AI Magazine Volume 11 Number 4 (1990) (© AAAI)

Articles

This article
proposes Design Reasoning
connectionism
as an Without Explanations
alternative to
classical
cognitivism
in under- R. D. Coyne
standing
design.

This article proposes connectionism as an alter-

The goals of design native to classical cognitivism in understanding the practice of
research are varied, design. It also considers the difficulties encoun- design? The most
and the sources of its tered within a particular view of the role of immediate impact
methods of inquiry explanations and typologies. Connectionism pro- would be expected
vides an alternative model that does not depend
are diverse, calling in the area of
on the articulation of explanations and typologies.
on such areas as improved tools and
history, hermeneu- environments for
tics, sociology, psy- designers. The tools
chology, philosophy, and computing. might ultimately include resource material to
Irrespective of the genre within which the facilitate decision making, such as textbooks,
research is undertaken, we can legitimately guides, codes, and computer programs. Envi-
ask, How does design research impinge on ronments might ultimately be systems of
organization and communication that are
tuned to the needs of designers.
However, the most important category
within tools and environments is essentially
of a different character. I refer here to the
contribution of design research to the intel-
lectual resources that facilitate discourse
about design. Thus, one of the most impor-
tant spin-offs from design research is the sets
of terms, models, intellectual structures and,
ultimately, the controversies that enable us to
talk about design.
We might suppose that design would still
successfully occur without these intellectual
structures, as it does in many cultures today.
Furthermore, insights could be expected to
arise from the deliberations of design practi-
tioners and critics whose activities do not
come under the heading of research. Howev-
er, there is obvious merit in articulating
design, separating it from other concerns,
and making it an object of study and careful-
ly reasoned reflection. The resulting intellectu-
al structures would be expected to benefit the
fostering of design ability, both formally in
design schools and as designers respond to
their environment and each other. Thus, the
products of design research are inexorably
linked with design education.

0738-4602/90/$4.00 ©1990 AAAI

Copyright, ©1990 American Association for Artificial Intelligence. All rights reserved. 0738-4602/90/$4.00
 Articles

There is little doubt that design research There are major differences of opinion
involves considerable controversy. As well as regarding the nature of design discourse. It is
inheriting the controversies from within the contended here that the advocacy of different
disparate genres of inquiry previously men- views on the role of explanations leads to dif-
tioned, design research has the potential to ferences in attitudes toward design teaching.
offend by impinging on our view of ourselves It is contended here that the content of the
as creative individuals. In a climate of contro- discourses held by partisans of different views
versy, it is appropriate to ask, What are the might not appear substantially different. That
benefits of the intellectual structures arising is, what is said at the drawing board or even
out of design research when there is so little in the lecture room might be similar. The dif-
agreement? The view propounded here is that ference is manifested primarily in policy for-
we are bound to make judgments on the rela- mulation and management: the discussion of
tive merits of particular models or points of design curricula, the way design programs are
view, and even the refusal to entertain a par- organized, and how the work of novice
ticular point of view can facilitate helpful dis- designers is evaluated.
course. Statements about what design is not These issues are brought into sharp relief in
can be as informing as views about what design teaching as well as when we consider
design is. how computers might be used in design.
The ultimate questions to be asked of a Apart from the practical uses that can be
view of design are, To what extent does the made of computer systems, there is the added
view facilitate discourse about design? and To benefit that the sharp focus provided by com-
what extent does a discussion of the differences, puter implementations allows us to critically
as well as the similarities, between this model test the consequences of our ideas. Comput-
and other models inform us about design? In ers allow us to take ideas to logical and, possi-
this way, one of the quests of design research bly, absurd limits. A clear example is the way
lies not only in discovering the truth about in which computers can be used to imple-
design but also in finding oppositions. We are ment theories about human problem solving
informed by the clash of views. and cognition. In the course of this discus-
The discussion to this point constitutes a sion, I look at two different approaches to
preamble to the presentation of two contrary understanding human cognition. I use simple
views that impinge on our thinking about explanations of computer implementations to
design: classical cognitivism and connection- demonstrate how the ideas work and high-
ism. Both largely call on the methods of light some implications for design teaching.
inquiry of cognitive science and computing. The two preeminent models of cognition
Both views are controversial, and both could are classical cognitivism and connectionism
be expected to be rejected outright by some (Clark 1989). Classical cognitivism focuses on
within the design community. It is contended the idea of symbols as mental representations.
here, however, that both views substantially Much AI work exploits the utility of this
contribute to discourse about design because approach, with its emphasis on rules and
of the terms and constructs they make avail- other explicit knowledge representation
able to us and by the nature of the paradoxes devices. Connectionism, however, focuses on
they bring to light. implicit knowledge representation. It is con-
cerned with modeling human reasoning at a
low level in an attempt to replicate the capa-
Design and Models of Cognition bility of human reasoning to transcend the
The predominant tradition of design teaching strictures of sharply defined categories and
(at least in the West) heavily depends on artic- formal logic.
ulating ideas through spoken and written From the point of view of design, the inter-
language. Drawings and the designed works esting aspect of these theories is that each can
themselves are considered insufficient as vehi- be used to account for different aspects of
cles for imparting design knowledge. Observ- design behavior, and each can be used to sup-
ing design and doing design are also insufficient. port different approaches to design. Classical
We require that they be supplemented with cognitivism supports an emphasis on design
words. Some of the features of the way we rules, hierarchies of types, and the articula-
talk about designing are addressed by Schön tion of the design process. However, connec-
(1982), with his characterization of reflection tionism can be said to emphasize experience
in action, and Rowe (1987). Although the and the emergence of design ideas without
popular idea of design discourse (or dialogue) attempts to articulate the process.
is not merely limited to words, the process is It is not the purpose of this article to advo-
imbued with a strong linguistic flavor. cate one view over the other or to preclude

