Automation in Construction 9 Ž2000.

197–215
www.elsevier.comrlocaterautcon

Automated facilities layout: past, present and future

)
Robin S. Liggett
Department of Architecture and Urban Design, UCLA, Box 951467, Los Angeles, CA 90095-1467, USA

Abstract

This paper reviews the history of automated facility layout, focusing particularly on a set of techniques which optimize a
single objective function. Applications of algorithms to a variety of space allocation problems are presented and evaluated.
Guidelines for future implementations of commercial systems are suggested. q 2000 Elsevier Science B.V. All rights
reserved.

Keywords: Automated space layout; Floor plan layout; Facility layout

1. Introduction most recently genetic optimization have been applied

to the problem.
Facility layout is concerned with the allocation of
With the growing demand for computerized facili-
activities to space such that a set of criteria Žfor
ties planning and management, there is the potential
example, area requirements. are met andror some
for automated space layout products to play a more
objective optimized Žusually some measure of com-
significant role. As interest in such products rekin-
munication costs.. This paper reviews alternative
dles and develops it seems appropriate to take an-
formulations of the problem Že.g., how space is
other look at the relatively long history of ap-
represented and methods of evaluating a plan. as
proaches to automated facility layout. The author last
well as existing solution algorithms. It identifies the
reviewed the field Žwhich had its origins in the early
specialized applications for which algorithms seem
1960’s. in 1985 presenting an overview of alterna-
particularly useful as well as the particular needs of
tive approaches to the layout problem and solution
facilities layout that must be considered when apply-
algorithms w39x. Since that time, commercial prod-
ing algorithms.
ucts have become available based on some of these
Some commercial facilities management systems
original algorithms and on the research side new
currently incorporate automated algorithms Žusually
solution techniques such as simulated annealing and
within an interactive framework. to solve facilities
planning problems of stacking and blocking activi-
ties. This paper also looks at the limitations of
current commercial space allocation products and
proposes a set of key requirements for implementing
)
Fax: q1-310-376-4936; e-mail: [email protected] the next generation of such systems.

0926-5805r00r$ - see front matter q 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 6 - 5 8 0 5 Ž 9 9 . 0 0 0 0 5 - 9
198 R.S. Liggettr Automation in Construction 9 (2000) 197–215

2. Overview and history the criteria used to generate, compare and evaluate
solutions.
Facility layout problems range in scale from the
assignment of activities to cities, sites, campuses or
buildings, to the location of equipment and personnel
groups on a single floor of a building. A layout 3. Representation of space
problem can surface in the design and allocation of
space in a new building or the reassignment of space All space planning problems consist of a set of
in an existing building. During the conceptual design activities to be located and a space in which to locate
phase, allocation of space within a new building can them. Space can be represented in different ways,
be used to test alternative options for building con- thus providing a method of classifying alternative
figuration. Plans can be evaluated with respect to types of layout problems:
best use of space in order to determine such things as Ø Space as discrete objects Žone-to-one assignment
the optimal number of floors, perimeter of the plan, problem..
etc. In an existing building, layout tools can be used Ø Space as area Žmany-to-one assignment problem,
for the on-going problem of space management. For for example, a stacking problem..
example, as project groups increase or decrease in Ø Space as area and shape Žblocking or floor plan
size, how should employees be located within an layout problem..
office so that group contiguity is maintained with a Both the problem formulation and solution tech-
minimum number of workspace moves? How can niques are impacted by the way activity and physical
unused space be consolidated effectively to minimize space are represented.
lease costs? More complex problems can involve The simplest layout problem is the assignment of
issues of time-phased layouts based on projected a set of discrete activities to a set of discrete loca-
changes in space needs w46x. tions in such a way that each activity is assigned to a
Since the early 1960s numerous computer pro- single location. This is called a one-to-one assign-
grams have been developed for the automated solu- ment problem Žalso known as an equal area layout
tion of such spatial layout problems. The objectives problem. and has some very interesting applications
and scope of these programs have varied widely. on both the micro and macro level. For example, the
Interest in this area has come from computer sci- assignment of buildings to sites or the assignment of
encerengineering researchers primarily looking at employees to preexisting offices or work stations can
problems of plant or production facilities layout Žor be a one-to-one assignment. The issues of size and
at the micro-scale, the layout of electronic circuits. shape do not enter into the layout process.
as well as from architects and interior designers Generally space planning applications are not as
interested in the design of large facilities such as straightforward as one-to-one assignment. The areas
office buildings, universities, hospitals, or depart- required by activities are not necessarily equal, so it
ment stores. More recently there has been interest on is not feasible to match activities and locations on a
the part of facilities managers concerned with reuse one-to-one basis. When assigning employees to ex-
and rearrangement of space. isting offices, we might want to consider multiple
Most of the research and development has fo- occupants. In stacking plan problems Žthe assignment
cused on what is known as the floor plan layout of activities to floors in a multi-storey building.,
problem, the physical arrangement of space on a plan more than one activity can be assigned to a single
Žreferred to as a block plan.. There are, however, floor or a single activity can occupy multiple floors
other applications of the space allocation problem— Žmany-to-one or one-to-many assignment.. How the
for example, an important commercial application area of an activity is apportioned among floors can
has been the assignment of activities to multiple be an important consideration in generating and eval-
floors of a building Žknown as the stack plan prob- uating a plan. In both of these examples, however,
lem.. Approaches to spatial allocation problems dif- activity size is still a relatively simple issue as actual
fer in terms of the type of problem addressed as well activity shapes are not considered.
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 199

The most difficult problems to represent are those

at the block-plan level. An activity is represented as
a polygon on the plan. This polygon should be able
to take on any shape and location while maintaining
the required activity area. The method for handling
unequal areas has a significant impact on the solu-
tion approach taken.

4. Approaches to automated layout

Automated space allocation algorithms require

some method of evaluation in order to guide the
layout process. There are three major paths that
solution techniques have followed. The first involves
the optimization of a single criterion function; specif-
ically the minimization of costs associated with com-
munication or flow of materials between activities.
While there are numerous drawbacks to such ap-
proaches, they have quite widespread application Fig. 1. A planar adjacency graph with corresponding dual w16x.
with respect to types of plans that can be generated.
This paper will focus primarily on this class of
solution techniques. but with finding an arrangement that satisfies a
A second path is based on a graph theoretic diverse set of constraints or relations. In this case the
approach. It is concerned primarily with generating a major criterion is feasibility. Early examples of this
layout that meets adjacency requirements between path are Eastman’s General Space Planner w12x and
activities. This approach requires the construction of Pfeffercorn’s Design Problem Solver w51x. Both
a planar adjacency graph where nodes represent ac- methods design layouts by placing objects so that
tivities to be located and edges Žor links. represent a they satisfy a set of constraints which involve such
direct adjacency requirement. The dual of a planar factors as position, orientation, adjacency, path, view,
graph determines the layout of the facility. A planar or distance. A more recent system of this type is the
graph with its corresponding dual is displayed in Fig. layout module of SEED w15x, a software system to
1. There is a long history of graph theoretic applica- support the early phases in building design under
tions to the layout problem Žsee Grason’s early work development at Carnegie Mellon. This module gen-
for example w20x.. Many approaches which follow erates schematic layouts of rectangular space under
this path are based on Richard Muther’s Systematic various constraints that include access, natural light
Layout Planning methodology w48x, the most fre- and privacy. It is based on Flemming’s LO-
quently used plant layout design methodology of the OSrABLOOS system w14x. Two other methods with
last 30 years. Muther’s methodology results in the a similar representation but with significantly differ-
generation of a space relationship diagram that is ent approaches to constraint satisfaction,
considered a ‘design skeleton’ from which a layout HeGeLrHeGeL-2 w2x and WRIGHT w6x, were also
can be generated. While graph theoretic approaches developed at Carnegie Melon Žsee Akin w3x and
differ from those that optimize a single criterion, Flemming et al. w14x for a comparison of these three
some of the heuristic solution techniques are similar. systems..
These will be noted later when such methods are While to date this type of approach has not been
discussed. the basis for commercial software, newer systems
Approaches following a third path are not con- such as SEED have considerable potential, particu-
cerned with optimizing a single measure or value, larly as interactive aids to the layout of schematic
200 R.S. Liggettr Automation in Construction 9 (2000) 197–215

