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

net/publication/265016049

A genetic algorithm approach for multi-product multi-period continuous

review inventory models

Article  in  Expert Systems with Applications · December 2014

DOI: 10.1016/j.eswa.2014.07.003

CITATIONS READS

31 151

3 authors:

Ilkay Saracoglu Seyda Topaloglu

Istanbul Atlas University Dokuz Eylul University
7 PUBLICATIONS   36 CITATIONS    28 PUBLICATIONS   730 CITATIONS   

SEE PROFILE SEE PROFILE

Timur Keskintürk
Istanbul University
18 PUBLICATIONS   167 CITATIONS   

SEE PROFILE

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

RG Academic Publishers & Reviewers View project

Workcell Formation - Worker Assignment- Constraint Programming View project

All content following this page was uploaded by Ilkay Saracoglu on 06 May 2020.

The user has requested enhancement of the downloaded file.

 Expert Systems with Applications 41 (2014) 8189–8202

Contents lists available at ScienceDirect

Expert Systems with Applications

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

A genetic algorithm approach for multi-product multi-period

continuous review inventory models
Ilkay Saracoglu a,⇑, Seyda Topaloglu b, Timur Keskinturk c
a
Graduate School of Natural and Applied Sciences, Dokuz Eylul University, Tinaztepe Campus, 35160 Izmir, Turkey
b
Department of Industrial Engineering, Dokuz Eylul University, Tinaztepe Campus, 35160 Izmir, Turkey
c
Department of Quantitative Methods, School of Business, Istanbul University, Avcilar, 34180 Istanbul, Turkey

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

Article history: This paper formulates an approach for multi-product multi-period (Q , r) inventory models that calculates
Available online 14 July 2014 the optimal order quantity and optimal reorder point under the constraints of shelf life, budget, storage
capacity, and ‘‘extra number of products’’ promotions according to the ordered quantity. Detailed
Keywords: literature reviews conducted in both fields have uncovered no other study proposing such a multi-
Multi-product multi-period policy product (Q , r) policy that also has a multi-period aspect and which takes all the aforementioned con-
Continuous review policy straints into consideration. A real case study of a pharmaceutical distributor in Turkey dealing with large
Inventory management
quantities of perishable products, for whom the demand structure varies from product to product and
Variable demand
shows deterministic and variable characteristics, is presented and an easily-applicable (Q , r) model for
distributors operating in this manner proposed. First, the problem is modeled as an integer linear
programming (ILP) model. Next, a genetic algorithm (GA) solution approach with an embedded local
search is proposed to solve larger scale problems. The results indicate that the proposed approach yields
high-quality solutions within reasonable computation times.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction the system. Harris (1913) first estimated the economic order quan-
tity to minimize the total cost of keeping stocks and placing orders.
Customer satisfaction has recently become one of the most A model was developed for a single product without any
important issues for companies as a consequence of globalization, constraints on the assumption that demand is known and lead
increased competition, and shrinking profit margins. The most fun- time is zero.
damental condition associated with customer satisfaction is the However, in many real inventory systems, there is usually
availability of products upon request. In order for companies to uncertainty with regard to demand and lead time. Continuous
satisfy this condition, excellent inventory management is essential. review inventory policy (Q , r) is one of the most common policies
Much research has been conducted on inventory management, for cases in which the demand is uncertain and a supply lead time
and results have shown that the type of inventory management exists. In this policy, an order quantity Q is placed when the inven-
needed varies according to the nature of the companies. To con- tory level is at or below reorder level r. The (Q , r) policies applied
struct an inventory management policy, it is therefore necessary on single items with fixed lead time have been studied by
to conduct an in-depth analysis of the nature of the target com- researchers such as Hadley and Whitin (1963), Federgruen and
pany. The inventory manager must determine (1) when an order Zheng (1992), Zheng (1992), Axsäter (1993), Kim and Benton
should be placed, and (2) how much should be ordered. Further, (1995), Axsäter (2007), Lau and Lau (2008), Mattsson (2010),
inventory, production, marketing, and finance managers should Hajiaghaei-Keshteli, Sajadifar, and Haji (2011), and Ang, Song,
operate in concert in order to reach an agreement on how to Wang, and Zhang (2013).
reduce production costs and inventory investments, and increase In actual inventory systems, however, multi-item storage is an
customer responsiveness. issue typically encountered. For instance, wholesalers, large
Good inventory management is closely linked to the inventory distribution chains, and department stores manage an inventory
policy implemented, and is based upon a profound analysis of system comprising a wide range of products. Garman (1976),
Gardner (1983), Hopp, Spearman, and Zhang (1997) formulated
⇑ Corresponding author. Tel.: +90 5324520069; fax: +90 2123665802. multi-product (Q , r) inventory policy with some constraints such
E-mail address: [email protected] (I. Saracoglu). as inventory investment, service level and order frequency. Jeddi,

http://dx.doi.org/10.1016/j.eswa.2014.07.003
0957-4174/Ó 2014 Elsevier Ltd. All rights reserved.
8190 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202

Shultes, and Haji (2004) analyzed multi-product stochastic inven- to the system as a base and provides purchase of needed amount
tory system with backorders, shortages, and budget constraints. and determines order time by considering lead time, whereas
Betts and Johnston (2005) constructed a multi-item (Q , r) model DRP system performs the distribution planning for the customers.
with the objective of determining an optimal replenishment policy While advanced planning and scheduling (APS) models are embed-
by reducing risks and increasing profit. ded to SAP to perform optimization using mathematical program-
As in most of the inventory management problems, unfortu- ming techniques, these models work under the assumption that all
nately, the model is naturally complex for the multi-product case parameter values are constant and known absolutely (Kallrath &
and it is intractable to control optimal policies analytically due to Maindl, 2006, Louly & Dolgui, 2013), they do not address the
the size of the solution domain. Therefore, it is observed that iter- specific inventory management problem in this study.
ative procedures such as heuristics and metaheuristics are used in The remainder of this paper is organized as follows: Literature
order to reach a solution in multi-product inventory systems. review is presented in Section 2. Problem description and assump-
Pasandideh, Niaki, and Tokhmehchi (2011) developed a multi-item tions, and our proposed mathematical model are given in Section 3.
(Q , r) inventory policy for a two-echelon inventory system by using A GA based solution approach is developed to solve the problem in
genetic algorithm (GA). Pasandideh, Niaki, and Nia (2011) Section 4. An illustrative computational case study is given in Sec-
proposed a multi-product EOQ model which does not only consider tion 5. In order to demonstrate how the proposed methodology is
the storage capacity, but also assumes the number of orders lim- applied, a real case study and sensitivity analysis are carried out in
ited, and they used GA to solve it. Mandal, Maity, Maity, Mondal, Section 6. Finally, conclusions and recommendations for future
and Maiti (2011) formulated multi-item multi-period production research are provided in Section 7.
problems with fuzzy-random parameters and space constraints
and solved using GA. Yang, Chan, and Kumar (2012) developed a
GA approach for solving the multi-retailer, multi-product and 2. Review of current literature
multi-period inventory system with backlogging and transporta-
tion capacity. Mousavi, Hajipour, Niaki, and Alikar (2013) In this paper, a multi-item perishable inventory management
suggested a heuristic using GA and simulated annealing for the system is proposed. Many products may lose their market value
multi-item multi-period inventory control problem with over time as they become outdated, called perishability. Perishabil-
discounts, time value of money, and inflation. ity is divided into two categories: fixed and random lifetime. Fixed
This paper analyzes a pharmaceutical distributor that buys lifetime is used for items that have a specific physical lifetime, such
medicines from pharmaceutical manufacturers and resells them as photographic film, drugs, vitamins, and ready packed food and
to pharmacies. The objective of the distributor is to meet the milk. Random lifetime is used for items whose spoilage time
demands of the pharmacies on time. Because a large number of cannot be estimated, such as fresh vegetables, fruits, and meats.
distributors is available in the sector, competition is intense. The fixed lifetime and exponential decay inventory study per-
Consequently, customer satisfaction is viewed with the utmost formed by Van Zyl (1964) is among the first perishable items study
priority. However, because distributors have to operate within conducted. Nahmias (1982), Nahmias (2011, 1991) and Goyal and
the confines of budget constraints, they need to make optimum Giri (2001) reviewed the current literature on ordering policies for
use of their budgets. In addition to the budget, shelf life and storage perishable items.
area – for drugs that require cold storage – are additional Continuous review policy (Q , r) has been used for nonperishable
constraints that have to be satisfied. Further, pharmaceutical man- items in a number of studies. The first such study was conducted
ufacturers may desire to deliver an additional amount of drugs for by Berk and Gurler (2008). The results of our literature survey per-
free to induce increases in the number of medicines sold in the formed to date show the (Q , r) inventory model being used as sin-
pharmacies, instead of offering discounts on the basis of quantity gle period in multi-product problems and for nonperishable
ordered. A further item to bear in mind is the fact that pharmaceu- products, as illustrated in Table 1.
tical manufacturers may provide convenient payment options to Zhao, Fan, Liu, and Xie (2007) tackled optimization with
the distributor who, in turn, may pass on these payment options storage-space constraint for the single-item (Q , r) policy, and pro-
to the pharmacies. Regarding the pharmaceutical sector, Bijvank posed an efficient algorithm with polynomial time computational
and Vis (2012) provided an overview on the literature for hospital complexity that obtains the optimal solutions. Then, they extended
inventory systems. Lapierre and Ruiz (2007) modeled the their proposed method to enhanced multi-item (Q , r) policies.
multi-product, multi-period logistics system as a mixed-integer Wang and Hu (2010) studied a continuous review (Q , r) model to
program under the limitation of storage and manpower capacities, find the optimal lot size and reorder point for a multi-item inven-
and developed a tabu search to solve the problem owing to the tory model with interactions among items under budgetary con-
large scale of the problem. straints. They formulated the problem as a nonlinear one and
This paper constructs a multi-item multi-period (Q , r) inventory constructed a simple heuristic approximating procedure to solve
policy considering items with variable demand and determines the it. Kundu and Chakrabarti (2012) analyzed uncertain lead time
optimal order quantity (Q) and reorder point (r) for each item to and demand condition for multi-item production. They created a
maximize the profit for the pharmaceutical distributor in question, mathematical model that can determine the safety factor and
considering all these constraints. order quantity and proposed a continuous review model for a
When we examine the companies in the pharmaceutical sector, multi-product system with both backorders and lost sales under
it can be seen that they are using a particular database and the budget constraints. Zhao, Qiu, Xie, and He (2012) investigated sin-
functionality of their systems is no more than keeping track of past gle- and multi-item (Q , r) inventory systems with a stochastic
sales and current stock information. Yet they are not capable of demand, limited resources, constant lead time, and backorders.
performing an optimization like the one we are suggesting in this We discovered from our literature review that the (Q , r) inven-
study. To give an example, among one of the most developed tory policy has not yet been carried out as multi-item multi-period.
Enterprise Resource Planning (ERP) programs, SAP uses Materials The multi-period studies reviewed primarily dealt with lot-sizing
Requirements Planning (MRP) and Distribution Requirements problems in production. Veinott (1965) considered a dynamic non-
Planning (DRP). MRP makes calculation for the dependent demand stationary multi-product inventory model in which the system was
by taking the input value (such as days of supply, lead time, fore- reviewed in equal-length time periods, unsatisfied demand was
cast, Master Production Scheduling (MPS), safety stock) introduced backlogged, and deterioration of stock did not occur in storage.
 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202 8191

