Advanced Engineering Informatics 26 (2012) 716–726

Contents lists available at SciVerse ScienceDirect

Advanced Engineering Informatics

journal homepage: www.elsevier.com/locate/aei

Comparative study of multiple criteria decision making methods for building design
Kristo Mela a,⇑, Teemu Tiainen b, Markku Heinisuo b
a
Tampere University of Technology, Department of Mechanics and Design, P.O. Box 589, FI-33101, Tampere, Finland
b
Tampere University of Technology, Department of Structural Engineering, P.O. Box 600, FI-33101, Tampere, Finland

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, multiple criteria decision making methods are studied in the context of building design. The
Received 30 September 2011 approach is to compare the functionality and the results provided by different methods on three test
Received in revised form 31 January 2012 problems that represent various design situations. The number of criteria in the test problems are two,
Accepted 2 March 2012
three and four. Multicriteria optimization is applied to generate the alternatives, among which a pre-
Available online 9 April 2012
ferred solution is to be searched by the decision making methods. Six methods have been selected for
comparison: the weighted sum method, the weighted product method, VIKOR, TOPSIS, PROMETHEE II,
Keywords:
and a procedure based on the PEG-theorem. The numerical study on the test problems indicate that in
Decision-making
Building design
most cases, the methods provide different solutions. The PEG-procedure tends to find a well-balanced
Multicriteria optimization solution, where none of the criteria is emphasized. While the ‘‘best’’ MCDM method is not discovered
in the study, information about the performance of the methods in building design problems is presented.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction disciplines involved in the design process in order to yield a compro-

1.1. Background
Construction management decisions typically involve several
conflicting aspects that need to be considered. These decision-mak-
ing situations can be formulated as multicriteria optimization prob-
lems, where the different aspects of a building project constitute the
conflicting criteria that are optimized simultaneously. It is widely
recognized that most of the total cost and the performance of the
building is determined by the decisions made in the conceptual de-
sign phase. Therefore, applying multicriteria optimization in this
early phase can lead to considerable savings in the building project,
see, for example, [1–3] for recent research and applications.
In the literature, a multitude of methods for computing Pareto
optimal designs that represent the compromise solutions between
the criteria, is available. As a result of multicriteria optimization, a
set of Pareto optimal solutions is obtained. However, usually a sin-
gle or few designs has to be chosen for further inspection and devel-
opment. In the context of building design, this decision-making task
can be very difficult, since due to the multidisciplinary nature of the
design problem, several non-commensurable criteria exist and of-
ten there are many Pareto optima to choose from. Multiple criteria
decision making (MCDM) methods provide a valuable tool for sup-
porting the choice of the preferred Pareto optimum. Together with
multicriteria optimization, the MCDM methods unite the different
⇑ Corresponding author. Tel.: +358 40 849 0563.
E-mail addresses: kristo.mela@tut.fi (K. Mela), teemu.tiainen@tut.fi (T. Tiainen),
markku.heinisuo@tut.fi (M. Heinisuo).
mise solution with a solid computational background.
Since the 1970s, literature on multiple criteria decision making
has been increasing tremendously [4]. Based on their properties,
problems and solution approaches can be categorized in various
ways (see, for example, [5–7]). For the purposes of this paper, we
call multicriteria optimization the process of generating the alter-
natives (Pareto optima), and multiple criteria decision making the
process of choosing a single preferred solution from among the
computed alternatives.
Interactive methods lie between these two categories. The prin-
ciple idea of these methods is that a single Pareto optimal solution
is computed and presented to the decision maker (DM), who either
accepts the given solution or guides the optimization process
based on her preferences and using the information included in
the current and previous Pareto optimal solutions. This process
forms an iterative loop, where new Pareto optima are generated,
until the DM is satisfied with the results. An interactive approach
might be preferable, if the DM is very active and if the computation
of a representative subset of Pareto optima is too expensive. A
number of interactive methods as well as a thorough discussion
about their philosophy and benefits can be found in [7].
The purpose of the present paper is to assess and compare the
results that MCDM methods give in building design problems.
The aim is to shed light on the usability of the different methods
and to study their fundamental differences from the point of view
of the decision maker. The approach chosen for carrying out the
comparison is to apply the methods to multicriteria optimization
problems that have been documented in the literature. This
approach imposes certain limitations on the conclusions that can

