See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/223004323

Supply chain network design using an integrated neuro-fuzzy and MILP

approach: A comparative design study

Article in Expert Systems with Applications · December 2009

DOI: 10.1016/j.eswa.2009.05.034 · Source: dx.doi.org

CITATIONS READS

82 897

3 authors, including:

Alev Taskin Ali Fuat Guneri

Yildiz Technical University Yildiz Technical University
71 PUBLICATIONS 3,027 CITATIONS 57 PUBLICATIONS 2,727 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

SPSA-FSR View project

Safety Engineering View project

All content following this page was uploaded by Ali Fuat Guneri on 18 March 2014.

The user has requested enhancement of the downloaded file.

 Expert Systems with Applications 36 (2009) 12570–12577

Contents lists available at ScienceDirect

Expert Systems with Applications

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

Supply chain network design using an integrated neuro-fuzzy and MILP approach:
A comparative design study
Alev Taskin Gumus *, Ali Fuat Guneri, Selcan Keles
Department of Industrial Engineering, Yildiz Technical University, 34349 Besiktas-Istanbul, Turkey

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

Keywords: In this study, an integrated supply chain (SC) design model is developed and a SC network design case is
Supply chain network design examined for a reputable multinational company in alcohol free beverage sector. Here, a three echelon SC
Neuro-fuzzy network is considered under demand uncertainty and the proposed integrated neuro-fuzzy and mixed
Mixed integer linear programming integer linear programming (MILP) approach is applied to this network to realize the design effectively.
Artificial neural networks
Matlab 7.0 is used for neuro-fuzzy demand forecasting and, the MILP model is solved using Lingo 10.0.
Then Matlab 7.0 is used for artificial neural network (ANN) simulation to supply a comparative study
and to show the applicability and efficiency of ANN simulation for this type of problem. By evaluating
the output data, the SC network for this case is designed, and the optimal product flow between the fac-
tories, warehouses and distributors are calculated. Also it is proved that the ANN simulation can be used
instead of analytical computations because of ensuring a simplified representation for this method and
time saving.
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction product, service and information. Initially an appropriate network

design is needed to control efficiently all the elements in all stages,
In today’s world, enterprises have to cope with the growing to be flexible against the changing situations, to provide coordina-
markets and with the increasing customer expectations. Because tion along the supply chain, and so be successful in SCM which is
of the customer expectations about obtaining the products at the very complicated (Cakravastia et al., 2002; Graves & Willems,
right time and quantity, and besides this the improvements against 2005). Therefore, SC network design is very important and at-
the risks created by the sudden fluctuations in local and global tempts to create the most efficient and effective SC for the com-
economies, companies need to analyse their working styles. Today, pany’s operating environment (Samadhi & Hoang, 1998).
the success measures for the companies are thought as lower costs, A SC network is commonly defined as the integrated system
shorter production time, shorter lead time, less stock, larger prod- encompassing raw material vendors, manufacturing and assembly
uct range, more reliable delivery time, better customer services, plants, and distribution centres, to ensure solutions for effectively
higher quality, and providing the efficient coordination between meeting customer requirements such as low costs, high product
demand, supply and production. For this reason, supply chain man- variety, quality and shorter lead times (Chauhan, Nagi, & Proth,
agement (SCM) concept is occurred and, SCM has become an 2004; Santoso, Ahmed, Goetschalckx, & Shapiro, 2005). The net-
important necessity. work is characterised by procurement, production, and distribu-
Huang, Sheoran, and Keskar (2005) describe a supply chain (SC) tion functions (Santoso et al., 2005). Leaving aside the
as a network of facilities that procure raw materials, transform procurement function (purchasing of raw materials), the SC net-
them into intermediate goods and then final products, and deliver work becomes a multi-echelon production/distribution system
the products to customers through a distribution system. SCM is (Altiparmak, Gen, Lin, & Karaoglan, 2009; Beamon, 1998; Bhaska-
strategically and systematically coordination of all functions of ran & Leung, 1997; Shapiro, 2001; Taskin Gumus & Guneri, 2007;
the companies in SC in order to increase long term performance Tsiakis, Shah, & Pantelides, 2001).
of both supply chain and the other companies in the chain (Cakra- The design of SC networks is a difficult task because of the
vastia, Toha, & Nakamura, 2002; Gunasekaran, 1999). intrinsic complexity of the major subsystems of these networks
SCM is the integration of key processes that is started by the and the many interactions among these subsystems, as well as
first supplier and continues until the end consumer to provide external factors such as the considerable uncertainty in product
demands (Santoso et al., 2005). In the past, this complexity has
* Corresponding author. Tel.: +90 2122597070/2242; fax: +90 2122585928. forced much of the research in this area to focus on individual com-
E-mail address: [email protected] (A.T. Gumus). ponents of supply chain networks. Recently, however, attention