Table 1
Published literature dealing with the multi-item (Q, r) inventory policy.

Author, year Budget Service Storage Inventory Order Backorder Resource Extra Lead Lifetime Demand Method
level investment frequency constraint quantity time
Multi item
Single period
Nonperishable item
Garman (1976) ⁄ F St L/H
Gardner (1983) ⁄ F D L/H
Hopp et al. (1997) ⁄ ⁄ F St L/H
Jeddi et al. (2004) ⁄ F St L/H
Betts and Johnston ⁄ F D/St A
(2005)
Zhao et al. (2007) ⁄ F St St/H
Wang and Hu (2010) ⁄ F St H
Pasandideh et al. ⁄ ⁄ ⁄ F D GA
(2011)
Kundu and ⁄ R St L
Chakrabarti
(2012)
Zhao et al. (2012) ⁄ F St I/LS
Multi period
Perishable item
Developed model in ⁄ ⁄ ⁄ F F St, D, Se GA/LS
this paper

St: Stochastic; Se: Seasonal; D: Deterministic; F: Fixed; A: Analytical; H: Heuristics; I: Iterative Procedure; L: Lagrangian; GA: Genetic Algorithm; LS: Local Search; R: Random.

The studies dealing with multi-item multi-period models are and ‘‘extra number of products’’ promotions constraints. Because
summarized in Table 2. Zhang and Eksioglu (2009) performed the of the large number of decision variables and constraints in the
sole multi-period study for perishable items. In their study, they proposed model, it is difficult to solve this model via exact meth-
presented a mixed-integer programming (MIP) model with ods, even for small problems. Consequently, to find a multi-item
multi-mode replenishment options and perishable inventory that multi-period (Q , r) inventory policy that tries to maximize the total
can be stored for a limited number of periods and considered a sin- profit, we propose a GA solution approach.
gle item.
The primary objective of this paper is to develop a novel integer 3. Problem statement and formulation
linear programming (ILP) model for the multi-product multi-per-
iod perishable products (Q , r) inventory problem with determinis- We focus on a pharmaceutical distributor that collects products
tic and variable demand and under budget, storage space, shelf life, from pharmaceutical manufactures and distributes them to

Table 2
Published literature dealing with multi-item multi-period inventory models.

Author, year Lifetime Demand Excess Method Policy Lead Budget Storage Quantity Resource Lifetime Extra
demand time discount constraint quantity
Multi-item + multi-
period
Nonperishable
Veinott (1965) St Backlogging St Optimal 0
ordering
Gen and Syarif D No HGA Optimal 0 ⁄
(2005) shortages ordering
Kannan et al. (2010) D No GA Supply F ⁄ ⁄
shortages chain
Mandal et al. (2011) D Shortages GA Production 0 ⁄
rate
Pasandideh et al. D No GA Optimal 0 ⁄ ⁄
(2013) shortages ordering
Mousavi et al. D No GA/SA Optimal 0 ⁄ ⁄ ⁄
(2013) shortages ordering
Single item + multi
period
Perishable
Zhang and Eksioglu D No H Lot-size F ⁄
(2009) shortages
Multi-item + multi-
period
Perishable
Developed model in F St, Se, D Shortages GA/LS (Q ,r) F ⁄ ⁄ ⁄ ⁄
this paper

St: Stochastic; Se: Seasonal; D: Deterministic; F: Fixed; H: Heuristics; GA: Genetic Algorithm; LS: Local Search; R: Random; HGA: Hybrid Genetic Algorithm.
8192 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202

pharmacies. This distributor wants to determine when to make an mdi expected mean demand during the
order and how much to order to meet the customer needs in a way lead time for product i.
that maximizes profit given certain budgetary constraints. Some li lead time for product i.
products require cold storage; therefore, storage area constraint spi sales price for product i.
exists for these products. Further, manufacturing companies may ppi unit purchasing price for product i.
offer extra quantities as promotion according to the given order csi ordering cost for product i, per order.
quantity because of competition. In addition, the products have sci unit shortage cost for product i.
specific lifetimes. As a result, the FIFO rule is applied, that is, the hi unit holding cost for product i, per
product already on hand is sold prior to the sale of more recently period.
delivered products. Because demand varies by product and the Ie interest rate per period.
demand for some products is uncertain or seasonal, a multi-prod- Mi the warehouse’s payment term offered
uct multi-period (Q , r) model has to be applied. Accordingly, a spe- by manufacturer in unit time for
cific order quantity (Qi) is given whenever the stock level for a product i.
product i is at or below level ri in this system. The order arrives Ni the customer’s payment term offered
at the storage with a delay of a certain lead time. If there is a pro- by distributor in unit time for product i.
motion on an order quantity Qi, the quantity that is added to our wi storage space required for product i.
storage is then ‘‘Qi + extra quantityi.’’. I1i,0 initial inventory level for product i.
We make the following assumptions in expressing our multi- shfi lifetime period for product i.
product multi-period (Q , r) inventory problem mathematically: upper_ri reorder point upper bound for product
i.
 There are multiple products. upper_Qi order quantity upper bound for
 There is only one warehouse. product i.
 Demand is deterministic, seasonal, or variable according to Uk upper break point for quantity range k.
products. Lk lower break point for quantity range k.
 Lead time is fixed and constant. EXi,k extra goods for quantity range k of
 There is a storage limit for products that require cold storage. product i.
 A finite budget limitation is enforced by the bank. B total available budget.
 Backorders are not allowed. storage total storage capacity for all products.
 Shortages are allowed. BigM big number.
 Depending on the size of the order, extra quantities are supplied Decision variables
by the manufacturers as ‘‘promotions.’’ Ii,t ending inventory level for product i in
 The planning horizon is finite and planning is periodical. period t.
 The manufacturers and customers both have payment terms, I1i,t positive inventory level for product i in
which are different for each product. period t.
 Lifetime is fixed and constant, and changes according to the I2i,t shortage quantity of product i at the
product. end of period t.
Qi,t order quantity for product i in period t.
3.1. Integer linear programming model for multi-product multi-period QSi order quantity for product i.
(Q , r) policy ri reorder point for product i.
Yi,t binary variable: equal to one if an order
The problem at hand is to formulate a model that maximizes is placed in period t; otherwise, zero.
the total profit subject to constraints on budget, storage area, shelf Ai,t binary variable: equal to one if the
life, and various promotions. The following notations are used to beginning inventory level of product i
derive the ILP formulation of the proposed multi-product multi- is less than the reorder point in period
period (Q , r) continuous review inventory model. t, and zero if the beginning inventory
level is greater than the reorder point,
3.1.1. Notations indicating whether an order should be
The sets, indices, parameters, and decision variables used in the placed.
model are given as follows: Bi,t binary variable: equal to one if the
ending inventory level of product i in
Sets and indices period t is positive, and zero if the
N number of products, where, i = 1, 2, . . ., inventory level is negative.
N. Di,t,k binary variable: equal to one if a
T number of planning periods, where, purchase is made in period t for
t = 1, 2, . . ., T. product i, and zero if no purchase is
K number of break points for extra made in period t for product i, with
quantity, where, k = 1,2, . . . ,K. quantity break k.
PC set of products that need cold storage
environment.
Parameters
Qi,t,E extra order quantity for product i in
3.1.2. Objective function
period t.
Our aim is to maximize profit; therefore, the objective function
xi,t expected demand for product i in
is the total of the sales revenue and money earned as a result of the
period t.
payment terms provided by the manufacturers, minus the total
 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202 8193