1474-0346/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.aei.2012.03.001
 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726
be drawn about the methods. Also, based on only few test prob-
lems, one method can hardly be stated as superior over the others.
However, as the test problems cover varying amount of criteria and
alternatives, the differences of the methods can be expected to
appear.
In the MCDM literature, the task of generating the alternatives
for the DM is often left for the designer, who may or may not be
the actual decision maker. The alternatives can be generated, for
example, by the designers involved in the design process, or by
computational optimization procedures. In this paper, we assume
that the building design problem has been formulated as a multi-
criteria optimization problem that has been solved by some appli-
cable method, which means that a finite number of Pareto optimal
(non-dominated) solutions has been obtained. The MCDM methods
operate on this set of alternatives, providing a ranking list of the
designs.
A selection of methods for comparison was chosen such that a
wide range of approaches would be covered. The considered meth-
ods are the weighted sum method (WSM), weighted product meth-
od (WPM), VIKOR, PROMETHEE, and TOPSIS. Additionally, a
method based on the PEG-theorem [8] is employed. This method
differs from the others in that it does not require any preference
information from the DM.
The paper is organized as follows. In Section 2, the basic con-
cepts of multicriteria optimization and MCDM relevant to the pa-
per are presented, along with a discussion on optimization
problems in building design. In Section 3, the methods to be com-
pared are briefly introduced. Then, in Sections 4–6, the methods
are applied to three building design case studies. The first two
cases are taken from the literature, and the third is formulated here
for the first time. Observations and discussion are presented in Sec-
tion 7, and final conclusions are given in Section 8.
1.2. Related work
Comparisons of MCDM methods have been presented in the lit-
erature. Zanakis et al. [6] compare eight methods, namely ELECTRE,
TOPSIS, weighted product method, weighted sum method and four
versions of analytical hierarchy process (AHP), using computer-gen-
erated data. The simulated problems have 5–20 criteria and 3–9
alternatives. The methods are compared against the weights and
ranks of the weighted sum approach by several measures. Salminen
et al. [9] compare the performance of ELECTRE II, PROMETHEE I, II
and SMART on application problems concerning environmental
decision-making situations. Again, the number of criteria is large,
and only few alternatives are available. The differences are analyzed
from the point of view of construction of criterion models, definition
of outranking relation and the ranking procedure. Opricovic and
Tzeng [10] present a comparison of VIKOR and TOPSIS methods,
which are based on distance measures. The focus of the comparison
is on the properties of the aggregating function as well as on the nor-
malization. Parkan and Wu [11] study the differences of the AHP,
data envelopment analysis (DEA) and operational competitiveness
rating (OCRA) methods by considering a process selection problem
related to electronics manufacturing.
A common outcome of the comparisons mentioned above is
that in many situations, the results of the different methods are
similar, but there are cases, where the results differ substantially.
Also, the authors of the comparisons avoid preferring one method
over another. This is understandable, since such a statement would
require either a solid theoretical rationale or comparison on a large
number of cases with actual decision makers, and typically neither
of these can be presented. Nevertheless, comparing MCDM meth-
ods on problems of different fields is still needed, since each field
has optimization problems with special characteristics that may
affect the choice of the MCDM method.
717
2. Multicriteria optimization in building design
Most design problems involve several conflicting aspects or cri-
teria that the designer tries to improve simultaneously. The final
result is a compromise, where human judgment and decision-mak-
ing is involved. Multicriteria optimization provides a valuable tool
for the designer (or decision-maker) to find the best compromise
solutions and to get quantitative information on the rate of conflict
of the criteria.
In general, a design problem can be formulated as a multicrite-
ria optimization problem as follows
min fðxÞ ¼ ff1 ðxÞ f 2 ðxÞ    f k ðxÞg ð1Þ
x2X
where f is the vector-valued objective function consisting of k crite-
ria, fi , that are mutually conflicting. The design variable vector, x,
must belong to the feasible set X that generally includes the con-
straints of the problem in form of inequalities or equalities:
X ¼ fx 2 Rn j g r ðxÞ 6 0; r 2 I; hq ðxÞ ¼ 0; q 2 Jg ð2Þ
where g r and hq are the inequality and equality constraint functions,
respectively, and I and J are the index sets containing as many ele-
ments as there are inequality and equality constraints, respectively.
The most fruitful definition for optimality in multicriteria prob-
lems has been proposed by Pareto [12], stating that a design is
optimal, if none of the criteria can be improved without deteriorat-
ing at least one criterion. To be more specific, a design x 2 X, is
Pareto optimal, if there does not exist another design x ^ 2 X such
^Þ 6 fi ðx Þ for all i ¼ 1; 2; . . . ; k and fj ðx
that fi ðx ^ Þ < fj ðx Þ for at least
one j ¼ 1; 2; . . . ; k (see, for example, [13]).
The literature on multicriteria optimization offers a multitude
of methods for computing Pareto optimal solutions. Theoretical re-
sults and classical methods are summarized in [7], and evolution-
ary solution techniques are presented in [14,15].
Two MCDM methods applied in this paper use the ideal solution
in their computations. The ideal solution is defined as the point in
the criterion space, whose components consist of the individual
ID
minima of the criteria. That is, let f denote the ideal solution.
Then, fiID ¼ minx2X f i ðxÞ. In the context of MCDM methods, the
ideal solution is taken from the set of computed alternatives in-
stead of actually computing k separate minimization problems.
For the results of the optimization to be practical, the formula-
tion of the optimization problem needs to include all essential fea-
tures of the design problem at hand. Common optimization criteria
in building design are various costs such as initial capital cost and
annual operating cost [1], and life cycle cost [2], energy consump-
tion [16–18] and recently environmental impact [2]. Most applica-
tions in the literature concern the early phase of the design, which
means that the design variables involve, for example, the choice of
materials and dimensions, number of storeys, wall types etc. The
constraints assure that the final design complies with building
codes and fulfils the functional requirements imposed by the
designer.
It is important to identify the type of optimization problem so
that an appropriate solution method can be chosen. First, the de-
sign variables are often both continuous (e.g. dimensions of the
building) and discrete (e.g. material type, number of columns), so
the problem is a mixed-integer problem. Second, most often there
are constraints or criteria that depend on the design variables non-
linearly. Therefore, most building design problems are multicriteria
mixed-integer nonlinear optimization problems, which are very diffi-
cult to solve in general. Most often these problems are solved by
genetic algorithms or other heuristic methods that overcome the
difficulties due to discrete variables.
In this paper, we assume that the number of computed Pareto
optima varies from tens to hundreds. On the other hand, the number
718 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726
of criteria is assumed to be pretty low, ranging from two to five. This
is in line with most building design problems that have been pre-
sented in the literature. It is worth noticing that in the comparison
studies cited in the previous section, the number of criteria is large,
ranging from 5 to 20, and the number of alternatives is low (e.g. 3–
30). Therefore, the conclusions presented in those studies may not
hold for the problems investigated in the present work.
3. Decision-making methods
In this section, the MCDM methods compared in this study are
briefly presented. More detailed descriptions can be found in the
references of each method.
Generally, the decision maker must provide her criteria prefer-
ences in a numerical form. Typical preference data are criteria
weights. Expressing the preferences numerically might be difficult,
especially if the number of criteria is large. On the other hand, the
data provided by the DM has a significant impact on the solution
the methods propose. If the DM hesitates in specifying the prefer-
ence data, several runs on different preference values and sensitiv-
ity analysis can be recommended.
Brans and Mareschal [19] propose seven requisites an MCDM
method should take into account. From the point of view of the
DM a key requisite is that the method is clearly understandable
and it does not use technical parameters that do not have clear
meaning for the method. The numerical aspects emphasize that
the results should not depend on the units in which the criteria
values are given, and that the magnitude of the difference of the
criteria between each pair of alternatives is taken into account.
3.1. Weighted sum method
Arguably the most widely applied approach, the weighted sum
method (WSM) is based on weights given by the decision maker.
Each criterion is given a non-negative weight, and the alternatives
are ranked by evaluating the weighted sum of the criteria [20].
It has been proposed, that WSM should be used as a standard
for evaluating MCDM methods [21]. The rationale for this is that
a multicriteria method should perform accurately also in single-
criteria problems, and the WSM gives the most acceptable results
in most single-criteria problems.
WSM suffers from the difficulty of choosing the proper weights
in the case where the dimensions of the criteria differ from each
other and their numerical values can be several orders of magni-
tude apart. This problem can be avoided, if the criteria are normal-
ized such that after normalization, each Pareto optimal solution
has criteria values from 0 to 1 for all criteria.
The weighted sum approach can also be used to generate Pareto
optimal solutions. Then, instead of calculating the weighted sum
for each alternative, an optimization problem is solved, and a de-
sign minimizing the weighted sum of the criteria is searched from
the feasible set.
3.2. Weighted product method
The weighted product method (WPM) creates a ranking of alter-
natives based on a multiplicative measure [22]. The DM specifies
weights for all criteria and each criterion value is raised to a power
corresponding the weight of that criterion. The ranking of the alter-
natives can be based on pairwise comparison or on the product
Q
k
w functions. Preference function values are multiplied by criteria
Pi ¼ fij j ð3Þ
j¼1
that is calculated for each alternative. In the above, fij is the value of
criterion j of alternative i. The method is often presented for
maximization problems, where the alternatives are ranked such
that the preferred solution has the greatest value Pi . In the maximi-
zation scheme, criteria to be minimized can be included by using
wj instead of wj . This is equivalent to using 1=fj instead of fj .
Eq. (3) can be applied without difficulty, if the criteria values are
positive. The impact of negative or zero criteria values must be in-
spected carefully, since they might distort the multiplicative rank-
ing measure. In general, negative values of the criteria should be
avoided, since exponentiating such a value by a rational exponent
leads to a complex number. If criterion fj to be minimized takes
only negative values, fj can be used in Eq. (3). Also, if fj < 0 is to
be maximized, then inserting 1=fj to Eq. (3) circumvents the
problem of negative criteria values. This is equivalent of minimiz-
ing fj , which is positive, since fj < 0, and then replacing wj by wj
as explained above. If criterion fj takes both negative and positive
values, then the above transformations do not fix the problem of
complex numbers appearing in the terms of Eq. (3). If for some
alternative fij ¼ 0, then the product of Eq. (3) equals zero regardless
of the values of the other criteria. This renders the ranking based
on Eq. (3) inapplicable.
We are not aware of procedures of handling the latter two is-
sues. However, in our test cases, the values of the criteria are al-
ways strictly positive or negative. Therefore, these issues
associated with the WPM described above do not appear in the de-
sign problems of this paper.
3.3. VIKOR
The VIKOR method [10] identifies a compromise design (or de-
signs) from among a set of alternatives by ranking them based on
their distance from the ideal solution. The distance is measured by
two instances of the weighted Lp -metric, namely p ¼ 1 and p ¼ 1.
The decision maker chooses the weights for the criteria according
to her preferences. In addition, the method uses a parameter v gi-
ven by the DM to control the balance between the two metrics. The
method uses linear normalization to eliminate the units of the cri-
teria functions.
A special feature of VIKOR is that it may give a set of preferred
alternatives instead of a single solution. This happens, if none of
the alternatives stands out, but there are several designs, that are
nearly as close to the ideal solution as the closest (in the sense of
L1 and L1 metrics).
Recently, VIKOR has been extended to a more interactive direc-
tion, where the DM can reset the weights using information pro-
vided by a trade-off analysis [23].
3.4. TOPSIS
TOPSIS [5] is another method based on distances from the ideal
solution. A design that is closest to the ideal point and the farthest
from the so-called negative-ideal point, whose components are the
worst criteria values appearing among the alternatives, is chosen
to be the best. The distances are measured by the Euclidean metric
and they are computed for normalized and weighted data, where
the weights are again determined by the DM.
3.5. PROMETHEE
PROMETHEE method family is based on pairwise comparisons
of the alternative designs. The comparisons are done by preference
functions. Brans and Vincke [24] present six main types of these
weights defined by the DM and summed to get aggregated prefer-
ence indices. Then positive and negative outranking flows are cal-
culated. PROMETHEE I gives a partial ranking by the outranking
flows and in PROMETHEE II the outranking flows are summed to
 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726 719

