Professional Documents
Culture Documents
Improved Genetic Algorithm For SCM
Improved Genetic Algorithm For SCM
a r t i c l e i n f o a b s t r a c t
Article history: Integrating inventory, location, and routing decisions in supply chains may considerably impact their per-
Received 16 February 2018 formance. In this paper, the Inventory Location Routing Problem (ILRP) is considered while adopting the
Revised 22 June 2018 Vendor Managed Inventory (VMI) strategy. A mathematical model is formulated to minimize the total
Accepted 4 September 2018
supply chain cost. Being NP-hard, an Improved Genetic Algorithm (IGA) is designed and used to solve
Available online xxxx
the problem. Two instances (10 and 30 customers) are solved; to study the effect of the total vehicles
capacity (number of available vehicles per depot and vehicle capacity), on the total supply chain cost.
Keywords:
The results show that, the IGA outperforms the GA in reaching lower cost, especially for high number
Inventory location routing problem
Genetic algorithms
of customers. The superiority of the obtained solution performance is basically achieved on the expense
Vehicle capacity of computational time. For the considered problem, the total cost decreases with the increase of vehicle
Supply chain cost capacity due to the usage of fewer depots.
Ó 2019 Production and hosting by Elsevier B.V. on behalf of Ain Shams University. This is an open access
article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction tors such as: FMCGs distribution, reverse logistics networks, solid
waste management, newspapers, . . . etc. An extension to the prob-
One of the key factors in achieving sustainable competitive lem, by determining the optimum inventory levels at the depot
advantage is supply chain network planning. Planning of a supply and/or customers, is made to the VPR resulting in the Inventory
chain distribution network reducing cost and/or improving respon- Routing Problem (IRP). Early tackling of IRP was done by Fisher
siveness is a difficult and dynamic task. The decisions considered et al. in 1982 [2] and Bell et al. in 1983 [3]. Another similar exten-
for planning distribution networks are either tactical or opera- sion was made by Balakrishnan et al., in 1987 [4], who introduced
tional decisions, such as selecting delivery frequency to the cus- an approach to solve the depot location decision integrally with the
tomers, setting the inventory levels at different facilities, routing decisions. This approached was named later by Laporte, in
quantities to be delivered, vehicle routing, determining dispatch- 1988 [5], as the Location Routing Problem (LRP). In 1986, the con-
ing times of vehicles, . . . etc. One of the operational decisions that cept of Supply Chain Management was introduced, introducing the
usually faces the supply chain mangers is the Vehicle Routing opportunity of collaboration and coordination between various
Problem (VRP). VRP is originally tackled in 1959 by Dantzig and supply chain members. The application of supply chain concepts,
Ramser [1] to design the most cost-effective delivery route from such as integrating the decision of supply chain logistical drivers:
a depot to a set of customers. This is considered one of the most Facilities – Inventory – Transportation, resulted in the Inventory
important problems in distribution management especially in sec- Location Routing Problem (ILRP). Liu and Lee, were one of the early
researchers considered inventory decision in LRP in 2003 [6]. ILRP
⇑ Corresponding author. is a mix of tactical and operational decisions are considered which
E-mail addresses: Ahmad_saif@eng.asu.edu.eg (A.S. Saif-Eddine), Mohamed.m. are determining inventory levels in the network, depot(s) location
mohamed@eng.asu.edu.eg (M.M. El-Beheiry), Amin_elkharbotly@eng.asu.edu.eg (s), and vehicles routes. In ILRP these decisions are optimized
(A.K. El-Kharbotly). simultaneously which improves the efficiency of the whole supply
Peer review under responsibility of Ain Shams University. network. In this paper, Vendor Managed Inventory (VMI) strategy
is adopted and integrated in the ILRP. A Heuristic is developed to
determine the delivery quantities to each customer and the deliv-
ery frequency. These values are used as input parameters to a
Production and hosting by Elsevier developed mathematical model; to determine the number and
https://doi.org/10.1016/j.asej.2018.09.002
2090-4479/Ó 2019 Production and hosting by Elsevier B.V. on behalf of Ain Shams University.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
2 A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx
locations of depots, vehicle routing, number of dispatched vehicles, the supplier to the distribution centres, inventory and safety stock
and vehicle dispatching times that minimize the total cost. One key holding costs at the distribution centres, holding inventory cost at
factor affecting the resulted total cost is the total vehicles capacity the customers (as the inventory at the end of period), and the
at each depot. This capacity can be expressed in the number of inventory waste cost. They solved the developed model by Lagran-
vehicles available (of the same capacity) at each depot, or the gian relaxation and introduced a heuristic to get upper bounds.
capacity of the vehicle for the same number of equal-capacity vehi- One of the cases of the ILRP is to deal with stochastic demands.