purchasing, holding, ordering, and shortage costs, and money lost If Qi,t > 0, then an order has been placed. Constraint (9) is used to
because of the payment terms provided to customers. determine whether an order has been placed or not. The total order-
Consequently, the objective function of the multi-product, ing cost (TOC) is the product of the number of orders given during
multi-period (Q , r) policy model is given in the equation below: the total planning period and the ordering cost.
Expected Total Shortage cost:
Maximize Total Profit ¼ Expected sales rev enue
þ Earnings resulting from payment term prov ided by manufacturer X
N X
T
ETSC ¼ sci  I2i;t ð11Þ
 Total cost  Loss resulting from payment term prov ided to customer
i¼1 t¼1
ð1Þ
Total Cost ¼ Purchasing cost þ Holding cost þ Ordering cost If the demand is higher than the level of the reorder point in period
t, then there can be a shortage quantity, in which case, the expected
þ Shortage cost ð2Þ
total shortage cost (ETSC) is the product of the expected shortage
MaxZ ¼ ðQ i ; r i Þ ð3Þ
quantity and the unit shortage cost.
The components of the objective function can be calculated as the
following relations: 3.1.3. Constraints
Expected sales revenue: The constraints of the model can be grouped into six sets:
budget, storage area, lifetime, inventory balance equations, and
X
T X
N X
T X
N
ETSR ¼ xi;t  spi  I2i;t  spi ð4Þ extra quantity. These constraints are explained in detail in the
t¼1 i¼1 t¼1 i¼1 ensuing subsections.
Inventory balance equations:
The total sales revenue obtained during the planned T periods is the
product of the demand quantity and unit sales price. If the incoming If
demand is not completely satisfied, sales loss will occur as a result Ii;t1 6 r i i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð12Þ
of the unmet demand. Expected sales revenue is calculated by
Then
subtracting this sales loss from the sales revenue.
Earnings resulting from the payment term provided by Q i;t þ Q i;t;E  I1i;t þ I2i;t þ I1i;t1  xi;t ¼ 0
manufacturer: i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð13Þ
X
N X
T Else
TPFS ¼ Q i;t  ppi  Ie  M i ð5Þ I1i;t1  I1i;t þ I2i;t  xi;t ¼ 0 i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð14Þ
i¼1 t¼1
and
Manufacturing companies may provide distributors with conve-
nient payment terms for purchased products. When the manufac- Q i;t ¼ 0 i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð15Þ
turer company provides the distributor with the ease to make the I1i;t1  r i þ 1 6 BigM  ð1  Ai;t Þ i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð16Þ
payment after the sales of the product, the distributor indicates ri  I1i;t1 6 BigM  Ai;t i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð17Þ
the interest income for not having to make the payment immedi- Q i;t þ Q i;t;E  I1i;t þ I2i;t þ I1i;t1  xi;t P BigM  ð1  Ai;t Þ
ately. The earnings that the distributor will obtain during the
i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð18Þ
planning period as a result of the payment terms is calculated as
in Eq. (5). Q i;t þ Q i;t;E  I1i;t þ I2i;t þ I1i;t1  xi;t 6 BigM  ð1  Ai;t Þ
Purchasing cost: i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð19Þ
 
X
T X
N I1i;t1  I1i;t þ I2i;t  xi;t P BigM  1  Ai;t
TPC ¼ Q i;t  ppi ð6Þ
i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð20Þ
t¼1 i¼1  
Each period, consideration is given to whether an order should be
I1i;t1  I1i;t þ I2i;t  xi;t 6 BigM  1  Ai;t
placed or not and, if an order is placed, the total purchasing cost i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð21Þ
that will accrue during the planning period is equal to the product Q i;t 6 BigM  Ai;t i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð22Þ
of the unit sales price and order quantity.
Loss resulting from the payment term provided to customer: According to the relations given in Constraints (12)–(15), if the
stock level at the beginning of a period is lower than the reorder
X
N X
T
TLPFC ¼ xi;t  spi  Ie  Ni ð7Þ point implied by Constraint (12), an order quantity Qi,t is placed
i¼1 t¼1 and the inventory balance equation is enforced by Constraint (13).
A distributor may also provide customers with convenient payment According to Constraint (13), for product i, inventory level Ii,t at
terms similar to those provided by the suppliers of the distributor. the end of period t is found by adding order quantity Qi,t and extra
In this case, the distributor’s loss is the product of the sales amount, quantity Qi,t,E to inventory level Ii,t1 at the beginning of the period,
payment term, and interest. and subtracting expected demand xi,t. Conversely, if the stock level
Holding cost: at the beginning of a period is greater than the reorder point, the
inventory balance equation is provided by Constraint (14), and no
X
N X
T
THC ¼ hi  I1i;t ð8Þ order is placed by Constraint (15). In order to model these relations
i¼1 t¼1 linearly, Constraints (16)–(22) were developed for the proposed ILP
The total holding cost (THC) during the planning period is the prod- model.
uct of the available stock quantity at the end of each period and the Order quantity constraints:
periodic holding cost.  
Ordering cost: Q i;t  QSi 6 BigM  1  Ai;t i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð23Þ
 
Q i;t  QSi P BigM  1  Ai;t i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð24Þ
Q i;t 6 BigM  Y i;t i ¼ 1; . . . ; N; t ¼ 1; 2; . . . ; T ð9Þ
X
N X
T Constraints (23) and (24) assert that the order quantities should be
TOC ¼ Y i;t  csi ð10Þ equal for all the periods if an order has been placed, which is indi-
i¼1 t¼1
cated by QSi for each product.
8194 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202