0957-4174/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2009.05.034
 A.T. Gumus et al. / Expert Systems with Applications 36 (2009) 12570–12577 12571

has increasingly been placed on the performance, design, and anal- from the view of its own vision. Nonetheless it is highly important
ysis of the supply chain as a whole (Tsiakis et al., 2001). for the company managers to understand the roles and the per-
Design of the chain should be able to integrate the various ele- spectives of the other companies along the supply chain, because
ments of the chain and should strive for the optimization of the every company is the participant of the next supply chain. The rea-
chain rather than the entities or group of entities. Information son behind is that the integration and management of the work
sharing and its control play a vital role in integration of the differ- processes in a firm will be successful when they are all important
ent elements of the chain and require highly coordinated efforts of from every firm’s perspective (Lambert & Cooper, 2000).
both engineers and managers (Fisher & Raman, 1996; Lee, Padma- The SC network design is one of the biggest strategic decision
nabhan, & Whang, 1997; Wang, Huang, & Dismukes, 2004). problems which are used for efficient long term operations in the
In the following section, SC network design concept is detailed. whole SC, and for this reason it needs optimization. The design fig-
Then, in the third and fourth sections, an integrated SC design ures out the numbers, capacity, layout and type of the factories,
model is developed and a SC network design case is examined warehouses and distribution centres. In addition it sets up the dis-
for a reputable multinational company in alcohol free beverage tribution channels and calculates the quantity of materials which
sector. Here, a three echelon SC network is considered and the pro- will be consumed in the production process, the quantity of mate-
posed integrated neuro-fuzzy and MILP approach is applied to this rials which will be transported from suppliers to customers, and
network to realize the design effectively. Matlab 7.0 is used for the quantity of materials which will be produced. The SC is divided
neuro-fuzzy demand forecasting and, the deterministic, static, into several stages to bear with the complexity of the designing
multi-echelon MILP model is solved using Lingo 10.0. Then Matlab and calculation problems. The number of the stages can be found
7.0 is used for ANN simulation to supply a comparative study and, due to the balance between the complexity and the integration
to show the applicability and efficiency of ANN simulation for this of the problem (Ballou, 2001; Tsiakis & Papageorgiou, in press).
type of problem. By evaluating the output data, the SC network for Basically, the SC network design is realized by three stages. These
this case is designed, and also the optimal product flow between stages are shown in Fig. 2.
the factories, warehouses and distributors are calculated. Also it First stage contains the studies which provide the supply chain
is proved that the ANN simulation can be used instead of analytical processes efficiently managed. There are three important factors in
computations because of ensuring a simplified representation for the first stage: (1) emphasizing the suppliers which provide direct
this method and time saving. input to this chain, (2) converting this integrated design to a mod-
ern business application, and (3) taking action by reason of opera-
2. Supply chain network design tional decisions are being effected by the suppliers directly. Second

Network design problem that is about determining the ele-

ments, numbers, locations and physical flow quantities has a stra-
tegic importance for SCM. In this section of the paper, SC network Decision maker
design concept is detailed.
The term of supply chain implies that there is only one player for
each stage. However in practice, it is possible for a manufacturer to Defining the business processes, inputs, outputs
supply material from different companies and working with differ- and the data
ent distributors. For this reason, actually most of the supply chains
First stage
are networks (Chopra & Meindl, 2004).
A SC network is a complicated whole that contains suppliers, Developing the performance of suppliers,
manufacturers, distribution centers, retailers and the systems, manufacturers and distributors
sub systems, operations, activities that develop the supply chain
and the relations among them (Shapiro, 2001). SC network is a ser-
ies of processes and stages, which starts with the material/infor- Defining capacity, efficiency and the
Second stage
mation suppliers and ends with the customer as shown in Fig. 1 constraints of facility layout
(Tsiakis et al., 2001). Every mid-stage is the customer of the next
stage and supplier of the previous stage. This means that the par-
ticipants have different roles in the network; however the basic Establishing optimal assignment plan and
Third stage
relationship is seen between suppliers and customers (Wang, facility layout plan for supply chain
2009). The SC seems different from the various perspectives of
the companies, because every company supposes itself at the cen-
ter and thinks the structure and the participants of the network Fig. 2. The stages of supply chain network design.

Fig. 1. The stages of a supply network.