WINTER 1990 73
Articles

Explanations are appeal to a commitment to the authority or

influence of someone else: “I did it this way
proffered to satisfy the because I am strongly influenced by the work
critics, promote decisions, of architect X who strongly favors roof lights.”
For most purposes, an unacceptable expla-
and persuade. nation is of the kind that appeals to an
unsupported preference: “I always do it this
way” or “I felt like a change.” Least accept-
able of all are explanations that appeal to
feelings that defy detailed explanation. I
other views. There are difficulties with both return to the unacceptable kinds of explana-
classical cognitivism and connectionism tions later. It is the acceptable explanations
(Minsky and Papert 1969; Dreyfus 1981). that I want to discuss here. They belong
This article mainly focuses on the problems more to the tradition of design education.
with classical cognitivism and how they are There are two approaches to explanations.
avoided (or possibly ignored) by connection- One is to accept that there is such a thing as
ism. Before considering these two approaches a true or false explanation and that if a par-
in detail, it is necessary to look at something ticular explanation is in error, then it can be
of the nature of design discourse. Because of proven to be so. The identification of a par-
the enormity of the subject, I focus on only ticular logic error should compel the designer
two aspects of this discourse: the role of to return to the task and revise the design. If
explanations and the use of categories. What it turns out that south-facing clerestory win-
is said about these two issues can be seen as dows do not produce sufficient surface illu-
representative of the problems encountered mination or that clerestory windows are not
more widely. part of the language of a particular style of
architecture after all, then there is the com-
pulsion to change the design.
The Role of Explanations There are difficulties with explanations
Explanations appear to feature prominently when taken in this way. Obviously,there is a
in design activity. We seek explanations of difficulty if explanations are seen to depend
the genesis and apparent evolution of on apparently controversial or unsound
designed objects. (Why does a doric column theory, which is currently the case in using
have a fluted shaft? How did the flying but- explanations that depend on theories about
tress come into existence? How do we human responses to spaces. However, expla-
account for the development of the Medieval nations also pose difficulties by the nature of
piazza?) Apart from requiring explanations to argument itself. On close inspection, we see
satisfy their curiosity and analyze others’ that explanations can involve us in an end-
designs, designers are called on to offer less chain of reasoning as we attempt to form
explanations for their own decisions. Expla- links between our conclusions and our
nations are proffered to satisfy the critics, assumptions, first principles, or empirical
promote decisions, and persuade. evidence. Furthermore, the details of expla-
There appear to be several categories of nations elude us as each reasoning step
explanation. Each enjoys varying degrees of demands further refinement. The problem is
acceptability depending on the context of well recognized in the Kantian critique of
the discussion. The following question is typ- Cartesian rationalism. The flaky nature of
ical of what we might ask a particular design- explanations is further aggravated when we
er (such as an architect): Why did you place a observe the human propensity to change
south-facing clerestory window above the explanations in preference to changing deci-
living room? The acceptable explanation is sions. We also appear to exercise the ability
usually of the kind that follows some line of to change our explanations according to the
argument, perhaps relating to lighting levels, audience for our ideas.
sun control, the brightness of the interior, Do these difficulties mean that explana-
and even the psychology of people respond- tions are invalid and undeserving of any
ing to and experiencing the space. A related credibility in discourse? The second
kind of argument pertains to style: an appeal approach to the value of explanations is
to the appropriateness of the architectural neatly summarized by a quotation pertaining
device (the south-facing clerestory window) to one of the primary roles of language: “In
to the particular architectural language. A using language we are not transmitting infor-
variation of this kind of argument is an mation or describing an external universe,

