Fuzzy Multi-Objective Optimization for

Network Design of Logistic and Production Systems

M. Dotoli, M.P. Fanti, A.M. Mangini, G. Tempone

Politecnico di Bari
Via Re David 200
70 125 Bari
Italy
{dotoli, fanti, mangini}@deemail.poliba.it, [email protected]

Abstract structure of both the logistic and production networks,

considering also the e-business relationships between
Global competition has given rise to Logistic and operators and the network environmental impact.
Production Systems (LPSs), that are distributed This paper builds upon the configuration strategy for
manufacturing systems integrating international LPSs proposed in [2], that describes the structure of the
logistics and information technologies with production. LPS by a digraph, where nodes are partners and edges
This paper builds upon an LPS network design model are links. Moreover, different costs are assigned to each
previously proposed by some of the authors. The link (edge), so that the performance indices can be
recalled technique formulates and solves a multi- obtained by the digraph structure. In order to optimize
criteria optimization problem to select the partners in the LPS structure, a multi-criteria objective problem is
the different stages of the production chain and the formulated and different Pareto-optimal LPS solutions
links connecting them. In this paper, in order to rank are obtained by an integer linear programming
the equally optimal Pareto solutions of such a problem, problem. However, selecting one of such LPS
we propose to employ fuzzy multi-criteria optimization. alternative configurations may be a complex task if the
Two fuzzification techniques and two different multi- dimension of the solutions set is very large. Hence, in
criteria methods are considered. In addition, the this paper the solutions are ranked using fuzzy multi-
methodology is illustrated by way of a case study. criteria optimization. Indeed, fuzzy logic provides a
Moreover, a discussion on the different advantages and natural framework to incorporate qualitative
limitations of the proposed techniques is provided. knowledge with quantitative information such as real
data. Therefore, fuzzy multi-objective optimization is
Keywords: logistic and production systems, network particularly suitable for choosing, on the basis of the
design, optimization, fuzzy logic, performance indices. subjective and qualitative knowledge provided by the
decision makers, the LPS configuration from the set of
1. Introduction available and equally optimal alternatives. In particular,
two fuzzification methods are taken into account and
Global competition imposes an improved efficiency two different fuzzy optimization strategies proposed in
on companies, that respond to this pressure by the related literature are considered. The methodology
reengineering their processes, integrating international is illustrated by way of a case study, and a discussion
logistics and information technologies with production. on the different advantages and limitations of the
This process has given rise to the formation of Logistic proposed techniques is provided.
and Production Systems (LPSs), defined as a collection The paper is organized as follows. Section 2 recalls
of independent companies, possessing complementary the digraph-based LPS model and Section 3 defines the
skills and integrated with streamlined material, optimization model and the fuzzy techniques to rank
information and financial flow [4, 9, 2]. the obtained Pareto optimal solutions. Moreover, in
Significant literature deals with the problem of LPS Section 4 the case study is analyzed and solved.
network design and a detailed survey can be found in Finally, Section 5 summarizes the conclusions.
[5]. However, although several conceptual models for
LPSs are proposed and discussed in the related 2. The Network Model
literature, research efforts are lagging behind in the
development of formal decision models for LPS design A Logistic and Production System (LPS) can be
[9]. In order to fill such a gap, we propose a defined as a hyper-network of material flows overlaid
configuration strategy for LPSs that determines the with an e-business information network. The LPS
contains different stages: raw material supply, M={M1,M2,…,MNM}, where each element Mq∈M
intermediate supply, manufacturing, distribution, retail, corresponds to a performance measure. Typical indices
customers, and de-manufacturing or re-cycling. After include cost, cycle time, product quality, energy
the de-manufacturing stage, recovered material, consumption and environmental impact [4]. These
components or energy feedback to suitable LPS stages performance indices encapsulate the characteristics and
are considered. We denote the LPS stages by the set the properties of each link. In particular, a performance
{ }
ST = P1,..., Pk ,..., PN , where NS is the number of
s
value is assigned to each link, considering m- and e-
stages. In particular, each stage Pk is described as a set links: Mq(mij) (Mq(eij)) with q=1,…,NM denotes the
of partners representing different actors of the value of the performance measure Mq associated with
production chain and we suppose that there are N link mij∈Lm (eij∈Le). Since v-links are considered only
partners in the system. Moreover, the Bill Of Material to impose that the optimal LPS sub-network includes a
(BOM) of stage Pk is a set of material and components certain number of producers and consumers, no
required for processes in the k-th stage and produced performance index is assigned to each v-link and it
by upstream stages. holds Mq(vi)=0 for each q=1,…,NM and vi∈Lv.
To describe the interactions among partners of Moreover, an e-link speeds up the communication
different stages of an LPS, two sets of links are process and thus reduces the response time affecting
introduced: material flow links (m-links) and performance measures such as costs and productivity.
information links (e-links) [2]. The m-links of stage Pk Hence, if two partners are connected both by an m-link
are denoted by the set Lm={mij}, where mij is an m-link and an e-link (e.g., mij and eij), the performance
starting from ni∈Pk and ending in nj∈Ph, with Pk, measures are suitably updated and are associated with
Ph∈ST and k≠h. Analogously, the e-link set is link mij only. On the other hand, if just an e-link
connects two LPS actors, the performance measure that
Le={eij}, where eij is an e-link starting from ni∈Pk and
is assigned to the e-link is the cost of information, such
ending to nj∈Ph, with Pk, Ph∈ST and k≠h. Figure 1
as Internet portals, websites and electronic databases.
depicts a generic LPS network.