12572 A.T. Gumus et al. / Expert Systems with Applications 36 (2009) 12570–12577

stage implements the design by comparing supply and demand at SONFIN, FINEST, EfuNN (evolving fuzzy neural network), dmE-
every demand points. In the second stage, the input from suppliers FuNN, NEFCLASS (neuro-fuzzy classification), evolutionary design
is transformed into final product or service. The third stage of the of neuro-fuzzy systems and the others (Abraham & Nath, 2000).
SC network design includes operational functions which consists In this paper, ANFIS is used to eliminate demand uncertainty. For
of optimal source and plans for all the network points in a mini- this reason, a stochastic-neuro-fuzzy model is proposed with de-
mum cost assignment problem (Talluri & Baker, 2002). mand forecasting using neuro-fuzzy calculations.
The ANFIS is a multilayer feed forward network which uses neu-
3. The integrated neuro-fuzzy and MILP approach ral network learning algorithms and fuzzy reasoning to map an in-
put space to an output space (Chang & Chang, 2006).
In this section, firstly our integrated neuro-fuzzy and MILP For simplicity, we assume the fuzzy inference system under
methodology is introduced. Next, neuro-fuzzy approximation con- consideration has two inputs, x and y, and one output, z. For a
cept and ANFIS (adaptive-network-based fuzzy inference system) first-order Sugeno fuzzy model (Takagi & Sugeno, 1985), a typical
architecture is detailed, because of using it in demand forecasting. rule set with two fuzzy if–then rules can be expressed as:
Subsequently, the integrated neuro-fuzzy and MILP model and its Rule-1 : If xA1 and yB1 Then f 1 ¼ p1  x þ q1  y þ r 1
analytical equations are presented, and then the alternative ANN ð1Þ
Rule-2 : If xA2 and yB2 Then f 2 ¼ p2  x þ q2  y þ r 2
simulation model is proposed.
where pi, qi and ri (i = 1 or 2) are linear parameters in the then-part
3.1. The integrated neuro-fuzzy and MILP methodology (consequent part) of the first-order Sugeno fuzzy model. The archi-
tecture of ANFIS consists of five layers, and a brief introduction of
In this paper, an integrated methodology is proposed containing the model is as follows (Chang & Chang, 2006; Escoda, Ortega, Sanz,
neuro-fuzzy and MILP calculations. This methodology has two & Herms, 1997; Esen, Inalli, Sengur, & Esen, 2008; Jang, 1993;
steps: (1) Demand forecasting using neuro-fuzzy approximation, Taskin Gumus & Guneri, 2009).
(2) network design using MILP. MILP uses the outputs of neuro- Layer 1: Input nodes. Each node of this layer generates member-
fuzzy approximation (demand forecasts) as demand inputs. Also, ship grades to which they belong to each of the appropriate fuzzy
there are cost and capacity inputs that are obtained from the SC sets using membership functions.
network under consideration. The flow of the methodology pro-
posed here is shown in Fig. 3. O1;i ¼ lAi ðxÞ for i ¼ 1; 2
ð2Þ
O1;i ¼ lBi2 ðyÞ for i ¼ 3; 4
3.2. Neuro-fuzzy approximation
where x, y are the crisp inputs to node i, and Ai, Bi (small, large, etc.)
Both neural networks and fuzzy systems are dynamic, parallel are the linguistic labels characterized by appropriate membership
processing systems that estimate input–output functions (Taskin functions lAi and lBi , respectively.
Gumus & Guneri, 2009). They estimate a function without any Layer 2: Rule nodes. In the second layer, the AND operator is ap-
mathematical model and learn from experience with sample data. plied to obtain one output that represents the result of the ante-
A fuzzy system adaptively infers and modifies its fuzzy associa- cedent for that rule, i.e., firing strength. Firing strength means
tions from representative numerical samples. Neural networks, the degrees to which the antecedent part of a fuzzy rule is satisfied
on the other hand, can blindly generate and refine fuzzy rules from and it shapes the output function for the rule. Hence the outputs
training data (Kosko, 1991). Fuzzy sets are considered to be advan- O2;k of this layer are the products of the corresponding degrees
tageous in the logical field, and in handling higher order processing from Layer 1:
easily. The higher flexibility is a characteristic feature of neural
O2;k ¼ wk ¼ lAi ðxÞ  lBj ðyÞ; k ¼ 1; . . . ; 4; i ¼ 1; 2; j ¼ 1; 2 ð3Þ
nets produced by learning and, hence, this suits data-driven pro-
cessing better (Takagi, 1990). Hayashi and Buckley (1994) proved
that (1) any rule-based fuzzy system may be approximated by a Layer 3: Average nodes. In the third layer, the main objective is to
neural net and (2) any neural net (feed forward, multilayered) calculate the ratio of each ith rule’s firing strength to the sum of all
may be approximated by a rule-based fuzzy system (Mitra & Hay- rules’ firing strength. Consequently, w  i is taken as the normalized
ashi, 2000). firing strength:
Basic studies about neuro-fuzzy integration are GARIC, FALCON, w
 i ¼ P4 i
O3;i ¼ w ; i ¼ 1; . . . ; 4 ð4Þ
ANFIS (adaptive-network-based fuzzy inference system), FUN,
k¼1 wk

Layer 4: Consequent nodes. The node function of the fourth layer

computes the contribution of each ith rule’s toward the total out-
Characterization of the existing SC network put and the function defined as:

 i fi ¼ w
O4;i ¼ w  i ðpi x þ qi y þ r i Þ; i ¼ 1; . . . ; 4 ð5Þ

Estimation of distributor demands Obtaining cost and capacity data from where w  i is the ith node’s output from the previous layer. As for
by neuro-fuzzy approximation existing SC network
fpi ; qi ; r i g, they are the coefficients of this linear combination and
are also the parameter set in the consequent part of the Sugeno fuz-
zy model.
Aggregation and integration of the Layer 5: Output nodes. The single node computes the overall out-
inputs of the MILP model
put by summing all the incoming signals. Accordingly, the defuzz-
ification process transforms each rule’s fuzzy results into a crisp
Running the MILP model and designing the most cost output in this layer:
efficient SC network via analytical calculations and ANN
simulation comparatively X
4 P4
wi fi
O5;i ¼  i fi ¼ Pi¼1
w 4
ð6Þ
Fig. 3. The integrated neuro-fuzzy and MILP methodology. i¼1 i¼1 wi
 A.T. Gumus et al. / Expert Systems with Applications 36 (2009) 12570–12577 12573

