igu r e 5 . The le arning me chanism of EFD .

Table 2 shows that E FD 1 ge ne rates se ven se ts of pie ce wise line ar


e quations at MR s 0.79 and « s 0.3, whe reas E FD ge ne rates only five
e quations at MR s 0.75 and « s 0.5. This indicate s that the mode ling
capacity of E FD 1 is supe rior. The cove rage rates of e quations from
E FD 1 are also highe r, indicating highe r pre cision of the results. B e -
cause E FD 1 doe s not require the input data sets to have a high de gre e
200 H. LI AND X. WU

1
Ta b le 2 . Comparis on of re sults from EFD and EFD

Results from EFD 1 Results from EFD

3 6.9 ? G H y M S s 0 s C O s 93 % . a 4 7 .6 ? G H y M S s 0 s C O s 86 % .
s 0.45 m ( G H ( 1 .3 2 m ; 48 m ( M S ( 73 m .
4 7.5 ? G H y M S s 0 s C O s 98 % .
s 1.32 m ( G H ( 1 .9 7 m ; 73 m ( M S ( 85 m .
4 3.2 ? G H y M S s 0 s C O s 84 % .
s 1.97 m ( G H ( 2 .0 5 m ; 85 m ( M S ( 10 7 m .
G H y 1.38 ? G W s 0 s C O s 97 % . G H y 1 .2 5 ? G W s 0 s C O s 77 % .
s 0.45 m ( G H ( 1 .6 3 m ; 0.8 m ( G W ( 1 .4 3 m .
G H y 1.17 ? G W s 0 s C O s 85 % .
s 1.63 m ( G H ( 2 .0 5 m ; 1.4 3 m ( G W ( 2.37 m .
D W y 1 0 .3 2 ? G W s 0 s CO s 84 % . D W y 10 .3 2 ? G W s 0 s C O s 93 % .
s 8.3 m ( D W ( 2 8 .4 m ; 0 .8 m ( G W ( 2 .3 7 m .
P H y 0 .95 ? C L s 0 s CO s 93 % . P H y 1.23 ? C L s 0 s C O s 82 % .
s 4.2 m ( P H ( 6 .7 m ; 2 .1 m ( C L ( 5.3 m .
P H y 1 .21 ? C L s 0 s C O s 89 % .
s 6.7 m ( P H ( 1 3 .7 m ; 5 .3 m ( C L ( 1 2 .6 m .
P H y 1 .23 ? C L s 0 s C O s 91 % .
s 13 .7 m ( P H ( 1 8.5 m ; 12 .6 m ( CL ( 16 .4 m .
P H y 1 .23 ? C L s 0 s C O s 88 % .
s 18 .5 m ( P H ( 2 7.4 m ; 16 .4 m ( CL ( 28 .4 m .
D W y 3 .67 ? T L s 0 s C O s 79 % . D W y 3.67 ? T L s 0 s C O s 89 % .
s 8.3 m ( D W ( 2 8 .4 m ; 2 ( T L ( 6 .
M S y 5 9.7 ? G W s 0 s C O s 81 % .
s 48 m ( M S ( 68 m ; 0 .8 m ( G W ( 1 .0 4 m .
MS y 3.67 ? GW s 0 s CO s 86% .
s 68 m ( M S ( 84 m ; 1 .0 4 m ( G W ( 1 .6 3 m .
M S y 3 .6 7 ? G W s 0 s C O s 93 % .
s 84 m ( M S ( 93 m ; 1 .6 3 m ( G W ( 1 .9 4 m .
M S y 3 .6 7 ? G W s 0 s C O s 85 % .
s 93 m ( M S ( 10 7 m ; 1.94 m ( G W ( 2.37 m .
B L y 2.14 ? M S s 0 s C O s 83 % .
s 98 m ( B L ( 2 45 m ; 4 8 m ( M S ( 8 4 m .
B L y 3.61 ? M S s 0 s C O s 88 % .
s 24 5 m ( B L ( 48 1 m ; 84 m ( M S ( 10 7 m .

a
C O , cove rage ra te .

of line arity, the limitation of E FD of only coping with line ar distribu-


tions is e liminate d. The disadvantage of E FD 1 is the incre ase of
computational tim e due to the re pe titive regressional analysis. Furthe r
re search is ne e de d to e ase the computational burde n of this le arning
module .
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 201

It is ne ce ssary to note that value s of the control criteria are


adjustable in making a trade -off be twe e n the cove rage rate and the
accre dibility. E xpe rim e nts show that lowe r MR and highe r « yie ld m ore
e quations and vice versa. For e xample , E FD 1 produce s 13 pie ce wise
line ar e quation se ts at MR s 0.79 and « s 0.5 and only 2 e quation se ts
at MR s 0.80 and « s 0.1.