74 AI MAGAZINE
 Articles

but are creating a cooperative domain of Another weapon in the

interactions” (Winograd and Flores 1986, p.
50). As part of this discourse, groups of people designer’s armory, one
exert influences on one another. In the con- that makes dialogue
text of professional design training, the flow
of the cooperative discourse is such that there possible, is the notion of
is generally some influence in favor of the categories.
conventions and norms of the particular
design discipline. The discourse produces
thinkers in harmony with the design disci-
pline or disciplinary matrix of the day (Kuhn
1970). There is insufficient space here to enter
into a full consideration of the challenge establish neat typologies than someone
posed by the hermeneutic movement invents an artifact that crosses the boundaries
(Gadamer 1975; Rorty 1979) within modern between several types.
philosophy to the instrumental view of lan-
guage. This challenge is well presented by
Winograd and Flores (1986). Designing
The use of explanations and types in design
discourse appears to extend beyond analysis
and criticism. Explanations and types feature
The Role of Categories in designing. Part of design discourse, particu-
Another weapon in the designer’s armory, larly in teaching, appears to involve turning
one that makes dialogue possible, is the explanations and categories around. They
notion of categories. Objects that are similar appear to act as generative devices for produc-
in some way bear the same word labels. ing design decisions. More precisely, they act
Again, there are penalties in claiming that the as prescriptions to others about how they
categories we use reflect some objective reali- might produce similarly appropriate design
ty. What actually constitutes particular cate- decisions. It is as if we dissect our explana-
gories is difficult to pin down, and definitions tions into logical modules. Each module con-
are notoriously elusive. In design, this diffi- stitutes a piece of knowledge; so, designers
culty is brought into sharp relief when we sometimes talk in terms of rules: If you want
talk of types (Moneo 1978), which can be seen lots of light, then put in a clerestory window;
as generic descriptions of artifacts. Thus, in if you want to avoid direct sunlight falling on
architectural design, we talk about the L- work surfaces, then orient the windows to
shaped kitchen, the courtyard house, the face south (in Australia).
mausoleum type, the civic building, and so Types similarly appear in talk about gener-
on. The identification and delineation of ating designs. Design can be characterized as
types is important. We create categories on identifying the appropriate type for the par-
the basis of certain similarities that are easy to ticular context and then instantiating (explor-
recognize. Having devised an object in terms ing the scope of variation allowed by the type
of a category, we can then talk of other prop- to arrive at a design instance).
erties known to belong to the type as a whole. We might suppose that these prescriptions
The use of categories introduces economies form part of design discourse. They feature
into the discussion. prominently in the advice given to novice
As with chains of explanations, the delin- designers. It is contended here that the dis-
eation of types leads us into difficulties, espe- course itself poses few problems, bearing in
cially if we declare that typologies describe mind its role within a “cooperative domain of
the reality of an external universe. Catalogs of interactions” (Winograd and Flores 1986, p.
types prove to be prohibitively large. The pat- 50). However, for some, it becomes important
terns of types and subtypes become complex to promote the appropriate prescriptions and
and fluid. Before long, we find that there is, excise those that are false. This view can be
after all, no satisfactory meaning to the L- realized as a search for design principles. For
shaped kitchen and its infinite variants, either those for whom the quest is less illusive, the
by the individual or society at large. Types design principles can even be taught or put
and subtypes are also nested in nonconform- into a computer. However, as previously indi-
ing ways. The boundaries between types are cated, if explanations and types are friable,
fuzzy (What is the real difference between a then their less certain use as prescriptions is
church and a chapel?). No sooner do we even more so.