Stage P1 Stage Pk Stage Pk+1 Stage PNs 3. The Optimization Model

n1 ni k nik +1 niN
s
3.1. The integer linear programming problem
n2 nik +1 nik +1 +1 ni N +1
..
…
.. ..
…
..
s
We now recall the LPS optimization model for
.
. . .
partner and link selection. More precisely, the aim of
n s1 n ik +s k −1 nN
n i k +1 + sk +1 −1
the model is to minimize a multi-criteria cost function
subject to a set of constraints that we characterize as
BOM, path, mutual exclusion and structural constraints
Figure 1. The structure of a generic LPS. [2]. The objective function and constraints are obtained
In addition, if the LPS design requires the selection on the basis of the analysis of the digraph describing
of several producers and/or several consumers in the the LPS structure.
LPS network, we introduce virtual links (v-links): such Let us consider the performance index Mq(mij)
links are a mathematical artifice to simplify the (Mq(eij), Mq(vi)) that is associated with the
selection of candidate producers and/or consumers [1, corresponding edge xh∈E. We indicate with cq=[cq1 cq2
2]. They are denoted by the set Lv={vi}, where vi is a v- … cqE]T the E-entry vector where the h-th entry is
link ending in ni∈Pk, with Pk representing the cqh=Mq(eij) and/or cqh=Mq(mij) associated with edge
manufacturing and/or purchasing stage, and starting xh∈E. Moreover, x=[x1 x2 … xE]T denotes an integer
from a virtual partner n fi . Hence, the number of v- vector where each element xh∈{0,1} with h=1,…,E
indicates the presence (xh=1) or the absence (xh=0) of
links is equal to the number of candidate producers link xh∈E in the optimal LPS configuration model. The
and/or consumers in the LPS, say NV.
optimization problem is defined as an Integer Linear
The interactions among the LPS stages are depicted
Programming (ILP) problem as follows [2]:
by a digraph D=(N,E). The node set N represents the
partner set of the network and each node ni∈N for
z=min f(x) (1)
i=1,…,N is associated with partner ni∈Pk for
k∈{1,…,NS} of the LPS network. Moreover, the edge
subject to
set E is such that an arc xh directed from ni to nj is in E
Ax≥B (2)
if there exists an m-link mij∈Lm, an e-link eij∈Le and/or
a v-link vi∈Lv. Finally, we denote with E the overall
xh∈{0,1} for h=1,…,E (3)
number of edges in D.
We complete the description of the LPS network by
introducing the set of performance indices where A is the r×E constraint matrix and B is an r-entry
vector of integers, r being the number of constraints.
 The objective function is the following [2]: minimization condition), i.e. the greater its fuzzy value.
Accordingly, for each j=1,…,qQ, element d’ij for
f(x)=Cx (4) i=1…m represents the fuzzy value of the j-th
performance index for the i-th LPS solution. For each
where C=[(cq1)T … (cqQ)T]T is a qQ×E criteria matrix j=1,…,qQ, the membership functions used to fuzzify
and cq1,…,cqQ are E-entry vectors associated with matrix DM and determine matrix D’ are alternatively:
performance indices Mq1,…,MqQ, respectively. 1. linear, defined by the function
Solving the ILP problem (1)-(4) for a particular
matrix C provides the maximal Pareto face of the  1 if d mij < d min j
solutions set [3]. More precisely, we obtain a set 
X*={x*i}, where each x*i∈X* is a Pareto optimal  d max j − d mij
solution selecting a sub-digraph D*i=(N*i,E*i) of D, d 'ij =  if d min j ≤ d mij ≤ d max j (5)
with N*i⊂N and E*i⊂E, respectively. If the h-th entry  d max j − d min j

