Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

CASE-BASED PRELIMINARY BUILDING DESIGN

By Simon F. Bailey ~ and Ian F. C. Smith,'- Member, ASCE

ABSTRACT; A team of civil engineers, architects, and computer scientists is using


Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

case-based design in the CADRE system as an approach for integrating preliminary


architectural and structural design. This approach focuses on dimensional and to-
pological adaptation of geometric models of existing buildings to find solutions for
new design problems. An advantage of such an approach is that a case is used in
which integration has already been achieved, and the design process is therefore
one of adapting the case to achieve new goals while maintaining the original building
features. Dimensional adaptation involves the solution of a set of linear and non-
linear constraints describing the geometry of structural and architectural abstrac-
tions of a building. Integrated solutions are produced because different abstractions
are related in terms of dimensions. The present paper describes how CADRE is
able to manipulate potentially thousands of geometric constraints that describe
complex objects such as buildings. Topological adaptation yields a new set of
dimensional constraints when the original set has no solution. The present work
concludes that cases consisting of a geometric building model both simplify design
knowledge acquisition and provide valuable initial integrated-design solutions that
serve as useful starting points for the design process.

INTRODUCTION

Computer-aided design (CAD) is currently one of the most active fields


of research and development in artificial intelligence. Many systems have
been developed to study human problem solving as well as to test the
potential for CAD in design practice. An examination of how these systems
have evolved and the difficulties that were encountered during their de-
velopment is given in the paragraphs that follow.
The integrated building design environment (IBDE) (Fenves et al. 1989)
was one of the first integrated CAD systems to be implemented. Design
knowledge from various building design domains is organized into inde-
pendent rule bases in separate system modules. The design of foundations,
the core, and column layout, for example, are treated as separate design
subtasks. A control module in a blackboard architecture (Hayes-Roth and
Hayes-Roth 1980) determines the solution sequence for each subtask. The
IBDE suffers from the problems of conflict and possible looping between
modules in a blackboard system (Schmitt 1989) caused by decomposing the
design process into subtasks that are in fact interdependent.
Prototypes (Gero et al. 1988) are parametric models of design objects
that are adapted to solve new design problems. The prototype approach
demonstrates the advantage of starting the design process with concrete
examples of solutions. Disadvantages are that parametric models must be
formulated by generalizing both design and domain knowledge, and this is
difficult to achieve for complex objects, such as buildings, which are char-
acterized by many thousands of parameters. Mutation (Zhao and Maher
~Res. Engr., ICOM (Steel Structures), Dept. of Civ. Engrg., Swiss Federal In-
stitute of Technology (EPFL), GCB-Ecublens, CH-1015 Lausanne, Switzerland.
2Assoc. Dir., Artificial Intelligence Lab. (LIA), Dept. of Comp. Sci., Swiss Fed-
eral Institue of Technology (EPFL), INR-Ecublens, CH-1015 Lausanne, Switzerland.
Note. Discussion open until March 1, 1995. To extend the closing date one month,
a written request must be filed with the A S C E Manager of Journals. The manuscript
for this paper was submitted for review and possible publication on January 4, 1994.
This paper is part of the Journal of Computing in Civil Engineering, Vol. 8, No. 4,
October, 1994. 9 ISSN 0887-3801/94/0004-0454/$2.00 + $.25 per page. Paper
No. 7702.