cles available at each depot. The objective of this paper is to opti- Shen and Qi [14] considered a three-tiered supply chain under cus-
mize the total supply chain cost in ILRP when VMI strategy is tomers’ stochastic demands, fulfilling a certain service level. They
adopted and investigate the effect of changing the total vehicles sought to select the distribution centres’ locations, assign the cus-
capacity at each depot on the total cost and the decision variables. tomers to the distribution centres, determine the ordering fre-
quency from the supplier, and set the safety stock levels at the
2. Literature review distribution centres. They formulated the problem as a Non-
linear Integer Programming model and solved large-size problems
Different cases of the ILRP were dealt with in the literature. The efficiently. Ahmadi-Javid and Azad [15] tackled a stochastic supply
simple ILRP is to select depot(s) location(s), vehicle routes, and chain, with uncertain customers’ demands. They aimed at optimiz-
inventory decisions for deterministic demands. Liu & Lee [6] and ing the depots’ locations, allocation, and capacity levels, as well as
Liu & Lin [7] studied a case of depot location and vehicle routing, the vehicles’ routing, and the order size, frequency of reordering,
taking some inventory control decisions into account. The formers and safety stock level for every distribution centre. They proposed
presented a two-phase heuristic to solve the problem. Their solu- a heuristic and inspected its effectiveness on different sizes and
tion is better than the solution of neglecting the inventory control structures. The conclusion was that the number of used distribu-
decisions. The latter proposed an efficient heuristic which is better tion centres increases by increasing the routing cost’s weight fac-
than the heuristics searching for local optima. They dealt with a tor, while it decreases by increasing the inventory cost’s weight
simple case of inventory; to determine the economical delivery factor. The inventory at the customers was not considered. Fuzzy
quantity, to repeat the routes exactly every certain period. demands were dealt with by Tavakkoli-Moghaddam and Raziei
Ahmadi-Javid and Seddighi [8] dealt with a 3-echelon capacitated [16]. A three-tiered, multi-product, multi-period supply chain
location, routing and inventory problem. They considered the hold- was studied. The fuzzy demands differ among the periods between
ing cost at the distribution centre only. A three-phase heuristic was the products. They sought to determine the distribution centres’
proposed, integrating Simulated Annealing with hybrid Ant Colony locations, the vehicles’ routing, as well as the quantities to be
System. It proved its effectiveness and achieved significant reduc- shipped directly to the distribution centres, and the delivery quan-
tion in the total cost. Guerrero et al. [9] studied a two-echelon sup- tities for the customers. Their objective was to minimize the sum
ply chain. The depots’ and customers’ inventory costs were of shortages at the customers, besides minimizing the total cost.
considered, for the products which are to be consumed in the suc- The considered holding cost dealt with the stored quantities from
ceeding periods, to fulfil deterministic demands. To calculate the a period to another. They formulated the problem as a Multi-
holding cost, they considered the ending inventory of every period. Objective Possibilistic Mixed-Integer Linear Programming model.
They presented a hybrid approach for solving the problem. It Other researchers considered the closed loop supply chain, like
showed a robust performance when tested by large benchmark Wang et al. [17] and Jiang and Ma [18]. They tackled a three-
instances of both the Location Routing Problem and the Inventory echelon closed loop supply chain, to reuse end-of-use products,
Routing Problem. It was able also to optimize the problem globally, having the collection and distribution demands following a Poisson
and achieved significant cost reductions, relative to the classical distribution. There are certain probabilities for the collected prod-
approach. Zhang et al. [10] tackled a two-echelon supply chain, ucts to be reused in a following period. New products are to be
having period-variable deterministic demands for a finite planning ordered and delivered from the plant to any depot, in case that
horizon. The customers’ holding cost was considered and deter- the total usable collected products by that depot are less than
mining the delivery quantities for every customer every period the total distribution demand in the following period. They aimed
was one of their aims. The periods’ ending inventory, as well as half at determining the logistic centres’ locations, vehicles’ routes, and
of the demand of the period, was considered when calculating the logistic centres’ inventory levels of both reused and new products.
holding cost. They developed a hybrid meta-heuristic, which The formers developed a two-phase heuristic algorithm; to solve
proved its effectiveness and efficiency. the stochastic dynamic model formulated. On the other hand,
As FMCGs is one of the main application of the ILRP, many Wang et al. [17] developed a hybrid genetic algorithm, which mod-
researchers dealt with perishable products. Hiassat and Diabat ified by Ahmad et al. [19] to consider selecting a customer as tran-
[11] tackled a problem of determining the warehouse location, shipment node; to decrease the total transportation cost. It was
vehicle routing, as well as the inventory of products at the cus- shown that the transportation cost can be reduced by considering
tomers. They sought for determining the delivery quantities to a transhipment node. Shariff et al. [20] dealt with the same prob-
the customers every period, with deterministic demands that differ lem, aiming at selecting one or more customers to act as the tran-
over the periods, for a perishable product. They considered the shipment node(s), and savings could be achieved. But both Ahmed
ending inventory (of every period) to calculate the inventory hold- et al. [19] and Shariff et al. [20] did not consider the inventory at
ing cost. The problem was formulated as a Mixed Integer Program- the transhipment node(s). These researchers [17–20] considered
ming model, and significant savings were obtained, relative to the inventory at logistics centres only.
multi-step optimization models. Hiassat et al. [12] developed a Another environment is the ILRP in conjunction of electronic
GA to solve a similar problem efficiently, where high quality solu- distribution networks. Li et al. [21] tackled an E-supply chain with
tions were obtained in reasonable time. Rafie-Majd et al. [13] con- returns. The returned merchandise does not have quality defects
sidered a three-echelon, multi-product, multi-period ILRP for and can be resold after a simple treatment at a merchandise centre.
perishable products, over a finite planning horizon using a hetero- The considered inventory cost includes the merchandise centres’
geneous fleet of vehicles. The customers’ demands are stochastic. inventory cost, of the returned merchandise, and the new products
They aimed at optimizing the total cost, which consists of: estab- ordered from the supplier, if the returned merchandise does not
lishing cost of the distribution centres, routing cost from the distri- satisfy the demand of the following period. A hybrid GA-SA algo-
bution centres to the customers, ordering and shipping costs from rithm was developed to solve the problem that outperformed the
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx 3
GA with respect to the optimal solution, computing time, and sta- travelling times and stochastic demands. Their objectives were:
bility. A B2C E-commerce distribution system utilizing VMI strat- minimizing the total cost and minimizing the maximum mean
egy, with fuzzy random demands was tackled by Chen et al. [22]. time spent to deliver to the customers. They included the distribu-
The inventory cost consists of the ordering costs, distribution cen- tion centres’ holding cost, and the ordering cost, in addition to the
tres’ holding costs, and backorder costs, while determining the safety stock cost, while considering the inventory cost. They pre-
order times and target distribution centres’ inventory levels. A sented a Multi-Objective Imperialist Competitive Algorithm, then
two-stage hybrid heuristic is developed, and the model’s effective- validated it, and compared it with three other algorithms. The pre-
ness, and the algorithm’s reliability were shown. sented algorithm showed its superiority. Again, no inventory was
ILRP is solved also under risk and disruptions. Seyedhosseini considered at the customers in this type of research.