get net outranking flow which is used as measure to rank the design in the feasible set. This is especially true for discrete multi-
alternatives. criteria optimization problems, whose Pareto fronts consist of dis-
PROMETHEE has been applied extensively in various fields, as crete points. In this case, the solution given by the PEG-procedure
shown in a recent literature review by Behzadian et al. [25]. can only be used as reference data. Then, for example, a design
In this study, we apply PROMETHEE II on the test problems. As nearest the Pareto-compromise design can be chosen. We adopted
described above, the method requires a choice of a preference this approach in the test problems of the present study.
function. We use three preference functions of the six described It is interesting to observe that PEG is the only method that as-
in [19]. The first is called the ‘‘Type 1’’ preference function, or the sumes that the alternatives are Pareto optimal. The other methods
‘‘Usual criterion’’. The second is ‘‘Type 3’’ preference function, or can be employed to any set of alternatives. However, it might be
the ‘‘V-shape criterion’’. Finally, the third is ‘‘Type 4’’ preference hard to justify the choice of an alternative that is dominated by
function, or the ‘‘Level criterion’’. some other alternative. Therefore, the assumption included in the
Let dj ða; bÞ ¼ fj ðaÞ  fj ðbÞ, where fj is the criterion j. Then the PEG procedure about the Pareto optimality of the alternatives can-
Type 1 preference function is defined as not be regarded as severe restriction.

0 dj 6 0
Pj ðdj Þ ¼ ð4Þ 3.7. Parameter values used in the computations
1 dj > 0
This means that if fj ðaÞ > fj ðbÞ, then a is strictly preferred over b (in In the following sections, all the MCDM methods described
the maximization case) with respect to criterion fj . This definition above are applied to three test problems. All methods except the
justifies the name ‘‘usual criterion’’ for this preference function. PEG-procedure need criteria weights. In all the computations, we
Type 3 preference function is defined by have chosen to give equal weights to all criteria. That is, if the num-
8 ber of criteria is k, then the weight wi ¼ 1=k.
>
> 0 dj 6 0 For the VIKOR method, the parameter value v ¼ 1=2 was used.
<
dj
Pj ðdj Þ ¼ pj
0 6 dj 6 pj ð5Þ All other parameter values are given as needed in the test
>
> problems.
:
1 dj > pj

Now if 0 < dj < pj , then a is ‘‘partially preferred’’ over b with respect 4. Office building problem
to fj . Alternative a is strictly preferred over b only when its criterion
value is the amount pj larger. This provides flexibility in the prefer- In a study by Wang et al. [2] a single-storey office building with
ence relation. 1000 m2 floor area located in Canada was considered. The two cri-
Finally, the Type 4 preference function is defined by teria are the life-cycle cost (LCC) and life-cycle environmental im-
8 pact (LCEI) of the building.
<0
> dj 6 qj
Considering only two criteria reduces the computational load in
1
Pj ðdj Þ ¼ qj < dj 6 pj ð6Þ generating the alternatives. On the other hand, less numerical
>2
:
1 dj > pj information about the performance of the building is provided,
and to take into account other aspects not present in the problem
This time two parameters, qj and pj are needed. The first indicates,
formulation requires careful human judgment. As there are only
that there must be a difference in the criteria value of at least qj be-
two criteria, an illustrative graphical presentation of the Pareto op-
fore one alternative is preferred over another, and one is strictly
tima in the criterion space is possible. The DM is then provided
preferred only when the difference is at least pj . Between those val-
with a clear picture of the rate of conflict between the criteria
ues, an equal ‘‘partial preference’’ is given.
and the location of the alternatives with respect to each other. In
In the above, it is assumed that all criteria are maximized. In the
this case, the MCDM methods may not be needed at all, if the
minimization case, fj can be replaced by fj , and the method can be
DM is able to make the decision based on the graphical presenta-
used with the preference functions defined above.
tion. If the DM hesitates, then the MCDM methods may help to
make the choice, for example, from among a cluster of designs that
3.6. PEG-procedure the DM is considering.
The life-cycle cost includes the building cost and the operating
Grierson [8] has approached the decision-making problem from cost for a 40-year life cycle. Life-cycle environmental impact is
the point of view of trade-off analysis. Supposing that each crite- measured as expanded cumulative exergy, which is the sum of
rion is advocated by a separate designer, the idea is that the
designers bargain to find a mutually acceptable design. It can be
Table 1
shown, that such a design exists and is unique. At this point the Pareto optimal solutions of the office building problem.
terms-of-trade lines of all designers coincide. This result is called
the PEG-theorem (PEG = Pareto–Edgeworth–Grierson), and a Design LCC (105 $) LCEI (107 MJ) Design LCC (105 $) LCEI (107 MJ)

numerical procedure for finding the unique Pareto-compromise de- 1 3.3520 4.5980 16 3.4610 4.0010
sign has been proposed in [8]. 2 3.3570 4.5950 17 3.4660 4.0000
3 3.3610 4.4910 18 3.4710 3.9990
Instead of a ranking list of the alternatives, the PEG-procedure
4 3.3660 4.4900 19 3.4850 3.9750
computes the Pareto-compromise design using the available Pare- 5 3.3680 4.4090 20 3.6220 3.9300
to optimal points. The method does not require any preference 6 3.3720 4.4010 21 3.6260 3.9240
information from the decision-maker. If the existence of terms- 7 3.3790 4.3290 22 3.6340 3.9210
of-trade lines and the concept of mutually acceptable design is ap- 8 3.3840 4.3220 23 3.6530 3.8970
9 3.3890 4.2360 24 3.6580 3.8570
proved, the solution provided by the PEG-procedure can be re- 10 3.3940 4.2320 25 3.6670 3.8510
garded as a balance between trade-offs, and as such, can be a 11 3.4360 4.1940 26 3.6870 3.8280
very appealing compromise. 12 3.4420 4.1840 27 3.6930 3.8210
For the Pareto-compromise design to lie on the Pareto front, it is 13 3.4460 4.0880 28 3.6980 3.8200
14 3.4510 4.0850 29 3.7040 3.8190
necessary that the front is continuous. Otherwise, a point in the cri-
15 3.4540 4.0100
terion space might be obtained that does not have a corresponding
720 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726

Table 3
Indices of the removed designs for the office building problem.