454

J. Comput. Civ. Eng., 1994, 8(4): 454-468


1992) is proposed as a means of increasing the creativity of prototypical
design. In this approach analogical reasoning is used to adapt prototypes
to include parts of others, thus making solutions more innovative.
The paradigm of case-based reasoning (CBR) originates from psycho-
logical models of human memory structure (Schank 1982). The underlying
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

concept is that humans use not only heuristic rules to solve problems, but
also refer to solutions of previous similar problems. For example, when
faced with the task of designing a new warehouse, the case of an existing
warehouse is adapted to suit new needs, rather than starting with a set of
rules that say that roofs are supported by beams, beams are supported by
columns, and columns must be founded, and so on.
Cases in some early CBR systems consist of the "ingredients and steps
to follow," as in C H E F (Hammond 1986) that produces new recipes from
existing ones. This type of case is a plan of the design process, formulated
at the knowledge-acquisition stage, as with prototypical design. Existing
solutions must be evaluated by a human designer and a case formulated to
suit the way it is used during reasoning. This complicates knowledge ac-
quisition.
CYCLOPS (Navinchandra 1988) is a landscape planning and design sys-
tem in which a case is both a previous problem and its solution. Analogical
reasoning is used to identify cases, not necessarily from the same domain,
that match the current design problem. The identification of relevant match-
ing cases from other domains necessitates a great deal of domain-inde-
pendent knowledge. Furthermore, cases must be formulated so that the
useful information is represented explicitly.
Due to the belief that a practical CBR system would need many hundreds
of cases from which to reason, much research has been concentrated on
indexing and retrieval of cases (Kolodner 1989). Systems were developed
that find the best case, rather than reason with the case once selected.
A R C H I E (Domeshek and Kolodner 1992) indexes and organizes decom-
posed cases (annotated building plans) in memory and provides a case
browser for the retrieval of stories, which are defined as "selective pre-
sentations about a case which have a lesson to teach." The user is left to
interpret a story, as A R C H I E does not reason, being a problem-sensitive
teaching tool rather than a design system.
CADSYN (Zhang and Maher 1993) is a case-based building design system
that considers architectural space planning, structural design, and services
design. The system assumes that design problems can be decomposed to
nearly independent subproblems, which can be solved independently and
then recomposed to provide a complete solution. As with a blackboard
system, however, ignoring the interdependence of subproblems can lead to
conflict and looping after recomposition of subsolutions.
The DDIS system (Wang and Howard 1991) combines CBR with design-
independent knowledge in a blackboard framework. The system represents
design knowledge with cases consisting of the problem specification, final
solution, intermediate propositions, design history, and so on. Cases are
retrieved by users who decide on similarity to the new design problem. The
system treats the design of steel columns, and is therefore yet to be tested
on complex tasks.

INTEGRATION
Buildings can be seen from many different points of view. For example,
to a civil engineer a building consists of structural elements, whereas an
455

J. Comput. Civ. Eng., 1994, 8(4): 454-468


Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

(a)

(b)

(c)

FIG. 1. Multiple Building Abstractions: (a) Structure, Load-bearing Elements; (b)


Space, Rooms; and (c) Envelope, Filter between Environment and Occupants

architect may view it as a collection of rooms or as an envelope providing


a filter between the surrounding environment and occupants (Fig. 1). We
refer to each of these interpretations as an abstraction. One of the most
difficult tasks in design is resolving conflicts that occur when integrating
solutions from multiple abstractions or decomposed subproblems. The con-
flict between architects and engineers is traditionally one of the greatest
problems in building design (Holgate 1986). In practice, designers rely
heavily on their experience of conflict resolution in order to make trade-
offs. In design practice, as in CAD, integration is the key to good design.
This is accepted by many researchers, but current CAD systems provide
little support. As mentioned above, blackboard systems may loop between
modules, unless the control module contains complete knowledge of all
possible conflicts as well as effective strategies to resolve them. The weak-
ness of blackboard systems is found precisely in the representation of knowl-
edge for conflict resolution. An improved model of the design process is
therefore needed.