WINTER 1990 75
Articles

Classical Cognitivism knowledge in automated design decision

making. A second option, favored by Wino-
Computer-based design research that focuses grad and Flores (1986) and Bijl (1989), is to
on the importance of explanations as a see the computer as a medium that facilitates
source of knowledge generally feeds on discourse. Thus, the emphasis is on the
models of thinking based on classical cogni- human design agent. A third option is to
tivism. This view is characterized by an focus on computational models that do not
emphasis on symbolic representations and rely on explanations as opposed to models
the manipulation of logical statements as that focus on the structure of explanations
capturing the essence of human reasoning. As and prescriptions. This latter provocative line
discussed previously, the language of design is the one that is pursued here.
discourse appears to trade heavily on logic
and its variants. If true, with the computer
medium, there is the evident possibility of Connectionism
the logic within human explanations and Because less attention has been directed at
prescriptions being rendered operable. Much the implications of a connectionist view of
of the AI field and, specifically, that of knowl- cognition on design, it is worth looking at
edge-based systems seek to operationalize the the ideas in some detail. Although the ideas
explanations, rules, type descriptions, and within connectionism are compelling (from
prescriptions of human discourse. Recent the point of view of their novelty if nothing
accounts of the role of formal systems in else), there is no intention here to advocate
design by Mitchell (1990), Coyne (1988), and its superiority to classical cognitivism. The
Coyne et al. (1990) make substantial use of intention is merely to present an opposing
this paradigm. view. A feature of connectionism is that it
Logical statements can be processed to skirts around the problems outlined in the
prove or disprove an assertion or to produce previous sections.
new assertions, or we can use some kind of In contrast to computer models that
symbol-manipulating algorithm to generate exploit the manipulation of symbols as a
combinations of design elements in keeping metaphor for human problem solving, con-
with a generative language. This approach is nectionist models emphasize the structure of
not entirely restricted to formal logic. Attempts the human cognitive hardware. In spite of
have been made to capture the fuzziness of the potential complexity of this approach,
human reasoning and the creative ways in certain models have been proposed that
which humans apparently handle contradic- render the general idea accessible (Rumelhart
tion (nonmonotonicity) (de Kleer 1986). and McClelland 1987b). In essence, the con-
These techniques generally operate within nectionist approach abandons the idea of
the realms of explicit knowledge representa- working with explicit knowledge representa-
tion as rules. Similarly, knowledge can be tions. Although rules and type definitions
brought to bear in selecting and instantiating serve in the logic of design discourse, they are
within catalogs and hierarchies of types. not at the heart of design processes. The con-
The apparent utility of these approaches nectionist models that are proposed rely
lies in the idea that knowledge is made heavily on the notion of memory and the
explicit and is open to scrutiny. The knowl- belief that decisions emerge from this
edge is of the kind that we invoke in much memory in response to some situation. The
design discourse. The disadvantage is that we mechanisms by which decisions emerge rely
find we have captured something of the logic on information-storage techniques and algo-
of the discourse but have captured nothing of rithms that are thought to model biological
the mediation of the human design agent. mechanisms, albeit in a pale way. (See Coyne
With the computer as design agent, the dis- and Postmus [1990] for an introduction to
course can be checked for consistency and connectionism in the context of design.)
developed to its logical conclusions. The diffi- A major assumption behind the approach
culty is that this checking takes place with a is that aspects of extremely large and unfath-
knowledge base derived from explanations, omably complex systems, such as the brain,
typologies, and prescriptions. These have can be studied to good purpose in a much
been shown to be extremely friable. Mean- scaled-down version. Thus, the connectionist
while, the tacit knowledge of the human approach involves the construction of com-
designer has eluded us. puter models that consist of networks (con-
There are at least three responses we can nected nodes [units] with certain numeric
make to this difficulty. The first is to make do attributes, generally, linkages and threshold
with the cautious application of explicit values). These models are intended to capture