Consider the assignment of N activities to N or

more sites, each of which can accommodate one and
only one activity. Associated with each pair of activi-
ties Ž i, j . is a measure of interaction QŽ i, j . Že.g.,
intensity of communication, level of traffic, etc...
Associated with each pair of sites Ž k,l . is a measure
of spatial separation C Ž k,l . Že.g., distance, travel
time, etc... In addition, a fixed cost, F Ž i,k ., may be
associated with the placement of activity i in site k.
If AŽ i . denotes the site to which activity i is as-
signed in a mapping A of activities to sites, the total
cost of a mapping Žsolution. can be given by:

Cost Ž A . s Ý F Ž i , AŽ i . .
activity i

Fig. 2. Display of access paths in a SEED layout w15x.

q Ý Ý Q Ž i , j . C Ž AŽ i . , AŽ j . . .
activity i activity j

plans in the design development phase of new facili- The objective is to find a mapping A, such that
ties. Some advantages of these systems are that they this cost function is minimized. The two parts of the
consider multiple criteria, maintain acceptable activ- function are known as the fixed cost term and the
ity shape, and can usually handle issues of circula- interactive cost term. The fixed cost term Žwhich is
tion space. Sample output from SEED with access equivalent to the criterion function of a linear assign-
paths displayed is shown in Fig. 2. A disadvantage ment problem. is concerned with the costs of assign-
of such systems is that they have not yet demon- ing a particular activity to a particular location. Fixed
strated the ability to handle large scale problems that costs might represent rent, special facilities construc-
are encountered in actual practice. tion requirements, or some measure of preference for
While a more detailed exploration of solution a particular site. The second term, which is the
techniques for layout problems presented in the re- quadratic portion of the objective function, intro-
mainder of this paper focuses primarily on the first duces costs caused by the interdependence of assign-
class of problems Žthe optimization of a single crite- ments. Interactive costs may represent a subjective
rion function., solution methods of the other classes judgment about grouping requirements or an actual
of problems will also be discussed as they relate to measure of flow of goods or employees.
the overall classification framework. The floor plan layout problem was first formu-
lated as a quadratic assignment problem by Armour
and Buffa in 1963. They considered the layout of a
5. One-to-one assignment: An early formulation manufacturing plant where the criterion to be mini-
as A Quadratic Assignment Problem mized was the cost of product flow between depart-
ments. Their work resulted in a computer program
One of the most popular approaches to automated called CRAFT ŽComputerized Relative Allocation of
facility layout was first formulated by Koopmans Facilities Technique. w9x.
and Beckmann w34x for problems concerned with the In the quadratic assignment formulation, floor
assignment of manufacturing plants to sites such that plan layout can be viewed as a combinatorial prob-
the cost of transportation of the flow of goods be- lem in which indivisible activities Že.g. departments
tween plants is minimized. Known as the Quadratic or individual employee work stations. are to be
Assignment Problem ŽQAP., it is concerned with assigned to fixed locations on a plan. In principle, it
finding optimal locations for a set of interrelated is possible to solve this problem by exhaustive enu-
objects. The problem can be described as follows. meration of all possible ways of assigning activities
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 201

to locations, and by selection of a plan which satis- assignment of activities to preexisting offices and
fies given constraints andror Žas in the case of the workspaces. Fig. 3 shows a typical office plan where
quadratic assignment formulation. yields the mini- space can be allocated on a one-to-one basis. Links
mum value for the criterion function. In practice this drawn on the plan represent adjacency requirements
turns out to be infeasible for problems of realistic between activities assigned to rooms.
size Žproblems of over 15 activities. since the num- We will first review a number of algorithms
ber of activityrlocation combinations involved is so which operate efficiently at the one-to-one level
vast. before considering the complexity added due to the
It can be shown that quadratic assignment prob- unequal area requirements inherent in floorplan lay-
lems belong to a class of mathematical problems out problems.
known as NP-complete. It is generally accepted that
the efficient solution of NP-complete problems is
impossible in principal. However a number of good 6. Solution procedures
approximate solution strategies do exist that produce
high quality solutions to realistically sized problems Existing approximate solution strategies can gen-
at acceptable cost. This is particularly true for the erally be classified into two categories: constructive
special case of the one-to-one problem. As men- initial placement strategies and iterative improve-
tioned previously, one-to-one formulations have a ment strategies. A constructive initial placement
number of commercial applications in terms of the strategy locates activities one by one, building a

Fig. 3. One-to-one office layout.

202 R.S. Liggettr Automation in Construction 9 (2000) 197–215

solution from scratch in a step-by-step fashion. An signment at any step in the decision tree, and select
iterative improvement Žalso known as hill climbing. the assignment which seems most likely to lead to an
strategy begins with some initial arrangement and optimal solution.
attempts to improve it incrementally. Simulated an- Constructive methods generally adopt either a
nealing, can be viewed as a variant of an iterative ‘local’ or ‘global’ orientation to a problem. Local
improvement strategy. More recently, genetic algo- methods consider only the assignments which have
rithms have become of interest for the solution of already been made; they tend to be less expensive
combinatorial problems such as the QAP. Genetic but yield poorer solutions. Global techniques attempt
algorithms begin with a set of possible solutions and to account for possible future moves in the evalua-
use mutation and crossover techniques to evolve tion of a particular assignment and, although more
existing solutions into better solutions. The next expensive, generally produce better solutions. Both
sections will briefly cover basic improvement and the Gilmore–Hillier and Graves–Whinston algo-
constructive techniques Žsee Refs. w33,37–39x for a rithms are examples of global techniques.
more detailed description and comparison of early
6.2. ImproÕement procedures
techniques. and then focus on newer approaches
which include simulated annealing and genetic algo- Improvement procedures start with a single solu-
rithms. tion and attempt to incrementally improve it. The
simplest version is the ‘pair-wise’ exchange. Starting
6.1. ConstructiÕe procedures
from an initial solution, the procedure consists of
Constructive procedures build a solution from systematically evaluating possible exchanges be-
scratch using an n-stage decision process. Some tween pairs of activities and making an exchange if
methods attempt to automate a set of ‘rules of it improves the value of the criterion. There are a
thumb’ for making intelligent assignments at each number of variants on the basic pair-wise exchange
stage, essentially modelling the thought process of a which focus on reducing the computational effort
human designer. A simple activity selection rule expended or on improving the quality of the solu-
might be: select the activity which has the highest tions generated. These variants generally involve the
connectivity to any activity already placed. A loca- method of selection of activities for possible ex-
tion can then be selected, either by a simple rule of change and which exchange to make Že.g., whether
thumb again Že.g., select the first empty location or not to make the first exchange that leads to an
adjacent to a placed activity starting at the top left improvement or to evaluate all possible exchanges
and working clockwise., or by more sophisticated and select the exchange that results in the maximum
criteria such as selecting that location which yields cost improvement.. Since this latter method can be
the minimum value of the criterion function consid- very costly with respect to computation time, meth-
ering only the activities already placed. ŽSee Refs. ods use different ways to limit evaluation of ex-
w11,42,49x for early constructive procedures.. changes. For example, Elshafei w13x only evaluates
The constructive decision process can be viewed all possible moves of a single activity. Hanan et al.
as a ‘tree search’ where at each branch the selection w22x limit exchanges to immediate neighborhoods or
of an activity–location assignment is made. Two activities. Volmann et al. w59x only considered the
more sophisticated and computationally intensive ap- exchange of the two activities which contribute the
proaches use mathematical bounds on the decision most to the total cost of the current solution at each
tree to guide the process. At each stage of the step. Other methods attempt to be intelligent about
decision tree, Gilmore w18x and Hillier w25x calculate the order in which they evaluate potential exchanges
a lower bound for the objective function for each w25,26x. A more expensive approach termed ‘biased
branch. The activity–location pair which yields the sampling’ selects randomly from the set of possible
minimum lower bound is selected for the next as- exchanges showing improvement. The probability of
signment. Graves and Whinston w21x use probability selection associated with each exchange is propor-
theory to calculate the expected value of the objec- tional to its corresponding cost reduction w50x. Re-
tive function for each possible activity–location as- sults have also been reported for experiments involv-
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 203