CADRE

In this project, we have implemented CBD in a system called C A D R E ,


which focuses on the representation and adaptation of cases rather than
their indexing and retrieval, as we believe that these are more important
issues for case-based building design. The advantage of our approach is that
a case is used in which integration has already been achieved, albeit for a
different problem. Beginning with such a case means that the design process
is now one of adapting the case to achieve new goals while maintaining the
456

J. Comput. Civ. Eng., 1994, 8(4): 454-468


original features of the building. These features include the trade-offs that
were made to achieve the integration of solutions from each abstraction.
Our approach is presented in the sections that follow. We begin with an
outline description of the paradigm. A presentation of the technical details
of each process follows, and finally a description of how CADRE has been
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

implemented is illustrated by a complete example of how the system has


been used to adapt one of the computer science buildings at the Swiss Federal
Institute of Technology (EPFL).

Paradigm
The cases used in C A D R E are geometric models of both structural and
architectural abstractions of a building. Users decide which building they
would like to adapt by selecting an appropriate case. The case is parame-
terized by CADRE and initial dimensional constraints describing both struc-
tural and architectural abstractions and their relationship to each other are
generated automatically. The user posts constraints that describe the new
design problem. C A D R E then attempts to solve the resulting system of
constraints during the dimensional adaptation process. If a solution is not
found, the user may ask C A D R E to attempt topological adaptations of the
building abstractions. A successful topological adaptation is followed by a
reparameterization of the building, constraint posting by the user, and finally
a dimensional adaptation in order to fix dimensions for the new topology.
This approach is illustrated in Fig. 2.

Cases
A case is a geometric model of both the structural system and the ar-
chitectural layout of spaces. Much design knowledge is implicit in the cases,
and therefore does not require explicit representation. Constraints used
during dimensional adaptation are generated automatically by generic pro-
cesses within the C A D R E system. The information stored in a case can
therefore be limited to a minimum, and cases can be created from building
plans without the addition of domain or design knowledge.

Caseselection) I LJserinteraction)

L ,Au~m~ic)!i
( Pormulationof PimensionalConstraints ~i)~
V

k ~>( ~opoloqicaladapk~ion~ii
@ qe5
~NP
FIG. 2. Case-Based Design in CADRE
457
J. Comput. Civ. Eng., 1994, 8(4): 454-468
Formulation of Dimensional Constraints
During dimensional adaptation, C A D R E attempts to solve a system of
constraints, which are derived at run-time from the geometrical model stored
in the case. For any topology in any abstraction the geometry of a building
can be expressed with constraints that, for example, describe the position
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

of a column or the area of a room. The geometries of different abstractions


are linked by using a common coordinate system. Constraints, such as span-
depth ratios for beams or relationships between the spacing of structural
frames and the overall length of a building, are generated automatically in
CADRE through context-sensitive evaluation of the existing case using
generalized domain knowledge. Generic domain knowledge is used to de-
termine supplementary constraints that define, for example, the maximum
span for a given type of floor slab. These constraints are often expressed
as inequalities, and it is therefore necessary to identify critical constraints,
which are then treated as equalities during dimensional adaptation. The
results of dimensional adaptation are checked with the remaining noncritical
inequality constraints.

Constraint Posting and Relaxation


A new problem is defined in C A D R E by the user posting additional
constraints on the original case, for example, by defining a new overall
dimension for a building or a new number of offices. Multiple design criteria,
for instance, structural safety and serviceability, mean that a building may
be overconstrained. The dimensional constraints that were generated for
each abstraction are then considered with the posted constraints during
dimensional adaptation. If the problem is overconstrained, the user is asked
to relax certain constraints. Similarly, if the problem is underconstrained
further constraints must be posted.