76 AI MAGAZINE
 Articles

aspects of the readily observable electrochem- This ability to learn and recall is interesting
and useful but becomes more significant
. . . connec-
ical behavior of groups of neurons. Rumelhart
and McClelland (1987b) present empirical when the system is taught a large number of tionist models
evidence that they argue strongly supports room examples (perhaps 60 or more). Some of emphasize the
the plausibility of the models. Connectionist these examples might be different from one
networks are claimed to exhibit properties another, some similar; so, there might be sev- structure of
that could be considered essential features of eral room examples with sofas, but each the human
human cognition: the idiosyncratic way they might display slightly different characteris-
learn, their reliance on memory, their ability tics. The weights and thresholds that make up cognitive
to deal with partial information, and even the memory of the system begin to interfere. hardware.
their pathologies. Here, we explore a further What happens now when the system is
possible claim: that they exhibit a propensity required to recollect a room with a sofa in it?
for behaving in a way that captures the A description is produced that resembles a
essence of design. They appear to synthesize, typical sofa-bearing room. The system has not
to innovate. produced a general sofa room description but
The idea is best explained with an example, a typical case. The algorithm that facilitates
using one particular connectionist network this recall can also be induced to summon
model. The example is trivial but demon- other room descriptions that are less typical.
strates the basic idea. The example has been According to the system, the most typical
implemented and is technically explained by room example might contain a sofa, an arm-
Rumelhart et al. (1987) and explored in the chair, and a carpet. If we tell the system that
context of design by Coyne, Newton, and the room also contains a bed, then we can
Sudweeks (1989). A connectionist network is induce the system to find a typical room that
trained about a set of rooms. In this model, contains a sofa and a bed. This room might
each of the possible features of a room is rep- be different from the one with the armchair
resented as a node on the network. (There is and carpet. This kind of recall facility is fun-
no suggestion that concepts map onto indi- damentally different from conventional
vidual neurons in real brains.) The features database approaches. The room descriptions
are descriptors, such as the fact that there is a are not locally stored in any particular part of
sofa in the room or that the room has a large the network but distributed across the whole
window, a carpet, or a refrigerator. There are network. As well as having a certain appeal
50 possible features and, therefore, fifty nodes on the basis of what is known about the dis-
in the network. Each room is presented to the tributed nature of human memory, this phe-
network in turn; that is, when the system is nomenon leads to certain generative
presented with room 1, its features are acti- properties.
vated on the network. An algorithm is This approach becomes extremely interest-
applied to make adjustments to the parame- ing when we force the system to produce a
ters of the network (actually weights on the description of a room that is not in accor-
connections between the units and threshold dance with any of the examples; for example,
values on the units). These adjustments we tell the system that the room has a sofa
ensure that if at a later time the network is and a sink. We find that the system generates
presented with a partial description of this a description of a room (a combination of fea-
room, then another algorithm will bring the tures) that we might recognize as a studio
original pattern of activations to life. The apartment. What has happened is that the
system has a rudimentary kind of memory. system has accounted for all the connections
Subsequent rooms are presented to the between descriptors to arrive at the most
system. The system learns these patterns as mutually compatible combination of descrip-
well. The memory capability of the system is tors. If we look at the operations of the
realized if we present the system with a fea- system at the algorithmic level, what has hap-
ture such as a sofa, and the system brings to pened is wholly unremarkable. The system
life other features such as an armchair, carpet, has not had to struggle to invent a new room
or standard lamp. If we activate the refrigera- type. The algorithm that simulates recall in
tor, then we would expect the stove, sink, and fact implements a kind of relaxation proce-
floor tile features to come to life. In other dure. The system settles into a state of activa-
words, the system is able to recollect exam- tion that is consistent with the parameters
ples given to it on the basis of fragments of (the weights and threshold values) set up
information. Already we see something that during the learning process. Recalling an
appeals to our way of thinking about cogni- imaginary room required as little effort as
tion: A simple idea can trigger a complex recalling a familiar room.
recollection. As stated previously, the example is a trivial