ing three-way, four-way and five-way exchanges of the system, T, is lowered in steps. The algorithm
w7,42x, however, the minor improvements to solu- terminates as T approaches zero. The use of simu-
tions have generally not balanced the additional cost lated annealing techniques makes it less likely to fall
of the generation process. into local optima, provided the annealing process is
Improvement procedures usually converge on lo- long enough.
cal optima. Since all improvement procedures re- Sharpe and Marksjo w53x show how an implemen-
quire an initial solution, a number of local optima tation of the simulated annealing method provides a
can be generated and compared by using different relatively simple but powerful approach to facility
starting configurations. Elshafei w13x experimented layout optimization. They have applied it success-
with a technique to retreat from a local optimum by fully to large scale problems with up to 200 loca-
selecting the move which results in the minimum tions. An additional advantage is that it can produce
cost increase. The exchange cycle is then repeated a number of near-optimal solutions from which de-
using this position as the starting solution which signers can select. Similar advantages are shown by
hopefully leads to a new local optima. A more recent the author of a program called CLASS, Computer-
solution technique taken from the area of statistical ized Layout Solutions using Simulated annealing
mechanics, simulated annealing, has been used to w30x.
‘back out’ of unattractive local optima.
6.4. Genetic algorithms
6.3. Simulated annealing Somewhat related to improvement procedures are
a class of algorithms which rely on analogies to
Simulated annealing techniques eliminate many natural processes. This type of algorithm has been
disadvantages of improvement Žhillclimbing. meth- described by Michalewicz w45x as follows:
ods. Solutions no longer depend on the starting point
Genetic algorithms are based on the principle of
and are more likely to converge on a good Žif not
evolution Žsurvival of the fittest.. In such algo-
optimal. solution. The main departure from tradi-
rithms a population of individuals Žpotential solu-
tional improvement methods is that changes accepted
tions. undergoes a sequence of unary Žmutation
at each stage of the optimization can actually in-
type. and higher order Žcrossover type. transfor-
crease the cost of the plan. In most hillclimbing
mations. As these individuals strive for survival: a
methods, a new solution is accepted only if it yields
selection scheme, biased towards fitter individu-
an improved value of the objective function. In
als, selects the next generation. After some num-
simulated annealing, an exchange can also be ac-
ber of generations, the algorithm converges with
cepted if the probability of the resulting cost increase
the best individual hopefully representing the op-
occurring is lower than a control parameter. This
timum.
technique is derived from a method that simulates
the cooling of a mass of vibrating atoms from a high Genetic algorithms share the following features
temperature T. The probability of acceptance Ž p . of Žsee Tate and Smith w56x.:
the exchange of a pair of activities equals one if the Ø An initial population of solutions Žcan be ran-
exchange provides a better value of the objective domly generated.
function. If, however, the cost change is positive Ø A mechanism for generating new solutions by
Ži.e., increases the cost., the probability of accep- combining features from solutions in the existing
tance p is a function of the difference in values of population Žreproduction..
objective function for the current solution and the Ø A mechanism for generating a new solution by
new solution Ž D ., and an additional control parame- operating on a single previously known solution
ter, T Žwhich represents temperature in the actual Žmutation..
annealing process.: p s ŽexpŽyDrT ... In general, Ø A mechanism for selecting the set of solutions
the lower the temperature T is, the smaller the from the populationŽs., giving preference to those
chances for the acceptance of a new solution are. with better objective function values Žselection.
During execution of the algorithm, the temperature and
204 R.S. Liggettr Automation in Construction 9 (2000) 197–215

Ø A mechanism for removing solutions from the tive as the mutation and reproductive mechanisms
population Žculling.. can be visualized in terms of human design pro-
Solutions can be selected to mutate or to repro- cesses. We can conceive of a designer working with
duce. Selection is performed with a bias towards a set of possible solutions which evolve toward one
choosing the better solutions in the current popula- or more preferred solutions. As these solutions evolve
tion. For the facilities layout problem mutation can there may be portions of the design which are com-
take the form of some variant of the pair-wise ex- mon to the best solutions. These are preserved when
change. The key feature of the reproduction process design options are combined to generate new solu-
used by Tate and Smith w56x is that any activity tions.
assigned the same location in both parents will oc- It is logical from a human perspective to think of
cupy that location in the offspring. For the remaining the solution process as a hybrid of a number of
locations, activity assignments are chosen randomly approaches. This is also true for automating the
from one or the other parent. Any unassigned activi- layout process and there are a number of hybrid
ties are then matched with the remaining unassigned solution techniques.
locations. As children are created, solutions with the
poorest values of the objective are eliminated Žculled. 6.5. Hybrid approaches
to keep the population the same size. As for simu-
lated annealing, excellent results have been reported Since both iterative improvement Žwhich includes
for this type of algorithm and there is considerable simulated annealing. and genetic algorithms require
interest in investigating further extensions. Any of initial solutions, it is generally preferable to begin
the reported improvement procedures are candidates with a reasonable solution rather than one which is
for mutation processes and there are numerous possi- randomly generated. On the other hand, while con-
bilities for reproductive transformations. structive procedures can produce good solutions,
Gero and Kazakov w17x, extending earlier work by there is almost always room for improvement. Using
Jo and Gero w29x, take advantage of what they term a constructive procedure to generate an initial solu-
‘superior’ evolved genes. Solutions, called geno- tion should reduce the number of iterations of the
types, are represented as a linear sequence of inte- improvement procedure and improve the quality of
gers which can be interpreted as the order activities the solution generated w38x. The coupling of a con-
are placed on the plan. After a fixed number of structive procedure with an improvement procedure
generations, the top ten percent of genotypes Žwith provides an effective combination of a global and
respect to a specified objective. and the bottom ten local approach to a problem. The constructive proce-
percent of genotypes are searched to identify groups dure sets the general tone of the solution while the
of genes that occur together Žcompact subsets of improvement procedure refines the details.
activities that appear together in the sequence.. Com- The author has shown good results with a hybrid
pact gene groups found almost exclusively in the approach which combines the Graves–Whinston
best solutions and almost never in the poorer solu- constructive procedure with a pair-wise exchange
tions are declared new ‘superior’ evolved genes. improvement algorithm w37x. More recently Huntley
These evolved genes are then represented as a single and Brown w27x have combined a high-level genetic
gene for further reproduction. Applications of this algorithm with a simulated annealing algorithm. Jo
approach show excellent results for an office and a and Gero w29x improve upon Liggett’s w39x solution
hospital layout problem found in the literature. The to an office layout problem by using it as an initial
authors have also shown that evolved genes tend to population for their genetic search algorithm, EDGE
represent design features that can be re-used later in ŽEvolutionary Design based on Genetic Evaluation..
similar layout problems. Heragu and Alfa w24x show that a hybrid method
Genetic search methods climb many peaks in which uses a modified penalty algorithm to generate
parallel making it more likely to settle on a global or an initial solution which is then improved using
near-global solution than constructive or improve- simulated annealing produces superior results when
ment procedures. Genetic algorithms are also attrac- compared to a two or three-way exchange procedure
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 205