Dimensional Adaptation
Dimensional adaptation involves the solution of a set of linear and non-
linear constraints on the parameters used to describe the building. The
constraints to be considered are generated as was described. Adaptation at
the dimensional level produces integrated solutions because constraints from
all abstractions are considered simultaneously. The key to dimensional ad-
aptation is dimensionality reduction, which we have implemented using the
REDUCE system (Hua et al. 1992). Dimensionality reduction is used to
identify the key parameters that define the possible adaptations. This is
different from prototypical design because the parameters we consider are
identified at run-time.
A factory and warehouse site is shown in l~ig. 3. (The numbers in the
figure correspond as follows: 1 -- warehouse, 2 = small-items storage, 3
= connecting structure, 4 = free storage, 5 = short-term storage, 6 =
truck access, and 7 = factory.) It is a gas- and water-pipe production and
storage facility in St. Gallen, Switzerland. The task is to redesign the build-
ings to suit the same production but with a larger pipe-storage area. This
is an example of how dimensional adaptation in CBD is effective for gen-
erating routine solutions.
As described, the following occurs. The original structure is evaluated
using generalized domain knowledge, and constraints governing the dimen-
sions of the structural elements are generated. Next, additional constraints
are posted by the user to do the following: (1) Define a new area for the
warehouse; (2) fix its width in order to use the same type of overhead crane;
458

J. Comput. Civ. Eng., 1994, 8(4): 454-468


Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

FIG. 3. Factory and Warehouse Design

and (3) fix the dimensions of the other buildings in order to use the same
production-line system.
The base parameters and constraints describing the dimensions of the
warehouse structure as well as the section properties of the crane rail beam,
columns, roof beams, and cladding are given in the following. They are a
representative subset of the constraints considered during dimensional ad-
aptation. The warehouse structure is shown in Fig. 4.
The warehouse geometry is described by the following base parameters:
the area A1, length L, half-width W1, frame spacing Ss, and puffin spacing
Sp. They are governed by the following dimensional constraints:
A1 = L ' 2 " W 1 (l)
Sy = s (2)
Sp = WJl7 (3)
W~ = 16.5 m (posted by the user) (4)
A1 > 1,400 m 2 (posted by the user) (5)
The crane rail beam is represented by the following base parameters:
Young's modulus E, and the second moments of area Iy and/~. Generalized
domain knowledge is used to formulate the constraints for the crane rail
beam, which is governed by limiting horizontal and vertical deflections Dh
and D,, due to crane live loads Och and Qcv
D~, < Sf/700 (6)
D~, = k," Oc~,'S}/(E'ly) (7)
D h < Sf/800 (8)
459

J. Comput. Civ. Eng., 1994, 8(4): 454-468


Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

FIG. 4. Warehouse Structure

D,~ = k2" Qct,'S~/(E'Iz) (9)


where kl and k2 -- a function of the continuity and fixity of the crane rail
beam at supports, and are determined from evaluation of the original case.
Cladding is governed by strength criteria; the plastic moment of resistance
must be greater than the applied moment due to snow load Q, and dead
load G. The cladding base parameters are the material yield strength fy,
and section plastic modulus Zy. The following constraint is expressed for a
unit width of cladding:
fy. Zy/,y, > ('yg'G + "yq'Qs)" S~/12 (10)
where ~q, ~/g and ~/r = partial load and resistance factors.
Inequality constraints that are critical are fixed as equalities, and C A D R E
uses dimensionality reduction on base parameters to identify key parame-
ters, reparameterize others, and determine which constraints can be ignored.
This has the following effect on the base parameters and constraints listed:

1. The frame spacing is identified by C A D R E as the key parameter for


dimensional adaptation of the warehouse area, and is solved as
460

J. Comput. Civ. Eng., 1994, 8(4): 454-468


Si = A , / ( I O ' W I ) = 8.5 m (11)

from constraints (1), (2), (4), and (5).