Inventory level constraints: Constraints (32)–(34) determine the quantity range within which
the given ordered quantity falls. Constraint (35) determines the
I2i;t 6 BigM  Bði; tÞ i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð25Þ
extra quantity in regard to the quantity purchased. Constraints
I1i;t 6 BigM  ð1  Bði; t ÞÞ i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð26Þ (36) and (37) assert that the ordered quantities and reorder points
Ii;t ¼ I1i;t  I2i;t ; i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð27Þ should be less than or equal to the upper bound values specified for
each product. Constraints (38) and (39) ensure that the integer
Constraints (25) and (26) define positive and negative inventory as
restrictions for all the variables are used in the model.
I1i,t = max{Ii,t, 0} and I2i,t = max{Ii,t, 0}, respectively. In period t, the
inventory holding cost is (hiI1i,t), and the inventory shortage cost is
(sciI2i,t). Constraint (27) finds the on-hand inventory level or short- 4. Proposed genetic algorithm approach
age quantity.
Storage space constraints: In this section, we present our proposed GA solution approach,
X  which has an embedded local search for determining the optimal
wi I1i;t1 þ Q i;t þ Q i;t;E 6 storage i ¼ 1; . . . ; N; t ¼ 1; . . . ; T
or near optimal multi-product multi-period (Q , r) inventory policy.
i2PC
GA is a stochastic search method proposed by Holland (1975) that
ð28Þ
simulates the development process of biological systems based on
This constraint ensures that the maximum inventory level is less Darwin’s ‘‘survival of the fittest’’ principle. It is one of the most
than or equal to the storage area capacity in each period. Because extensively used algorithms in inventory problems with respect
some products in the storage need a cold environment, storage con- to other heuristics. GA is a population-based adaptive search tech-
straint has to be taken into account for such products. nique which is effective when the search space of the problem is
Budget constraints: large. GA is one of the best solution approaches among other meta-
heuristics when large-scale problems are considered (Soleimani,
X
N X
T
Q i;t  ppi 6 B ð29Þ Seyyed-Esfahani, & Shirazi, 2013). It has been applied to our prob-
i¼1 t¼1 lem as well as to other inventory management related problems in
the literature (Gupta, Bhunia, & Goyal, 2009, Pasandideh et al.,
The product of the order quantity planned for the entire planning
2011, Yang et al., 2012) successfully. Besides, GA facilitates an easy
period and the purchasing price should be below the budget. The
solution representation for our problem requiring a large number
distributor should complete the purchasing process without
of variables and constraints and an easily applicable neighborhood
exceeding this credit limit.
solution generation due to its population based search in compar-
Lifetime constraints:
ison to single solution search based metaheuristics such as simu-
I1i;t1 þ Q i;t þ Q i;t;E lated annealing and tabu search.
6 shfi i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð30Þ
mdi Gen and Syarif (2005) proposed a hybridized tree-based GA to
solve multi-period production/distribution and inventory prob-
Products have specific lifetimes; the products have to be destroyed
lems. They adopted a fuzzy logic controller to improve the perfor-
if these lifetimes expire. Therefore, the time obtained by dividing
mance of the algorithm, and considered multiple products with
the total inventory at the beginning of the period and the incoming
independent demand, multiple shared resources, and multi-peri-
orders by the average demand should be shorter than the lifetime.
ods. Kannan, Sasikumar, and Devika (2010) proposed a GA to solve
Extra quantity constraints:
the mixed-integer linear programming model for a multi-echelon,
8
> EX i;1 ; L1 6 Q i;t < U 1 multi-period, multi-product closed-loop supply chain network
>
>
>
< EX i;2 ; L2 6 Q i;t < U 2 model containing storage space and processing time restrictions
Q i;t;E ¼ .. ð31Þ in order to minimize the total cost. Pasandideh, Niaki, and
>
> . Mousavi (2013) developed a random demand and limited storage
>
>
:
EX i;k ; Lk 6 Q i;t < U k capacity multi-period multi-item inventory control problem with
the aim of determining the optimal order quantities of products
In an effort to stay ahead of the competition, manufacturing compa- in different periods and minimizing the total inventory cost. They
nies may send various extra quantities, EXi,k, of a product according considered the all-unit and incremental quantity discount options.
to the quantity purchased by the distributor. Constraint (31) shows The modern field of genetics provides a further explanation of
the extra quantities offered by a manufacturing company in return evolution and the natural selection involved in the survival of the
for the quantity ordered. If the distributor places an order for i prod- fittest. GAs start with an initial set of random solutions, called a
uct between L1 and U1 in t period, s/he will obtain an extra quantity population, and then choose parents to reproduce.Each individual
of EXi,1. These products arrive in storage at no cost and are sold at in the population is called a chromosome. A simple GA that renders
the normal sales price to the customer. good solutions in many practical problems is composed of three
Q i;t  Lk P BigM  Di;t;k i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ; k ¼ 1; . . . ; K ð32Þ operators: selection, crossover, and mutation. The efficiency of
  GAs depends on many parameters, such as the initial population,
Q i;t  U k þ 1 6 BigM  1  Di;t;k
the representation of individuals, the selection strategy, and the
i ¼ 1; . . . ; N; t ¼ 1; . . . ; T; k ¼ 1; . . . ; K ð33Þ recombination of crossover and mutation operators.
X
K
Di;t;k ¼ 1 i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð34Þ
k¼1 4.1. Chromosome representation
X
K
Q i;t;E ¼ EX i;k  Di;t;k i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð35Þ The first and most important step in the GA approach is decid-
i¼1 ing on the chromosome structure. The GA approach begins with a
Q i 6 upper Q i i ¼ 1; . . . ; N ð36Þ numerical demonstration of the decision variables and data
r i 6 upper ri i ¼ 1; . . . ; N ð37Þ involved in the chromosome structure. In this paper, the decision
Ai;t ; Bi;t ; Y i;t ; Di;t;k 2 f0; 1g i ¼ 1; . . . ; N; t ¼ 1; . . . ; T; k ¼ 1; . . . ; K ð38Þ variables of the proposed model are reorder point, r, and order
quantity, Q. Matrix notation is used for the demonstration of a
r i ; QSi ; Q i;t ; Ii;t ; I1i;t ; I2i;t P 0; integer i ¼ 1; . . . ; N; t ¼ 1; . . . ; T ð39Þ
chromosome. The first and second rows of the matrix contain the
 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202 8195

order quantity (Qi) and reorder point (ri) variable, respectively chromosome
(i = 1, . . . ,N; the number of products). The columns of the matrix
show the products. Consequently, a chromosome comprises a 2 …
 N matrix, including two rows and as many columns as the
number of products. The structure of such a chromosome is …
depicted in Fig. 1.
It is remarkable that binary coding chromosome structures are
predominantly used in lot-size scheduling problems in the litera- gene
ture, whereas real numbers are used in inventory problems. In
Fig. 1. Gene string and chromosome representation of 2  N matrix.
most cases, binary coding is not used to solve large decision
variable optimization problems (Gupta, Bhunia, & Goyal, 2007).
Further, in our problem, the chromosome structure is formed using objective is to increase the penalty value as the number of iterations
real values. increases in order not to choose infeasible solutions.

4.2. Initial population generation 4.3.1. Calculation of the fitness values for variable demand
The fitness value calculation procedure for products with vari-
GA starts a search with an initial population. It is well known able demand structure is different from those of the deterministic
that the performance of metaheuristics is affected by the selection structure. If the demand of a product is variable, J demand sets are
of the initial solution (s). If the initial solution is sufficient, the pos- generated to represent the variability inherent in that demand. To
sibility of reaching better solutions increases and convergence to calculate the fitness values of products with variable demand, the
the near optimal or optimal solution will be faster. In order not corresponding (Q , r) values in the chromosome are put into the
to depend solely on the initial population, a local search embedded fitness function for each demand set and the fitness value calculated
in the GA generates the neighborhood solutions of the best by taking the average of the fitness values for each demand set.
chromosomes in the initial population and returns the best one
to the population. The initial population is determined randomly 4.4. Proposed local search
so that the decision variables Q and r are between the lower and
upper limits. The lower limit is assumed to be zero for each prod- To avoid local optimum during the GA process, the process is
uct, whereas the upper limit is determined to be four times the cal- combined with different procedures. In this paper, a local search
culated values of the economic ffi order quantity,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi EOQi and ri. For each approach, proposed by Zhao et al. (2012), is used to find neighbor
P
product, EOQ i ¼ 2csi  X i =hi , where X i ¼ Tt¼1 xi;t , and reorder solutions of the current (Q , r) solution by increasing and decreasing
point is calculated as ri = mdi⁄li. Therefore, all the chromosomes Q and/or r in succession. Thus, neighbor solutions of (Q , r) are
in the initial population are generated randomly as the size of obtained by changing Q and r; i.e., (Q , r) to (Q , r + 1), (Q , r  1),
the population (pop_size). (Q + 1, r), (Q  1, r), (Q + 1, r + 1), (Q  1, r  1), (Q + 1, r  1),
(Q  1, r + 1). If a chromosome that has a better fitness value is
4.3. Calculating the fitness values obtained in each iteration at the local search, that chromosome is
included to the population. This procedure is iterated as many
Each chromosome in the initial population is evaluated by its times as the determined number of iterations. The general struc-
fitness value. Because the procedure used to select the parent chro- ture of the GA with local search procedure is depicted in Fig. 2.
mosomes for reproduction depends on the fitness values, it is
important to calculate them properly. The individuals with the 4.5. Selection
worst fitness values die, whereas the individuals with the better
fitness values survive. A penalty function is deployed to address In this paper, roulette wheel selection (Gen & Cheng, 1997) is
the problem constraints. Since the aim is to maximize profit, the utilized in the selection of parent chromosomes for reproduction.
fitness function is calculated by subtracting the penalty function The roulette wheel selection method is a probabilistic selection
from the objective function as follows: approach that depends on the performance of the chromosomes.
The fitness value, fitnessvalue (vj), of each chromosome in the pop-
Fitness function ¼ objective function  penalty function P size
ulation and the total feasibility value, F ¼ pop
j¼1 fitnessv alue ðv j Þ,
Consequently, an individual that has maximum profit in a for the population are calculated, where pop_size is the number of
population has maximum fitness value. The following penalty func- chromosomes in the population. The probability value, pj = fitness-
tion, developed by Joines and Houck (1994), is used in this paper: value (vj)/F, for each chromosome and the cumulative probability
Ppop size
value, qj ¼ j¼1 pj , are found. A random number (rand) is
ev alðxÞ ¼ f ðxÞ þ pðt; xÞ ð40Þ derived from the uniform [0, 1] range and the chromosome that
X
m
b coincides with the range including that random number is selected
pðt; xÞ ¼ p/t di ðxÞ ð41Þ
to produce new individuals.
i¼1