et al. [23] dealt with a three-echelon supply chain, having stochas- It is obvious from the literature that many efforts were directed
tic demands. Random disruptions occur at the distribution centres, to solve different configurations of the ILRP, yet it seldom tackled
which can cause lost-sales. Customers, who need the product the inventory at the customers. Even when tackled, the holding
urgently and there is not enough quantity will source it from cost was roughly estimated, as the holding cost over the whole per-
another supplier(s) i.e. it will be lost sales, while backordering is iod for example. Also, it was rarely tackled to introduce inventory
considered for other customers. They considered the distribution management strategies to the ILRP such as Quick Response, VMI,
centres’ holding cost of the order quantities, as well as the safety Distribution Integration . . . etc. Dealing with continuous demand
stocks. A Mixed Integer Non-Linear Programming model was for- at the customers in conjunction with periodic decision making
mulated, and a meta-heuristic method was implemented to solve for ILRP was not frequently tackled in literature. In this paper,
it, then a sensitivity analysis was conducted on the important VMI strategy is adopted to be used at the customers, when having
parameters. Multi-objective optimization was dealt with by deterministic continuous demand. A mathematical model is pro-
Tavakkoli-Moghaddam et al. [24]. They considered a three-level posed for the ILRP, where the holding cost at the customers is mod-
supply chain having retailers’ uncertain demands. The aim was to elled more accurately. The mathematical model is solved with an
determine the distribution centres’ locations and capacities, allo- Improved Genetic Algorithm (IGA). The remainder of this paper is
cate retailers to the distribution centres, which have to be allocated organized as follows: the problem is defined, and the model is
also to the suppliers, decide the transportation decisions for both developed in Section 3, then Section 4 presents the solution
levels, which include selecting the vehicle type and its routing, as methodology. Section 5 provides the numerical experiments, while
well as the inventory control decisions, such as ordering quantities the reached results and discussion are presented in Section 6, and
from the supplier to the distribution centres, the distribution cen- finally, conclusions and future work are covered in Section 7.
tres’ safety stock levels, and their delivery quantities to their retail-
ers. The objectives were: minimizing the total costs minimizing the 3. Problem description and model formulation
transportation time of the product along the supply chain. The
holding cost at the distribution centres is considered and calcu- A Third-Party Logistics (3PL) provider is taking charge of the
lated for the average ordered quantities, in addition to the safety logistics activities of a manufacturer (factory) producing a single
stock holding cost. Nekooghadirli et al. [25] considered a bi- product by shipping goods from its facility to one or more depot(s),
objective two-echelon multi-product ILRP, having probabilistic keeping inventory at the depot, if needed, then delivering these
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
4 A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx
goods to customers having constant demand rates. The depots are 3.1.3. Auxiliary sets and variables
already existing and the service is offered by the 3PL provider. A
schematic drawing of the supply chain network is given in Fig. 1.
The 3PL provider adopted the VMI strategy, by keeping more Vkp: Set of customers to be visited k 2 K , p = 1 , 2 , . . ., P.
inventory at the customers’ facilities and incurring a part of their by vehicle ‘k’ in period ‘p’,
holding cost; to be able to make less frequent deliveries to the cus- Skp: Non-empty sub-set of Vkp, k 2 K, p = 1, 2, . . ., P.
tomers. The 3PL provider must take decisions about the frequency stpijk: Start time for traversing arc i, j 2 NC [ ND, i – j,
of deliveries to customers, then based on them, decisions about (i, j) by vehicle ‘k’ in period ‘p’, k 2 Ko, o 2 ND, p = 1, 2,
which depots will be used, inventory levels at the depots, number ... P
of vehicles dispatched, vehicles dispatching times and vehicles ndi: No. of deliveries to be made to i 2 NC
routing are made. The vehicles used from the factory to depot(s) customer ‘i’ within the
are large enough to deliver the quantity needed at each depot in repeated period, indexed by n,
one shipment, and the delivery will be made, such that, the prod- EIni : Inventory level of product at n = 1, 2, . . ., ndi, i 2 NC
ucts can be delivered to customers at the beginning of each period. customer ‘i’ just before the ‘nth’
Also, the depot’s capacity is large enough to accommodate any delivery,
quantity, the number of vehicles available at each depot is limited tdni : Time between delivery ‘n’ at n = 1, 2, . . ., ndi, i 2 NC
and constant, and all vehicles have the same capacity, i.e., homoge- customer ‘i’ and the following
neous fleet. delivery ‘n + 1’,
The following assumptions are considered on modelling the
problem: no transhipment is considered between the depots, no
ending inventory at the depot in each period, i.e., products arrive
and are dispatched to customers in the same period, vehicles’ 3.1.4. Decision variables
routes start and end at the same depot before the end of the period,
deterministic and known travelling times are considered, no wait-
ing is allowed at the customers, the customer has an unlimited ri: Period multiplier, defining the i 2 NC
storage capacity, and no shortage is allowed. number of periods, every
The problem is formulated; to minimize the total cost, aiming at which, customer ‘i’ is visited,
determining the following: the delivery quantities for each cus- qi: Delivery quantity for customer i 2 NC
tomer and the periods in which they should by delivered, the loca- ‘i’,
tions of used depots in each period, and the number and routes of ypi : A binary variable, defining i 2 ND, p = 1, 2, . . ., P
the vehicles dispatched from each depot, as well as their dispatch- whether depot ‘i’ is used in
ing times. period ‘p’,
xpijk: A binary variable, defining i, j 2 NC [ ND, i – j,
3.1. Nomenclature whether vehicle ‘k’ passes k 2 Ko, o 2 ND, p = 1, 2,
through arc (i, j) in period ‘p’, . . ., P
3.1.1. Indices and sets stpijk: Dispatching time for vehicle ‘k’ i 2 ND, j 2 NC, k 2 Ki,
from depot ‘i’ to pass through p = 1, 2, . . ., P
arc (i, j) in period ‘p’,
NC: Set of customers, indexed by i, and j
ND: Set of potential depots, indexed by i, j, and o
Ko: Set of available vehicles at depot ‘o’, indexed by k o 2 ND
P: Total number of periods, making the repeated 3.2. The period multiplier and delivery quantity heuristic
period, indexed by p
The heuristic proposed, to assign customers to periods and
decide the delivery quantity, is presented in this subsection.