50% removed, 14 designs

3 4 5 7 8 12 14
19 21 23 25 26 27 29
10% removed, 2 designs
14 29

The PROMETHEE II method showed peculiar behaviour. When

the Type 1 preference function was employed, the method failed
to give a single preferred solution. This is due to the fact that all
alternatives are non-dominated and there are only two criteria
with equal weights. In this case, every alternative is better in one
criterion and worse in the other than the other alternatives. With
the Type 1 preference function, this leads to a situation, where
the positive outranking flow is equal to the negative outranking
flow for all alternatives, and none of them becomes the preferred
solution.
We then switched to Type 3 preference function with the
threshold of strict preference value p ¼ 0:01. In this case, the meth-
od gives the same solution as WPM. We also tried Type 4 prefer-
Fig. 1. Pareto front and compromise solutions of the once building problem.
ence function with the parameter values p ¼ 0:3 and q ¼ 0:1.
This time the solution coincided with the one given by the WSM.
It has been observed in the literature, that the number of alter-
Table 2
natives influences the results of the MCDM methods [6]. To study
Office building problem solutions of the MCDM methods. The labels of the columns this phenomenon, the methods were applied twice more on two
indicate the percentage of Pareto optima of the original data that were used in the subsets of the Pareto data of Table 1. In the two cases, 50% and
calculations. The numbers correspond to the labels of the Pareto optima in Table 1. 10% of the Pareto optima were randomly removed, respectively.
Method 100% 50% 90% The indices of the removed points are given in Table 3. The solu-
tions of the MCDM methods are given in Table 2 in the 3rd and
WSM 15 15 15
WPM 16 16 16 4th columns, labeled 50% and 90%, respectively. It can be seen that
TOPSIS 19 16 19 for this test problem, the methods are not very sensitive to the
VIKOR 15a 15a 15a number of alternatives. Only the solution of TOPSIS changes, when
PROMETHEE II
the number of Pareto optima is reduced to half of the original
Type 1 – – –
Type 3 16 16 16
amount. This is due to the fact, that the preferred solution of TOP-
Type 4 15 15 15 SIS was among the removed Pareto optima.
PEG 15 15 15
a
Indicates that more than one solution was proposed. 5. Design of a hall

exergy consumption due to resource inputs and abatement exergy
required to recover the negative impacts due to waste emissions.
The variables are both discrete and continuous. They include
the building orientation, aspect ratio, the ratio between the win-
dow area and the total area for a given facade, window type, wall
structure type, and wall and roof insulation thicknesses and type.
By applying a multiobjective genetic algorithm, 29 non-domi-
nated solutions shown in Table 1 are presented to the decision ma-
ker (for details, see Wang et al. [2]). The MCDM methods described
in the previous section were applied to this data. In Fig. 1, the non-
dominated solutions are plotted along with the preferred designs
given by the MCDM methods. The results are also given in Table
2, where the second column shows for each method the preferred
design number corresponding to the data in Table 1.
It can be seen that the methods give nearly identical results. All
of them manage to reflect the equal importance of the criteria set
by the weighting, by providing compromise solutions where nei-
ther criterion is favored. The compromise solution of TOPSIS devi-
ates from the rest, but is still in the same group of designs. The
Pareto-compromise design zPEG ¼ f3:4539 4:0444g produced by
the PEG-procedure lies in between the actual Pareto optima. The
solution that is closest to this point was taken as the compromise
design. This coincides with the solution of WSM and VIKOR.
The VIKOR method proposed a more than one solution. The first
alternative of this set of solutions was chosen as the preferred
solution. This is indicated in Table 2 by the superscript a.
In this section, we consider the multicriteria optimization of a
hall building as presented in [26,27]. The building is optimized
for costs, resource efficiency and the aesthetic preference of the de-
signer. The criteria are evaluated using utility functions that take
values ranging from 0 to 1. The utility functions have been con-
structed such that the lower the cost and resources criteria values
are, the closer the utility functions get to 1. Therefore, the minimi-
zation of costs and resources actually corresponds to maximization
of the utility functions.
With three criteria, the need for MCDM methods becomes more
evident. Even though the Pareto optima can be plotted in the crite-
rion space, it can be very difficult to interpret the three-dimen-
sional figure and to compare the alternatives. Now each Pareto
optimum provides the DM with more computed information about
the design than in the bicriteria case. On the other hand, it is more
difficult to compare the alternatives. It might be especially difficult
to compare alternatives with respect to the aesthetic preference
function, since it is not clear, how a small difference in the value
representing the aesthetics of the building should be interpreted.
This difficulty is present, when qualitative design aspects are re-
duced to mathematical functions. When such functions are in-
cluded as optimization criteria, the results of the MCDM methods
should be viewed with a critical eye.
The cost function is the average cost per square metre over the
lifespan of the building. The resource function includes the amount
of energy from non-renewable and renewable sources as well as
 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726 721

the emission of gases with global warming potential, each term gi- Table 5
ven per square metre per year. A weighted sum of the three terms Hall design problem solutions of the MCDM methods.

is computed and a utility function is employed to give a value for Method 100% 50% 90%
resource efficiency. Finally, the designer preference function drives WSM 4 14 4
the solution towards the ideal building as defined by the designer. WPM 4 14 4
In the current problem, the designer has considered six aesthetic TOPSIS 4 5 4
aspects. For example, the designer has specified ideal values for VIKOR 9 9a 9
PROMETHEE II 39 48 46
the slenderness of the load-bearing structures and for the distance PEG 14 9 9
between frames. For more details, see [26,27].
a
The design variables are the number of frames and horizontal Indicates that more than one solution was proposed.

load-bearing structures, the type of horizontal structures (truss

or beam), material of the structures (horizontal and vertical, steel
Table 6
or wood), the U-value for the facade and the type of ground sup-
Indices of the removed designs for the hall design problem.
port (clamped or jointed).
The optimization has been performed using a multiobjective ge- 50% removed, 31 designs
2 4 7 8 11 13 15 18
netic algorithm. Detailed analysis and discussion can be found in
25 26 27 28 29 33 35 37
[26], with 82 Pareto optimal solutions reported (Appendix C, pp. 39 40 41 42 44 45 46 47
241). Our inspection revealed that for the numerical accuracy pro- 51 54 57 58 59 60 62
vided, 20 of the solutions were actually dominated, if only slightly. 10% removed, 6 designs
We then removed these solutions, and applied the MCDM methods 20 24 29 40 45 52
for the remaining 62 designs. The criteria values are presented in
Table 4.
The results of the MCDM methods are shown in Table 5, similar and PEG give a more balanced solution, where none of the criteria
to the office building case in the previous section. Again, 50% and is clearly preferred. Finally, PROMETHEE II (with a Type 1 prefer-
10% of the alternatives were randomly removed to investigate ence function) finds a solution, where two criteria nearly achieve
the influence of the number of criteria. The indices of the removed their maximum value at the expense of the remaining criterion. In
designs are given in Table 6. this case, criterion 1 performs quite poorly compared with the other
It can be seen that in this case there is more variation in the re- two.
sults than in the previous test problem. When the complete data is When 10% of the Pareto optima are removed, it can be seen that
involved, WSM, WPM and TOPSIS give equal results. The other the solution provided by PEG changes, even though the original
methods produce different solutions. solution remains in the set of alternatives. This time PEG gives
To examine the results more closely, we computed the follow- the same result as VIKOR (still a well-balanced alternative),
ing ratio for each criterion and all alternatives: whereas WSM, WPM, and TOPSIS provide the same solution as
before. PROMETHEE drives the solution even closer to the point
fji  minffji g
i where both criteria 2 and 3 are maximized. This is due to the pref-
r ji ¼ ð7Þ
maxffji g  minffji g erence function that was used, namely the Type 1 function. Other
i i
preference functions give more balanced solutions, where the
where fji is the value of criterion fj for alternative i, and max and min weights are reflected better.
are taken over the alternatives. The closer r ji is to 1, the closer the When half of the Pareto optima are removed, the original solu-
alternative i is to the best value of criteria j. The results are tions of WSM, WPM, TOPSIS and PROMETHEE II are among the re-
r4 ¼ f0:97;0:65;0:82g;r9 ¼ f0:78;0:77;0:76g;r14 ¼ f0:70;0:79; 0:83g, moved alternatives. In this case, the solution of TOPSIS differs from
and r39 ¼ f0:31;0:93;0:95g. Thus, it can be seen that WSM, WPM WSM and WPM, which, again produce equal preferred solutions.
and TOPSIS give a solution that actually prefers criterion 1. VIKOR The solution of the PEG changes again from the original point,

Table 4
Pareto optimal solutions of the hall design problem.