where eval(x) is the evaluated fitness function value, p(t, x) is the 4.6. Genetic operators: crossover
variable penalty factor, t represents the generation of the GA, di(x)
is the penalty term for single constraint, and a and b are parameters The design of the genetic operators has an important effect on
used to adjust the scale of the penalty value. pt = C  t, where C is a stabilizing the performance of GA. Genetic operators are grouped
constant. In this penalty function, it is essential that the parameter into two types: crossover and mutation. Crossover performs the
values be chosen correctly. The recommended values are 1, 1 or 2, mating process of the individuals that will produce new offspring
and 0.5 for the parameters a, b, and C, respectively. However, we for the future new population. A mutation operator is used in the
used b = 5, a = 1, and C = 0.5 in our problem because the recom- process to randomly change the genetic value in the chromosome.
mended parameter values did not result in high-quality solutions: In the proposed GA approach, since our chromosome structure is a
di(x) is calculated as the penalty term for each iteration. The 2  N matrix, the crossover operator is applied in three different
8196 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202

Fig. 2. General structure of the GA with local search.

ways. Information is given about the application of the proposed column is made with a 50% probability by selecting a random
crossover procedure related to our example problem as follows: number (crosspoint) in the [1, N1] range. This crossover, Type
2, is depicted in Fig. 4.
1. Changing just one row: If the random number we have derived in 3. Crossover on the right or left hand side of a selected row: If the
the [0, 1] range is smaller than the first crossover type rate Pc,1 derived random number is smaller than the determined third
that we have determined, then a new offspring is generated crossover type rate, then a new offspring is obtained by replac-
by completely replacing the randomly selected row with the ing the right or left hand side of the selected row with a 50%
row of the other parent. This crossover, Type 1, is depicted in probability. This crossover, Type 3, is depicted in Fig. 5.
Fig. 3.
2. Changing the right or left hand side of both rows from the deter- 4.7. Genetic operators: mutation
mined point: If the derived random number is smaller than
the determined second crossover type rate Pc,2, then both rows This operator arbitrarily selects genes to counteract premature
are replaced together. The determination as to whether a cross- convergence and to maintain enough diversity in the population.
over will be conducted from the right or left hand side of the Mutation is employed to maintain the diversity of the entire
 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202 8197

that virtually all children are different from their parents. The
genes selected for the mutation operator applied in this paper were
randomly changed between the lower and upper limits. The appli-
cation of the mutation operator is discussed by Eiben and Smith
(2007) and Goldberg (1989).

4.8. Termination criteria

In general, two methods are used for terminating a GA. In the

first of these methods, the algorithm is terminated after a prede-
fined number of generations (iter_num), whereas in the second
method, the algorithm is terminated when no further development
can be obtained in the solution after a predefined number of
successive generations. In our proposed GA approach, we use the
former method.

4.9. Parameter optimization

Fig. 3. Only one row applied for crossover.
We stated above that generation number (iter_size), population
number (pop_size), crossover type rate (Pc,1, Pc,2), and mutation rate
(Pm) are the most important parameters in GA. In order to calibrate
the proposed GA, the Taguchi method is used in this paper to sta-
tistically design the experimental investigation. Traditional exper-
imental design procedures are in essence too complicated and
difficult to use because a great deal of experimental work has to
be carried out as the number of process parameters increases. To
solve this problem, the Taguchi method uses a special orthogonal
arrays design to study the entire parameter space with only a small
number of experiments. We employ Taguchi’s orthogonal arrays,
the signal-to-noise (S/N) ratio, and analysis of variance (ANOVA)
to find the appropriate levels and to analyze the effect of parame-
ters on the objective values obtained. The S/N ratio indicates the
amount of variation present in the response variable. The aim is
to maximize the signal-to-noise ratio. Taguchi categorizes objec-
tive functions into three groups: the smaller-the-better type, the
larger-the-better type, and nominal-is-best type. Our objective
function is to maximize profits; therefore, we choose the larger-
the-better type S/N ratio and calculate as follows:
!
Fig. 4. Crossover applied on the column.
1X n
1
S=N ¼ 10  log ð41Þ
n i¼1 ðobjectiv e functionÞ2i

where n is the replication number and objective functioni is the

result of experiments in replication i.
We defined four design factors that are important in GA perfor-
mance for a three-product eight-period problem type: (1) iteration
size (iter_size), (2) population size (pop_size), (3) type rate (Pc,1,Pc,2),
and (4) mutation rate (Pm). The possible three levels are considered
for each design factor, as shown in Table 3. While 34 = 81 experi-
ments should be made for four factors with three levels, Taguchi
analysis enables us to conduct only 27 experiments by choosing
the orthogonal array L27. By executing the GA 10 times for each
experiment, we obtained average, minimum, and maximum fitness
values for each experiment.
On obtaining the results of the Taguchi experiment, we trans-
formed them to S/N ratios. To explore the relative significance of
individual factors in terms of their main effects on the objective

Table 3
Experimental factors and their levels.

Factors Iteration Population Mutation Crossover

Fig. 5. Crossover applied on both row and column. number size rate type
Coding iter_num pop_size Pm (Pc,1, Pc,2)

population by changing individuals bit by bit with a small proba- Level 1 200 20 0.05 (0.25, 0.50)
Level 2 500 50 0.15 (0.33, 0.66)
bility Pm 2 [0, 1]. Mutation prevents the algorithm from escaping
Level 3 1000 100 0.4 (0.50, 0.75)
the local minima trap. Crossover and mutation operators ensure
8198 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202

Table 4
ANOVA table of means.

Source DF Seq SS Adj SS Adj MS F P

Main effects 4 1492494 1802638 450660 6.58 0.002
iter_num 1 455951 455951 455951 6.66 0.019a
pop_size 1 28267 36512 36512 0.53 0.475
Pm 1 965861 1183356 1183356 17.28 0.001a
Pe 1 42416 146676 146676 2.14 0.162
2-Way interactions 4 522674 316515 79129 1.16 0.365
iter_num  pop_size 1 112564 112564 112564 1.64 0.217
iter_num  P m 1 137754 137754 137754 2.01 0.174
pop_size  P m 1 27378 15521 15521 0.23 0.640
Pm  Pc 1 244978 66176 66176 0.97 0.339
3-Way interactions 1 227173 227173 227173 3.32 0.086
pop_size  Pm  Pc 1 227173 227173 227173 3.32 0.086
Residual error 17 1164360 1164360 68492
Total 26 3406701
a
Factors effective at 95% confidence level.

function, ANOVA was conducted: the results are shown in Table 4. have the same rank values. The mutation parameter has the great-
As can be seen in the table, factors with p 6 0.05 at 95% confidence est effect on both the S/N ratio and the average. Iteration number
level, i.e., iteration number and mutation rate, are effective on the has the second- greatest effect, followed by crossover and popula-
results. In addition, it can be seen that the interaction between the tion, respectively. As seen in Table 5, the levels with S/N ratio that
four factors does not cause a significant difference in the results. give the greatest value are the levels that are effective in the
Further, the most effective factor is the mutation rate that has solution. Finally, the parameter values that we should use are as
the highest F value. We can obtain information about which factor follows: {Pm = 0.05, iter_num = 1000, Pc,1, Pc,2 = (0.33, 0.66), and
is more effective with Taguchi analysis, as shown in Table 5. pop_size = 50}.
Table 5 shows the average of the response characteristics (S/N
ratios, means) for each level of all factors. It includes the ranks 5. Computational study
based on delta statistics comparing the relevant sizes of the effects.
Delta statistics is the difference between the greatest and the In this section, various test problems with varying numbers of
smallest average values of each factor. S/N ratio and mean tables products and periods are solved and results compared with the
results of the ILP model in order to evaluate the performance of
the proposed GA. We used LINGO 9.0 software to solve the ILP
Table 5
model and terminated the execution of the model after five hours.
Response table for signal-to-noise ratios (larger is better). The optimality of the solutions in the rows marked with asterisks
was not confirmed because LINGO could not complete its search
Level Iteration Population Mutation Crossover
for an optimal solution within the specified time limit. The GA
1 86.77 86.81 86.94 86.80 solutions were obtained from a MATLAB 7.11 implementation.
2 86.86 86.90 86.85 86.92
All test problems were solved on an IntelÒ Core ™ Duo PC with a
3 86.91 86.85 86.76 86.83
Delta 0.14 0.09 0.18 0.13 [email protected] GHz CPU and 4.00 GB RAM.
Rank 2 4 1 3 Table 6 shows the total profit value and CPU time (seconds)
results of the ILP model and the proposed GA. The proposed GA

Table 6
Results obtained by ILP and the proposed GA.