or just to simulated annealing alone. Kaku et al. w31x jacency. values can be specified between modules of
propose a hybrid heuristic that consists of three parts. the same activity to encourage contiguity of activity
In the first part, several partial assignments are gen- space. Heuristics in constructive procedures can also
erated for use as starting points for a constructive be used to ensure contiguity of space. For example,
heuristic Ža breadth first search tree is used to enu- after the initial module is placed for an activity in the
merate a set of good partial assignments.. In the Liggett implementation of the Graves–Whinston al-
second part, these starting points are used to con- gorithm w41x, locations evaluated for the placement
struct complete assignments. They experimented with of subsequent modules of the same activity are lim-
both the Gilmore w18x and Graves–Whinston w21x ited to adjacent modules. The order of location mod-
constructive approaches for completing the solution. ule selection is based on the expected value of the
Finally, attempts at improving the constructed solu- objective function. For constructive procedures there
tions are made by the application of both pair-wise is no guarantee, however, that activities will not be
and triple-exchange routines. split Žunless backtrack strategies are employed at
perhaps a prohibitive cost. and modularization can
produce irregular shapes that are undesirable or un-
7. Unequal areas workable Žsee Fig. 4 for the effect of modularization
on a block plan..
Most of the approaches reviewed above can be Some approaches do not require Žor allow. a
used to generate acceptable solutions to the one-to- predefined plan perimeter making it easier to control
one assignment problem. However, space planning activity shape. A key feature of a constructive proce-
problems are generally more complex than the clas- dure, SHAPE w23x, is the order candidate location
sical quadratic assignment formulation due to the modules are considered. The shape of the layout
imposition of activity area requirements. Since areas evolves as departments are placed in the modules
required by activities are not necessarily equal, it is surrounding those already occupied. The collection
not always feasible to match activities and locations of modules adjacent to a partial layout constitutes a
on a one-to-one basis. Note, this is not true if potential activity location. The set of modules that
location perimeters Že.g., existing offices. have been minimizes the current value of the objective function
predefined since the assignment of activities to loca- is chosen for activity assignment. Earlier versions of
tions can be limited by size constraints. Such con- this type of program include CORELAP w35x and
straints are generally handled in the quadratic assign- ALDEP w52x. While such approaches guarantee activ-
ment formulation with the fixed cost function. Fixed ity contiguity, they are not generally effective given
costs can be set prohibitively high for the assignment existing plan perimeters. These programs can, how-
of activities to offices of unacceptable size. At the ever, be valuable tools for exploring variations in the
block plan level, however, the problem becomes size and shape of a building during the schematic
more difficult since we are locating activities on a design phase. For the algorithms that require preset
plan where location boundaries have not been pre- perimeters, layouts free from shape constraints can
fixed. be generated by specifying a large square perimeter
or a perimeter the size and shape of the site.
7.1. Modular approach Buffa and Armour’s early approach to the unequal
area floorplan layout problem limited the exchange
A typical approach to the unequal area block plan of activities to pairs of equal area w9x. This constrains
problem Žfor the class of algorithms discussed above. the problem significantly in terms of exploration of
is to partition the plan into equal size modules. Each the possible solution space. Other improvement pro-
activity is then partitioned into modules of the same cedures exchange individual activity modules, but
size according to required floor area. The problem is must include constraints to maintain contiguity of
then one of assigning activity modules to location activity shape. Exchange procedures can also be used
modules in a one-to-one fashion. For the quadratic to improve shapes resulting from constructive proce-
assignment problem, artificially high interaction Žad- dures.
206 R.S. Liggettr Automation in Construction 9 (2000) 197–215

Fig. 4. Effect of modularization on a block plan.

The method of plan modularization has important ities with high interactions in close proximity to each
implications for the automated algorithm. If an exist- other. The leaves on the tree represent activities to be
ing building has not been designed on a square placed, and interior nodes represent the slicing opera-
planning grid or if the perimeter has other than right tion Žleft cut, right cut, bottom cut or top cut as
angled edges, the modularized plan will only approx- shown in Fig. 5.. A simulated annealing procedure is
imate the actual perimeter. The smaller the grid size, used to exchange slicing operators in the tree, thus
the better the approximation of the existing plan. generating different rectangular partitioning schemes.
However, a small grid size increases the number of The objective function used to drive the annealing
modules. This leads to increased computation time in process includes the quadratic term of the QAP and a
generating a layout. It may also cause fragmented penalty term for geometric constraints. Tam shows
activity shapes. good results with 20 and 30 activities.

7.2. Arrangement of rectangles

Other approaches to the unequal area layout prob-

lem partition rectangular shapes into smaller rectan-
gles for assignment. Tam w54,55x represents the lay-
out as a slicing structure that is constructed recur-
sively by partitioning a rectangular block. Each rect-
angular partition in the slicing structure corresponds
to the space allocated to an activity. Tam uses cluster
analysis to generate a slicing tree which places activ- Fig. 5. Tam’s slicing free and associated layout w54x.
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 207