of x*i is x*h=1 and xh is an edge from ni∈Pk to nj∈Ph,  0 if d mij > d max j
then the solution selects edge xh∈E*i and nodes ni,
nj∈N*i; in addition, if the h-th entry of x*i is x*h=1 and
2. sigmoidal, defined by the function
xh is an edge corresponding to a v-link associated to
producer or consumer nj∈Ph, then the solution selects
partner ni∈N*i. We call m the cardinality of the set of 1
d 'ij = (6)
 d max j + d min j 
Pareto optimal solutions X*. d min j  d mij − 
 2 
1+ e  
3.2. The fuzzy multi-objective optimization
After the ILP problem is solved, the decision makers
have to evaluate and validate the Pareto optimal with i=1,…,m and (
d max j = max d m1j ,..., d mmj , )
solutions, representing each a possible LPS
configuration. Two factors may make this phase ( )
d min j = min d m1j ,..., d m mj for each j=1,…,qQ.
difficult for the decision makers: 1) the number of
solutions can be very large and validating each of them 3.2.1. FTOPSIS
is time consuming; 2) the solutions suitability depends The Fuzzy Technique for Order Preference by
on the importance assigned to performance indices. Similarity to Ideal Solution (FTOPSIS) [7] is a fuzzy
Hence, we propose the fuzzy multi-objective multi-objective decision technique based on simple
optimization for ranking the equally optimal solutions geometric concepts: the best alternative exhibits the
of problem (1)-(4) and choosing, on the basis of the shortest distance from the Best Ideal Solution (BIS)
subjective and qualitative knowledge provided by the and the farthest distance from the Worst Ideal Solution
decision makers, the best LPS configuration. The fuzzy (WIS) in a geometrical (i.e., Euclidean) sense. The
multi-objective techniques considered in this work are FTOPSIS technique consists of the following steps.
the following well-known methods [7]: 1) the Fuzzy Step 1. Joining weights to the performance indices.
Technique for Order Preference by Similarity to Ideal Determine the fuzzified decision matrix D’ and
Solution (FTOPSIS); 2) the Fuzzy Analytic Hierarchy associate to each performance index j=1,…,qQ a weight
Process (FAHP). qQ
wj, with j=1,…,qQ and ∑ w j = 1 , modelling the
Input data. Both techniques require as input data an j=1
mxqQ decision matrix DM, where we recall that m is the importance degree of the j-th performance index in the
number of solutions of the ILP problem (1)-(4) and qQ ranking of the different LPS configurations.
is the number of performance indices on the grounds of Step 2. Constructing the normalized decision matrix.
which the solutions ranking is performed. Hence, the Determine each element n’ij of the mxqQ normalized
generic element d mij of DM, with i=1,…,m and fuzzified decision matrix N’ as follows:
j=1,…,qQ, represents the j-th performance value of the d 'ij
i-th LPS solution x*i. The fuzzification process n 'ij = , i=1,…,m, j=1,…,qQ. (8)
m
associates to each value d mij of DM a fuzzy value d’ij,
∑ d 'ij2
with 0≤d’ij≤1, defining the mxqQ fuzzified decision i =1
matrix D’, that depends on the alternatives satisfaction Step 3. Constructing the weighted normalized
degree with respect to the performance indices. Since decision matrix. Determine the mxqQ weighted
these have to be minimized, the lower a performance normalized decision matrix W’, where w’ij=n’ij⋅wj, for
value of an alternative, the greater its satisfaction i=1,…,m and j=1,…,qQ.
degree has to be (so that it better satisfies the
 Step 4. Determining the best and worst ideal eigenvalue of the so-called comparison matrix. The
solutions. Determine the BIS as the ideal solution with technique consists of the following steps.
performance indices given by the row vector Step 1. Structuring the decision problem as a
G = [G1 ... G qQ ] , where Gj=max(n’1j,…,n’mj) with hierarchy. Select the first level of the hierarchical
structure as the overall goal “LPS Efficiency”. Define
j=1,…, qQ. Moreover, determine the WIS as the ideal
the second level, that is composed by the qQ considered
solution associated to performance indices of the row
criteria contributing to the goal. Moreover, determine
vector H = [H1 ... H q Q ] , where Hj=min(n’1j,…,n’mj)
the third level as the alternative LPS configurations to
with j=1,…, qQ. rank in terms of the criteria defined in the second level.
Step 5. Calculating the separation measure. Step 2. Joining weights to the criteria. Determine
Calculate the separation distance SG,i from the BIS of the fuzzified decision matrix D’ and associate a weight
each alternative x*i with i=1,…,m as follows: qQ
wj to each index j=1,…,qQ, with ∑ w j = 100 .
qQ j=1
SG,i = ∑ (n 'ij − G j ) 2 . (9) Step 3. Constructing the pairwise comparison
j=1 matrix CM. Compare the qQ criteria that define the
second level of the hierarchical structure with each
Moreover, determine the separation distance SH,i of other and construct the qQxqQ pairwise comparison
x*i with i=1,…,m from the WIS as follows: matrix CM by Saaty’s original AHP scale in Table 1.
More precisely, determine each element cmij of CM
qQ
with i=1,…,qQ and j=1,…,qQ, representing the relative
SH,i = ∑ (n 'ij − H j ) 2 . (10)
importance of the i-th performance index compared to
j=1
the j-th criterion, by evaluating the difference |wi-wj| of
the performance indices weights and associating an
Step 6. Calculating the relative closeness of
integer value from 1 to 9 to it according to Table 1.
alternatives to the ideal solution. Determine the
Step 4. Determining the eigenvector associated to
closeness Ci to the WIS of each alternative x*i with the maximum eigenvalue of the comparison matrix.
i=1,…,m as follows: Calculate the eigenvalues set { λ 1, λ 2,…, λ R} of CM,
where R is the matrix rank. Let λ max be the maximum
SH,i
Ci = . (11) eigenvalue of CM, then determine its eigenvector vmax.
SG,i − SH,i Compute the priority vector:

Step 7. Ranking alternatives. The ranked vector of P = vmax ⋅ qQ = [p1 ... pq Q ]T , (12)
alternatives is Ω=[Ω1 … Ωm]T, where Ωi for i=1,…,m
is the generic i-th LPS configuration and exhibits a
where each element pj with j=1,…,qQ of P represents
closeness Ci>Ci+1: hence, Ω1 is the best LPS
the importance degree of the j-th performance index
configuration and Ωm is the worst one.
associated to the j-th column of D’: the greater pj, the
Table 1. Saaty’s original AHP scale. more important the j-th performance index.
Step 5. Raising alternatives to the criteria power.
0 5 15 25 35 45 55 65 75 Determine the alternative values associated to each j-th
Pairwise
÷ ÷ ÷ ÷ ÷ ÷ ÷ ÷ ÷ performance index as follows:
differences
5 15 25 35 45 55 65 75 100
AHP scale 1 2 3 4 5 6 7 8 9 CRITj = [d '1j ... d 'mj ] (13)

3.2.2. FAHP
for each j=1,…,qQ. Determine the following vectors:
The Fuzzy Analytic Hierarchy Process (FAHP) [6,
10, 7] is the second fuzzy multi-objective decision
( ) j = ( d '1j ) j ... ( d 'mj ) j 
p p p
technique considered in this work. All the elements Cj=[c1j … cmj]= CRITj (14)
involving a decision problem (overall goal, criteria,
alternatives) are arranged in a hierarchical structure and
objectives are of varying degrees of importance. The for each j=1,…,qQ.
ranking is achieved by assigning to each of the equally Step 6. Determining the decision model. For each
optimal LPS configurations a power indicative of its alternative x*i with i=1,…,m, determine:
importance and then raising each fuzzy value to the
appropriate power. Such powers are obtained by
determining the eigenvector of the maximum
(
ALTi = min ci1 ,..., ciq Q ) (15)
 Supplier Manufacturer Distributor Retailer Consumer Recycler
Stage P 1 Stage P2 Stage P3 Stage P 4 Stage P 5 Stage P 6
to n10 to n8 to n15 to n5
to n12 x2
to n9 to n12 2

8
to n13 to n16
x1

x1
n15

x8
x2 to n5 x9 x19 x32
x 41 n12

6
X3

x1
x20 to n6
n1 X24 to n6 n5 to n14

4
x44
n10

x6
x47 to n13

1
x33 x10 to n17

x6
x12

x2
x1 n8 to n4

1
to n7 1 to n9

x4
to n5 to n18 to n7

8
X4 to n11 to n14
n16
n2 X25
x34
to n6
3 3
to n15
to n8 x5
to n7 x1 to n16
to n12 x55
to n5 to n12 x6
5 n13
8
x2 x29 to n13
x57
x5 x42 4
n3 n6 x1 to n4 n17

x5
x26 to n6 x45 to n17
n9

9
x35 x3 to n14 x51 to n13
x3 0 to n12
to n7 1 to n18
to n9 x1
to n5 5 to n15

x5
x6

7
to n11 4

x1
x5

2
n4 x27 to n6 to n8 x66 to n16
x36 to n12 n11 x49 to n13 to n4 n14 x56 n18
8

x40
x3

x7 to n14 x58
to n7 n7 x43 to n13

x5
x6 63

x46 3
x2
x

x6
0
2

to n12 x3 to n17

0
x3

to n13 7 to n14 to n14 to n3

9

to n14 to n18
to n11 to n9

Figure 2. The digraph model of the case study ( m-link, e-link, e- and m-link).