The power-of-two rule is used to assign customers to periods,
3.1.2. Parameters where the period multiplier is limited to be a power-of-two inte-
ger. This ensures that every customer is visited in even or odd
periods only.
Di: Demand rate of customer ‘i’ per i 2 NC Since the objective is to minimize the supply chain cost, then it
period, is preferred to visit far customers, having low demand rate, as well
C: Capacity of any vehicle in number as low unit inventory holding costs less often. Based on these con-
of units, siderations, a heuristic is developed to specify the period multiplier
dco: Transportation cost from factory to o 2 ND for every customer. The heuristic extends the heuristic developed
depot ‘o’ per period, by El-Shawarby et al. [26], which was used to determine the peri-
fck: Dispatching cost of vehicle ‘k’, k 2 Ko, o 2 ND odicity of production for every product in a flow-shop scheduling
tc: Transportation cost per unit problem, to be used in ILRP. The heuristic is as follows:
distance,
hdo: Unit inventory holding cost at o 2 ND Step 1: Determine the ratio ‘Ti’ for every customer:
depot ‘o’ per period, min fdij g
hci: Unit inventory holding cost at i 2 NC Ti ¼ j
; i 2 NC; j 2 ND.
h:Di
rffiffiffiffiffiffiffiffiffiffiffiffiffi
customer ‘i’ per period,
Step 2: Calculate Ri = log2 minTfi T i g and round it to the nearest
dij: Distance of arc (i, j), i, j 2 NC [ ND, i – j i
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx 5
Step 4: Determine the delivery quantity for each customer: dispatched from a depot, as given in Eq. (3). On the other hand,
qi = ri Di. the vehicle’s routing cost is a linear function of the travelled dis-
The delivery quantity is inspected whether it violates the vehi- tance, as given in Eq. (4).
cle capacity. If it violates the vehicle capacity, divide the period X X XX
multiplier by 2. The modified delivery quantity is re-inspected. xpojk fck ð3Þ
p2P o2ND k2K o j2J p
Step 5: Determine the total number of periods, making the
repeated period: P ¼ max fr i g. X X X X X
i
xpijk tc dij ð4Þ
Step 6: Determine the number of deliveries made for each cus- p2P o2ND k2K o j2J p [fog=fig i2J p [fog
tomer: ndi ¼ rPi
3.3. Assigning customers to periods heuristic 3.4.3. Customers’ vendor managed inventory holding cost modelling
According to the literature surveyed, the holding cost during the
After specifying the periodicity of delivering to each customer, period was never considered. To approach the real case, the hold-
the first period at which delivering starts to that customer should ing cost is considered within every period, taking into considera-
be specified. A heuristic is developed to assign customers to peri- tion the arrival time of the product at every customer. The
ods. The heuristic is as follows: holding cost is calculated by multiplying the holding cost per unit
per unit time, by the average inventory level, then, by inventory
Step 1: Start assigning all customers having a period multiplier holding time between replenishments. Since it is assumed that
r = 1 to every period. Then, for every period ‘p’, update ‘Jp’, which there is no opening inventory at the customer before the first deliv-
is the set of customers assigned to period ‘p’, p = 1, 2, 3, . . . P. ery, then the first repeated period is not representing the steady
Step 2: Consider the next period multiplier r = 2 r. state and its costs should be discarded. The steady state inventory
Step 3: Get all unassigned customers having a period multiplier levels are considered to calculate opening inventory level for the
ri = r. first steady state repeated cycle.
Step 4: Assign the unassigned customer that has the maximum To determine the number of periods after which the steady
delivery quantity to the first period found to have the least total state is reached, four cases, that may occur at any customer until
delivery quantity ‘fri’. Assign it also to the following periods reaching the steady state, are discussed and given in Fig. 2, where
spaced apart from that period by the period multiplier ‘ri’ as P = 8 periods, ri = 2, and qi = 10 units:
well as its multiples (‘fri’, ‘fri + ri’, ‘fri + 2ri’, . . ., ‘fri + (n-1) ri’,
. . ., ‘fri + (ndi-1) ri’), where n = 1, 2, 3, . . . ndi. Then update ‘Jp’. a. Deliveries are done exactly at the same times for every visit
(Fig. 2 (a)).