Tate and Smith w57x use a genetic algorithm with 8. Constraint based methods for unequal area
a flexible bay structure. The prespecified rectangular layout
area is divided in one direction into bays of varying
width. Each bay is then divided into rectangular Methods which do not focus on a single objective
partitions of equal width but different lengths. The but rather seek to generate solutions that meet a
bays are flexible in that their widths will vary with number of different Žpossibly conflicting. constraints
their number and contents. In the Tate and Smith generally deal with the unequal area layout problem
equal-area genetic algorithm discussed earlier w56x, a in terms of arrangements of rectangles. Three meth-
solution was represented by a permutation of inte- ods mentioned earlier, LOOS w14x, WRIGHT w6x and
gers Žone through n, where n is the number of HeGeL w2x, all formulate the problem as one of
activities. and the locations of the integers in the arrangement of rectangles with sides parallel to axes
sequence correspond to locations on the plan. In this of an orthogonal system. They all attempt to satisfy
implementation, the locations correspond to bay par- two types of constraints: one set that is dependent on
titions and an additional sequence is required to the structure or topology of the problem such as the
represent a solution. This second sequence defines requirement that the rectangles not overlap and fit
the number of bays and the number of partitions per within a given boundary; and a second set of con-
bay by indicating where breakpoints exist. Changing straints which are independent of structure and con-
the location of break points can change the number sider attributes such as area, dimension, orientation,
of rectangles per bay andror the number of bays. and adjacency requirements.
Solutions are evaluated and compared using the ob- LOOS and HelGeL are hierarchical generate-and-
jective function of the quadratic assignment problem test methods that incrementally construct solutions
with the addition of adaptive penalty functions to by adding one rectangle at a time to a partial solu-
handle area and proportion requirements for activi- tion, testing for constraint satisfaction at each step.
ties. LOOS w14x employs a breadth first search, generating
Van Camp et al. w58x use a nonlinear model. candidate solutions at each stage by enumerating all
Activities are considered to be of fixed area but of possible ways of adding a new rectangle to a partial
variable dimensioned rectangular shape. Using a sim- solution. Each intermediate state is examined for
ilar measure as the QAP, the objective function constraint violation. If a structure dependent con-
represents the cost of flow of material between activ- straint is violated the intermediate solution is pruned.
ities multiplied by the distance between activity cen- A count of other constraints violated is made for any
troids. An additional term measures distance between partial state and only those with the lowest counts
activities and the outside wall. Constraints require are selected for expansion. By sequentially expand-
that no two departments overlap, that departments ing partial solutions and pruning less promising can-
must be contained in the facility and meet area as didates, LOOS commits to a set of current globally
well as shape constraints Žagain using the penalty best candidate solutions and avoids backtracking.
function approach.. Nonlinear optimization is used to HeGeLrHeGeL2 w2x also constructs a solution in
generate an initial feasible solution Žnote activity a step-by-step fashion, however, it follows a depth-
centroids and dimensions are viewed as continuous first search. The solution procedure is based on a
variables which can vary over the dimensions of the protocol analysis of the problem-structuring and
exterior boundaries.. The nonlinear optimization problem solving behavior of designers. Layout re-
generally results in local optima. The procedure then quirements are expressed as relationships between
applies a pair-wise exchange improvement procedure objects to be located. These relationships, which are
to reduce the value of the interactive portion of the called ‘predicates’, are used as generative constraints
objective function. At this stage infeasible solutions or evaluative criteria. A generative predicate results
Žwhere an activity is too large for the available in the selection of a design unit to be placed Žfor
location. are accepted. A return is then made to the example, a direct access requirement.. Alternative
nonlinear algorithm to generate a new feasible solu- locations for the design unit are then generated based
tion. on the predicate. These locations are tested against
208 R.S. Liggettr Automation in Construction 9 (2000) 197–215

the other predicates associated with the design unit. Designers can choose between minimizing empty
If none of the generated locations meets the test space Žmaking the most use of space. or maximizing
criteria, HeGeL will backtrack. If a single location empty space by compacting the layout. Jacobs uses a
meets the criteria, the placement is made. If there are two-stage solution procedure where each stage com-
multiple locations, they are presented to the user for bines both a constructive and improvement proce-
selection. The user selects a final location and others dure. In the first stage units are ranked for order of
are stored for possible later backtracking. While the placement by boundary preferences, adjacency pref-
original approach relied on the user to direct the erences, and the distance measure. A solution is then
search in terms of selection of the next generative constructed incrementally based on the priority or-
predicate or unit to place, HeGeL2 implements an dering by generating all feasible locations for the
optimization methodology to make the best place- next design unit to be placed. A location is selected
ment decision at each stage. for placement based on the objective function. If a
WRIGHT w6x implements a constraint-directed unit will not fit on the layout, all units are removed
search called disjunctive constraint satisfaction in and the process is restarted. Once a solution is
which constraints are incrementally satisfied. The constructed it is improved with a pair-wise exchange
layout is represented using algebraic equations and procedure. Since there is an element of randomness
inequalities of variables that represent the border in the initial placement ordering as well as in the
lines, dimensions, areas and orientation of the design selection of a location for placement Žif two loca-
units. These are called atomic constraints. Disjunc- tions yield the same value for the objective, the
tive constraints are Boolean combinations of atomic selection is made randomly., a different solution will
constraints and specify the ‘‘structural alternatives result each time the process is restarted.
considered by WRIGHT for satisfying the
constraint.’’ For example, a disjunctive constraint
might represent the four alternatives of placing one
unit directly adjacent to another Žit can be north, 9. Graph theoretic approaches
south, east or west of the second.. Solutions are
generated by sequentially instantiating disjunctive Graph theoretic approaches also handle the un-
constraints and solving the current constraint satis- equal area block plan. In these approaches a block
faction problem ŽCSP.. A CSP is consistent if there plan is constructed as the dual of a planar graph
exists values for all variables that simultaneously where nodes represent spaces and links represent
satisfy all constraints. If a CSP is inconsistent, the required adjacencies. While it is always possible to
propagation algorithm backtracks and selects an al- construct a block plan from a planar graph which
ternative disjunct. The full solution process will find meets the given adjacency requirements between
all significantly different solutions. Since a problem spaces and between spaces and the outside area, the
can be underconstrained or overconstrained, the de- resulting plan may not meet size and shape require-
sign process is characterized by the addition or relax- ments imposed on each space. Constructing a block
ation of constraints by the designer. plan that meets size and shape requirements is a
Rather than formulating competing criteria as a nontrivial problem.
diverse set of constraints, Jacobs w28x combines them A graph theoretic approach is a two stage process.
into a single weighted objective function. The crite- In the first stage a planar graph which corresponds to
ria considered include the distance between design the adjacency requirements is generated. A planar
units with respect to frequency of interaction Žthe graph is one which can be drawn so that no two
typical single objective function criteria considered edges intersect. A planar graph is maximal if no
by the first class of solution techniques., as well as edges can be added without losing planarity. It is
direct adjacency requirements. He also includes cri- possible that the adjacency requirements cannot be
teria related to the alignment of spaces with a goal of represented by a planar graph. In this case the prob-
keeping the structure of the layout as simple as lem is overconstrained and the solution procedure
possible. A final consideration is the use of space. becomes one of generating maximal planar graphs
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 209