Generalizing Design Patterns by Empirical Pattern Recognition.


The ge ne ralization of de sign patterns de als with re cognizing e mpirical
conne ctions in qualitative de sign data. It is also applie d in e ach conce pt
space . In this ge ne ralization, variable s with qualitative value s are e x-
tracte d and re pre se nte d as node s in an n -space. De sign e xamples are
fed into the n -space , and conne ctions are e stablishe d through values of
the attribute s. E ach conne ction has a we ight that indicate s the rate of
the e xample numbe r posse ssing the value pair. After all e xample s have
be e n e xe cute d, conne ctions with we ights ove r a prede te rmine d thresh -
old are e xtracte d as de sign patte rns.
The algorithm for ge ne rating de sign patterns comprise s two phases:
structure le arning to se t up the basis for positioning de sign patterns and
link ge ne ration in orde r to asse mble de sign patte rns. The two phase s
are furthe r de scribe d as follows.

Computing the Frequency Distribution: A group of qualitative at-


tributes are the by-product of le arning de scriptive attributes for de sign
conce pts. Assume we have ide ntified n qualitative attribute s from the
de scriptive attribute s for a de sign concept; the se qualitative attributes
formulate the basis of pattern ge ne ration. E xample s are used again to
provide input information for le arning de sign patterns. In e ach e xample ,
qualitative attribute s, toge the r with the ir discrete value s, represe nt a
feasible state within which e ach pair of value s contribute s one unit to a
link.
The le arning me thod e stablishe s and update s the we ight of the links
using e ach feasible state . The re are two possible case s. First, if a link
matching a value pair in a feasible state alre ady e xists in the patte rn,
the n the we ight is increase d by 1. Se cond, if the link doe s not e xist, the n
the link will be e stablishe d and its we ight will be indicated as 1. As such,
all feasible state s can be proce ssed to ge ne rate the we ight of e ach link,
and the we ights are converted to freque ncies by ave raging we ights with
the total numbe r of feasible state s.
202 H. LI AND X. WU

Extracting Design Patterns: W e have com puted the freque ncy distri-
bution Pi s i s 1, . . . , k ., whe re k is the numbe r of fe asible state s. The
distribution is nonne gative and attains the value 0 if the re is no link at
all among the value s of qualitative attribute s be ing conside re d. To
de te rmine de sign patterns, the links with the freque ncies that e xcee d a
pre de termine d thre shold are conside red as strong links, and value s of
qualitative attribute s associated with strong links are e xtracte d. The
e xtracte d attribute -value pairs form a de sign pattern.

Coordination of the Generalization


The combination of the re sults ge ne ralize d from the thre e proce sse s
above produce s an e mpirical ne twork, as shown in Figure 3. The thre e
proce sse s are coordinate d in the following way to yie ld a cohe re nt
ge ne ralization.

Step 1: V ariable s, de fault values, and value range s gene ralize d from the
first m odule constitute the node s and structure of the e mpirical
ne twork.
Step 2: Run the se cond module to ge ne rate e mpirical formulas among
quantitative variable s.
Step 3: Run the third m odule to ge ne rate de sign patterns among
qualitative attribute s.

As me ntione d e arlier; the e mpirical formulas and de sign patte rns


are an integral part of the e m pirical ne twork.
The system for ge ne ralizing de sign conce pts has bee n pre sented.
This system has e xercised all thre e tasks ide ntified e arlie r. It illustrate d
how the comple xitie s of de sign affect the de ve lopme nt of e mpirical
ne tworks as the re pre sentation for de sign conce pts. It also illustrate d
how the de scription language is de te rmine d so that information on
de sign is e xpresse d. Most important, the illustration has shown that the
ge ne ralization proce ss is large ly controlle d and directe d by e mpirical
ne tworks as the re prese ntation for de sign conce pts. The structure of the
e mpirical ne tworks provides a common thread among ge ne ralization
module s. Individually, e ach module ope rate s on spe cific de sign data and
ge ne ralize s part of the knowle dge require d in the structure of e mpirical
ne tworks; colle ctively, the thre e module s inte grate and ge ne ralize the
ne ce ssary knowle dge for assem bling e mpirical ne tworks.
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 203