Design f1 f2 f3 Design f1 f2 f3 Design f1 f2 f3

1 0.9330 0.7270 0.6840 22 0.6730 0.9040 0.8270 43 0.5740 0.9360 0.8340
2 0.9330 0.7320 0.5860 23 0.6480 0.8650 0.8570 44 0.5740 0.9470 0.7850
3 0.9210 0.7360 0.3590 24 0.6230 0.9120 0.8260 45 0.5720 0.9330 0.8600
4 0.9190 0.8720 0.7890 25 0.6230 0.9130 0.7710 46 0.5710 0.9360 0.8620
5 0.8800 0.8870 0.7300 26 0.6120 0.9200 0.7420 47 0.5710 0.9340 0.8630
6 0.8700 0.9070 0.5980 27 0.6090 0.9130 0.8200 48 0.5700 0.9390 0.8310
7 0.8310 0.8770 0.7510 28 0.6080 0.9160 0.7640 49 0.5670 0.9330 0.8640
8 0.8230 0.8800 0.7680 29 0.6070 0.9260 0.7360 50 0.5660 0.9470 0.7870
9 0.8190 0.8980 0.7590 30 0.6060 0.9270 0.6830 51 0.5630 0.9220 0.8660
10 0.7960 0.8760 0.7740 31 0.6010 0.9350 0.6890 52 0.5620 0.9470 0.7990
11 0.7900 0.8740 0.8170 32 0.6010 0.9340 0.8220 53 0.5610 0.9350 0.8640
12 0.7870 0.8970 0.7870 33 0.5950 0.9310 0.8440 54 0.5560 0.9330 0.8750
13 0.7810 0.8850 0.8120 34 0.5870 0.9470 0.7770 55 0.5530 0.9340 0.8710
14 0.7810 0.9030 0.7960 35 0.5870 0.9380 0.7860 56 0.5410 0.9380 0.8400
15 0.7780 0.9060 0.7620 36 0.5860 0.9450 0.8160 57 0.5340 0.9460 0.8190
16 0.7770 0.8730 0.8200 37 0.5800 0.9200 0.8470 58 0.5030 0.9290 0.8830
17 0.7750 0.9030 0.7980 38 0.5780 0.9310 0.8560 59 0.4750 0.9300 0.8800
18 0.7740 0.9060 0.8010 39 0.5770 0.9340 0.8580 60 0.4660 0.9480 0.7810
19 0.7340 0.9290 0.6030 40 0.5760 0.9380 0.8270 61 0.4310 0.9490 0.7760
20 0.7200 0.9070 0.7720 41 0.5750 0.9390 0.8210 62 0.4200 0.9480 0.7860
21 0.6890 0.9120 0.6540 42 0.5750 0.9460 0.8120
722 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726
Fig. 2. Single family residential house.
Table 7
Continuous variables.
Variable Lower bound Upper bound
Length L 10 15
Width B 7 10
Height H 2.4 3.5
being the same as in the case where 10% of the alternaties were re-
moved. The preferred solution of PROMETHEE II is still in the re-
gion, where two of the three criteria are close to their maximum
values. Thus, we can conclude that for the PROMETHEE II method
to better respond to the preference of the DM, the associated pref-
erence function needs to be selected carefully.
6. Single family residential house
In this section the multicriteria optimization of a single-family
house as shown in Fig. 2 is considered. The task is to minimize
the construction work cost, construction material cost, thermal
transmittance of the building, and to maximize a customer prefer-
ence function. The two costs are held apart such that the DM can
study the amount of work and material costs directly by looking
at the obtained Pareto optima. The thermal transmittance is re-
lated to the need for heating the building which is directly con-
nected to the annual energy consumption in regions of cold
climate such as Finland. Finally, the customer preference function
transfers the qualitative aspect of the building related to the sup-
posed comfortability of the customer to a mathematical function,
thus being similar to the third criterion of the hall design problem
of the previous section.
The problem has both continuous and discrete design variables.
The dimensions B, L, and H are the continuous variables, and their
bounds are given in Table 7. The remaining variables are discrete,
and they are related to the walls, roof, and windows. For wall types
there are two discrete choices as depicted in Fig. 3. The construc-
Fig. 3. Two wall types.
tion work cost and construction material cost are taken from
[28] and linear interpolation is employed to obtain values for the
different properties when the insulation thickness is varied. The
layer stretched when altering the insulation thickness is number
4 in type I wall and 5 in type II wall. Wall and roof insulation thick-
nesses are both chosen from eight possible values. Corresponding
costs and U-values are given in Table 8. Window size is chosen
from the seven values given in Table 9. Finally, the number of
windows is also a discrete variable. The house has two doors, with
material cost 250€, work cost 36:18€, and area 2 m each. The
U-value of both windows and doors is 1:4 W=mK.
The constraints for optimization are that the window area Aw is
at least 10% of the living area Al and that the window area does not
exceed the wall area Awalls .
Then the optimization problem is
min fðxÞ ¼ ff1 ðxÞ f 2 ðxÞ f 3 ðxÞ  f4 ðxÞg
x2X
such that Aw ðxÞ P 0:1Al ðxÞ ð8Þ
Aw ðxÞ 6 Awalls
where f1 is the construction work cost (€), f2 is the construction
material cost (€), f3 is the thermal transmittance (W/K), and f4 is
the customer preference. Since the problem is written as a minimi-
zation problem, f4 must be equipped with the minus sign. The set X
includes the bounds for the continuous variables and possible val-
ues for the discrete variables.
Construction work cost is calculated by multiplying the building
part areas with respective costs per square metre. For windows
and doors the costs are for piece, so the unit cost is multiplied by
the number of items. Total work cost is then calculated by
P P
f1 ðxÞ ¼ Ai ðxÞC i;w ðxÞ þ nj ðxÞC j;w ðxÞ ð9Þ
i j
where Ai is the area of building part i and C i;w is the work cost per
square metre for respective part, nj is the number of item j and
C j;w is the work cost for one item.
 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726 723

Construction material cost is calculated similarly but indices w The thermal transmittance is calculated by taking all the areas
referring to work is replaced by m, so of the parts of the envelope and multiplying with their respective
P P thermal transmittance coefficients
f2 ðxÞ ¼ Ai ðxÞC i;m ðxÞ þ nj ðxÞC j;m ðxÞ ð10Þ
i j P
f3 ðxÞ ¼ Ai ðxÞU i ðxÞ ð11Þ
i