which maximize the number of adjacencies or the ing row of the hexagonal graph. A feasible layout
weighted adjacencies met by the graph. Once a graph with respect to area is determined by adjusting the
has been formed, the second stage involves generat- height of a row. Fig. 6 shows a hexagonal adjacency
ing the actual layout. See Baybars and Eastman w5x graph with the resulting layout. Goetschalckx’s algo-
and Foulds w16x for a more detailed discussion of the rithm, SPIRAL, has been implemented in a commer-
graph theoretic approach and early applications. cial product called FactoryOPT by CIM TECH-
A typical graph theoretic heuristic for the layout NOLOGIES of Ames, Iowa. Montreuil et al. w47x
problem consists of the following steps: uses a linear programming model to generate a block
Stage 1: Generating a planar graph layout of rectangular spaces from a planar adjacency
Ø Form a weighted graph of the relationships graph. A limitation of both Goetschalckx’s and Mon-
between facilities treuil et al.’s approach is that the building perimeter
Ø Identify a maximal planar subgraph of rela- must be rectangular.
tively high weight The graph theoretic formulation differs from the
Stage 2: Generating a block plan traditional quadratic assignment approach to the lay-
Ø Construct the dual from the planar subgraph. out problem in a number of ways. The fundamental
The dual represents a layout apart from the fact difference is that the graph theory approach consid-
that shapes and areas have not been taken into ers only direct adjacency requirements. No consider-
account. ation is given to nonadjacent pairs of facilities with
Ø Attempt to accommodate shapes and areas in respect to communication costs, even if they are
forming a block plan from the dual. relatively close together on the plan. In addition,
Much research focuses on just the first stage of
the process-generating a maximal planar graph. A
planar graph can be generated by adding or subtract-
ing edges following a step-by-step process. Typically
edges are added in a greedy fashion Žlocal construc-
tive method. where planarity is tested after each
edge addition. Leung w36x presents such a construc-
tive procedure which capitalizes on the fact that
triangulated graphs are maximally planar. The
method starts with a planar subgraph which is gener-
ated by enumerating all possible groups of four
vertices. The group with maximum weight is se-
lected. At each subsequent step either a single vertex
or a triple of vertices which maximizes the additional
edge-weight per vertex is added to a face. Once a
maximal planar graph is constructed, many methods
apply some type of improvement procedure such as
pair-wise exchange or simulated annealing to im-
prove the adjacency score while maintaining pla-
narity.
By requiring a hexagonal structure for the adja-
cency graph, Goetschalckx w19x has developed an
efficient method for generating a rectangular block
plan that meets area requirements from the dual of a
planar graph. The rectangular floor plan is divided
into rows based on the number of rows of the
hexagonal graph. Each row in the plan is then parti- Fig. 6. Goetschkhs’s hexagonal adjacency graph with resulting
tioned into the number of spaces in the correspond- layout w19x.
210 R.S. Liggettr Automation in Construction 9 (2000) 197–215

fixed costs are not included in the graph theoretic require multiple floors. This can be modeled as a
formulation nor are preassigned spaces accommo- one-to-one module placement problem, where the
dated. Foulds w16x views the graph theoretic ap- floors and activities are comprised of equal area
proach as more appropriate for the design of a new modules. However, it is more attractive to think in
facility where there is more design freedom, while terms of assigning an entire activity to a floor in a
the QAP formulation is more useful in a structured single step Žnote, activities may also be allocated
situation. over multiple floors.. Both constructive and im-
provement procedures can be used to generate a
plan. Simple rules of thumb for locating groups on
10. Many-to-one and one-to-many assignment floors Žlocal heuristics. can produce decent starting
solutions to be modified by improvement procedures.
A problem that has not been addressed as much in A simple pair-wise exchange requires the exchange
academic research but is probably the most widely of relatively equal area activities unless there is free
used commercial application is the stack plan. In space on a floor. Alternatively, clusters of groups of
stacking problems activities and locations may have similar size can be identified for exchange. It is also
unequal areas, however, there is not the added com- possible to exceed available floor area during the
plexity activity shape brings to block plan problems. exchange process using some form of adaptive
Locations in a stack plan generally represent floors penalty function to ultimately converge on an accept-
of a building Žphysically shown as a bar.. Activities able solution. While we have not found applications
are represented by areas, colored bars assigned to the of some of the newer approaches such as genetic
floors Žsee Fig. 7.. Multiple activities can be as- algorithms to the stack problem, there could be a
signed to a single floor or a single activity can great deal of potential in this area.

Fig. 7. A typical stack plan.

 R.S. Liggettr Automation in Construction 9 (2000) 197–215 211

Mahdavi et al. w43x, Zhang presents a new ap-

proach to the stacking problem. She clusters what are
termed functional units Žactivities to be located. into
groups and assigns groups to floors such that the
weight between floors is minimized. To begin the
process functional units are sorted by their total
connected weights in descending order. Each of the
first nf y 1 units is then assigned to a separate group
Žwhere nf represents the number of floors in the
stacking problem.. The remaining units are assigned
to the last group. Each group has an area constraint
corresponding to one floor of the building. Func-
tional units are moved from the last group to other
groups such that intergroup weights are minimized
and area requirements are satisfied. When no move
will produce a gain or improve the area balance, the
process is over. Groups are now assigned to actual
Fig. 8. Using spacefiling curves to construct layouts w8x.
floors. If the floors are of different size than it is
obvious which cluster is assigned to which floor. If
tween floors or within floors. The criterion function
they are the same size, a dynamic programming
used to drive the improvement process is a function
algorithm is used to assign groups to floors with the
of horizontal travel between locations on the same
objective of maximizing adjacent floor weights.
floor and vertical travel which is a combination of
Another similar problem is the assignment of
horizontal travel from each activity to the lift and the
multiple occupants to a single office. In this case as
travel time of the lift. The use of space filling curves
well, an algorithm need not worry about the actual
ensures activity contiguity is always maintained and
shape of space, only the activity area assigned to
can help manage activity shape.
each location. The same solution techniques can be
Kaku et al. w32x use a K-median heuristic to
applied to the stack and the many-to-one office
cluster departments into groupings in such a way that
assignment problem. Only the graphic representation
inter-group interaction is minimized where the num-
is different.
ber of floors determines the number of groups. This
is similar to Mahdavi et al.’s w43x stacking process.
11. Multi-floor layout problems Once clusters are created they are assigned to floors
using a quadratic assignment objective function.
Block and stack problems are considered simulta-
Block plans for each floor are then generated based
neously with multi-floor algorithms. Bozer et al. w8x
on distances between locations on the floor and
have developed an algorithm called MULTIPLE
ŽMULTI-floor Plan Layout Evaluation. which adopts between these locations and the elevator. Interactions
with activities on other floors are assumed to flow
the grid cell representation for the unequal area
through the elevator. Thus activities with strong
block layout problem. Each floor of the building is
connections to other floors will be placed next to the
divided into grid modules. Spacefilling curves are
elevator. This algorithm is limited to equal area
used to layout activities on the grid. A space filling
activities and a single elevator per floor.
curve is a way of visiting neighbors on a grid by
taking horizontal, vertical or diagonal steps to adjoin-
ing grid cells Žsee Fig. 8.. The layout is controlled by 12. Expert systems
the order activities are placed in the grid. An initial
layout assigns activities to floors. A simulated an- While not true expert systems, many of the early
nealing improvement algorithm is used to modify the heuristics for automated layout implemented rules of
order of layout by exchanging activity locations be- thumb that a designer might follow in generating a
212 R.S. Liggettr Automation in Construction 9 (2000) 197–215