Problem type Product no Period no Total profit ($) GAP (%) Computational time CPU (sec) GAP (%)
ILP Proposed GA-best ILP Proposed GA
2P_4T 2 4 9792.86 9624.49 1.72 5.00 22.80 356.00
2P_8T 2 8 19510.75 19221.80 1.48 69.00 29.87 56.71
2P_10T 2 10 27606.49 27606.49 0.00 756.00 32.51 95.70
2P_11T 2 11 30467.76 30466.99 0.00 1457.00 35.27 97.58
2P_12T 2 12 33870.36 33630.79 0.71 13698.00 35.45 99.74
2P_14T 2 14 40200.92a 40200.92 0.00 18000.00 39.00 99.78
2P_15T 2 15 43840.68a 43858.74 0.04 18000.00 40.73 99.77
3P_4T 3 4 10566.50 10566.50 0.00 26.00 36.44 40.15
3P_8T 3 8 23531.93 23531.93 0.00 17259.00 41.82 99.76
3P_10T 3 10 29122.90a 32004.85 9.90 18000.00 53.77 99.70
4P_4T 4 4 10987.09 10385.54 5.48 54.00 44.45 17.69
4P_8T 4 8 22621.60a 23029.78 1.80 18000.00 57.43 99.68
5P_4T 5 4 10984.86 10947.90 0.34 162.00 55.31 65.86
5P_8T 5 8 24034.04a 23828.11 0.86 18000.00 83.67 99.53
6P_4T 6 4 11115.78 11071.98 0.39 596.00 68.17 88.56
6P_8T 6 8 22237.50a 23852.91 7.26 18000.00 98.73 99.45
7P_4T 7 4 11146.60 11146.60 0.00 2812.00 83.64 97.03
7P_8T 7 8 21534.83a 23493.55 9.10 18000.00 111.51 99.38
8P_4T 8 4 11162.11a 11149.30 0.11 18000.00 104.92 99.42
a
Interrupt after 5 h so the solution may not be optimal.
 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202 8199

terminated when the iteration number reached 1000, and 20 rep- met for five times by the distributor. We tried to estimate the
lications were applied in order to find the best value for the profit. stockout cost by using Type I service level as indicated in
The table shows that better results can be obtained with GA in Nahmias (1993). A one-year planning horizon is handled as 52
much shorter computation times. periods (weeks) in our study. If a customer is lost when it is
We calculated GAP (%) values for the total profit and solution rejected for 5 times by the distributor throughout the planning
times using Eq. (42) in order to clearly compare the solution results period, it is expected that the demand is met for 52  5 = 47 times.
and times achieved by the ILP model and the GA approach. The Therefore, a is calculated as a = 47/52 = 0.90. This value corre-
values computed are shown in Table 6. sponds to the F(r) value and as a result, stockout cost is computed
as in Eq. (43), where F(r) represents the average percentage of the
ðGA Solution  ILP SolutionÞ demands which are satisfied as they occur.
GAPð%Þ ¼  100 ð42Þ
ILP Solution
While an optimal solution for a two products 12 periods [2P_12T] sci ¼ Q i hi =½ð1  FðrÞÞX i  ð43Þ
problem could be obtained in approximately 3 h and 48 min by
the ILP model, a solution only 0.71% worse was reached in 35.45
The stockout cost is calculated as sc2 = (190)(3.45)/
CPU seconds by the proposed GA, incurring 99.74% less time. As
[(1  0.90)(6228)] = $1.09 for Pro2.
the number of products and periods increases, reaching an opti-
The lifetime of a drug begins as soon as its production is com-
mum solution with ILP becomes much more difficult. Consequently,
plete and the expiration date is indicated on the package. The
it may not be possible to reach even an integer feasible solution.
amount of time that elapses during the transportation from manu-
facturer to the distributor is unknown. Pharmacies are generally
6. Case study of a pharmaceutical distributor in Turkey reluctant to buy products that only have a short lifetime left. Prod-
ucts with a lifetime equal to or less than one year are categorized
The company addressed in this paper is a pharmaceutical dis- as short lifetime products. In this paper, the lifetimes of the drugs
tributor performing drug distribution with a single regional stor- are taken as 1/4 of their real lifetimes, thus guaranteeing that they
age. Because the pharmaceutical distributor stores approximately are sold before their lifetime expires. Pharmaceutical distributors
5000 products and distributes to 900 pharmacies, the products can obtain customer satisfaction initially by providing requested
are classified according to ABC analysis. The demand structure of products and through good communication and reliability. Further,
the products is decided on the basis of the percentage of the coef- the convenient payment terms provided to the pharmacy is an
ficient of variation V (standard deviation/mean demand) values important factor in this selection. Because of the intense competi-
calculated for a given number of periods, as described by Taha tion in the pharmaceutical industry, manufacturing companies
(2007). The products that have V values below 20% are considered may offer extra quantities or various promotions according to the
deterministic products, whereas those with V values greater purchased quantities. An extra quantity can be offered in particular
than 20% are considered within the context of variable products. for equivalent and generic drugs. The extra quantities provided for
A quarter-based comparison is made for the products whether they the selected nine drugs according to the ordered quantity range are
have a variable demand structure or a seasonal one. In the end, shown in Table 9. These promotions may vary for each product and
nine products were chosen to represent the deterministic, manufacturer.
seasonal, and variable product groups from each ABC class, as The distributor has a yearly budget of $1,200,000 for the
shown in Table 7. selected nine products, with the planning horizon taken as one
The study has been initiated in 2012. The input values such as year with weekly planning periods. Some products have to be
initial stock, purchasing and sales prices, demand have been stored under cold storage conditions. Pro5 is an example of such
obtained from the company’s database. The sales data of the prod- a drug. The storage capacity for this drug is taken as 1500 contain-
ucts for the last 4 years have been obtained from the Interconti- ers/period. Because the storage cost is higher for this type of prod-
nental Marketing Services [IMS] Health company in Turkey and uct, inventory holding cost percentage is taken as 10% higher.
the variability of the demand structure of the products has been We generated demand sets for Pro3, Pro6, and Pro9 using prob-
examined. ANOVA analysis has been performed to examine ability distributions Weibull (66.5, 0.459), Beta (87, 0.702, 1.51),
whether there is difference in the demand of products between and Normal (42.1, 30.2), respectively. The representation of the
the periods. Inventory holding cost has been calculated using variability of these product demands by the given probability func-
warehouse rent, personnel maintenance, and utilities data of the tions were already confirmed via the statistical test.
company along with the ratios given in Greene (1997). Ordering Next, we determined the lower and upper bound values for our
cost includes the costs of purchasing department and that of the (Q , r) decision variables. Our chromosome structure is represented
processes of product purchasing, acceptance and control. Shortage by a 2  9 (Q , r) decision variables matrix. The lower (lower_Qi,
cost has been calculated as given in Nahmias (1993). The distribu- lower_ri) and upper bound values (upper_Qi, upper_ri) are listed in
tor may face customer loss if the order of a pharmacy cannot be Table 10.

Table 7
Products selected to represent the various groups along with their descriptions.

Product name ABC class Demand type Product description

Pro1 A Deterministic Helps treat asthma
Pro2 A Seasonal Treats many different infections caused by bacteria
Pro3 A Variable Relaxes muscles and increases blood flow to particular areas of the body
Pro4 B Deterministic Antibiotics for surgery infection
Pro5 B Seasonal Breaks disulfide bonds in mucus and liquefies it
Pro6 B Variable Inhibits gastric acid secretion
Pro7 C Deterministic Treats chronically elevated levels of iron in the blood
Pro8 C Seasonal Alleviates itching pain related to insect sting and simple skin irritations
Pro9 C Variable Alleviates nausea
8200 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202

Table 8
Input data for the products.