Table 8
Wall and roof data. where Ai is area of a part i of envelope and U i is the thermal trans-
mittance coefficient of respective part. Thermal transmittance coef-
Type Insulation U-value (W/ Material cost Work cost
ficients are calculated according to [29,30].
thickness (mm) m2 K) ð€=m2 Þ ð€=m2 Þ
The customer preference criterion depends on the aesthetic
I 148 0.28 53.78 39.14 taste of each customer, and such a function has to be constructed
I 173 0.25 56.38 42.55
I 198 0.22 58.97 45.96
separately for each instance. Also, sometimes it is impossible to ex-
I 223 0.2 61.57 49.37 press the customer preference as a mathematical function. In this
I 248 0.18 64.17 52.78 case it was decided that the criterion favours large living area, high
I 298 0.16 69.36 59.60 rooms and large windows. This leads to the following expression
I 348 0.14 74.55 66.42
I 398 0.12 79.74 73.24 f4 ðxÞ ¼ wA Al ðxÞ þ wH HðxÞ þ ww Aw ðxÞ ð12Þ
II 96 0.33 50.65 54.54
II 146 0.25 54.74 58.38 where wi are weights for different measures and Al is the living area
II 171 0.22 56.78 60.29
of the building defined by
II 196 0.2 58.82 62.21
II 246 0.17 62.91 66.05
Al ¼ ðB  2tÞðL  2tÞ ð13Þ
II 296 0.15 66.99 69.88
II 346 0.13 71.08 73.715
where t is thickness of the wall. Weights chosen now are
II 396 0.11 75.16 77.55
Roof 350 0.12 77.09 42.98 wA ¼ 1 m2 ; wH ¼ 5 m1 and ww ¼ 5 m2 . Thus, the function f4 is
Roof 400 0.11 78.85 42.98 dimensionless.
Roof 450 0.10 80.61 42.98 The optimization was performed using a multiobjective genetic
Roof 500 0.09 82.37 42.98 algorithm. As a result, 116 non-dominated solutions in the four-
Roof 550 0.08 84.13 42.98
Roof 600 0.07 85.89 42.98
dimensional criterion space were obtained, see Table 12. Again,
Roof 650 0.07 87.65 42.98 the MCDM methods were applied in order to determine the
Roof 700 0.06 89.41 42.98 preferred solution.
The results of the MCDM methods are given in Table 10. Also for
this case, 10% and 50% of the alternatives were removed to study
the effect of the number of alternatives. The indices of the removed
Table 9 alternatives are given in Table 13.
Window alternatives. Ratios similar to Eq. (7) were computed for the preferred solu-
tions to study how they emphasize the different criteria. This time,
Width (m) Height (m) Work cost (€/m2) Material cost (€/m2)
the closer the ratio is to 0, the closer the criteria is to its most de-
0.3 0.6 23.91 161.3
sired value. The results are r2 ¼ f0; 0; 0:15; 0:94g; r32 ¼ f0:46; 0:43;
0.6 0.6 23.91 161.53
0.6 1.2 23.91 212.53 0:40; 0:40g; r38 ¼ f0:46; 0:45; 0:41; 0:43g; r81 ¼ f0:22; 0:20; 0:22; 0:63g
1.2 1.2 23.61 282.68 and r126 ¼ f0:32; 0:42; 0:47; 0:33g. Thus, it can be seen that the PEG-
1.5 1.4 33.36 363.72 procedure (r32 and r38 ) and VIKOR (r126 ) give balanced solutions.
1.5 1.6 33.36 397.89
The preferred solution of TOPSIS (r81 ) emphasizes the criteria f1 ,
2.1 1.4 38.24 442.06
f2 and f3 , but the solution is still fairly balanced, in contrast to
the solution provided by WSM, WPM and PROMETHEE (r2 ) which
is less balanced, since all but one criteria are close to their mini-
Table 10
mum (f1 and f2 actually achieve their minimum value). When alter-
Family house design problem solutions of the MCDM methods. natives are removed, only the result of PEG changes slightly. In
both cases the solutions chosen by the methods were also included
Method 100% 50% 90%
in the reduced set of alternatives.
WSM 2 2 2 The solutions provided by the methods differ from each other
WPM 2 2 2
TOPSIS 81 81 81
substantially. The reason for this can be found by looking at the
VIKOR 126a 126a 126a problem formulation and the non-dominated solutions. The con-
PROMETHEE II 2 2 2 flict between the cost criteria (f1 and f2 ) is very mild. This has a sig-
PEG 38 32 38 nificant impact on the performance of the methods. Small
a
Indicates that more than one solution was proposed. dimensions B and L lead to small wall and roof areas and thus small
values of costs f1 and f2 . Also small houses lead to relatively small

Table 11
Summary of the MCDM methods used.

Method Weights Parameters Comments

WSM Yes – Normalization of criteria recommended
WPM Yes – Zero or negative criteria values need special attention
TOPSIS Yes –
VIKOR Yes v Might produce several preferred designs
PROMETHEE II yes Type of preference function, q; p; s Choice of preference function is important
PEG No – The solution may be unattainable
724 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726

thermal transmittance (value of f3 ) even if the insulation thickness
is low. On the other hand, the customer preference criterion of
such designs tends to be far from its maximum value. Thus, f4 is
in greater conflict with f1 and f2 than f3 . The WPM and WSM
methods favour solutions, where three of the four criteria have
low values. The theoretical basis of PROMETHEE II explains why
it behaves similarly, although the method is completely different
from WPM and WSM. The pairwise comparison favours solutions
performing well in more than one criteria and that is the case with
designs having small B and L, where f1 and f2 (also f3 ) are small.

Table 12
Pareto optimal solution of the single-family house design problem.