3.3. The integrated MILP model 3.4. The ANN simulation model

This model is based on Yan, Yu, and Cheng’s (2003) model. This We simulate the proposed integrated model, including the
is a deterministic, static, multi-echelon MILP model with linear neuro-fuzzy module computing demand, to design and calculate
constraints, while Yan et al.’s (2003) contains logical constraints the approximate minimum total cost of the SC network with an
concerning material requirements, in addition to linear ones. Also, artificial neural network (ANN) here. The analytical computations
differently, the proposed model uses neuro-fuzzy approximation of the integrated neuro-fuzzy and MILP model are very complex
outputs as demand inputs. and time-consuming. Also, in analytical solution the neuro-fuzzy
The objective function to be minimized includes both the trans- forecast values are considered deterministically as the inputs of
portation costs from factories to warehouses and from warehouses the MILP model, while neuro-fuzzy forecasting can be realized
to distributors, fixed costs for factories and warehouses. The objec- during the ANN model simulation in an integrated manner.
tive function is set by the Eq. (7). The nomenclature is given in A quick estimate can be obtained by simulation. We use
Appendix A. ANNs here, because simulation can serve as a simplified repre-
The objective function: sentation of analytical model while neural network can serve
" # " # as a simplified representation of simulation model. Also, conven-
XX XX X X
Min C ij X ij þ C jk Y jk þ /i Ui þ d j Dj ð7Þ tional simulation software (ARENA, SLAM II, etc.) could not be
i j j k i j used because our model incorporated neuro-fuzzy forecasting.
ANNs are heavily used in the engineering and scientific fields to
The constraints of the model and their definitions are as below: model systems ranging from control systems to artificial intelli-
The constraints: gence (Guneri & Taskin Gumus, 2008; Taskin & Guneri, 2006).
X ANNs are networks of simple processing elements capable of pro-
1: X ij 6 ai Ui ; 8i ð8Þ
cessing information in response to external inputs (Badiru, 1992;
j
X Freeman & Skapura, 1991; Haykin, 1999; Hecht-Nielsen, 1989).
2: Y jk 6 bj Dj ; 8j ð9Þ The ANN we used is a multi layer perceptron (MLP) network,
k
X the most common neural network model. The network consists
3: Ui 6 F ð10Þ of an input layer, one or more hidden layers, and an output layer.
i
X Each layer computes a nonlinear activation function of a weighted
4: Dj 6 D ð11Þ sum of the layer’s inputs. The learning algorithm is the general-
j
X X ized delta rule, which ‘‘learns” by performing gradient descent
5: X ij  Y jk ¼ 0; 8j ð12Þ on the error surface (Jondarr, 1996; Rumelhart, Hinton, & Wil-
X
i k liams, 1987; Vysniauskas, Groen, & Kröse, 1993; Zurada, 1995).
6: Y jk P ck ; 8k ð13Þ The solution structure of the proposed ANN simulation model
j can be seen from Fig. 4.
7: Ui ; Dj ¼ f0; 1g; 8i;j ð14Þ
8: X ij ; Y jk P 0; 8i;j;k ð15Þ
4. A practical design case in alcohol free beverage sector

Here, a SC network design case is presented for a reputable

1. The first constraint is given by Eq. (8) and about factory capacity. multinational company in alcohol free beverage sector using
It implies that the product quantity which is distributed from a the proposed integrated neuro-fuzzy and MILP approach. First
factory to the warehouses cannot be more than the current of all, the existing supply network is described. Then, the multi
capacity of the factory. stage MILP model is run by using the inputs calculated via neu-
2. This is warehouse capacity constraint (Eq. (9)). Here, the product ro-fuzzy calculations and obtained from the existing network for
quantity which is distributed from a warehouse to the distribu- this case.
tors can’t be more than the current capacity of the warehouse.
3. According to number of factories constraint the number of the 4.1. Characterization of the existing supply network
opened factories can’t be more than 2, and it is shown by the
Eq. (10). The core business of the company observed is production, sales
4. This constraint, presented by Eq. (11), is about the number of and distribution of the carbonated and non carbonated soft drinks.
warehouses. Likewise, the number of the opened warehouses The sales network in Turkey is divided into 14 parts (e.g. Istanbul
cannot be more than 3. Asia, Istanbul Europe). Sales centres are established in these 14
5. Balance constraint for the first echelon mean that the total sales areas. There are 44 area sales managers (ASM) who manage
product quantity from factories to warehouses must be equal sales centres. Planning is realized for 16 different brands and with
to the total product quantity from warehouses to distributors their 223 stock keeping units (SKU).
(see Eq. (12)). Here, the SCM department is in coordination with all other
6. According to the balance constraint for the second echelon departments. The demand of distributors, key accounts and cold
(demand constraint) given by the Eq. (13), the capacity of the drink points are informed to the SC managers by ASMs. Then they
warehouses should be equal to or more than the demand of gather the information and evaluate them. This information is
the distributor calculated by neuro-fuzzy approximation. turned into production plans. After production, these plans become
7. Constraint of 0–1 integers. If the factories/warehouses are open, sales plans. Predictions are revised in 24 h to meet the goals of the
the value should be 1, otherwise it should be 0 (Eq. (14)). company. The structure of SC network is suitable for sales trends.
8. Constraint of cardinal numbers is given by the Eq. (15) and Also SC specialists make suggestions to sales department in order
implies that the quantity of the products that will be sent from to work with suitable profit margins.
the factories to the warehouse and from the warehouses to the When the company’s SC network is designed, SCM department
distributors should be cardinal numbers and also they should concentrates on the transportation costs, the demands of the dis-
be integers. tributors, the quotas of the areas and the decision of selection
12574 A.T. Gumus et al. / Expert Systems with Applications 36 (2009) 12570–12577