so that ALTi provides information about the industrial experience. In addition, the NV=6 v-links
satisfaction of alternative x*i with respect to the edges x67, x68, x69, x70, x71 and x72 in Table 2 are
performance indices and their importance degree. distinguishable by the italic font and exhibit a zero
Step 7. Ranking alternatives. The ranked performance measure. Figure 2 depicts the digraph
alternatives are collected in the ordered vector Ω=[Ω1 D=(N,E) describing the LPS, exhibiting N=18 nodes
… Ωm]T, where Ωi for i=1,…,m is the generic i-th LPS and 66 edges associated to m- and/or e-links, while the
configuration with ALTi>ALTi+1: hence, alternative Ω1 six v-links are omitted for the sake of simplicity. Edges
is the best LPS configuration, Ωm is the worst one. x1, x2, x7, x13, x14, x15, x16, x17, x47, x48, x49, x50, x51, x52,
x62 and x63 correspond both to m- and e-links, edges
Table 2. Data for the case study links. x64, x65, x66 are associated with e-links and the
remaining edges of the digraph in Figure 2 are
Total costs Energy CO2 emission associated with m-links only. In addition, v-links are
Links (M1) (M2) (M3) selected as follows: x67 is associated to node n5, x68 to
in US$ in MJ in kgCE n6, x69 to n7, x70 to n12, x71 to n13 and x72 to n14.
x1 (m-and e-link) 41.80 359.00 0.87
x2 (m-and e-link) 46.70 332.00 0.74 4.2. The ILP problem for the case study
x3 (m-link) 319.00 1479.0 2.21
… … … …
4.2.1. The BOM constraints
x64 (e-link) 0.30 0.00 0.00
x65 (e-link) 0.40 0.00 0.00 The components supplier constraints are obtained
x66 (e-link) 0.20 0.00 0.00 assuming that the BOM of the manufacturers stage (P2
x67 (v-link) 0.00 0.00 0.00 in Figure 2) is as follows: computer (C), hard-disk-
… … … … driver (H), monitor (M), and keyboard/mouse (K). We
x72 (v-link) 0.00 0.00 0.00 assume that C is produced by n1 and n2, H is produced
by n1, n2 and n3, M is produced by n2, n3 and n4, and K
4. The Case Study is produced by n3 and n4 [4]. Hence, with reference to
Figure 2, the constraints imposed on the edges are:
4.1. The case study network model
We consider a case study inspired by an example x3 + x4 – x67 ≥ 0
proposed in [4]. The target product is a typical desktop x3 + x4 + x5 – x67 ≥ 0
computer system consisting of the computer, hard disk x4 + x5 + x6 – x67 ≥ 0
driver, monitor, keyboard and mouse. The LPS is x5 + x6 – x67 ≥ 0
composed of NS=6 stages: four suppliers, three x24 + x25 – x68 ≥ 0
manufacturers, two distributors, two retailers, three x24 + x25 + x26 – x68 ≥ 0 (16)
consumers and four recyclers, for a total of N=18 x25 + x26 + x27 – x68 ≥ 0
candidate partners. Some selected data for the case x26 + x27 – x68 ≥ 0
study are reported in Table 2, that shows the values of x33 + x34 – x69 ≥ 0
performance indices total costs (M1), energy (M2) and x33 + x34 + x35 – x69 ≥ 0
CO2 emission (M3), associated with the links of the x34 + x35 + x36 – x69 ≥ 0
considered LPS and determined on the basis of x35 + x36 – x69 ≥ 0
4.2.2. The path constraints x7 + x14 + x16 + x17 – x70 ≥ 0
It is necessary to select a path that starts from a node x62 + x47 + x51 + x49 – x71 ≥ 0
of the manufacturer stage P2 and ends to a node of the x63 + x48 + x52 + x50 – x72 ≥ 0
consumer stage P5. Consequently, the constraints x1 – x16 – x47 – x48 ≤ 0
derived from the digraph structure are as follows: x2 – x13 – x15 – x14 – x51 – x52 ≤ 0
x7 – x70 ≤ 0
x9 + x29 + x40 + x16 + x14 + x17 – x70 ≥ 0 x62 – x71 ≤ 0
x41 + x42 + x43 + x47 + x51 + x49 – x71 ≥ 0 x63 – x72 ≤ 0
x44 + x45 + x46 + x48 + x52 + x50 – x72 ≥ 0 x7 – x64 ≤ 0
x8 + x9 + x41 + x44 + x10 + x11 – x67 = 0 x62 – x65 ≤ 0
x28 + x29 + x42 + x45 + x30 + x31 – x68 = 0 x63 – x66 ≤ 0
x38 + x40 + x43 + x46 + x37 + x39 – x69 = 0
x12 – x8 – x28 – x38 = 0 (17) Table 3. Optimal solutions for objective
x13 + x14 + x51 + x52 + x15 – x10 – x30 – x37 ≥ 0 function costs, energy and CO2 emission.
x16 + x47 + x48 – x12 ≥ 0 CO2
x17 + x49 + x50 – x11 – x31 – x39 ≥ 0 Optimal Costs Energy Indices of selected
in
x16 + x47 + x48 – x13 ≥ 0 solutions in US$ in MJ edges
KgCE
x17 + x49 + x50 – x15 ≥ 0 2,4,5,10,23,25,27,32,45,51,
x*1 1873.54 13771.0 365.13 54,59,63,66,67,68,71,72
Hence, the first line of (17) means that if node n12 2,4,5,10,23,25,27,32,45,51,
x*2 1873.54 13771.0 365.13 53,60,63,66,67,68,71,72
(x70=1) is selected, then at least an edge between x9, 2,4,5,10,23,25,27,30,32,51,
x29, x40, x16, x14 and x17 has to be selected. In addition, x*3 1883.54 12991.0 69.03 52,53,60,63,66,67,68,71,72