Design f1 f2 f3 f4 Design f1 f2 f3 f4
1 24197.5 49716.9 119.358 282.564 68 22525.8 46316.9 97.3287 241.868
2 12231.1 27215.6 57.6029 111.106 69 15806.5 30113.4 44.9681 105.315
3 19684.7 45570.7 104.083 242.808 70 12354 28074.5 54.7804 107.704
4 21918.4 42682.2 68.5682 181.34 71 22718.6 45150.9 89.068 239.917
5 25625.3 52375.2 123.189 289.381 72 30885.5 59705.3 101.159 276.457
6 19893.7 43820.7 88.8027 210.38 73 15985.7 31837.3 57.1165 135.086
7 23079 50341.7 99.7443 248.866 74 14098.4 29252.8 61.2021 139.46
8 17939.6 38485 76.0123 179.812 75 12335.3 28230.7 55.5395 111.304
9 20740 40123.4 64.9475 157.854 76 27818.3 56836 122.055 298.34
10 16067.3 30081.1 43.9593 99.205 77 13263.8 30108.4 63.7796 142.793
11 27645.3 56265.8 129.417 301.454 78 17304 33188.8 48.1363 114.553
12 15292.8 32510.7 55.0963 118.422 79 25199.4 49436.1 87.8209 229.243
13 24711.8 49028.8 83.9768 226.783 80 21024.1 43390.7 68.7838 176.996
14 18086.7 34751.9 58.3197 140.058 81 16547.5 33944.7 65.2114 173.193
15 19381.6 37154.9 63.1813 152.367 82 16411.1 31285.5 46.1901 108.384
16 18469.3 38765.6 74.657 187.283 83 23969.2 50922.6 110.315 283.121
17 27347.4 55553 128.668 299.056 84 28854.6 58177 96.1298 266.83
18 27710.7 56231.5 129.815 301.138 85 22463.6 46033.8 81.3335 220.12
19 12668.8 27633.8 55.9927 110.791 86 26769.6 54623 119.281 291.397
20 13510.9 30328.4 67.0599 150.595 87 27094.6 52730.2 93.5357 256.939
21 30437 59473.7 104.368 289.721 88 16483.6 36756.7 69.4545 161.535
22 29869.2 57944.7 95.7013 259.235 89 14748.6 30447 60.4446 141.858
23 17021.1 32503.8 55.9922 135.112 90 22949.7 47219 108.993 269.874
24 23143.6 47813 114.371 272.179 91 14002.5 31507.7 56.9397 118.93
25 20543.5 39285.4 60.676 144.93 92 15686.5 35016 69.8576 153.702
26 27408 56039.1 127.361 297.782 93 13713.6 30856.2 65.6158 144.833
27 18009.3 36913.4 88.7494 202.578 94 24618.7 49827.1 114.148 276.006
28 29762.5 58250.3 103.019 286.343 95 23701.6 48493.6 111.135 272.12
29 17500.7 32137.5 50.6941 112.219 96 29846.7 60747 102.3 288.526
30 23487.5 46914.2 88.951 226.265 97 15917.1 35927.1 70.9066 158.095
31 25909 53938.4 139.552 295.965 98 31135.9 60153.1 102.669 288.689
32 21195 42117.8 82.9863 220.127 99 13893.7 30991.4 65.7612 153.315
33 14977.6 30103.6 67.2055 150.758 100 28135.7 57268 123.11 299.418
34 15526.9 30855.2 65.6004 150.357 101 31754.4 60332.9 94.6676 260.125
35 22493.7 45988.6 103.211 250.956 102 22918.6 46247.7 105.331 252.34
36 22219.3 44238.6 81.9587 205.229 103 29781.5 56598.9 85.4799 229.273
37 24088.4 50344.2 132.892 284.024 104 29121.9 58004.2 107.327 292.141
38 21115.3 42521 83.2377 213.185 105 13910.2 30340.2 59.9782 119.354
39 22374.5 46202 107.697 253.691 106 14253.8 29419.2 51.5872 115.137
40 25448.8 52992.3 137.963 292.84 107 12924.5 27389.8 52.822 104.555
41 16884.4 31771.7 52.6977 122.6 108 14733.3 27993.5 46.5845 101.274
42 17593.2 32478.9 50.899 113.203 109 18244.8 37147.3 80.118 192.997
43 18493.9 37272.6 85.9473 200.86 110 15246.2 29979 57.8189 130.089
44 18713.3 35374.7 51.7886 125.126 111 28641.8 55797.4 97.6005 267.474
45 23852.9 49156.3 107.66 264.487 112 17634.3 35256.6 63.7487 153.438
46 25907.1 53544.1 140.832 297.625 113 12910.4 27408.5 52.7256 104.676
47 23357.8 45341.9 81.2699 207.928 114 19169.2 38965.6 84.6537 207.464
48 28739.7 56532.5 103.412 275.483 115 24417.7 47713.8 83.1482 222.988
49 24194.7 47811.3 86.5024 226.54 116 27690.1 56335.8 117.575 283.446
50 16761.5 32710.2 55.6382 139.575 117 20413.1 40002.8 74.7216 192.618
51 28985.3 56595.1 102.263 282.938 118 12744.9 28096.5 54.3195 110.671
52 27027.6 54858.8 127.776 296.79 119 15439.5 29224.5 43.4138 98.0063
53 20131.5 36693.8 61.007 154.25 120 14032.1 27684.4 56.9718 116.597
54 27212.7 55068.9 122.108 295.298 121 31314 59430.7 94.7645 259.053
55 15232.3 35093 83.2591 188.504 122 18104 36047.9 69.1124 179.593
56 13684.8 29252.3 51.1407 110.232 123 23567.4 54504 129.792 283.589
57 17248.8 35753.3 95.8774 195.251 124 29449 58044.8 107.106 289.695
58 28064.4 55567.1 107.631 276.689 125 29241.8 56793.1 98.7788 270.016
59 19951.1 38661.3 58.0975 143.059 126 18520 41573.5 89.5969 234.642
60 16138.3 31469.6 52.7525 130.54 127 27657.5 56054.9 122.564 296.867
61 19851.1 39806.2 77.5963 204.145 128 22403.7 47048.4 89.4658 229.956
62 24655.4 46775.4 76.644 195.834 129 22645.1 45999.5 100.028 250.793
63 29760.6 57352 100.758 272.595 130 30607.7 61530 103.533 289.368
64 25986.4 53627.4 141.297 297.839 131 21652.1 42070.4 74.6411 190.514
65 30798.9 59169.2 101.83 285.438 132 13444.5 28806.9 50.4399 109.006
66 15736.3 30197 46.0862 107.475 133 13856.6 30415.3 59.9406 123.04
67 14474.8 30062.7 48.9296 109.651 134 26122.8 54708.1 139.623 300.584
 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726 725