Delivery ‘n’ for customer ‘i’ is done in period ‘p’, where: b. Deliveries are done at different times for every visit, where
shortage happens at first period (Fig. 2 (b)).
p ¼ fri þ ðn 1Þ ri c. Deliveries are done at different times for every visit, where
surplus inventory remains from the previous delivery at
the arrival of the succeeding delivery (Fig. 2 (c)).
Step 5: Repeat steps 3 and 4, till all customers having a period d. Deliveries are done at different times for every visit, where
multiplier of r are assigned. If r = P, then END, if else, GOTO step 2. there are both shortage and excess inventory in the first
repeated period (Fig. 2 (d)).
3.4. Cost modelling
Since there is no opening inventory, shortage occurs until the
3.4.1. Modelling of depot related costs first delivery is done. Hence, the inventory level is zero until the
There are two costs related to the depot. First, the transporta- first delivery.
tion cost to depot is given by Eq. (1), which represents which depot
will be used in period ‘p’, multiplied by the transportation cost to EI1i ¼ 0 ð5Þ
that depot. The other depot related cost is the inventory holding
After the delivery is made, the available inventory is consumed
cost of products, which is the result of multiplying the holding cost
continuously, with a fixed demand rate, the inventory holding
per unit per unit time, by the quantity of products at the depot, by
time, based on which the holding cost is calculated, is considered
the time this quantity stays at the depot. Since the whole quantity
from the first delivery, till either the inventory level reaches zero,
arrives at the beginning of any period, and is dispatched at the
after ‘qi/Di’ time, or another delivery is made, after ‘td1i 0 time. The
same period, then the quantity dispatched by any vehicle ‘qi’ is
inventory holding time is given by min. {qi/Di, td1i }
multiplied by the dispatching time of the vehicle ‘stpojk’, which rep-
The ending inventory before the following delivery ‘EI2i 0 is either
resents the inventory holding time of this quantity, i.e., the time
zero, or the difference between the available quantity after the pre-
this quantity stays at the depot, multiplied by the holding cost
vious delivery ‘EI1i + qi’ and the consumed quantity ‘td1i Di’. That is
per unit per unit time ‘hdo/pt’, as given by Eq. (2).
represented by:
X
P X þ
p
ypi dco ð1Þ 1
EI2i ¼ EI1i þ qi tdi Di
p¼1 o2ND
00 1 0 11 Where:
X
P X X X X X p X X X X X X
@@ hd 1
tdi ¼ r i pt þ
fr þr i fr þr i
st ijki
fr fr
xijki stijki
xijk qi A @ AA xijki
p p o
xojk st ojk ð2Þ
p¼1 o2ND k2K o j2J p [fog=fig i2J p j2J
pt o2ND k2K o j2J p [fog o2ND k2K o j2Jp [fog
p
The time between the first and second deliveries ‘td1i 0 equals the
3.4.2. Vehicle cost modelling number of periods between the two deliveries, multiplied by the
The vehicle cost includes dispatching cost and routing cost. The period time, in addition to the difference between the arrival time
vehicle dispatching cost is a fixed amount paid for every vehicle of the second delivery minus the arrival time of the first delivery,
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
6 A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx 7
X X
xpijk ¼ xpjik 8i 2 ND [ NC; o 2 ND; k 2 K o ; p ¼ 1; 2; ; P customer, i.e., waiting is forbidden, as indicated by constraints
j2ND[NC=fig j2ND[NC=fig (18). Constraints (19) ensure that every vehicle should return to
ð13Þ its dispatching depot before the period’s end. Constraints (20)
X X and (21) are the integrality constraints, while constraints (22) are
xpijk Skp 18Skp # V kp ; Skp –;; o 2 ND; k 2 K o ; p ¼ 1; 2; ; P the non-negativity constraints.
i2Skp j2Skp
The objective function (8) seeks to minimize the total cost. Con- 4.3. Creation of initial population
straints (9) imply that each vehicle is dispatched at most once dur-
ing every period. Constraints (10) dictate ending every vehicle’s Step 1: Generate random fractional numbers for the dispatching
route at the same depot it was dispatched from, if it is used in this times’ sub-chromosome, from a U(0, 1) distribution. Then
period. Constraints (11) ensure that only one vehicle should reach actual dispatching times are calculated as follows:
any customer assigned to the period only once. Constraints (12) Actual Dispatching time = start of period + fractional number * pt.
ensure that every customer, which is not assigned to the period, Step 2: Set p = p + 1, and Select randomly a depot that has at
is not visited, and ensure that any vehicle dispatched from depot least one unused vehicle.
‘o’ cannot reach another depot. Constraints (13) guarantee that, if Step 3: Select randomly an unassigned customer of period ‘p’,
a vehicle reached any node ‘i’, it should leave it, and if it does and assign it to the first unused vehicle, then check if it is fea-
not, it cannot leave it. Constraints (14) are the sub-tour elimination sible to visit the customer and return to the depot before the
constraints. Constraints (15) and (16) ensure that a depot is used if end of the period, then confirm the assignment, otherwise,
and only if at least a single vehicle is dispatched from it during that select another customer. If all customers to be served during
period, and that a vehicle can’t be dispatched from a closed depot. this period are assigned, go to step 8.
Constraints (17) indicate that the vehicles’ capacities are not Step 5: Find the customer(s) that can be inserted in the route of
exceeded. When the vehicle reaches a customer, products are selected vehicle without violating its maximum capacity to
immediately unloaded, then, the vehicle moves directly to the next form set C.
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
8 A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx
Fig. 3. Flow chart for the application of improvement heuristics to the GA.