WINTER 1990 77
Articles

. . . the value of these models is that each provides us with

vocabularies and conceptual structures for enlivening design
discourse.

one, and here we do not have the persuasive- In summary, the connectionist models pro-
ness of the detailed implementation at our vide structures for considering ideas impor-
disposal. (See Coyne [1990] for an example tant to design: the use of precedence and the
pertaining to simple foundation design in emergent nature of much design reasoning.
buildings and Coyne, Newton, and Sudweeks The connectionist models provide ways of
[1989] for details on the furniture layout accounting for this behavior without recourse
example.) Further levels of sophistication can to the explicit representation of rules and
be contemplated if we extrapolate this experi- type hierarchies.
ment to allow units in the connectionist net-
work to pertain not to features describing
rooms but to any combination of concepts. Implications for
We soon run into the limitations of this par- Understanding Design
ticular model. However, if we assume that
How do the implications of connectionist
this model bears some resemblance to aspects
models of cognition impinge on how we
of human cognition, then several valuable
understand the design process? Leaving aside
observations can be made:
any argument about the superiority or veraci-
First, the system does not require explicit
ty of either classical cognitivism or connec-
type definitions or type boundaries to
tionism, we can see that a commitment to
demonstrate generalizing behavior. At no
one or the other could promote tendencies
stage was the system told any of the labels we
toward particular approaches to design
normally attach to room types (such as bed-
understanding. Thus, the value of these
room, bathroom, or kitchen). However, the
models is that each provides us with vocabu-
system behaved as if it was governed by typo-
laries and conceptual structures for enliven-
logical knowledge.
ing design discourse. As indicated in the
Second, no rules are presented to the system,
following discussion, connectionism provides
yet it can behave as if there are rules: All rooms
a structure for talking about aspects of design
with beds have a wardrobe. Further experi-
that might previously have eluded us.
ments have indicated that as well as making
generalizations, such systems retain informa-
tion about exceptional cases (Rumelhart and Classical Cognitivism
McClelland 1987a; Coyne and Newton 1990). The approach of classical cognitivism provides
Third, significant from the point of view of a structure for considering the following
design, such systems can cross the boundaries aspects of design: (1) the importance of rules,
between implied type descriptions to produce (2) the study of typologies, (3) the importance
novel but consistent combinations of fea- of explanations, (4) the establishing of evalu-
tures. This faculty appears to be a major ation criteria, and (5) the use of computers.
human quality not addressed by symbolically The first aspect of design emphasizes the
oriented models. importance of rules. This view takes seriously
Fourth, the process by which descriptions the notion of explanations as a source of gen-
are produced, except at the trivial computa- erative knowledge. We should not only teach
tional level, is distinctly void of logical expla- the theories that pertain to the effective anal-
nations. If we see the system as a rudimentary ysis of designs but also decision-making prin-
kind of problem solver (what goes with a bed ciples and procedures. In light of the elusive
and a sofa?), then the solution simply emerges. nature of rules and their heuristic nature, this
In some cases, the explanation would have to view can be modified as a quest by individu-
be, “I remember such a combination of ele- als to discover their own rules and methods.
ments.” Because the system might never have There is also considerable benefit in making
been exposed to its own solution before, the this knowledge explicit as tables, diagrams,
more general explanation is, “The answer is reports, and flowcharts.
consistent with my experience.” Second is the study of typologies. The defi-