the fourth line of (17) means that if node n5 (x67=1) is 2,3,5,10,23,25,27,30,32,51,
x*4 1894.54 12694.0 69.05 52,54,59,63,66,67,68,71,72
selected, then one and only one edge between x8, x9, 2,3,5,10,23,25,27,30,32,51,
x41, x44, x10 and x11 has to be selected. x*5 1894.54 12694.0 69.05 52,53,60,63,66,67,68,71,72
2,3,5,7,10,14,18,23,25,27,
x*6 1896.34 12678.1 60.03 30,32,51,59,64,67,68,70,71
4.2.3. The mutual exclusion constraints 2,3,6,7,10,14,18,22,25,27,
We assume that the LPS network has to include two x*7 1914.34 12542.7 58.04 30,32,51,53,64,67,68,70,71
and only two partners both in the manufacturer and x*8
2,3,6,7,10,14,18,22,24,27,30
1949.34 12327.7 58.14 ,32,51,53,64,67,68,70,71
consumer stages (respectively stages P2 and P5 in
2,3,6,7,10,14,18,22,33,36,
Figure 2). In addition, we suppose that at most two x*9 1967.34 12312.7 58.31 37,51,53,61,64,67,69,70,71
partners have to be selected in the recycler stage (stage 2,3,6,7,10,15,17,18,22,33,
x*10 2023.24 12291.5 56.45 36,37,51,53,61,64,67,69,70,71
P6 in Figure 2). Furthermore, we assume that two and
2,4,5,10,23,25,27,30,32,51,
only two m- and e-links have to connect the first stage x*11 1883.54 12991.0 69.03 52,54,59,63,66,67,68,71,72
and the others. Hence, with reference to Figure 2, the 2,4,5,7,10,14,21,23,25,27,
x*12 1885.34 12975.1 60.01 30,32,51,53,64,67,68,70,71
mutual exclusion constraints are as follows:
2,4,5,7,10,14,18,23,25,27,
x*13 1885.34 12975.1 60.01 30,32,51,59,64, 67,68,70,71
x67 + x68 + x69 = 2 1,2,4,5,10,13,16,21,23,25,
x70 + x71 + x72 = 2 x*14 1939.54 13036.8 29.76 27,30,32,51,53,67,68,70,71
x18 + x19 + x20 + x21 – x70 ≤ 0 (18) 1,2,4,5,10,13,16,18,23,25,
x*15 1939.54 13036.8 29.76 27,30,32,51,59, 67,68,70,71
x53 + x55 + x57 + x59 – x71 ≤ 0 1,2,4,6,10,13,16,18,22,25,
x54 + x56 + x58 + x60 – x72 ≤ 0 x*16 1957.54 12901.4 27.77 27,30,32,51,53,67,68,70,71
x1+ x2 + x7 + x62 + x63 = 2 1,2,4,6,10,13,16,25,27,30,
x*17 1994.70 12912.8 25.59 51,67,68,70,71
1,2,4,6,10,13,15,25,27,30,
4.2.4. The structural constraints x*18 2050.40 12894.6 23.73 47,50,67,68,71,72
The constraints derived from the digraph structure x*19 2052.40 12892.0 23.69
1,2,4,6,10,13,15,16,25,27,
30,50,67,68,70,72
are as follows: 1,2,13,15,17,25,27,30,34,36,
x*20 2067.50 12902.3 23.67 37,48,68,69,70,72
x23 – x5 – x26 – x35 ≤ 0 x*21 2079.24 12353.6 26.16
1,2,3,6,10,13,15,16,18,22,33
x23 – x21 – x59 – x60 = 0 36,37,50,54,61,67,69,70,72
1,2,3,6,10,13,15,16,18,33,
x22 – x67 ≤ 0 x*22 2096.82 12359.8 25.05 36,37,50,61,67,69,70,72
x32 – x68 ≤ 0 1,2,3,6,10,13,15,16,33,36,
x*23 2115.40 12367.0 23.88 37,50,67,69,70,72
x61 – x69 ≤ 0
x18 – x70 ≤ 0
x53 – x71 ≤ 0 4.2.5. Solution of the ILP problem
x54 – x72 ≤ 0 The ILP problem (1)-(4) is solved considering the
x22 + x32 + x61 – x18 – x53 – x54 = 0 (19) following performance indices: costs, energy and CO2
emission (objective function f). In particular, the Pareto DM exhibits m=23 rows, corresponding to the number
optimal solutions are obtained by the well-known two- of optimal solutions, and qQ=3 columns, related to the
phases simplex method in the Matlab framework [8]. number of performance indices costs, energy and CO2
Table 3 reports for each optimal solution of objective emission. Hence, matrix DM is composed by the
function f the corresponding performance indices and second, third and forth column of Table 3.
the selected arcs of the LPS network. As an example, Subsequently, we determine the fuzzified decision
the digraph corresponding to solution x*4 in Table 3 is matrix D’: Tables 4 and 5 represent D’ when the linear
depicted in Figure 3. Note that the digraph is composed and sigmoidal fuzzification (5) and (6) are respectively
by two parallel productive chains and a remarkable used. Note that the lower the crisp value of a solution
advantage is observed: if a transporter is temporarily referring to a performance index, the higher its fuzzy
unavailable and/or a communication way cannot be value (compare Table 3 with Tables 4 and 5).
momentarily employed, the productive cycle does not We assume that the decision makers have to rank
stop. The digraphs of the other equally optimal the equally optimal solutions when some criteria are
solutions in Table 3 may easily be obtained from the more important than others, according to the following
last column of the table. cases:
a) costs weigh 50% in the decision, while energy
x2 x32 and CO2 emission weigh 25% each;
n1 x3 n15
n5 b) CO2 weighs 50% in the decision, while costs
n2 x25 x5 and energy weigh 25%.
x1