Step 6: For every customer belong to C find all feasible insertion 4.4. Improvement procedure
(s) in the selected vehicle route, and find the corresponding cost
(dispatching + routing) of each, then select the least possible Two approaches are applied in the improvement procedure, and
insertion. If no feasible insertion found then go to step 3, Other- their resulting chromosomes are compared, and the one with the
wise for other feasible insertion go to step 5. least total cost is chosen. The chromosome undergoes the three
Step 8: If p < P, go to step 2. Otherwise, proceed to step 9. improvement stages successively in the first approach, while in
Step 9: Convert the dispatching times to fractional numbers by: the second approach, the cost of both chromosomes (before and
after each improvement stage) will be evaluated, and the one
Fractional number ¼ ðdispatching time start of periodÞ=
having the minimum cost will go through the next stage of
ðmax: feasible dispatching time for the route start of periodÞ improvement.
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx 9
4.4.1. Assign-to-Cheapest-Depot Improvement (ACDI) route ‘i’ from depot ‘j’. If it is timely infeasible to perform the
route from depot ‘j’, the corresponding cost is ‘M’.
Step 1: Set p = 1. Step 4: For each row, subtract the cost of dispatching and rout-
Step 2: Find the used vehicles in each depot in period ‘p’. ing the route from the currently dispatching depot from the
Step 3: Construct a cost matrix CDM. The rows represent the whole row. A zero indicates the currently dispatching depot,
routes of the used vehicles, and the columns represent the or an alternative. A negative value indicates a saving, and a pos-
depots. An entry ‘cdmij’ represents the cost of dispatching and itive value indicates a cost increase.
routing of route ‘i’ from depot ‘j’. For any route ‘i’, if it is infea- Step 5: If there is any negative value in the matrix, proceed to
sible to be originated from depot ‘j’, then, the corresponding step 6. Otherwise, go to step 13.
cost is set to indefinitely large number ‘M’. Step 6: Identify the depot ‘dn’ that corresponds to the min. neg-
Step 4: If the number of columns > 1, proceed to step 5. If the ative value throughout the matrix, to be the new dispatching
number of rows = 1, assign the remaining route to the least cost depot. Identify the corresponding route ‘rc’ to be the considered
dispatching depot, and go to step 8. If the number of col- route.
umns = 1, assign the remaining routes to be originated from this Step 7: If the depot ‘dn’ has an unused vehicle, assign route ‘rc’
remaining depot, and go to step 8. to be dispatched from that depot, subtract the min. negative
Step 5: Calculate a penalty for each row and column, that is the value of the row from the whole row, and go to step 5. Other-
difference between the minimum two costs in that row or col- wise, proceed to step 8.
umn, then find the maximum penalty. Step 8: Identify the candidate routes ‘cr’, that are dispatched
Step 6: Assign the route/depot with the maximum penalty to from depot ‘dn’, to have a route moved from depot ‘dn’, to free
the least cost depot/route. a vehicle for route ‘rc’.
Step 7: Remove the row of the assigned route from the matrix. If Step 9: Identify the candidate depots ‘dd’, which are the unused
the depot has all its vehicles used, remove its column from the depots that has an unused vehicle at least each, in addition to
matrix. Go to step 4. the depot that currently dispatched route ‘rc’.
Step 8: If p < P, set p = p + 1 and go to step 2. Otherwise, stop. Step 10: Find the min. cost of dispatching routes ‘cr’ from
depots ‘dd’. Identify the corresponding candidate route and
4.4.2. Depot closing improvement (DCI) depot.
Step 11: If the summation of the min. negative value through-
Step 1: Set p = 1. out the matrix and the min. cost of dispatching routes ‘cr’ from
Step 2: Find the used vehicles in each depot in period ‘p’. depots ‘dd’ is negative, proceed to step 12. Otherwise, set the
Step 3: Close the depot that has the least number of used vehi- min. negative value throughout the matrix to be ‘M’, and go to
cles, and consider its routes as unassigned routes. step 5.
Step 4: Construct a cost matrix DCM. The rows represent the Step 12: Move the candidate route to the candidate depot, and
unassigned routes, and the columns represent the used depots the considered route ‘rc’ to the new dispatching depot ‘dn’. Sub-
that have unused vehicles. An entry ‘dcmij’ represents the cost tract the min. negative value throughout the matrix from its
of dispatching and routing of route ‘i’ from depot ‘j’. If it is corresponding row in the matrix. Subtract the min. cost of dis-
timely infeasible to perform the route from depot ‘j’, the corre- patching routes ‘cr’ from depots ‘dd’ from the corresponding
sponding cost is ‘M’. row. Go to step 5.
Step 5: If the number of columns > 1, proceed to step 5. If the Step 13: If p < P, set p = p + 1 and go to step 2. Otherwise, stop.
number of rows = 1, assign the remaining route to the least cost
dispatching depot, and go to step 11. 4.5. Crossover operator
Step 6: Calculate a penalty for each row (column), that is the
difference between the minimum two costs in that row 4.5.1. 2-Point crossover for the dispatching time fractions’ Sub-
(column). Chromosome
Step 7: Find the max. overall penalty among all rows and col- The fractional numbers are multiplied by 1024, to be integer
umns. Assign the corresponding route to the corresponding numbers, and then, transformed to the equivalent binary from.
depot, which cause the least cost in the row/column of the The 2-point crossover is conducted on the binary form. The
max. overall penalty. resulted binary form of the offspring is transformed back to its
Step 8: Remove the row of the assigned route from the matrix. If integer form, and then to its fractional number.