Product i 1 2 3 4 5 6 7 8 9
Lead time (li) (week) 1 1 1 1 1 1 1 1 1
Economic order quantity (EOQi) (unit) 102 206 132 462 159 95 5 294 308
Purchasing cost (ppi)($) 59.70 17.26 55.39 3.40 8.05 17.30 1209.79 3.40 2.72
Sales price (spi)($) 63.88 18.64 59.27 3.70 8.80 18.70 1233.99 3.70 3.00
Inventory holding cost (hi)($) 0.21 0.06 0.20 0.01 0.03 0.06 4.31 0.01 0.01
Ordering cost (csi)($) 10 10 10 10 10 10 10 10 10
Shortage cost (sci)($) 2.14 1.09 1.64 0.47 0.66 2.27 43.14 0.74 0.71
Interest rate (Ie)(%) 15 15 15 15 15 15 15 15 15
Payment term for distributor (Mi) (week) 12 16 17 3 12 17 12 17 8
Payment term for pharmacy (Ni)(week) 8 12 10 3 8 12 8 12 4
Demand type (Di) 0 1 1 0 1 1 0 1 1
Shelf life (shfi) (week) 12 12 24 16 10 36 12 36 36
Storage limit (wi) (unit) 0 0 0 0 1500 0 0 0 0
Initial invenory (I0i) (unit) 156 186 412 187 753 26 2 49 43

Table 9
Extra promotional quantities offered with some products.

Pro2 Pro3 Pro4 Pro5 Pro6

Lower Upper Extra Lower Upper Extra Lower Upper Extra Lower Upper Extra Lower Upper Extra
limits limits quantities limits limits quantities limits limits quantities limits limits quantities limits limits quantities
10 49 0 0 4 0 0 4 0 0 4 0 0 4 0
50 99 4 5 9 1 5 9 5 5 9 0 5 8 1
100 149 5 10 19 3 10 49 15 10 49 5 9 15 2
150 199 10 20 49 8 50 99 90 50 99 10 16 25 4
200 249 15 50 249 30 100 249 240 100 249 100 26 62 6
250 399 18 250 499 250 250 499 750 250 499 100 63 95 10
400 599 39 500 999 500 500 749 1500 500 and 200 96 160 24
more
600 1199 119 1000 1499 1000 750 999 2250 161 241 34
1200 and 500 1500 2999 1500 1000 1999 3000 242 338 54
more
3000 and 3000 2000 2499 6000 339 and 78
more more
2500 and 7500
more

With a $1,200,000 budget and all other constraints, the total

Table 10 profit that the pharmaceutical distributor will obtain at the end of
The lower and upper bounds for the (Q, r) decision variables. the planning horizon is estimated as $334,642.19. This profit
Products Pro1 Pro2 Pro3 Pro4 Pro5 Pro6 Pro7 Pro8 Pro9 level can be reached by placing orders of Q i = {241, 600, 500,
121, 100, 96, 6, 133, 207} order quantities at r i ¼ f101; 203;
lower_Qi 0 0 0 0 0 0 0 0 0
upper_Qi 400 760 480 1700 1212 360 20 1080 1120 392; 182; 237; 30; 5; 59; 73g inventory levels. Ten independent
lower_ri 0 0 0 0 0 0 0 0 0 replications were applied for the proposed GA approach; the details
upper_ri 400 480 640 480 964 100 20 192 180 of the best solution obtained are presented in Table 11. The results
were obtained at 90% service level and with the stock holding,

Table 11
Solution results for the Pharmaceutical Distributor.

Products Pro1 Pro2 Pro3 Pro4 Pro5 Pro6 Pro7 Pro8 Pro9 Total
Q⁄ 241 600 500 121 100 96 6 133 207
r⁄ 101 203 392 182 237 30 5 59 73
Expected total sales revenue 332170.8 127178.72 383785.40 23125.44 134714.72 23087.81 320834.80 8230.65 6919.84 1360048.20
(ETSR) ($)
Total earning from payment 10514.53 4564.92 5541.76 60.84 668.43 809.14 13577.47 360.89 150.21 36248.23
term TPFS ($)
Total purchasing cost (TPC) ($) 302141.70 98382.00 191095.5 6993.80 57624.00 17438.40 312125.82 7777.84 6474.96 1000054.02
Total losing from payment term 7706.36 4425.81 11129.77 201.19 1562.69 803.45 13956.31 286.42 80.27 40152.30
(TLPFC) ($)
Total holding cost (THC) ($) 1309.56 1517.20 8155.28 100.63 497.0025 216.43 568.92 43.39 68.19 12476.62
Total ordering cost (TOC) ($) 210 95 69 170 420 105 430 172 115 1786.00
Expected total shortage cost 0 389.574 3089.644 0 3500.53 53.13 0 151.18 1.224 7185.28
(ETSC) ($)
Expected total shortage 0 282.3 796.3 0 3182.3 38.5 0 487.7 5.1 4792.2
quantity (ETSQ) (Unit)
Expected total profit (ETPR) ($) 31317.70 26934.04 175787.97 15720.66 71778.93 5280.54 7331.21 160.69 330.41 334642.19
 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202 8201

In accordance with actual company values, the inventory hold-

ing cost was taken as 18.51% of one unit purchasing cost. For this
. reason, we researched how our expected maximum profit value
will change for a 40% change in the inventory holding cost. If
the inventory holding cost decreases by 40%, then the expected
maximum profit value increases by 1.47%. If the inventory holding
.
. cost increases by 40%, then the expected maximum profit value is
. . . . . observed to decrease by 4.37%. The effect of inventory holding cost
. change on the expected maximum profit value is depicted in Fig. 6.
We estimated ordering cost at $10 per order. When the ordering
cost decreases by 50%, the change in the expected maximum profit
. value is only 0.69%. Further, when the ordering cost increases by
50%, there is a 0.52% decrease in the profit value. The change for
Fig. 6. Profit versus inventory holding cost % change. the ordering cost and the expected maximum profit value is shown
in Fig. 7.
In accordance with actual company values, we assumed stock-
. out cost as a loss from the profit. The obtained stockout cost corre-
sponded to 90% customer service level (CSL). If CSL is chosen as
99%, the profit value decreases by 11.82%, whereas if CSL is chosen
as 99.5%, the profit value decreases by 20.07%. Thus, we can assert
. that CSL has an important effect on the profit value. Stock holding
cost increases as the service level increases. The effect of service
. . . . level change on the expected maximum profit value is shown in
.
Fig. 8.

. 7. Conclusions

The relevant literature review shows that the (Q , r) inventory

Fig. 7. Profit versus ordering cost % change.
model has been only used as single period in multi-product prob-
lems and for nonperishable products. It is also found out that, for
multi-product multi-period case apart from the continuous review
. . inventory models, it is primarily dealt with lot-sizing issues in pro-
duction planning and there is no study for perishable items as well.
In this paper, we formulated a novel approach for multi-product,
. multi-period (Q , r) inventory models, with the objective of maxi-
mizing the profit under constraints such as storage area, budget,
. shelf life, and various promotions. As the number of products and
periods increase, we found that an optimal solution cannot be
. reached with ILP in a reasonable computational time. For this rea-
. . .
son, we proposed a GA solution approach since the problem of the
pharmaceutical distributor case studied is a larger sized problem.
Fig. 8. Percentage change in profit for customer service level. To determine the efficacy of the proposed GA, we compared the
solution results obtained for problem instances of various sizes to
ordering, and stockout cost data given in Table 8. The computa-
those obtained from the ILP model. The analysis results showed
tional time for an average replication was 2,737.07 s.
that the proposed GA approach can produce efficient and effective
According to the current policy adopted by the company, they
solutions. Finally, we gave real results for our case study.
examine the stocks one by one every week and place an order if
In future research, this model can further be generalized to
their budget is available considering the promotions and sales
include safety stock in the chromosome structure proposed for
within the last three months. Moreover, the calculation of reorder
the GA solution approach. In addition, we will generalize the model
point and fixed-order quantity in the sector has been determined
to allow partial backlogging and/or not allow shortages to increase
to be very useful as a result of the negotiations. We are in constant
the service level. We also plan to handle supply chain management
cooperation with the company and in case the purchases suggested
problems including multi-suppliers and multi-customers in the
by us are applied, the following benefits can be observed compared
model. Finally, the performance of the proposed GA approach can
to their own purchases. For example; for Pro2, whereas the profit
be examined with other metaheuristic methods and besides, other
of the company at the end of its purchases is 32,456.81$, it would
inventory control policies such as (s, S)- and (Q, r, T) can be applied
be possible to obtain a profit of 26,934.04$ if the purchase policy
on the problem.
suggested by us were used.