2. C A D R E determines that the puffin spacing Sp and distributed loads
Qs and G are redundant parameters because they are independent of the
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

warehouse length. Therefore, the cladding constraints are independent of


adapted parameters and are ignored. The original section properties are
maintained.
3. The crane-rail-beam deflection constraints are fixed as equalities, and
they are reduced to
I, = 700.k,.Qc,,.S~/E (12)
from constraints (6) and (7) and

/~ = 800.k2-Q~h.S~/E (13)
from constraints (8) and (9).
4. Q,, and Qch are determined to be independent of the adaptation; they
are dependent on the type and size of gantry crane, which is fixed by
constraint (4). The values of kl" Qc~, and k2" Qch are evaluated from the
original case and substituted into constraints (12) and (13). These constraints
are then solved for the adapted value of Si
ly = 700. 557. S~/E = 0.000134 m 4 (14)

/~ = 800" 171-S~/E = 0.000047 m 4 (15)


Since in this example there is enough knowledge to fix the adaptation,
CADRE automatically resizes structural elements. These new dimensions
are then checked with the constraints that were previously identified as not
being critical. If there were remaining degrees of freedom, the adaptation
would be performed in collaboration with the user.

Topological Adaptation
In the event that a solution cannot be found by dimensional adaptation,
a topological adaptation may be tried. The aim is to adapt the original
topology to a new one that results in a new dimensional model and set of
dimensional constraints that are then treated by dimensional adaptation
processes.
A rule-based approach is used for the topological adaptation of the struc-
ture. Commonsense domain-independent rules are used to change the ge-
ometry. For example, if spans are to be increased, then the number of
frames must be reduced. Domain knowledge is used for deciding when the
type of construction needs to be changed. For example, a flat slab floor
system is suitable up to spans of 8 m, but beyond that, a beam and slab
system is preferable. The function of structural elements is also considered
during topological adaptation. Elements that contribute to the overall sta-
bility of the building (e.g., bracing, shear walls) are identified automatically
and maintained during adaptation. Flemming's (1986) algorithm for gen-
erating alternative arrangements of rectangles is used for room layout ad-
aptation.
The need for topological adaptation is illustrated by the example shown
in Fig. 5. The problem is to adapt the room and structural-element layout
461

J. Comput. Civ. Eng., 1994, 8(4): 454-468


[]

[]
tour,~
[]

17[NIN~i

[] It}
~I"Ct~N

m
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

[]

k IOta

(a)

5pace~ ~c~are

+ g
(b)

(c)

+ g

(d)

FIG. 5. Topological Adaptation in CADRE: a) Original Case; (b) Original Topol-


ogies; (c) Adapted Topologies; and (d) New Geometries

so that columns can be concealed in walls. To begin with, there is a conflict


in the dining room [Fig. 5(a)].
It is not obvious from the graphs of the two topologies [Fig. 5(b)] that
the central column is not coincident with a wall. This is due to difficulties
in relating topologies from different abstractions to each other in order to
detect conflicts. At the dimensional level [Fig. 5(a)] the problem is obvious.
For the overall dimensions shown in Fig. 5(a), dimensional constraints on
minimum room dimensions and maximum span lengths [(16) and (17)] can-
not be solved without topological adaptation.
room width, room length > 3.0 m (16)
column spacing < 5.5 m (17)
The task is then to direct this adaptation, without resorting to an arbitrary
462

J. Comput. Civ. Eng., 1994, 8(4): 454-468


generate and test approach. C A D R E uses information related to how the
solution failed at the dimensional level. For example, we know that the
conflict comes from the width of the rooms and the distance between col-
umns in the x-direction. We therefore need to adapt topologies in order to
change these parameters, that is, alter the number of column bays or rooms
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

in the x-direction. Fig. 5(c) illustrates possible adapted topologies.


Fig. 5(d) shows the corresponding geometries realized by dimensional
adaptation using the new dimensional constraints.

Implementation
The C A D R E system is implemented using Lisp and C. A user interface
has been created within AutoCAD, which facilitates case creation and allows
visualization of a case during the design process. Menus within AutoCAD
are used to control C A D R E (Fig. 6). Constraint posting by the user is also
carried out graphically.
Cases can be created through graphical input of structural elements and
rooms, which are then stored as objects, in order to build up geometric
models of each abstraction. For example, a beam is represented as an object
having a start point, end point, depth, width, and material type. The struc-
tural abstraction of a case is therefore a collection of many such elements.
No further description of the structural elements in a building need to be
input. The layout and dimensions of spaces can be defined in a similar way.
Services, such as power, water, heating, and ventilation, can also be rep-
resented, but for the moment these are not included.
Selected cases can be viewed before and after adaptation through the use
of standard AutoCAD commands. Abstractions may be viewed simulta-
neously or independently, as in Fig 7.