Table 13 regarded as being advocated by different designers, and a solution

Indices of the removed designs for the single-family house. that is mutually agreeable is sought, then prescribing unequal
50% removed, 67 designs weights or other preference data should somehow be justified.
4 5 6 7 9 11 14 15 16 17 Based on our test problems, we can observe that as the number
19 20 21 22 23 24 26 27 31 35 of criteria increases, the results of the methods vary more. In the
36 37 41 43 44 45 47 49 53 54
56 57 59 60 61 62 69 71 72 74
bicriteria case, all methods gave nearly identical preferred solu-
76 80 82 83 89 91 92 93 96 97 tions. We argue that the need for MCDM methods for problems
99 101 102 103 104 105 106 108 112 115 with two criteria is rather limited, since the alternatives can be
119 120 121 124 131 132 134 plotted in the criterion space and this visualization displays very
10% removed, 11 designs clearly the conflict and trade-offs between the criteria. However,
5 40 52 62 63 70 72 87 95 122 as the number of criteria increases, the need for MCDM methods
127 133 134
is evident.
Our study motivates us to propose the following procedure for
the application of MCDM methods in building design problems.
The PEG procedure gives a well-balanced solution as in the First, the alternatives should be non-dominated (Pareto optimal).
problems of the two previous sections. As the ranking of VIKOR For bicriteria problems, plotting the alternatives in the criterion
and TOPSIS is based on a combination of distance measures rather space might give enough information for the DM to reach a deci-
than a single value as in WPM and WSM, these methods are able to sion. Otherwise, and especially if there are more than two criteria,
give more balanced solutions. the PEG procedure is carried out, even if other MCDM methods are
employed. This will give a fairly balanced solution without any
7. Discussion preference data and it can serve as a starting point for the DM. If
the PEG solution is not satisfying for the DM or if she wishes to fur-
Based on the three case studies presented in the previous sec- ther explore the alternatives, an MCDM method is employed based
tions, some general remarks about the decision-making process on the preference data the DM is able to provide and how well she
in building design can be made. understands the idea of the method. It is to be expected that the
In order to apply the MCDM in a reliable way, the characteris- DM wants to try the method with different values of the preference
tics of the optimization problem that forms the basis for generating data. In this case, several preferred solutions are provided. Finally,
the alternatives should be carefully studied. More specifically, the the DM must make a choice from among these solutions. We argue
degree of conflict of the criteria can have a significant impact on that in most cases a human decision maker does not want the com-
the behaviour of the MCDM methods. This was demonstrated by puter to provide a definite solution but rather wants to have con-
the single-family house case in Section 6. Also, it is important to trol over the process. In these situations, it is better to allow the
understand the theoretical basis of the MCDM methods to avoid DM to explore the alternatives than to propose a single solution
possible pitfalls, such as the inability of the PROMETHEE II method based on the first choice of weights.
to find a preferred alternative for bicriteria problems, when Type 1
preference function is used. Finally, if qualitative aspects of the
8. Conclusions
building are taken as optimization criteria, it might be a good idea
to explore the set of alternatives by applying the MCDM methods
This comparative study has presented our experiences with six
with various weights and other preference data.
well-known MCDM methods as a part of a design procedure that
A summary of the methods is presented in Table 11, with some
employs multicriteria optimization. Based on the test problems
observations made during the theoretical study and the computa-
of the present paper and the results of other comparisons, it seems
tions. To discuss the requisites of an MCDM method proposed in
that the best MCDM method can hardly be found. It is difficult to
[19], we can first conclude that all the methods are understandable
make definite conclusions based on numerical studies, but they
and pretty easily implemented. Most of them have issues that may
can discover special instances where the methods act strangely.
lead to unexpected behaviour in the numerical computations (see
Also, tendencies in the behaviour of the methods can be uncovered.
the comments in Table 11). However, these issues can usually be
As the decision making related to building design problems in-
avoided by appropriate modifications to the basic procedures.
volve several quantitative and qualitative aspects of different disci-
From the point of view of additional information required, the
plines, the MCDM methods can be very helpful in finding an
simplest method for the user is PEG, since it does not need any
agreeable compromise. If there are several decision makers in-
preference data from the decision maker. All other methods re-
volved, the decision-making process becomes substantially more
quire weights for the criteria. This is also the only preference data
difficult. The study of group decision making for building design
for the WSM, WPM and TOPSIS. In addition to weights, VIKOR
problems would be an interesting extension to the MCDM theory
needs the parameter v which controls the balance of the two met-
applied in this paper.
rics used for measuring the closeness to ideal solution. PROM-
ETHEE II requires the choice of the preference function and the
related parameter values. If the DM is able to provide this data, Acknowledgements
PROMETHEE gives some flexibility to the dominance relation.
However, the preference function needs to be selected carefully. This paper is dedicated to the fond memory of Prof. Donald E.
Type 1 preference function is unable to provide a preferred solu- Grierson, who participated in the writing of the earlier version of
tion for bicriteria problems and in our test cases with more than the paper presented in the 2011 EG-ICE Workshop.
two criteria the solutions provided by PROMETHEE II with the Type The research was funded by the Finnish Funding Agency for
1 preference function did not reflect the equal importance of the Technology and Innovation (Tekes).
criteria as designated by the weights.
Weights and other preference data allow the decision maker to
References
guide the selection of the compromise solution. However, for prob-
lems with more than two or three criteria, prescribing weights or [1] D. Grierson, S. Khajehpour, Method for conceptual design applied to office
criteria preference becomes more difficult. If the criteria are buildings, Journal of Computing in Civil Engineering 16 (2002) 83–103.
726 K. Mela et al. / Advanced Engineering Informatics 26 (2012) 716–726
[2] W. Wang, R. Zmeureanu, H. Rivard, Applying multi-objective genetic
algorithms in green building design optimization, Building and Environment
40 (2005) 1512–1525.
[3] Y. Rafiq, M. Rustell, Building information modelling driven by the evolutionary
computing, in: 2011 EG-ICE Workshop, University of Twente, The Netherlands,
2011.
[4] J. Wallenius, J.S. Dyer, P.C. Fishburn, R.E. Steuer, S. Zionts, K. Deb, Multiple
criteria decision making, multiattribute utility theory: recent
accomplishments and what lies ahead, Management Science 54 (2008)
1336–1349.
[5] C. Hwang, K. Yoon, Multiple attribute decision making, Lecture Notes in
Economics and Mathematical Systems, vol. 18, Spinger-Verlag, 1981.
[6] S.H. Zanakis, A. Solomon, N. Wishart, S. Dublish, Multi-attribute decision
making: a simulation comparison of select methods, European Journal of
Operations Research 107 (1998) 507–529.
[7] K. Miettinen, Nonlinear Multiobjective Optimization, Kluwer Academic
Publishers., 1999.
[8] D.E. Grierson, Pareto multi-criteria decision making, Advanced Engineering
Informatics 22 (2008) 371–384.
[9] P. Salminen, J. Hokkanen, R. Lahdelma, Comparing multicriteria methods in the
context of environmental problems, European Journal of Operations Research
104 (1998) 485–496.
[10] S. Opricovic, G.-H. Tzeng, Compromise solution by MCDM methods: a
comparative analysis of VIKOR and TOPSIS, European Journal of Operational
Research 156 (2004) 445–455.
[11] C. Parkan, M.-L. Wu, Comparison of three modern multicriteria decision-
making tools, International Journal of Systems Science 31 (2000) 497–517.
[12] V. Pareto, Cours d’Economie Politique, F. Rouge, 1896.
[13] W. Stadler, Multicriteria Optimization in Engineering and in the Sciences,
Plenum Press, 1988.
[14] K. Deb, Multi-Objective Optimization using Evolutionary Algorithms, John
Wiley & Sons, 2002.
[15] A. Osyczka, Evolutionary Algorithms for Single and Multicriteria Design
Optimization, Physica-Verlag, 2002.
[16] N. D’Cruz, A.D. Radford, J.S. Gero, A Pareto optimization problem formulation for
building performance and design, Engineering Optimization 7 (1983) 17–33.
[17] K. Shea, A. Sedgewick, G. Antonuntto, Multicriteria optimization of paneled
building envelopes using ant colony optimization, in: Intelligent Computing in
Engineering and Architecture, Ascona, Switzerland, pp. 627–636.
[18] P. Geyer, Models for multidisciplinary design optimization: an exemplary
office building, in: K. Gürlebeck, C. Könke (Eds.), 17th International Conference
on the Applications of Computer Science and Mathematics in Architecture and
Civil Engineering, Weimar, Germany.
[19] J.-P. Brans, B. Mareschal, PROMETHEE methods, in: J. Figueira, S. Greco, M.
Ehrgott (Eds.), Multiple Criteria Decision Analysis: State of the Art Surveys,
Springer Science + Business Media Inc., 2005, pp. 163–186.
[20] P.C. Fishburn, Additive utilities with incomplete product set: applications to
priorities and assignments, Operations Research 15 (1967) 537–542.
[21] E. Triantaphyllou, S.H. Mann, An examination of the effectiveness of multi-
dimensional decision making: a decision-making paradox, Decision Support
Systems 5 (1989) 303–312.
[22] D. Miller, M. Starr, Executive Decisions and Operations Research, Prentice-Hall,
1969.
[23] S. Opricovic, G.-H. Tzeng, Extended VIKOR method in comparison with
outranking methods, European Journal of Operational Research 178 (2007)
514–529.
[24] J.-P. Brans, P. Vincke, A preference ranking organisation method: (the
PROMETHEE method for multiple criteria decision-making), Management
Science 31 (1985) 647–656.
[25] M. Behzadian, R. Kazemzadeh, A. Albadvi, M. Aghdasi, PROMETHEE: a
comprehensive literature review on methodologies and applications,
European Journal of Operational Research 200 (2010) 198–215.
[26] P. Geyer, Multidisziplinäre Entwurfsoptimierung für das Bauwesen, Ph.D.
thesis, Technische Universität, Berlin, 2009 (in German).
[27] P. Geyer, Component-oriented decomposition for multidisciplinary design
optimization in building design, Advanced Engineering Informatics 23 (2009)
12–31.
[28] Mittaviiva Oy, Rakennusosien kustannuksia 2010, Rakennustieto, 2010
(General Cost Guide for Buildings, in Finnish).
[29] A.F. Mills, Basic Heat and Mass Transfer, Prentice Hall, 1999.
[30] EN 6946, Building Components and Building Elements, Thermal Resistance
and Thermal Transmittance, Calculation Method, CEN, 1997.