FUTURE RESEARCH AND CONCLUDING REMARKS


In this pape r ge ne ralization proble ms in de sign have be e n frame d into
three tasks: de te rm ining a knowle dge repre sentation for ge ne ralization
re sults, de te rmining a de scription language for de sign e xample s, and
se le cting or de ve loping ge ne ralization te chnique s to produce the knowl-
e dge re pre sentation. The knowle dge re pre sentation is ide ntified as
the pe rce ption of an intelligent obse rver who takes into account how
the comple xitie s of de sign can be computationally accommodated by the
re prese ntation and de sign m ode ls cre ate d upon the re prese ntation. The
de scription language , on the othe r hand, repre sents an abstraction of
information in de sign. The se le ction and de velopme nt of ge ne ralization
technique s aim at mapping de sign e xamples into the knowle dge re pre -
se ntation.
In orde r to provide a continuous vie w of ge ne ralization problems in
de sign, it is ne cessary to note seve ral issue s relate d to significant ope n
proble ms in the re search on ge ne ralizing de sign knowle dge .

The Sample Complexity and the Open-Ended Nature


The study of sample comple xity s Natarajan, 1991 . is done to e stablish a
quantitative relationship be twe e n the number of e xample s re quire d in
ge ne ralization and the appropriatene ss of the conce pt. A conce pt ge ne r-
alized form two de sign e xample s m ay we ll be re garde d as a coincide nce .
How many e xample s are ne e de d before we can de clare the ge ne ral
applicability of a de sign conce pt? This is a que stion that ne e ds to be
inve stigate d and answe re d in the use of gene ralization strate gies for
de sign knowle dge acquisition.
It is ne cessary to re ite rate the ope n-ende d nature of ge ne ralization.
This implie s that whe n one make s a ge ne ralization about som e aspe ct
of de sign, the re is no crite rion to indicate to what e xte nt the gene raliza-
tion re sults will re main valid. The re is also no natural lim it to the le ve l
of de tail in which knowle dge of de sign may be de scribed, be cause a
ge ne ralization is always drawn from a limite d numbe r of e xample s. The
ope n-ende dne ss is one of the reasons why it is important to de fine
care fully a re pre se ntation through which the ge ne ralization re sults are
e xpre sse d and why it is important to de fine a de scription language so
that crucial de sign prope rtie s are re prese ntable , while trivialitie s are
ignore d.
204 H. LI AND X. WU

The Time Effect in Generalization


A ge ne ralization program is re quired to look continually for late r de sign
e xample s be cause late r de sign e xample s re prese nt rece nt advance s of
technologies and unde rstandings in de sign. The program ne e ds to
de te ct and de le te the obsole te parts in de sign conce pts and replace
the m with information from ne w e xample s. O ne way to re alize this is to
assign differe nt we ighting ratios to de sign e xamples cre ated at differe nt
time s: the late r e xample s are give n more we ight than the othe rs, so that
later e xamples have more influe nce in contributing to the ge ne ralization
re sults. Future re search is ne e de d to imple me nt and test the tim e e ffect
in ge ne ralization.

Inconsistency
It is ne ce ssary to note that an e sse ntial re quireme nt for ge ne ralizing
de sign knowle dge is that the represe ntation, the de scription language ,
and the ge ne ralization te chnique s be consiste nt. Inconsistency m ay arise
if e ithe r s 1 . the de scription language is insufficie nt to prese nt de sign
information for ge ne ralizing the targe t repre sentation or s 2 . the ge ne r-
alization te chnique s are unable to shape the de sign e xample s into form s
de fine d in the knowle dge repre sentation. In ge ne ral, the re are no
facilities to e lude the inconsistency or recove r it during the gene raliza-
tion. Although gene ralization technique s can e asily incorporate proba-
bilistic and statistical me thods and the se m e thods can be e rror re silient
and can improve ge ne ralizations over a pe riod of time , ge ne ralization
technique s do not provide any fe e dback on how to re vise the knowle dge
re prese ntation or the de scription language to im prove the use fulne ss of
the ge ne ralization re sults in de sign. The difficulty in avoiding inconsis-
tency makes the de velopme nt of ge ne ralization systems for de sign an
e xpe rime ntal proce ss.
In concluding, this pape r vie ws a ge ne ralization problem in de sign
as an induction proble m, which involve s the de te rmination of a target
knowle dge re prese ntation for gene ralization and a de scription language
for de sign e xample s, as we ll as the se le ction or de ve lopm e nt of gene ral-
ization technique s to gene rate the knowle dge repre se ntation from de -
sign e xample s. The ge ne ralization proce ss is large ly guide d by the target
re prese ntation. The repre sentation de line ates a ``filtere d’’ pe rce ption of
an inte llige nt obse rver, the de scription language repre senting a le ve l of
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 205