FIG. 6. CADRE Graphical User Interface

463

J. Comput. Civ. Eng., 1994, 8(4): 454-468


EXAMPLE SESSION WITH CADRE
A complete example showing adaptation for one of the computer science
buildings at EPFL (the INR building) is now given. This is one of six
multipurpose buildings recently constructed at EPFL for use by the De-
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

partments of Electrical Engineering and Computer Science. The future oc-


cupants of the INR building decided to change a room, originally intended
to be a laboratory, into a classroom. Unfortunately, a column is situated in
this lab, which is unwanted in a classroom. The following example dem-
onstrates how C A D R E is used to adapt the INR building to avoid this
conflict.
Fig. 8(a) shows the conflict of the column in the classroom. A constraint
is posted by the user to bind the x-coordinates of a wall and the conflicting
column (menu "constraint-pos" in Fig. 6). C A D R E attempts to solve the
set of dimensional constraints. To model the four floors of the building in
three dimensions requires more than 4,000 parameters. Dimensionality re-
duction simplifies the problem to a set of eight constraints on eight key
parameters, but finds that no solution exists, and notifies the user. The
failure of the initial dimensional adaptation is due to conflicting constraints
on the minimum area of the classroom and maximum spans for the floor
slabs. This is a typical situation that could lead to looping in a blackboard
system.
The user decides to try a topological adaptation of the structure, selecting
this option from the CL-CMD menu. C A D R E uses the knowledge that
dimensional adaptation failed because column spacing was too small and
therefore adapts the structural topology to increase column spacing in one
direction. The solution is to change the floor-slab system from a flat slab
to a one-way slab and to reduce the number of building frames from six to
four. New dimensional constraints are generated for this topology, and
subsequent dimensional adaptation finds a solution [Fig. 8(b)].

FIG. 7. Viewing Structural Abstraction of INR Building

464

J. Comput. Civ. Eng., 1994, 8(4): 454-468


, I
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

1 llIl
(a) (b)

l[lllll 11 11 llJlfJ
(c) (d)

FIG. 8. Room Layout and Structure in Plan: (a) Column in Second-Floor Class-
room; (b) Structure Adapted to Room Layout on Second Floor; (c) New Conflict
on First Floor; and (d) Room Layout Adapted to Structure on First Floor

The user then visually checks the effect of this new structure on the floor
plans for other floors. Unfortunately, there is a conflict on the first floor
[Fig. 8(c)]. The user decides to try a topological adaptation of the first-floor
room layout.
A topological adaptation of the first floor produces an alternative room
layout, and a corresponding set of dimensional constraints are generated.
A further dimensional adaptation finds a solution and it is shown to the
user [Fig. 8(d)].

DISCUSSION

A complete model of a building may involve many thousands of param-


eters and constraints, and it is therefore desirable to reduce these to a
minimum. In the example of the INR building, each floor is composed of
more than 30 rectangular spaces, and there are 39 columns, four shear walls
and a floor slab. More than 150 parameters are needed to represent the
geometry of these objects in two dimensions. To model the four floors of
the building in three dimensions requires more than 2,000 parameters. If
the positions of windows and doors and the dimensions of structural elements
are included, this number more than doubles. In the example illustrated in
Fig. 8, dimensionality reduction simplifies the problem to a set of eight
constraints on eight key parameters. For complex, nonlinear problems, this
approach is more efficient than that of prototype-based systems, which
always have to consider all parameters and constraints to instantiate solu-
tions from parametric models. However, the reduction process can take
hours if the problem is large. The use of specialized constraint solving
methods specific to those forms of nonlinearity that occur in building design
constraints (first- and second-order polynomials) can increase the speed of
this process.
A problem with case adaptation is that if it diverges too far from the
original case, implicit features may be lost. For example, a change in ori-
465