layout. Modelling the solution process of the human closeness relationships between departments. A stan-
designer is dealt with more explicitly in the work by dard constructive algorithm is linked with an im-
Akin et al. w2x on layout protocol analysis which provement component to generate the layout. Cle-
resulted in the HeGeL system discussed earlier. land and Hills w10x use a simulated annealing ap-
A more traditional expert system approach devel- proach for the actual layout, but use a knowledge-
oped by Malakooti and Tsurushima w44x combines based system as an intelligent editor to guide the
multiple-criteria decision making with an expert sys- designer in problem formulation and as an intelligent
tem. The expert system has four parts: critic to assess the quality of the layout and to
Ø A data base which expresses the problem to be suggest ways to improve it.
solved. All raw data is treated as facts such as the
number of activities, size, and flow.
Ø A knowledge base which stores domain-specific 13. Commercial applications
problem-solving knowledge such as rules of
thumb for generating the layout. While there seems to be considerable interest in
Ø A priority base, which contains priorities for rules, computer programs for facility layout, there are sur-
adjacency, the order of assignment, etc. prisingly few commercially available products. The
Ø An inference engine, which controls the so-called ‘layout’ features of many CAD systems
problem-solving structure. simply provide a graphic interface for the user to
The expert system interacts with the decision layout a plan in manual mode with little or no access
maker through the inference engine allowing the to information concerning the layout criteria. On the
decision maker to change the priorities or rules. An other hand, a solution generated by an automated
interpretation of the layout, which includes all the algorithm that is based on a single cost function
rules that have been used to assign activities to sites, captures only one aspect of a designer’s concerns in
is displayed so that the decision maker can see why any realistic context. A system which meets com-
individual assignments have been made. A what-if mercial needs of today should provide interface ca-
analysis module allows the decision maker to change pabilities ranging from complete user interaction,
information in the data, the knowledge or priority where the user interactively specifies the location of
bases and see the results. By giving priority to each activity, to complete automation, where an
different criteria and comparing the resulting layouts, algorithm generates an initial solution w40x. Or as
the system can automatically update priorities based desired, a designer should be able to interactively
on the decision maker’s choices. locate some activities and use an algorithm to locate
Recent references to expert systems tend to focus or suggest locations for others. Rather than generat-
more on the integrated problem solving experience ing a single least-costly plan, the designer with the
than on the actual layout process. Many expert sys- aid of automated algorithms can make tradeoffs be-
tems use existing tools for the actual layout which tween competing criteria and converge on a solution
include both constructive and improvement proce- that responds to a broad spectrum of complex and
dures mentioned earlier. Abdou and Dutta w1x use an often ill-defined issues.
expert system to derive the relationship chart from a In order to meet the needs of facility designers
set of multiple criteria that are fuzzy, non-quantifia- and managers a number of factors must be present in
ble and apparently conflicting. Once the relationships a commercial product:
are derived a standard layout generation package Ø The ability to handle large scale problems
Žsuch as ALDEP or CORELAP. is used to derive a Ø A modern interactive interface
suitable layout. The expert system is then used to Ø Support for an iterative design process
examine the feasibility of the result. Ø Links to CAD and Facilities Management
FLEXPERT w4x, a facility layout expert system Databases
based on the theory of fuzzy logic follows a similar It is clear that realistic space allocation problems
scenario. It uses the expert system to generate a can involve the assignment of space for very large
relationship chart to combine criteria on flow and organizations. A single problem can include multiple
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 213

buildings, numerous floors and hundreds if not thou- quirements. Output resulting from a layout planner
sands of activities. Automated solution procedures to can be used to update the databases and even gener-
date have not been tested for problems of this size. ate transactions such as move orders.
The program interface should be able to prepare In spite of the long research history associated
and present data at any desired level of aggregation with automated layout and space allocation systems,
and use output from one stage of the design process in practice these systems have not been utilized to
to generate subproblems at the next stage Žfor exam- their full potential. We would expect this to change
ple, from stacking a multi-storey building to block in the near future given the increasing interest in
plans of individual floors.. An early commercial facilities planning and management, the increasing
implementation, the Calcomp Facilities Planning and use of computer-aided design tools in the building
Management Application Package, allowed designers design and management industry, and the improve-
to select both the level of space aggregation and ments in computer hardware and software which
activity aggregation from a graphically displayed make the solution of larger scale problems possible
organization chart. Such an approach is an effective as well as facilitate human–computer interaction.
way of reducing the size of large scale problems to
make both human and algorithmic problem solving
possible. Newer drag and drop graphic interfaces, References
now expected by users, can be used to move activi-
ties from an organization chart to the graphic repre- w1x G. Abdou, S. Dutta, An integrated approach to facilities
sentation of a plan or to shift activities around a plan. layout using expert systems, Int. J. Prod. Res. 28 Ž4. Ž1990.
685–708.
Experience has shown that layout tools are most w2x O. Akin, B. Dave, S. Pithavadian, Heuristic generation of
effective when employed in an iterative fashion. For layouts ŽHeGeL.: based on a paradigm for problem structur-
example, in a typical layout problem, the design ing, Environ. Plan. B: Plan. Des. 19 Ž1992. 33–59.
process might start by automatically generating a w3x O. Akin, R. Sen, Navigation within a structure search space
plan where only information on activity interactions in layout problems, Environ. Plan. B: Plan. Des. 23 Ž1996.
421–442.
is considered. Generally the result produced will be w4x A. Badiru, A. Arif, FLEXPERT: facility layout expert system
unsatisfactory, prompting the designer andror client using fuzzy linguistic relationship codes, IIE Trans. 28 Ž4.
to make their design requirements more explicit. Ž1996. 295–309.
This can be accomplished by adding information on w5x I. Baybars, C. Eastman, Enumerating architectural arrange-
activity–location preferences and activity preassign- ments by generating their underlying graphs, Environ. Plan.
B 7 Ž1980. 289–310.
ments. Location preferences, for example, can be w6x C. Baykan, Formulating spatial layout as a disjunctive con-
added to the quadratic assignment objective function straint satisfaction problem, Doctoral Dissertation, Depart-
in the form of fixed costs. Most systems should also ment of Architecture, Carnegie Mellon University, 1991.
have the capability for preassigning activities to par- w7x T. Block, PLOP—Plant layout optimization procedure, Uni-
ticular locations to account for preexisting condi- versity of Melbourne, Melbourne, 1978.
w8x Y. Bozer, R. Meller, S. Erelebacher, An improvement-type
tions. Appropriate trade-offs can then be made as the layout algorithm for single and multiple-floor facilities, Man-
problem is gradually transformed from one in which age. Sci. 40 Ž7. Ž1994. 918–932.
few locations are fixed to a complete solution. w9x E. Buffa, G. Armour, Allocating facilities with CRAFT,
One of the most important applications of com- Harvard Business Rev. 42 Ž2. Ž1964. 136–159.
w10x G. Cleland, W. Hills, A knowledge-based systems approach
puterized facility layout is in the area of ongoing
to the layout design of large made-to-order products, in: J.S.
space management. Here the link between a facilities Gero, F. Sudweeks ŽEds.., Artificial Intelligence in Design
management database and the layout program is ’94, Kluwer, The Netherlands, 1994, pp. 257–274.
critical. An inventory database of personnel, equip- w11x H. Edwards, B. Gillett, M. Hale, Modular allocation tech-
ment and space provides information on the current nique ŽMAT., Management Sci. 17 Ž3. Ž1970. 161–169.
w12x C. Eastman, Automated space planning, Artificial Intelli-
layout of space in the building which can be dis-
gence 4 Ž1973. 41–64.
played and evaluated by the layout program. The w13x A. Elshafei, Hospital layout as a quadratic assignment prob-
layout program can then be used to generate and test lem, Operations Res. Q. 28 Ž1. Ž1977. 167–179.
alternative configurations meeting new space re- w14x U. Flemming, C. Baykan, R. Coyne, Hierarchical generate-
214 R.S. Liggettr Automation in Construction 9 (2000) 197–215