x51 n13
4
0

x5

The LPS configurations ranked as top three by the

x5

n6
x3
9

n3 0
7 n9 x52 n14
two considered techniques using the linear
x2
n4 n18 fuzzification method are reported in Table 6. In
x63
x66
particular, for case a) the FTOPSIS and FAHP methods
x23 provide the same results and choose alternative x*6 as
the best solution. Indeed, this LPS configuration is
Figure 3. The digraph of solution x*4 ( characterized by one of the lowest values of costs;
m-link, e-link, e- and m-link). moreover, among the low cost solutions, it exhibits
Table 4. Matrix D’ for the case study with acceptable values of the less important criteria energy
linear membership functions. and CO2 emission. For case b), the FAHP method
selects configuration x*7 as the best solution, while the
Optimal CO2 FTOPSIS technique ranks this alternative at the second
Costs Energy position. Indeed, such an LPS configuration exhibits
solution emission
x*1 1 0 0 one of the lowest values of CO2 emission; moreover,
x*2 1 0 0 among the low CO2 emission solutions, it features
x*3 0.958900 0.527550 0.86759 acceptable values of costs and energy.
… … … …
Table 6. Best alternatives for cases a) and
x*21 0.153020 0.958100 0.99280
b) with linear membership functions.
x*22 0.080708 0.953900 0.99600
x*23 0 0.948990 0.99940 Case a) Case b)
Position FTOPSIS FAHP FTOPSIS FAHP
Table 5. Matrix D’ for the case study with
1 x*6 x*6 x*8 x*7
sigmoidal membership functions.
2 x*5 x*5 x*7 x*6
Optimal CO2 3 x*4 x*4 x*6 x*8
Costs Energy
solution emission
x*1 1 0 0 Table 7. Best alternatives for cases a) and
1 0 0 b) with sigmoidal membership functions.
x*2
x*3 0.996700 0.553560 0.965000 Case a) Case b)
… … … … Position FTOPSIS FAHP FTOPSIS FAHP
x*21 0.046905 0.996500 0.999900 1 x*7 x*6 x*7 x*7
x*22 0.013001 0.995800 1 2 x*6 x*5 x*6 x*6
x*23 0 0.994800 1 3 x*5 x*4 x*5 x*5