J. Comput. Civ. Eng., 1994, 8(4): 454-468


entation of the building may mean that windows no longer provide good
natural lighting.
CBD produces primarily routine solutions, and some that are innovative.
However, creative design is not supported by this approach. The fact that
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

new solutions are being derived from previous solutions in the same domain
means that the process is not creative, and is the reason why many re-
searchers are trying to use analogical reasoning to be able to consider so-
lutions from other domains. For the moment, creativity should probably be
left to humans. It is after all what they enjoy most in designing, whereas
the routine tasks, such as sizing structural elements, are tedious. When
designers are freed from such tasks, more time can be devoted to innovation
and creativity.
CADRE integrates the adaptation of building abstractions at the dimen-
sional level. However, treating multiple abstractions at the topological level
is more complicated than at the dimensional level. This is because it is
difficult to relate the topologies of different abstractions to each other. At
the dimensional level all objects are represented by their dimensions and
positions; thus, a common reference is more readily available than at the
topological level. This revives problems that are similar to those associated
with blackboard systems, since adapting one abstraction may lead to conflict
with another. We are now working to improve integration of the topological
adaptation processes by increasing the amount of information that one ab-
straction shares with another. Situations in which conflicts could arise as a
result of topological adaptation can thus be avoided. For example, the
architectural adaptation process must acount for the position of shear walls
so as to ensure natural lighting in rooms where it is needed (storerooms
and bathrooms, for example, do not need large windows). Additional im-
provements are possible when topological adaptation is limited to a part of
a building. For example, in the problem of the column in the room shown
in Fig. 8, it is better to adapt the layout of rooms local to the column. In
this way, adaptation is simplified and integration elsewhere in the building
is maintained.
Finally, current work is focusing on case combination as a method for
increasing the power of topological adaptation. This work is expected to
provide better support for innovative design.

CONCLUSIONS
The use of geometric models of buildings as cases facilitates knowledge
acquisition. Deriving dimensional constraints and symbolic representations
of topology at run-time restricts the amount of information used and eases
case creation. This is an advantage over approaches that require that more
information is included in the case, or that cases be formulated to suit the
design process.
Cases provide the design process with a starting point at which solutions
in all abstractions are integrated. An advantage of our approach to CBD
is that the trade-offs made during design are implicit in cases, and conflict-
resolution knowledge need not be represented explicitly in the system.
Dimensional adaptation involving nonlinear constraints is simplified by
dimensionality reduction. Topological adaptation increases innovation in
CBD by changing the design solution space.
466

J. Comput. Civ. Eng., 1994, 8(4): 454-468


ACKNOWLEDGMENTS

The development of C A D R E has been funded by the Swiss National


Research Foundation. The writers would also like to acknowledge G. Schmitt
and S-G. Shih at the Laboratory for Computer Aided Architectural Design
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

(CAAD) at the Swiss Federal Institute of Technology in Zurich, as well as


B. Faltings and K. Hua at the Laboratory for Artificial Intelligence (LIA)
in Lausanne. They have helped develop and implement most of the ideas
described in the present paper. Recent contributions of Bharat Dave and
Laurent Bendel at C A A D , and Kim Jent and Jean-Marc Ducret at ICOM
are also gratefully acknowledged.