and-test versus constraint-directed search, in: J. Gero ŽEd.., w34x J. Koopmans, M. Beckmann, Assignment problems and loca-
Proceedings of the Artificial Intelligence in Design Confer- tion of economic activities, Econometrica 25 Ž1967. 53–76.
ence ’92, Kluwer, Dordrecht, 1992, pp. 817–838. w35x R. Lee, J. Moore, CORELAP—computerized relationship
w15x U. Flemming, R. Coyne, S. Fenves, J. Garrett, R. Woodbury, layout planning, J. Ind. Eng. 18 Ž3. Ž1976. 195–200.
SEED—software environment to support the early phases in w36x J. Leung, A new graph-theoretic heuristic for facility layout,
building design, Proc. IKM94, Weimar, Germany, 1994, pp. Manage. Sci. 38 Ž4. Ž1992. 594–605.
5–10. w37x R. Liggett, The quadratic assignment problem: an analysis of
w16x L. Foulds, Techniques for facilities layout: deciding which applications and solution strategies, Environ. Plan. B 7 Ž1980.
pairs of activities should be adjacent, Management Sci. 29 141–162.
Ž12. Ž1983. 1414–1426. w38x R. Liggett, The quadratic assignment problem: an experimen-
w17x J. Gero, V. Kazakov, Space layout problems using evolved tal evaluation of solution strategies, Manage. Sci. 27 Ž1981.
design genes, Artificial Intelligence in Eng. 12 Ž3. Ž1998. 442–460.
163–176. w39x R. Liggett, Optimal spatial arrangement as a quadratic as-
w18x P. Gilmore, Optimal and suboptimal algorithms for the signment problem, in: J. Gero ŽEd.., Design Optimization,
quadratic assignment problem, J. Soc. Ind. Appl. Math. 10 Academic Press, 1985, pp. 1–40.
Ž2. Ž1962. 305–313. w40x R. Liggett, A designer-automated algorithm partnership, an
w19x M. Goetschalckx, An interactive layout heuristic based on interactive graphic approach to facility layout, in: Y. Kalay
hexagonal adjacency graphs, Eur. J. Operational Res. 63 ŽEd.., Evaluating and Predicting Design Performance, Wiley,
Ž1992. 304–321. 1992, pp. 101–124.
w20x J. Grason, An approach to computerized space planning w41x R. Liggett, W. Mitchell, Optimal space planning in practice,
using graph theory, Proceedings of the Design Automation Computer-Aided Des. 13 Ž5. Ž1981. 277–288.
Workshop, June 28–30, Atlantis City, NJ, IEEE, New York, w42x M. Los, The Koopmans–Beckmann problem: some computa-
1971, pp. 170–179. tional results, Universite de Montreal, Centre de Recherche
w21x G. Graves, A. Whinston, An algorithm for the quadratic sur les Transports, 1976.
assignment problem, Manage. Sci. 17 Ž3. Ž1970. 453–471. w43x A. Mahdavi, O. Akin, Y. Zhang, Formularization of concur-
w22x M. Hanan, P. Wolff, B. Agule, Some experimental results on rent performance requirements in building problem composi-
placement techniques, ACM Des. Automation Conf. Proc. 13 tion, Working Paper, School of Architecture, Carnegie Mel-
Ž1976. 214–224. lon University, 1998.
w23x M. Hassan, G. Hogg, D. Smith, SHAPE: a construction w44x B. Malakooti, A. Tsurushima, An expert system using priori-
algorithm for area placement evaluation, Int. J. Prod. Res. 24 ties for solving multiple-criteria facility layout problems, Int.
Ž5. Ž1986. 1283–1295. J. Prod. Res. 27 Ž5. Ž1989. 793–808.
w24x S. Heragu, A. Alfa, Experimental analysis of simulated an- w45x Z. Michalewicz, Genetic AlgorithmsqData Structuress
nealing based algorithms for the layout problem, Eur. J. Evolution Programs, Springer-Verlag, Berlin, 1992.
Operational Res. 57 Ž1992. 190–202. w46x B. Montreuil, A. Laforge, Dynamic layout design given a
w25x F. Hillier, Quantitative tools for plan layout analysis, J. Ind. scenario tree of probable futures, Eur. J. Operational Res. 63
Eng. 14 Ž1. Ž1963. 33–40. Ž1992. 271–286.
w26x F. Hillier, M. Conners, Quadratic assignment problem algo- w47x B. Montreuil, U. Venkatadri, H.D. Ratliff, Generating a
rithms and location of indivisible facilities, Manage. Sci. 13 layout from a design skeleton, IIE Trans., January, 1993.
Ž1. Ž1966. 42–57. w48x R. Muther, Systematic Layout Planning, Cahners Books,
w27x C. Huntley, D. Brown, A parallel heuristic for quadratic Boston, 1973.
assignment problems, Computers Ops. Res. 18 Ž3. Ž1991. w49x R. Muther, K. McPherson, Four approaches to computerized
275–289. layout planning, Ind. Eng., February 1970, 39–42.
w28x F. Jacobs, A layout planning system with multiple criteria w50x C. Nugent, T. Vollmann, J. Ruml, An experimental compari-
and a variable domain representation, Manage. Sci. 33 Ž8. son of techniques for the assignment of facilities to locations,
Ž1987. 1020–1034. Operations Res. 16 Ž1. Ž1968. 150–173.
w29x J. Jo, J. Gero, Space layout planning using an evolutionary w51x C. Pfefferkorn, The design problem solver: a system for
approach, Architectural Sci. Rev. 36 Ž1. Ž1995. 37–46. designing equipment or furniture layouts, in: C. Eastman
w30x S. Jojodia, I. Minis, G. Harhalakis, J. Proth, CLASS: Com- ŽEd.., Spatial Synthesis in Computer-Aided Building Design,
puterized Layout Solutions using Simulated annealing, Int. J. Wiley, New York, 1975, pp. 98–146.
Prod. Res. 30 Ž1. Ž1992. 95–108. w52x J. Seehof, W. Evans, Automated layout design program, J.
w31x B. Kaku, G. Thompson, T. Morton, A hybrid heuristic for the Ind. Eng. 18 Ž12. Ž1967. 690–695.
facilities layout problem, Computers Ops. Res. 18 Ž3. Ž1991. w53x R. Sharpe, B. Marksjo, Solution of the facilities layout
241–253. problem by simulated annealing, Comput. Environ. Urban
w32x B. Kaku, G. Thompson, I. Baybars, A heuristic method for Syst. 11 Ž4. Ž1986. 147–154.
the multi-story layout problem, Eur. J. Operational Res. 37 w54x K. Tam, Simulated annealing algorithm for allocating space
Ž1988. 384–397. to manufacturing cells, Int. J. Prod. Res. 30 Ž1991. 63–87.
w33x A. Kusiak, S. Heragu, The facility layout problem, Eur. J. w55x K. Tam, Genetic algorithms, function optimization and facil-
Operational Res. 29 Ž1987. 229–251. ity layout design, Eur. J. Operational Res. 63 Ž1992. 322–346.
 R.S. Liggettr Automation in Construction 9 (2000) 197–215 215

w56x D. Tate, A. Smith, A genetic approach to the quadratic w58x D. van Camp, M. Carter, A. Vannelli, A nonlinear optimiza-
assignment problem, Computers Ops. Res. 22 Ž1. Ž1995. tion approach for solving facility layout problems, Eur. J.
73–83. Operational Res. 57 Ž1991. 174–189.
w57x D. Tate, A. Smith, Unequal-area facility layout by genetic w59x T. Vollmann, C. Nugent, R. Zartler, A computerized model
search, IIE Trans. 27 Ž4. Ž1995. 465–473. for office layout, J. Ind. Eng. 19 Ž7. Ž1968. 321–329.