Design ANN simulation network structure

Ask the current values

of variables
Define the Define the
inputs and proposed
outputs of the integrated
ANN Call ANFIS forecast model
values for demand

Change the ANN parameters (eg. acitivation and transfer

functions, Iearning rate, etc.) and validate the results

Design optimal SC network and, calculate product quantities

flowing through the network and min. SC cost

Fig. 4. The structural steps of the proposed ANN simulation model.

the factory and the product which will be produced. Except the main computer in the same day at evening, invoices for the next
unusual circumstances, the factory or the warehouse that feeds day’s products which will be distributed are prepared and the pro-
an area is obviously known. However, when it is difficult to supply duction and distribution plans are evaluated. Moreover, dispatch-
the demand of the areas or when a problem occurs in the produc- ers are responsible for informing the distributors about price
tion, products are sent from another factory or from another amendments, handling the procedures against broken or defective
warehouse. products and delivering the new ones to the customers.
The quotas of the areas are constant. When these quotas are
multiplied by estimated demand, the quantity that will be sent 4.2. Running the model
to the warehouses is found. Products are tried to be sent to the dis-
tributors directly, instead of sending them firstly from factories to To run the model, firstly the transportation cost and capacity
city warehouses and then to the distributors in order to minimize data of existing SC network members are presented. Then distrib-
the costs. utor demand is forecasted using neuro-fuzzy approximation. And
Due to the estimated demand values are not exact; the best way finally, analytical and ANN simulation model results are compared
is working with a safety stock after defining the risk levels and and discussed via sensitivity analysis.
then the quantity for each warehouse. The relationship between
the production plan and demand and safety stock is calculated as:
4.2.1. Obtaining cost and capacity data from existing SC network
Demand from ASMs þ Safety Stock  Current Stock Here, a deterministic, static, multi-echelon MILP model is used
¼ Production Plan as mentioned before. 2 factories (F1, F2), 3 warehouses (W1, W2,
W3) and 6 distributors (D1, D2, D3, D4, D5, D6) are selected from
There are dispatchers who work for the company, and their job the company’s system in order to explain the existing design of
is determining the sales route, so they are helpful to enhance the the network.
SC’s impact. Also distribution routes are determined according to During this model application the product flow is followed by
the incoming orders. In addition to this, dispatchers visit the key considering only one product of the company. It can be seen from
account points periodically and take the orders with their barcode Fig. 5 that the product flows through 2 factories, 3 warehouses and
readers in their hands. Then, these orders are downloaded to the 6 distributors in our model.

Fig. 5. The SC network that consists of factories, warehouses and distributors.

 A.T. Gumus et al. / Expert Systems with Applications 36 (2009) 12570–12577 12575

Table 1 IF Price is low AND Quality is high AND NtwFlex is high THEN
The transportation costs from the factories to the warehouses (cent/case). Demand is very high,
Factories Warehouses IF Price is normal AND Quality is normal AND NtwFlex is high
W1 W2 W3 THEN Demand is normal,
............
F1 0.01 1.15 0.41
F2 1.15 0.01 0.74
The forecast values of ANFIS and ARIMA for each distributor de-
mand, and their MSE (mean square error) values are compared in
Table 5. The MSEs of ANFIS forecast values are found to be lower
Table 2 than ARIMA’s for all distributors in this case, and this can be seen
The transportation costs from the warehouses to the distributors (cent/case). from Fig. 6.
Warehouses Distributors
4.2.3. Analytical and simulation model results
B1 B2 B3 B4 B5 B6
During the implementation process 2 factories, 3 warehouses
W1 0.48 0.65 0.63 0.71 0.45 0.42 and 6 distributors of the reference company’s SC network are con-
W2 0.60 0.36 0.55 0.52 0.72 0.76
sidered in the model. An integrated neuro-fuzzy and MILP ap-
W3 0.13 0.25 0.17 0.39 0.06 0.05
proach is proposed to this network to realize the design
effectively. The neuro-fuzzy outputs are used in MILP model as in-
puts together with other input data. The model is solved in two
Table 3
ways; analytically and via ANN simulation, to show the applicabil-
The capacities and fixed costs of the factories and warehouses. ity and efficiency of ANNs in this type of design problem. The pur-
pose of this approach is making open/close decisions of the
Factories/Warehouses Capacity Fixed Costs (Cent/Case)
factories and warehouses, and calculating the quantity of produc-
F1 3,011,970 4.0 tion that flows from factories to warehouses and from warehouses
F2 1,298,716 3.9
to distributors with minimum cost.
W1 3,785,630 2.6
W2 1,564,479 2.4 The model is run using Lingo 10.0 for analytical computations
W3 346,094 2.3 and Matlab 7.0 for ANN simulation. The model outputs and results
can be seen from Tables 6 and 7. The results of the two methods are
observed in a comparative way, and it is decided that ANN simula-
The following tables (Tables 1–3) indicate us the costs, capaci- tion can be used to come out with reliable conclusion, although it is
ties and demands of the different elements in our model. We can seen that Lingo has better performance in this kind of model.
see the transportation costs from the factories to the warehouses Table 6 implies that the quantity of product that flows from F1
in Table 1, and from the warehouses to the distributors in Table to W1 should be 255.162. As it is seen in Tables 6 and 7, U1 equals
2. Capacities and fixed costs are given in Table 3 for the factories to 1, and this means that the first factory should be in progress in
and warehouses. the process.
According to the result of ANN simulation, the first and second
4.2.2. Estimation of distributor demands via neuro-fuzzy factories, and the first and second warehouses are open, but the
approximation
The network is trained with 84 demand data and then the de-
mand forecasts for each of the distributors are realized. According Table 5
to the test results, the outputs are close to real values with rather The forecasted demand values of the distributors and their MSEs.
low error. Also ARIMA is used to forecast demand with the same Distributors ANFIS MSEANFIS ARIMA MSEARIMA
data to make a comparison and test the success of neuro-fuzzy
D1 116,803 0.00298 120,573 0.00689
method.
D2 55,425 0.00387 54,380 0.00450
For neuro-fuzzy computations Matlab 7.0 Fuzzy Logic Toolbox D3 74,668 0.00305 73,965 0.00392
and ANFIS module, and for statistical analysis Eviews 3.0 are used. D4 9,660 0.00450 10,250 0.00597
Demand data membership functions for neuro-fuzzy computations D5 81,539 0.00359 82,832 0.00443
D6 56,820 0.00380 56,305 0.00389
can be seen from Table 4. There are 12 rules set as below:

0.008
Table 4
0.007
Demand data membership functions.
0.006
Membership functions
0.005
Demand Very low - [5000] (unit)
Low - [10,000] 0.004
Normal - [50,000] 0.003
High - [85,000]
Very High - [120,000] 0.002
Product unit price (Price) Low - [0 0.5 1.1] ($) 0.001
Normal - [0.95 1.25 1.3]
High - [1.25 1.35 1.4] 0
Product quality (Quality) Low - [0 2 4] D1 D2 D3 D4 D5 D6
Normal - [3 6 7]
High - [6 8 10]
MSEANFIS MSEARIMA
Network flexibility (NtwFlex) Low - [0 2 4]
High - [3 7 10]
Fig. 6. The MSEs of ANFIS and ARIMA forecast values.
12576 A.T. Gumus et al. / Expert Systems with Applications 36 (2009) 12570–12577

Table 6 0.05000
Analytical solution outputs (using Lingo 10.0).
0.04000
Variable Value Variable Value
0.03000
Flow of products
X 11 255,162 Y 21 0 0.02000
X 12 0 Y 22 0
0.01000
X 13 55,425 Y 23 19,243
X 21 0 Y 24 9660 0.00000
X 22 28,903 Y 25 0 D1 D2 D3 D4 D5 D6
X 23 0 Y 26 0
Y 11 116,803 Y 31 0 AnalyticalCalc . ANN Sim.
Y 12 0 Y 32 0
Y 13 0 Y 33 55,425
Fig. 7. The sensitivity analysis results for analytical and ANN simulation methods.
Y 14 0 Y 34 0
Y 15 81,539 Y 35 0
Y 16 56,820 Y 36 0
U1 1 D1 1
needed to be minimized are increased, too. All experiments have
U2 1 D2 1 global optimum after running the revised model. As seen from
D3 1 Fig. 6, maximum increase in the objective value is realized by the
Objective value 167,231 increase of the demand value for the first distributor demand con-
straint, for two of the methods. It means, the first constraint is the
most sensitive one over against changes and affects the objective
value intensively. The least increase is seen by the second distrib-
Table 7
ANN simulation outputs (using Matlab 7.0). utor demand value increase in analytical method, while it is by the
fourth distributor demand increase in ANN simulation. There are
Variable Value Variable Value
differences between objective value change amounts for two
Flow of products methods, but they are not far out from each other.
X 11 255,162 Y 21 0
Our model analyses the production flow among the members of
X 12 0 Y 22 75,450
X 13 0 Y 23 44,568 supply chain and aims network cost minimization. It should be
X 21 0 Y 24 9660 emphasized that our model contains limited echelons, data titles
X 22 129,753 Y 25 0 and decision criteria to realize the implementation process. But it
X 23 0 Y 26 0 is possible to expand the concept of SC and also to obtain more
Y 11 116,803 Y 31 0
Y 12 0 Y 32 0
beneficial information by increasing the amount of data and diver-
Y 13 0 Y 33 0 sifying the decision criteria.
Y 14 0 Y 34 0
Y 15 81,513 Y 35 0
5. Conclusion
Y 16 56,820 Y 36 0
U1 1 D1 1
U2 1 D2 1 It is known that now the competition is not between the com-
D3 0 panies, competition is between the SCs that consist of several com-
Objective value 182,021 panies. As we think that the supply network efficiency is related to
the weakest participant’s efficiency, the system should be im-
proved by handling the problems with an integrated point of view
third warehouse is closed. On the other hand analytical method for flexibility and efficiency. The basic priority for managing the SC
gives us a solution in which all the factories and the warehouses in the right way is designing the SC network properly.
are open. The cost of Warehouse 3 seems less than the others. In this study, an integrated SC design model is developed and a
However the cost of transportation from the factories to the ware- SC network design case is examined for a reputable multinational
house is higher and it has a lower capacity. Because of being lo- company in alcohol free beverage sector. Here, a three echelon SC
cated near to the factories for Warehouse 1 and Warehouse 2, network is considered and the proposed integrated neuro-fuzzy
the transportation cost from factories to the warehouses costs low- and MILP approach is applied to this network to realize the design
er. For this reason, it is considered to close the Warehouse 3. While effectively. Matlab 7.0 is used for neuro-fuzzy demand forecasting
ANN simulation finds 182,021 $ for the minimum cost, analytical and, the deterministic, static, multi-echelon MILP model is solved
method’s result is 167,231 $. It should be remembered that the using Lingo 10.0. Then Matlab 7.0 is used for ANN simulation to
alterations in the fixed costs or in capacities change the model supply a comparative study and, to show the applicability and effi-
results. ciency of ANN simulation for this type of problem. By evaluating
If it is needed to make a sensitivity analysis to detail the MILP the output data, the SC network for this case is designed, and also
model structure, a variable must be changed while the others are the optimal product flow between the factories, warehouses and
constant. Demand is considered and emphasized as to be uncertain distributors are calculated. Also it is proved that the ANN simula-
in this paper. For this reason, the right-hand side of the constraints tion can be used instead of analytical computations because of
under ‘‘balance constraint for the second echelon (demand con- ensuring a simplified representation for this method and time
straint)” title is changed alternately, and then the changes in the saving.
objective value are observed. There are six demand constraints be- Our implementation considers only a part of the reference com-
cause of six distributors. So there are six sensitivity analysis results pany’s supply chain system; however it can enlighten us about the
that are gained by changing the demand values on right-hand side whole system. To conclude, it can be said that, it is possible to ex-
of the constraints, consecutively. The objective value changes are pand the supply chain concept and having more accurate results
shown in Fig. 7 by demand constraint changes, in a comparative about the whole system by increasing the amount of data and
manner. diversifying the decision criteria. Also, there is uncertainty for
The demand values at the right-hand sides of the constraints are the cost and capacity variables in addition to demand. So, these
increased by the rate of 10%. Then the objective values that are variables can be calculated by benefiting from statistics, neuro-fuz-
 A.T. Gumus et al. / Expert Systems with Applications 36 (2009) 12570–12577 12577