6.1. Sensitivity analysis References

Ang, M., Song, J.-S., Wang, M., & Zhang, H. (2013). On properties of discrete (r, Q) and
In this section, because inventory holding cost, ordering cost,
(s, T) inventory systems. European Journal of Operational Research, 229, 95–105.
and stockout cost are difficult to decide, sensitivity analysis is Axsäter, S. (1993). Exact and approximate evaluation of batch-ordering policies for
performed in order to see the effect of these costs on the profit two-level inventory systems. Operations Research, 41(9), 777–785.
value. The sensitivity analysis is shown by increasing or decreasing Axsäter, S. (2007). A heuristic for triggering emergency orders in an inventory
system. European Journal of Operational Research, 176, 880–891.
cost parameters, taking one at a time and keeping the others at their Berk, E., & Gurler, U. (2008). Analysis of the (Q , r) inventory model for perishables
own values. In each case, the best value of optimal profit is found. with positive lead time and lost sales. Operations Research, 56(5), 1238–1246.
8202 I. Saracoglu et al. / Expert Systems with Applications 41 (2014) 8189–8202

Betts, J. M., & Johnston, R. B. (2005). Determining the optimal constrained multi- Lapierre, S. D., & Ruiz, A. B. (2007). Scheduling logistic activities to improve hospital
item (Q , r) inventory policy by maximizing risk-adjusted profit. IMA Journal of supply systems. Computers & Operations Research, 34, 624–641.
Management Mathematics, 16(4), 317–338. Lau, AH.-L., & Lau, H-S. (2008). An improved (Q , r) formulation when the stockout
Bijvank, M., & Vis, I. (2012). Inventory control for point-of-use locations in hospitals. cost is incurred on a per-stockout basis. International Journal of Production
Journal of the Operational Research Society, 63, 497–510. Economics, 111, 421–434.
Eiben, A. E., & Smith, J. E. (2007). Introduction to evolutionary computing (2nd ed.). Louly, M-A., & Dolgui, A. (2013). Optimal MRP parameters for a single item
Berlin Heidelberg: Springer-Verlag. inventory with random replenishment lead time, POQ policy and service level
Federgruen, A., & Zheng, Y. (1992). An efficient algorithm for computing an optimal constraint. International Journal of Production Economics, 143, 35–40.
(r, Q) policy in continuous review stochastic inventory systems. Operations Mandal, S., Maity, A. K., Maity, K., Mondal, S., & Maiti, M. (2011). Multi-item
Research, 40(4), 808–813. multi-period optimal production problem with variable preparation
Gardner, E. S. (1983). Approximate decision rules for continuous review inventory time in fuzzy stochastic environment. Applied Mathematical Modelling, 35,
systems. Naval Research Logistics Quarterly, 30, 59–68. 4341–4353.
Garman, M. B. (1976). Multi-product economic order quantity analysis under Mattsson, S- A. (2010). Inventory control in environments with seasonal demand.
minimum inventory valuation constraints. Operational Research Quarterly, 27(4), Operations Management Research, 3, 138–145.
983–989. Pergamon Press. Mousavi, S. M., Hajipour, V., Niaki, S. T. A., & Alikar, N. (2013). Optimizing multi-
Gen, M., & Cheng, R. (1997). Genetic algorithms and engineering design. John Wiley item multi-period inventory control system with discounted cash flow and
and Sons. inflation: Two calibrated meta-heuristic algorithms. Applied Mathematical
Gen, M., & Syarif, A. (2005). Hybrid genetic algorithm for multi-time period Modelling, 37, 2241–2256.
production/distribution planning. Computers & Industrial Engineering, 48, Nahmias, S. (1982). Perishable inventory theory, a review. Operations Research,
799–809. 30(4), 680–708.
Goldberg, D. E. (1989). Genetic algorithms in search optimization and machine Nahmias, S. (1993). Production and operations analysis (2nd ed.). Richard D. Irwin,
learning. Addison Wesley Longman, Inc. Inc.
Goyal, S. K., & Giri, B. C. (2001). Recent trends in modeling of deteriorating Nahmias, S. (2011). Perishable inventory systems. Springer Science+ Business Media.
inventory. European Journal of Operational Research, 134, 1–16. Pasandideh, S. H. R., Niaki, S. T. A., & Nia, A. R. (2011). A genetic algorithm for vendor
Greene, J. H. (1997). Production & inventory control handbook (3rd ed.). A division of managed inventory control system of multi-product multi-constraint economic
the McGraw-Hill Companies. order quantity model. Expert Systems with Applications, 38, 2708–2716.
Gupta, R. K., Bhunia, A. K., & Goyal, S. K. (2007). An application of genetic algorithm Pasandideh, S. H. R., Niaki, S. T. A., & Mousavi, S. M. (2013). Two metaheuristics to
in a marketing oriented inventory model with interval valued inventory costs solve a multi-item multiperiod inventory control problem under storage
and three-component demand rate dependent on displayed stock level. Applied constraint and discounts. International Journal of Advanced Manufacturing
Mathematics and Computation, 192, 466–478. Technology, 69, 1671–1684.
Gupta, R. K., Bhunia, A. K., & Goyal, S. K. (2009). An application of genetic algorithm Pasandideh, S. H. R., Niaki, S. T. A., & Tokhmehchi, N. (2011). A parameter-tuned
in solving an inventory model with advance payment and interval valued genetic algorithm to optimize two-echelon continuous review inventory
inventory costs. Mathematical and Computer Modelling, 49, 893–905. systems. Expert Systems with Applications, 38, 11708–11714.
Hadley, G., & Whitin, T. M. (1963). Analysis of inventory systems. Englewood Cliffs, Rafaat, F. (1991). Survey of literature on continuously deteriorating inventory
New York: Prentice-Hall. models. The Journal of the Operational Research Society, 42(1), 27–37.
Hajiaghaei-Keshteli, M., Sajadifar, S. M., & Haji, R. (2011). Determination of the Soleimani, H., Seyyed-Esfahani, M., & Shirazi, M. A. (2013). Designing and planning a
economical policy of a three-echelon inventory system with (r, Q) ordering multi-echelon multi-period multi-product closed-loop supply chain utilizing
policy and information sharing. International Journal of Advanced Manufacturing genetic algorithm. International Journal of Advanced Manufacturing Technology,
Technology, 55, 831–841. 68, 917–931.
Harris, F. W. (1913). How many parts to make at once? Factory. The Magazine of Taha, H. A. (2007). Operations research an introduction (8th ed.). Pearson Education,
Management, 10, 135–152. Inc.
Holland, J. (1975). Adaptation in natural and artificial systems. Ann Arbor: University Van Zyl, G. J. J. (1964). Inventory control for perishable commodities, unpublished
of Michigan Press. (Ph.D. dissertation). University of North Carolina.
Hopp, W. J., Spearman, M. L., & Zhang, R. Q. (1997). Easily implementable inventory Veinott, A. (1965). Optimal policy for a multi-product, dynamic, non-stationary
control policies. Operations Research, 45, 327–340. inventory problem. Management Science, 12, 206–304.
Jeddi, B. G., Shultes, B. C., & Haji, R. (2004). A multi-product continuous review Wang, T.-Y., & Hu, J-M. (2010). Heuristic method on solving an inventory model for
inventory system with stochastic demand, backorder, and a budget constraint. products with optional components under stochastic payment and budget
European Journal of Operational Research, 158, 456–469. constraints. Expert Systems with Applications, 37, 2588–2598.
Joines, J. A., & Houck, C. R. (1994). On the use of non-stationary penalty functions to Yang, W., Chan, F. T. S., & Kumar, V. (2012). Optimizing replenishment polices using
solve nonlinear constrained optimization problems with GA’s. Fogel, 132, genetic algorithm for single-warehouse multi-retailer system. Expert Systems
579–584. with Applications, 39, 3081–3086.
Kallrath, J., & Maindl, T. I. (2006). Real optimization with SAPÒ APO. Berlin Heidelberg: Zhang, S., & Eksioglu, S. D. (2009). Economic lot-sizing problem with multimode
Springer-Verlag. replenishment and perishable inventory. In Proceedings of the 2009 industrial
Kannan, G., Sasikumar, P., & Devika, K. (2010). A genetic algorithm approach for engineering research conference.
solving a closed loop supply chain model: A case of battery recycling. Applied Zhao, X., Fan, F., Liu, X., & Xie, J. (2007). Storage-space capacitated inventory system
Mathematical Modelling, 34, 655–670. with (r, Q) policies. Operations Research, 55(5), 854–865.
Kim, J. S., & Benton, W. C. (1995). Lot size dependent lead times in a (Q , r) inventory Zhao, X., Qiu, M., Xie, J., & He, Q. (2012). Computing (r, Q) policy for an inventory
system. International Journal of Production Research, 33, 41–58. system with limited sharable resource. Computers and Operations Research, 39,
Kundu, A., & Chakrabarti, T. (2012). A multi-product continuous review inventory 2368–2379.
system in stochastic environment with budget constraint. Optimization Letters, Zheng, Y.-S. (1992). On properties of stochastic inventory systems. Management
6, 299–313. Science, 38(1), 87–103.

View publication stats