abstraction of de sign information. Ge ne ralization te chnique s e xtract


re le vant information from de sign e xample s to fit into the targe t re pre -
se ntation.
From the choice of a knowle dge re prese ntation to the sele ction of a
de scription language , the de signe r of a ge ne ralization system builds in
biases conce rning useful and irre le vant information in de sign. The se
biases constitute both the strength and the we akne ss of the gene raliza-
tion system . If the biases are appropriate , the y can provide the basis for
e nsuring the usefulne ss of the ge ne ralization results; othe rwise, the y
can pre vent the system from inferring any useful ge ne ralizations. The re-
fore, along with issue s m e ntione d above, future rese arch is re quired to
de ve lop me thods that will de te ct possible inade quacie s in the thre e
tasks we have ide ntifie d and automatically re pair the m, so that the
ge ne ralization re sults will be m ore use ful in de sign.

REFERENCES
Arciszewski, T., M. Mustafa, and W. Z iarko. 1987. A methodology of de sign
knowle dge acquisition for use in learning e xpert syste ms. In t. J. Ma n -
Ma ch in e Stu d ie s 27:23 ] 32.
Brachman, R. J. and H. J. Le ve sque , e ds. 1985. Read in gs in kn ow led ge represen ta -
tio n . Los Altos, CA: Morgan Kaufmann.
Coyne , R. D. 1992. Mode lling the eme rge nce of design de scription across
sche mes. E nv iro n . Plan n in g B 18:427 ] 458.
Coyne , R. D., M. A. Rose nman, A. D. Radford, M. Balachandran, and J. S.
Ge ro. 1990. Kn o w led ge-b ased d esign system s . Reading, MA: Addison-W e s-
le y.
Craw, S., and D. H. Slee man. 1995. Refine ment in response to validation. E xp ert
Syst. Appl. 8 s 3 .:343 ] 349.
Dasgupta, S. 1989. The structure of design proce ss. In Adv an ces in co m pu ters,
V ol. 28, ed. M. Yovits. New York: Acade mic Pre ss.
Duffy, A. H. B., and K. J. MacCallum. 1989. Computer re prese ntation of
num erical expe rtise for preliminary ship design. Ma r. Tech n o l. 26:289 ] 302.
Dym, C. L., R. P. Henche y, E. A. De lis, and H. Gonick. 1988. A knowle dge -base d
system for automated archite ctural code che cking. Co m pu t.-A id ed D esign ,
20 s 3. :137 ] 145.
Eastman, C. 1969. Cognitive proce sses and ill-structure d proble ms: A case study
for de sign. In IJCAI-69 , Washington, DC, pp. 669 ] 690.
Finge r, S., and J. R. Rinde rle . 1990. Transforming behavioral and physical
re presentations of me chanical de sign. Pro ceed in gs of th e First In tern a tio n al
206 H. LI AND X. WU

Wo rksh o p on Fo rm al Meth o d s in E n gin eerin g D esign , Ma n u fa ctu rin g an d