zy approximation or another high-performance heuristic to ensure Gunasekaran, A. (1999). Agile manufacturing: A framework for research and
development. International Journal of Production Economics, 62(1–2), 87–105.
more realistic results.
Guneri, A. F., & Taskin Gumus, A. (2008). The usage of artificial neural networks for
finite capacity planning. International Journal of Industrial Engineering – Theory,
Appendix A. List of notation Applications and Practice, 15(1), 20–33.
Hayashi, Y., & Buckley, J. J. (1994). Approximations between fuzzy expert systems
and neural networks. International Journal of Approximate Reasoning, 10, 63–73.
ai the capacity of factory i Haykin, S. (1999). Neural networks: A comprehensive foundation (2nd ed.). New
bj the capacity of warehouse j Jersey: Prentice Hall.
Hecht-Nielsen, R. (1989). Neurocomputing. MA: Addison-Wesley, Reading.
ck the demand of distributor k Huang, S. H., Sheoran, S. K., & Keskar, H. (2005). Computer-assisted supply chain
C ij the cost of transportation from factory i to warehouse j configuration based on supply chain operations reference (SCOR) model.
C jk the cost of transportation from warehouse j to distributor k Computers and Industrial Engineering, 48(2), 377–394.
Jang, J. S. R. (1993). ANFIS: Adaptive-network-based fuzzy inference system. IEEE
D maximum total number of warehouses Transactions on Systems, Man and, Cybernetics, 23(3), 665–685.
F maximum total number of factories Jondarr, C. G. H. (1996). Back propagation family album. Technical Report C/TR96-05,
X ij the quantity of the product from factory i to warehouse j Department of Computing, Macquarie University.
Kosko, B. (1991). Neural networks and fuzzy systems. NJ: Prentice Hall, Englewood
Y jk the quantity of the product from warehouse j to distributor k
Cliffs.
ui fixed cost of factory i Lambert, D. M., & Cooper, M. C. (2000). Issues in supply chain management.
dj fixed cost of warehouse j Industrial Marketing Management, 29(1), 65–83.
Lee, H. L., Padmanabhan, V., & Whang, S. (1997). Information distortion in a supply
chain: The bullwhip effect. Management Science, 43(4), 546–558.
 Mitra, S., & Hayashi, Y. (2000). Neuro-fuzzy rule generation: Survey in soft
1; if factory is in use
Ui ¼ computing framework. IEEE Transactions on Neural Networks, 11(3), 748–768.
0; other Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1987). Learning internal
representations by error propagation. In D. E. Rumelhart & J. L. McClelland
(Eds.), Parallel distributed processing (pp. 318–362). Cambridge, MA: MIT Press.
vol. 1.
Samadhi, T. M. A. A., & Hoang, K. (1998). Partner selection in a shared partner
 selection in a shared CIM system. International Journal of Computer Integrated
1; if warehouse is in use Manufacturing, 11(2), 173–182.
Dj ¼
0; other Santoso, T., Ahmed, S., Goetschalckx, M., & Shapiro, A. (2005). A stochastic
programming approach for supply chain network design under uncertainty.
European Journal of Operational Research, 167, 96–115.
Shapiro, J. F. (2001). Modeling the supply chain. CA: Duxbury Thomson Learning Inc..
Takagi, H. (1990). Fusion technology of fuzzy theory and neural network – Survey
References and future directions. In Proceedings of international conference on fuzzy logic
neural networks, Iizuka, Japan (pp. 13–26).
Abraham, A., & Nath, B. (2000). Designing optimal neuro-fuzzy systems for Takagi, T., & Sugeno, M. (1985). Fuzzy identification of systems and its applications
intelligent control. In Proceedings of the sixth international conference on to modeling and control. IEEE Transactions on Systems, Man and Cybernetics, 15,
control automation robotics computer vision (ICARCV 2000), Singapore. 116–132.
Altiparmak, F., Gen, M., Lin, L., & Karaoglan, I. (2009). A steady-state genetic Talluri, S., & Baker, R. C. (2002). A multi phase mathematical programming approach