the depot has all its vehicles used, remove its column from the
matrix. If there is any remaining column with feasible dis- 4.5.2. Best-Route-Best-Feasible-Insertion (BRBFI) crossover for the
patches, go to step 5. Otherwise, go to step 9. routes’ sub-chromosome
Step 9: For each closed depot, find the cost of dispatching and The proposed crossover operator depends on the Best Cost-Best
routing all the unassigned routes. Timely infeasible dispatched Route Crossover (BCBRC), which was used by Ghoseiri & Ghannad-
are considered as ‘M’. pour [27]. Each period is dealt with separately. The route, having
Step 10: Select the depot that has the least total dispatching and the cheapest route cost per customer, is chosen in every parent,
routing cost of the remaining unassigned routes. Assign the separately. The route cost per customer consists of the vehicle dis-
remaining routes to that depot. patching, routing, and its share of the transportation cost from
Step 11: If p < P, set p = p + 1 and go to step 2. Otherwise, stop. manufacturer to depot(s) (the transportation cost is divided by
the number of dispatched vehicles in that period). The sum is then
4.4.3. Exchange-Depots Improvement (EDI) divided by the number of its customers to yield the route cost per
customer. The customers of the cheapest route cost of the first par-
Step 1: Set p = 1. ent are omitted from their routes in the second parent, and then re-
Step 2: Find the used vehicles in each depot in period ‘p’. inserted in the cheapest routing cost feasible insertion, sequen-
Step 3: Construct a cost matrix EDM. The rows represent the tially, in the descending order of their delivery quantities. The
used vehicles, and the columns represent the used depots. An same happens for the customers of the cheapest route cost of the
entry ‘edmij’ represents the cost of dispatching and routing of second parent when omitted from their routes in the first parent.
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
10 A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx
This crossover operator seeks for minimizing the number of dis- The values in bold are the default values, which are fixed while
patched vehicles, the number of used depots, besides the total cost. other parameter(s) changes. The holding cost at the customers is
be taken as ten times that at the depot, while the holding cost at
4.6. Mutation operator the depot is one tenth of the transport cost from the depot to the
customers. The transportation cost from the depot to the cus-
4.6.1. Inversion mutation for the dispatching time fractions’ sub- tomers per unit distance is 1/400 from that from the factory to
chromosome the depot. The transportation cost from the factory is not related
The fractional numbers are multiplied by 1024, to be integer to the distance as no routing occur and the factory is far enough
numbers, and then, transformed to the equivalent binary from. than the proposed depot locations, therefore the cost is not chan-
The inversion mutation is conducted on the binary form after ged with the change of the depot location. The GA parameters used
selecting two points randomly. The resulted binary form is trans- are: population size: 40, number of generations between
formed back to its integer form, and then to its fractional number. improvements: 25, total number of generations: 250, crossover
fraction: 0.8, number of elite chromosomes: 1. Each of the three
considered instances with default parameters were solved 20
4.6.2. Best/Worst-Insertion-Of-Random-Number-Of-Worst/Best-
times to stand on the precision and convergence of the IGA. The
Routes (BWIRNWBR) mutation operator for the routes’ sub-
maximum deviation is found to be 0.3%, 2.2% and 3.3% for the 10,
chromosome
30 and 50 customers instances, respectively. This implies that
A random number of routes ‘nr’ is chosen. The customers of the
the proposed IGA has a good precision, therefore, on considering
‘nr’ cheapest route costs per customer, are omitted and re-inserted
any other parameter, the instance is solved only 3 times and the
in the largest routing cost feasible insertions, sequentially, in the
best solution is considered for analysis.
descending order of their delivery quantities. In contrast, the cus-
tomers of the ‘nr’ largest route costs per customer, are omitted
and re-inserted in the cheapest routing cost feasible insertions,
6. Results and discussion
sequentially, in the descending order of their delivery quantities.
6.1. The developed improvement effect on the solution on the ILRP
5. Numerical experiments
The first set of instances are solved; to show how the proposed
In order to evaluate the effectiveness of the proposed IGA and to improvements to the IGA affect the solution value, convergence
determine the parameters affecting the decisions related to the and time. In Figs. 4–6, the best reached solution is plotted at each
vehicles at each depot, a benchmark instance (as quoted from generation in case of solving using IGA, and on solving the
Canada Research Chair in Distribution Management web-site instances using GA (no Improvement is made to the GA), for
[28]) is modified from the Multi-Depot Vehicle Routing Problem 180 min as a stoppage criterion (which represents 4 times the
(MDVRP) literature; to cope with the studied case. The extracted
instance, namely p01, was first solved to test the effectiveness of
the developed GA operators and improvements, as no benchmark 1100
instance was found for the ILRP. The problem was solved for three
1050
times and the best reached solution was found to be higher than
the best-known solution by 7.5%, which is considered acceptable
1000
deviation as the IGA is designed to solve ILRP.
Total Cost per Period
Table 1
The parameters of the different sizes.
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx 11
2500 It is obvious that the IGA may result in a better solution, especially
at larger problems on the expenses of the solution time. In case of
2300
IGA, the solution time is longer than the GA by about 4 min. but
2100 yields to lower cost by 10.18%. The results of these runs, given in
table 2, show that even the run time for the 10, 30, and 50 cus-
Total Cost per Period
1100 6.2. The effect of vehicle capacity on the total cost per period
900
Increasing the vehicle capacity allows decreasing the total cost.