Assem bly, Colorado, pp. 133 ] 151.
Fishe r, D. 1987. Knowledge acquisition via incre mental conce ptual cluste ring.
Ma ch in e L earn . 2:139 ] 172.
Gero, J. S. 1990. Design prototype s: A knowle dge repre sentation sche ma for
de sign. AI Ma g. Winter:26 ] 36.
Ginsberg, A., S. M. Weiss, and P. Politaki. 1988. Automatic knowledge base
re fineme nt for classification syste ms. Artif. In tell. 35:197 ] 226.
Herrmann, J., and R. Be ckmann. 1994. LE FT } a system that learns rule s about
V LSI de sign from structural de scriptions. Appl. Artif. In tell. 8:85 ] 108.
Kamal, S. Z ., F. Mistree , J. Sorab, and W. E. V anArsdal. 1989. The de velopme nt
of an inductive le arning system for de sign using e xperimental information.
In Artific ia l in telligen ce in d esign , e d. J. S. Ge ro, pp. 521 ] 538, Springe r-
V erlag, Ne w York.
Li, H. 1993. Le arning de sign conce pts to assist pre liminary de sign. Ph.D. thesis.
De partme nt of Archite ctural and De sign Scie nce , University of Sydne y,
Australia.
Li, H. 1994. Ma ch in e lea rn in g o f design co n cepts . Computational Me chanics
Pre ss.
Lu, S. C-Y., and D. K. Tche ng. 1991. Building layered mode ls to support
engine ering de cision making: A machine le arning approach. J. E n g. In d .
113 s 1 . :1 ] 9.
Mahe r, M. L., and H. Li. 1994. Le arning de sign conce pts using machine
le arning te chnique s. AI E D AM , pp. 95 ] 111.
Michalski, R. S. 1975. A variable -value d logic and its application to pattern-
re cognition and machine le arning. In Mu ltip le-v alu ed lo gic an d co m pu ter
scien ce , e d. D. Rine , pp. 204 ] 213, Morgan Kaufmann, Ne w York.
Michalski, R. S. 1983. A the ory and methodology of inductive le arning. AI Ma g.
20:111 ] 116.
Mitche ll, T. M. 1982. Generalization as search. AI Ma g. 18:203 ] 226.
Natarajan, B. K. 1991. Ma ch in e lea rn in g: A th eo retic al approach . Los Altos, CA:
Morgan Kaufmann.
Pahl, G., and W. Beitz. 1984. E n gin eerin g d esign . Berlin: Springe r-V e rlag.
Polya, G. 1954. Ma th em atic s an d pla u sible reason in g. Vo l. 1: In d u ctio n an d
an alo gy in m a th em atic s . Prince ton, NJ: Prince ton University Pre ss.
Re ich, Y. 1993. The de velopme nt of Bridge r: A methodological study of
re se arch on the use of machine le arning in de sign. Artif. In tell. E n g.
8 s 3 .:65 ] 181. Spe cial issue on Machine Le arning in De sign.
Riddle, P. J. 1989. Reformulation from state space to reduction space . Pro ceed -
in gs o f th e Sixth In tern a tio n al Wo rksh o p on Ma ch in e L earn in g , New York:
Morgan Kaufmann.
LEARNING DESIGN KNOWLEDGE AS GENERALIZATION 207

Saxena, S. 1989. E valuating alternative instance repre se ntations. Pro ceedin gs of


th e Sixth In tern atio n al Wo rksh o p on Ma ch in e L earn in g. Ne w York: Morgan
Kaufmann.
Simon, H. A. 1973. The structure of ill-structured proble ms. AI Ma g.
4 s 3 ] 4 .:181 ] 201.
Sriram, D., R. D. Logche r, N. Groleau, and J. Cherneff. 1989. DICE: An
obje ct-orie nte d program ming e nvironm ent for coope rative e ngine ering de-
sign. Te chnical Report No. IESL-89-03. Inte lligent Engine e ring Systems
Laboratory, Department of Civil Engine e ring, MIT, Cambridge, MA.
Sriram, D., R. D. Logche r, A. Wong, and S. Ahmed. 1991. An obje ct-oriente d
framework for coope rative e ngine ering de sign. In Com pu ter-a id ed co o pera -
tiv e prod u ct d ev elo pm en t, e d. D. Sriram, D. Logche r, and S. Fukuda. Ne w
York: Springe r-V erlag.
Thom as, D. E . 1985. Artificial inte lligence in design and test: Gue st editor’s
introduction. IE E E D es. Test 2 s 4 .:21 ] 34.
White hall, B. L., S. C-Y. Lu, and R. Ste pp. 1990. CAQ : A machine le arning tool
for engine e ring. Artif. In tell. E n g. 5 s 4. :189 ] 198.
Wu, X. 1993. Inductive learning: Algorithms and frontie rs. Artif. In tell. Rev .
7 s 2 .:93 ] 108.
Wu, X. 1997, in press. Ma ch in e lea rn in g , Albe x, Me lbourne , Australia.

View publication stats

Anda mungkin juga menyukai