78 AI MAGAZINE
 Articles

nition of terms used by designers is important The fourth aspect is the belief that new
so that we have a common base for discus- ideas can emerge from prosaic ideas. The abil-
sion, which extends to the definition and ity to create is inherent within the human
study of typologies. For example, the study of cognitive hardware. One of the prerequisites
building types and their evolution is an for a successful creative endeavor is a thor-
important part of architectural history, primar- ough grounding in the conventional.
ily as source material for our own designing.
The third aspect involves the importance of
explanations. Design decisions must be justi-
fied. Designs should be modified in light of
proven inconsistencies in explanations. There Conclusion
is a tendency to take explanations given by The jump from a theory of cognition to its
successful designers at face value. practical outworking in understanding design
The fourth aspect involves the establishing is bound to be hazardous. Here, I chose the
of evaluation criteria. The presuppositions on safe course of maintaining that different
which explanations and decision are based paradigms of cognition enrich the way we
should be made explicit. Making presupposi- talk about design. Their influence on how we
tions explicit is important to establish where actually do design poses even greater difficul-
an argument begins and ends when all the ties. The language of the connectionist
logical statements are strung together. paradigm allows such ideas as emergence to
Fifth is the use of computers to support this be discussed within a framework. The lesson
process. The knowledge by which design deci- from connectionism is that computational
sions are made can be put into a computer models exist by which we can describe appar-
and made operable. Shortcomings in such a ently informal operations. Connectionism
knowledge-based system are addressed by pro- challenges the advocacy of formal rigor in
viding the system with more knowledge and design by offering a formal model that in fact
more sophisticated control structures. supports an informal view of the design pro-
cess. This challenge presents us with an
attractive basis for a deconstruction. Like
Connectionism
most interesting ideas, these propositions
The approach of connectionism provides a inevitably contain the seeds of their own
structure for considering the following destruction.
aspects of design: (1) the importance of prece-
dence, (2) intuition, (3) the articulation of
design knowledge, and (4) the belief that new Acknowledgments
ideas can emerge from prosaic ideas. This work is supported by a special project
The first aspect of design is the importance grant from the University of Sydney and a
of precedence. Exposure to events and project grant from the Key Centre of Design
instances is important. A rich experiential Quality, University of Sydney. Comments on
base is required to facilitate design reasoning. earlier drafts of this article from Fay Sud-
Learning to design by doing and observing is weeks, Adrian Snodgrass, Sid Newton, and
important. Observing without generalizing Mike Rosenman were much appreciated.
has a role in education, as does copying.
Familiarity is the best teacher.
The second aspect is intuition. This view References
accepts that certain design activities cannot Bijl, A. 1989. Computer Discipline and Design Prac-
be externalized. It accepts the fickle nature of tice—Shaping Our Future. Edinburgh: Edinburgh
explanations as they are used to justify design University Press.
decisions. It accepts that aspects of design and Clark, A. 1989. Microcognition: Philosophy, Cognitive
design teaching defy traditional academic and Science, and Parallel Distributed Processing. Cam-
scientific treatment. It gives credit to the bridge, Mass.: MIT Press.
power of persuasion that extends beyond the Coyne, R. D. 1990. Modelling the Emergence of
compulsions of logic. Design Descriptions across Schema Boundaries. In
Third is the articulation of design knowl- Proceedings of the First Australian Conference on
edge. It accepts that much is imparted in Neural Networks, 41–42. Sydney: University of Sydney.
design education that cannot be made explic- Coyne, R. D. 1988. Logic Models of Design. London:
it. It is pluralistic and accommodating to dif- Pitman.
ferent views and coteries of expertise that Coyne, R. D., and Newton, S. 1990. Design Reason-
enliven and extend the cooperative domain ing by Association. Environment and Planning B(17):
of interactions. 39–56.