700 This is clear in Fig. 7 for the 10 customers instance, and Fig. 8 for
0 500 1000 1500 2000 2500 3000 3500 4000 4500 the 30 customers instance. As the number of available vehicles at
Generation each depot increases, the total cost decreases in a stepped manner
IGA GA in both instances. This is due to that, as the number of vehicles per
depot increases, the need to use additional depot demolishes, as
Fig. 5. The best solution reached at each generation for the 30 customers problem.
given in Figs. 9 and 10 for both instances, respectively. The trans-
portation cost from factory to the depot(s) may represent 20%
70% of the total cost, hence, it is the most governing cost parame-
4200 ter, which governs the pattern of the total cost, although the rout-
ing cost increases with the decrease of the number of depots
3700 (driven by increasing the number of vehicles per depot), as given
in Figs. 11 and 12. The routing cost increases because customers
3200 are randomly located with respect to the potential depots loca-
Total Cost per Period
2700
2000
2200 1800
1600
1700
1400
Total Cost/Period
1200
1200
700 1000
0 500 1000 1500 2000 2500 3000
800
Generation
IGA GA 600
Fig. 6. The best solution reached at each generation for the 50 customers problem. 400
200
Table 2
Comparing the performance of GA and HGA in solution value and time.
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
12 A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx
3500 300
3000
250
2500
Total Cost/Period
200
RoutingCost/Period
2000
150
1500
100
1000
50
500
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
No. of Vehicles/Depot No. of Vehicles/Depot
40 60 80 120 200 40 60 80 120 200
Fig. 8. Total cost at different vehicles capacities with the increase of number of Fig. 11. Routing cost at different vehicles capacities with the increase of number of
vehicles per depot (30 customers instance). vehicles per depot (10 customers instance).
3.5 700
3 600
2.5 500
No. of Used Depots
Routing Cost/Period
2 400
1.5 300
1 200
0.5 100
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
No. of Vehicles/Depot
No. of Vehicles/Depot
40 60 80 120 200
40 60 80 120 200
Fig. 12. Routing cost at different vehicles capacities with the increase of number of
Fig. 9. Number of used depots at different vehicles capacities with the increase of
vehicles per depot (30 customers instance).
number of vehicles per depot (10 customers instance).
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx 13
References
0.2
[1] Dantzig GB, Ramser JM. The truck dispatching problem. Manage Sci
1959;6:81–91.
0 [2] Fisher ML, Greenfield A, Jaikumar R, Kedia P. Real-time scheduling of a bulk
0 50 100 150 200 250
delivery fleet: Practical application of Lagrangean relaxation. Technical report,
Vehicle Capacity The Wharton School, University of Pennsylvania; 1982.
10 30 [3] Bell W, Dalberto L, Fisher ML, Greenfield A, Jaikumar R, Kedia P, et al.
Improving the distribution of industrial gases with an on-line computerized
Fig. 13. Vehicle utlization at different vehicles capacities when four vehicles are routing and scheduling optimizer. Interfaces 1983;13(6):4–23.
[4] Balakrishnan A, Ward J, Wong R. Integrated facility location and vehicle
available per depot.
routing models: Recent work and future prospects. Am J Math Manage Sci
1987;7:35–61.
[5] Laporte G. Location-routing problems. In: Golden B, Assad A, editors. Vehicle
16 Routing: Methods and Studies. Amsterdam: North-Holland; 1988. p. 163–98.
[6] Liu SC, Lee SB. A two-phase heuristic method for the multi-depot location
14 routing problem taking inventory control decisions into consideration. Int J
Adv Manuf Technol 2003;22(11):941–50.
[7] Liu SC, Lin CC. A heuristic method for the combined location routing and
12
inventory problem. Int J Adv Manuf Technol 2005;26(4):372–81.
No. Of Dispatched Vehicles
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002
14 A.S. Saif-Eddine et al. / Ain Shams Engineering Journal xxx (xxxx) xxx
[24] Tavakkoli-Moghaddam R, Forouzanfar F, Ebrahimnejad S. Incorporating Mohamed M. El-Beheiry is an assistant professor the
location, routing, and inventory decisions in a bi-objective supply chain Department of Design and Production Engineering,
design problem with risk-pooling. J Indus Eng Int 2013;9(1):19. Faculty of Engineering, Ain Shams University, Cairo,
[25] Nekooghadirli N, Tavakkoli-Moghaddam R, Ghezavati VR, Javanmard S. Solving Egypt. He has a degree in Mechanical Engineering
a new bi-objective location-routing-inventory problem in a distribution (1994) and an MSc in Production Planning and
network by meta-heuristics. Comput Ind Eng 2014;76:204–21. Scheduling (1999) from the Ain Shams University. He
[26] El-Shawarby I, Sarin SC, El-Kharbotly AK. Multi-stage Common Cycle also has a PhD degree in Supply Chain Management
Scheduling Problem Ph.D. Thesis. Cairo, Egypt: Ain Shams University; 2000. (2004). His research areas include production schedul-
[27] Ghoseiri K, Ghannadpour SF. Multi-objective vehicle routing problem with ing, inventory control, supply chain dynamics, and
time windows using goal programming and genetic algorithm. Appl Soft
simulation.
Comput 2010;10(4):1096–107.
[28] Canada Research Chair in Distribution Management web-site, http://neumann.
hec.ca/chairedistributique/data/mdvrp/old/, last accessed: 22 July 2017.
Ahmad S. Saif-Eddine is a lecturer assistant and a Ph.D. Amin K. El-Kharbotly is a professor at the Design and
student at the department of Design and Production Production Engineering Dept., Faculty of Engineering,
Engineering, Faculty of Engineering, Ain Shams Univer- Ain Shams University, Cairo, Egypt since 1993. He has
sity, Cairo, Egypt. He has a B.Sc. in Mechanical Engi- received his Ph.D. in 1977 from Ain Shams University.
neering (2005), and a M.Sc. in Production Planning and
Scheduling (2011) from Ain Shams University.
Please cite this article as: A. S. Saif-Eddine, M. M. El-Beheiry and A. K. El-Kharbotly, An improved genetic algorithm for optimizing total supply chain cost in
inventory location routing problem, Ain Shams Engineering Journal, https://doi.org/10.1016/j.asej.2018.09.002