APPENDIXI. REFERENCES
Domeshek, E. A., and Kolodner, J. L. (1992). "A case-based design aid for archi-
tecture." Artificial intelligence in design '92, Kluwer Academic Publishers, Boston,
Mass., 497-516.
Fenves, S. J., Flemming, U., Hendrickson, C., Maher, M. L., and Schmitt, G.
(1989). "An integrated software environment for building design and construc-
tion." CIFE Tech. Rep. No. 010, Center for Integrated Facility Engineering (CIFE),
Stanford University, Stanford, Calif.
Flemming, U.. (1986). "On the representation and generation of loosely-packed
arrangements of rectangles." Envir. and Ping., B. Envir. and Design, 13, 189-
205.
Gero, J., Maher, M. L., and Zhang, W. (1988). Chunking structural design knowledge
as prototypes. The Architectural Computing Unit, Department of Architectural
Science, University of Sydney, Sydney, Australia.
Hammond, K. (1986). "CHEF: a model of case-based planning." Proc., AAAI-86:
Fifth Nat. Conf. on Artificial Intelligence, Morgan Kaufmann, Los Altos, Calif.,
267-271.
Hayes-Roth, B., and Hayes-Roth, F. (1980). A cognitive model of planning. Morgan
Kaufmann, San Mateo, Calif.
Holgate, A. (1986). The art in structural design. Oxford University Press, New York,
N.Y.
Hua, K., Smith, I., Faltings, B., Shih, S., and Schmitt, G. (1992). "Adaptation of
spatial design cases." Artificial intelligence in design '92, Kluwer Academic Pub-
lishers, Boston, Mass., 559-575.
Kolodner, J. (1989). "Judging which is the best case for a case based reasoner."
Proc., Darpa Workshop on Case-Based Reasoning, Morgan Kaufmann, San Mateo,
Calif., 77-81.
Navinchandra, D. (1988). "Case based reasoning in CYCLOPS, a design problem
solver." Proc., Darpa Workshop on Case-Based Reasoning, Morgan Kaufmann,
San Mateo, Calif., 286-301.
Schank, R. (1982). Dynamic memory--a theory of reminding and learning in com-
puters and people. Cambridge University Press, London, England.
Schmitt, G. (1989). "Architectural pre-processor for engineering expert systems."
IABSE Colloquium, Expert Systems in Civil Engineering, International Association
for Bridge and Structural Engineering, Bergamo, Italy, 291-302.
Wang, J., and Howard, H, C. (1991). "A design-dependent approach to integrated
structural design." Artificial intelligence in design '91, Butterworth-Heinemann,
Oxford, England, 151-170.
Zhang, D. M., and Maher, M. L. (1993). "Using case-based reasoning for the
synthesis of structural systems." 1ABSE Colloquium, Knowledge-Based Systems
in Civil Engineering, International Association for Bridge and Structural Engi-
neering, Beijing, China, 143-152.
Zhao, F., and Maher, M. L. (1992). "Using network-based prototypes to support
creative design by analogy and mutation." Artificial intelligence in design '92,
Kluwer Academic Publishers, Boston, Mass., 773-793.
467

J. Comput. Civ. Eng., 1994, 8(4): 454-468


APPENDIX II. NOTATION

The following symbols a r e u s e d in this p a p e r :

warehouse area;
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/19/18. Copyright ASCE. For personal use only; all rights reserved.

A 1
Oh = horizontal deflection of crane rail beam;
D v ~-- vertical deflection of crane rail beam;
E = Young's modulus;
L = yield strength of steel;
G = self weight of cladding;
second moment of area about major axis;
second moment of area about minor axis;
k 1 = elastic deflection coefficient;
k z = elastic deflection coefficient;
L = length of warehouse;
Qch z horizontal live load due to gantry crane;
Qcz, vertical live load due to gantry crane;
Qs = distributed snow load;
st= spacing between building frames;
spacing between roof purlins;
W 1 = half-width of warehouse;
Z y ~-- section plastic modulus;
"yg = dead-load factor;
"yq = live-load factor; and
"~r ~ resistance factor.

468

J. Comput. Civ. Eng., 1994, 8(4): 454-468

You might also like