algorithm for multi-product supply chain network design. Computers and for effective supply chain design. European Journal of Operational Research,
Industrial Engineering, 56(2), 521–537. 141(3), 544–558.
Badiru, A. B. (1992). Expert systems applications in engineering and manufacturing. NJ: Taskin Gumus, A., & Guneri, A. F. (2009). A multi-echelon inventory management
Prentice Hall. framework for stochastic and fuzzy supply chains. Expert Systems with
Ballou, R. H. (2001). Unresolved issues in supply chain network design. Information Applications, 36(3), 5565–5575.
Systems Frontiers, 3(4), 417–426. Taskin Gumus, A., & Guneri, A. F. (2007). Multi-echelon inventory
Beamon, B. M. (1998). Supply chain design and analysis: Models and methods. management in supply chains with uncertain demand and lead times:
International Journal of Production Economics, 55(3), 281–294. Literature review from an operational research perspective. Proceedings of
Bhaskaran, K., & Leung, Y. T. (1997). Manufacturing supply chain modelling and the Institution of Mechanical Engineers Part B: Journal of Engineering
reengineering. Sadhana, 22, 165–187. Manufacture, 221(10), 1553–1570.
Cakravastia, A., Toha, I. S., & Nakamura, N. (2002). A two-stage model for the design Taskin, A., & Guneri, A. F. (2006). Economic analysis of risky projects by ANNs.
of supply chain networks. International Journal of Production Economics, 80(3), Applied Mathematics and Computation, 175(1), 171–181.
231–248. Tsiakis, P., & Papageorgiou, L. G. (in press). Optimal production allocation and
Chang, F. J., & Chang, Y. T. (2006). Adaptive neuro-fuzzy inference system for distribution supply chain networks. International Journal of Production
prediction of water level in reservoir. Advances in Water Resources, 29(1), 1–10. Economics.
Chauhan, S. S., Nagi, R., & Proth, J.-M. (2004). Strategic capacity planning in supply Tsiakis, P., Shah, N., & Pantelides, C. C. (2001). Design of multi-echelon supply chain
chain design for a new market opportunity. International Journal of Production networks under demand uncertainty. Industrial and Engineering Chemistry
Research, 42(11), 2197–2206. Research, 40, 3585–3604.
Chopra, S., & Meindl, P. (2004). Supply chain management: Strategy, planning and Vysniauskas, V., Groen, F. C. A., & Kröse, B. J. A. (1993). The optimal number of learning
operation. Prentice Hall. samples and hidden units in function approximation with a feedforward network.
Escoda, I., Ortega, A., Sanz, A., & Herms, A. (1997). Demand forecast by neuro-fuzzy Technical Report CS-93-15, University of Amsterdam, November.
techniques. IEEE International Conference on Fuzzy Systems, 3, 1381–1386. Wang, G., Huang, S. H., & Dismukes, J. P. (2004). Product-driven supply chain
Esen, H., Inalli, M., Sengur, A., & Esen, M. (2008). Modelling a ground-coupled heat selection using integrated multi-criteria decision-making methodology.
pump system using adaptive neuro-fuzzy inference systems. International International Journal of Production Economics, 91(1), 1–15.
Journal of Refrigeration, 31(1), 65–74. Wang, H. S. (2009). A two-phase ant colony algorithm for multi-echelon defective
Fisher, M. L., & Raman, A. (1996). Reducing the cost of demand uncertainty through supply chain network design. European Journal of Operational Research, 192(1),
accurate response to early sales. Operations Research, 44(1), 87–99. 243–252.
Freeman, J. A., & Skapura, D. M. (1991). Neural networks: Algorithms, applications, and Yan, H., Yu, Z., & Cheng, T. C. E. (2003). A strategic model for supply chain design
programming techniques. MA: Addison-Wesley, Reading. with logical constraints: Formulation and solution. Computers and Operations
Graves, S. C., & Willems, S. P. (2005). Optimizing the supply chain configuration for Research, 30, 2135–2155.
new products. Management Science, 51(8), 1165–1180. Zurada, J. M. (1995). Introduction to artificial neural systems. Boston: PWS Publishing.

View publication stats