4.3. Ranking the Pareto optimal solutions Table 7 reports the three best LPS configurations
Before applying the proposed fuzzy multi-objective when the sigmoidal membership function is employed.
techniques to the case study, the input data have to be Comparing Tables 6 and 7, we remark that using the
determined. Referring to Table 3, the decision matrix same fuzzy multi-objective technique (FTOPSIS or
FAHP) with a different fuzzification method results in 5. Conclusions
a different ranking. For example, when case a) is
considered, the FTOPSIS technique chooses x*6 as the Logistic and Production Systems (LPSs) are a
best solution when a linear membership function is business strategy integrating international logistics and
used, while it selects x*7 when a sigmoidal information technologies with production. Using an
membership function is used. The different results approach previously proposed by some of the authors,
obtained with the two fuzzification methods may be this paper models the LPS by a digraph describing the
clarified remarking that the linear and sigmoidal stages and the links among them. Moreover, an integer
functions are similar in the intermediate part of their linear programming problem with suitable constraints
curve, but they differ significantly at the extremes of is stated. In order to rank the equally optimal
the universe of discourse. alternative solutions, the paper proposes to employ
fuzzy multi-criteria optimization. Hence, some LPS
4.4. Discussion of the results configurations are selected for evaluation and
The obtained results demonstrate the effectiveness validation by the decision makers. The methodology is
of fuzzy multi-criteria optimization in overcoming the applied to a case study inspired by the related
limitations of Pareto optimality. Particularly, the results literature. The obtained results prove the effectiveness
show that the considered techniques are all successful, of fuzzy multi-criteria optimization in overcoming the
but they are characterized by some differences. limitations of Pareto optimality. Moreover, a
First, while the FTOPSIS method is based on the discussion on the advantages and limitations of the
determination of the best and worst solution from the proposed techniques is provided.
set of alternatives, the FAHP technique relies on
pairwise comparisons between solutions. Hence, while References
FTOPSIS employs the solutions set as a whole to
determine two ideal alternatives by way of simple [1] A. Agnetis, C. Arbib, M. Lucertini, S. Nicoloso, Il
geometric concepts, FAHP faces the decision problem Processo Decisionale, La Nuova Italia Scientifica,
in a more complex and accurate way, providing an 1992 (in Italian).
approach to rank alternatives based on their reciprocal [2] M. Dotoli, M.P. Fanti, C. Meloni, M.C. Zhou, “A
assessment. Thus, FAHP may be preferred for its multi-Level Approach for Network Design of
accuracy. On the other hand, the pairwise comparisons Integrated Supply Chains”, Int. J. Prod. Res., 2005
may be unmanageable, especially when the solutions (in print).
set is very large. Similar concerns arise with respect to [3] M. Ehrgott, Multicriteria Optimization, Springer
the FAHP limitations when a novel candidate is added Verlag, 2000.
to a pre-existent set of alternatives: applying the [4] Y. Luo, M.C. Zhou, R.J. Caudill, “An Integrated
method may be inconvenient, since the ranking is e-Supply Chain Model for Agile and
based on pairwise comparisons, so that the whole Environmentally Conscious Manufacturing”,
decision process has to be re-designed in such a case. IEEE/ASME Trans. Mechatron., Vol. 6, pp. 377-
Second, with regard to the difference between the 386, 2001.
use of linear or sigmoidal membership function, we [5] P. Pontrandolfo, O.G. Okogbaa, “Global
can infer that the sigmoidal one involves a larger Manufacturing: a Review and a Framework for
number of alternatives with low values of the Planning in a Global Corporation”, Int. J. Prod.
performance indices in the choice of the best solution, Res., Vol. 37, pp. 1-19, 1999.
so that more alternatives are in competition to be [6] T.L. Saaty, “How to Make a Decision: the
selected. In addition, with respect to the linear function Analytic Hierarchy Process”, Europ. J. Oper. Res.,
the sigmoidal one better filters out the solutions that Vol. 48, pp. 9-26, 1990.
exhibit a low satisfaction degree. [7] E. Triantaphyllou, C.-T. Lin, “Development and
Third, it is important to remark the inherent Evaluation of Five Fuzzy Multiattribute Decision-
robustness of fuzzy multi-criteria optimization: despite Making Methods”, Int. J. Approx. Reas., Vol. 14,
the explained differences among the discussed pp. 281-310, 1996.
methods, the corresponding results are similar, i.e., the [8] P. Venkataraman, Applied Optimization with
obtained subsets of top LPS configurations are alike. MATLAB Programming, Wiley Interscience, 2001.
Hence, we infer that fuzzy multi-criteria optimization [9] N. Viswanadham, R.S. Gaonkar, “Partner
is quite tolerant of the particular technique as well as of Selection and Synchronized Planning in Dynamic
the specific shape of the membership functions. Manufacturing Networks”, IEEE Trans. Robot.
Finally, note that the optimal solution can be Automat., Vol. 19, pp. 117-130, 2003.
validated by comparison with a base case LPS defined [10] R.R. Yager, “Fuzzy Decision Making Including
using known parameters and obtained via traditional Unequal Objectives”, Fuzzy Sets Syst., Vol. 1, pp.
business practices [4]. 87-95, 1978.