WINTER 1990 79
Articles

Coyne, R. D., and Postmus, A. G. 1990. Spatial Distributed Processing: Explorations in the Microstruc-
Applications of Neural Networks in Computer- ture of Cognition, Volume 2: Psychological and Biologi-
Aided Design. Artificial Intelligence in Engineering cal Models, eds. J. L. McClelland and D. E.
5(1): 9–22. Rumelhart, 216–271. Cambridge, Mass.: MIT Press.
Coyne, R. D.; Newton, S.; and Sudweeks, F. 1989. Rumelhart, D. E., and McClelland, J. L., eds. 1987b.
Modelling the Emergence of Schemas in Design Parallel Distributed Processing: Explorations in the
Reasoning. In Proceedings of Modeling Creativity Microstructure of Cognition, Volume 1: Foundations.
and Knowledge-Based Creative Design, eds. J. S. Cambridge, Mass.: MIT Press.
Gero and M. L. Maher, 41–42. Sydney: University Rumelhart, D. E.; Smolensky, P.; McClelland, J. L.;
of Sydney. and Hinton, G. E. 1987. Schemata and Sequential
Coyne, R. D.; Rosenman, M. A.; Radford, A. D.; Bal- Thought Processes in PDP models. In Parallel Dis-
achandran, M.; and Gero, J. S. 1990. Knowledge- tributed Processing: Explorations in the Microstructure
Based Design Systems. Reading, Mass.: of Cognition, Volume 1: Foundations, eds. D. E.
Addison-Wesley. Rumelhart and J. L. McClelland, 7–57. Cambridge,
Mass: MIT Press.
de Kleer, J. 1986. An Assumption-Based Truth
Maintenance System. Artificial Intelligence Schön, D. 1982. The Reflective Practitioner. Cam-
28:127–162. bridge, Mass.: MIT Press.
Dreyfus, H. 1981. Misrepresenting Human Intelli- Winograd, T., and Flores, F. 1986. Understanding
gence. In Artificial Intelligence: The Case Against, ed. Computers and Cognition: A New Foundation for
R. Born, 41–54. London: Croom Helm. Design. Reading, Mass.: Addison Wesley.
Gadamer, H.-G. 1975. Truth and Method. London:
Sheed and Ward.
Kuhn, T. S. 1970. The Structure of Scientific Revolu- Richard Coyne is a senior lec-
tions. Chicago: University of Chicago Press. turer in the Department of
Architectural and Design Sci-
Minsky, M., and Papert, S. 1969. Perceptrons. Cam- ence at the University of Sydney,
bridge, Mass.: MIT Press. where he teaches and undertakes
Mitchell, W. J. 1990. The Logic of Architecture: research in the areas of comput-
Design, Computation, and Cognition. Cambridge, er-aided design, computer
Mass.: MIT Press. graphics, knowledge-based sys-
Moneo, R. 1978. On Typology. Oppositions tems, and design theory and
13:23–45. methods. He also teaches in the
design studio and in the area of design communica-
Rorty, R. 1979. Philosophy and the Mirror of Nature.
tions. Coyne’s recent publications include the book
Princeton, N.J.: Princeton University Press.
Logic Models of Design (Coyne 1988) and the co-
Rowe, P. 1987. Design Thinking. Cambridge, Mass.: authored book Knowledge-Based Design Systems
MIT Press. (Coyne et al. 1990). He is an architect and landscape
Rumelhart, D. E., and McClelland, J. L. 1987a. On architect and is interested in design in all its facets,
Learning the Past Tense of English Verbs. In Parallel including design discourse as a social phenomenon.

1991 AAAI Events Deadlines The CSNET electronic addresses

for AAAI are as follows:
January 18 — Innovative Applications of Artificial Intelligence (IAAI) paper
AAAI membership inquiries: membership@aaai.org
submissions due. Send to AAAI
AI Magazine subscriptions: membership@aaai.org
January 30 — National Conference on Artificial Intelligence (NCAI) paper
submissions due. Send to AAAI AI Magazine editorial queries & letters:
engelmore@sumex-aim.stanford.edu
February 15 — AAAI-91 brochure mailed. Allow three weeks for delivery
AI Magazine announcements: aimagazine@aaai.org
March 1 — Notification of change of address for Joint Membership Directory AI Magazine advertising & press releases:
March 15 — IAAI Brochure and Preliminary Program mailed. Allow three aimagazine@aaai.org
weeks for delivery AAAI national conference: ncai@aaai.org
March 26-28 — Spring Symposium Series, Stanford University AAAI Press: press@aaai.org

April 19 — AAA1-91 preliminary conference program mailed Workshop information: workshops@aaai.org

Innovative applications conference: iaai@aaai.org
June 15 — AAAI-91 and IAAI 91 housing deadlines
Spring symposium series: sss@aaai.org
July 14-19 — AAA1-91, Anaheim, California
Other inquiries and requests: admin@aaai.org
July 15-17 — IAAI, Anaheim, California

80 AI MAGAZINE