Int. J.

Production Economics 182 (2016) 263–275

Contents lists available at ScienceDirect

Int. J. Production Economics

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

Multi-period inventory routing problem under carbon emission

regulations
Chun Cheng a,b, Mingyao Qi a,n, Xingyi Wang a,b, Ying Zhang a,b
a
Research Center on Modern Logistics, Graduate School at Shenzhen, Tsinghua University, Shenzhen 518055, China
b
Department of Industrial Engineering, Tsinghua University, Beijing 100084, China

art ic l e i nf o a b s t r a c t

Article history: This paper analyzes the impacts of carbon emission regulations on the traditional inventory routing
Received 21 January 2015 problem (IRP). We first present the traditional IRP model, which is an inbound commodity collection
Received in revised form system consisting of one assembly plant and a set of geographically dispersed suppliers. At the beginning
8 January 2016
of each period, a fleet of capacitated identical vehicles depart the depot to pick up products from sup-
Accepted 12 August 2016
pliers to meet the assembly plant’s demand, which is deterministic and time-varying. We use fixed
Available online 4 September 2016
transportation cost, fuel consumption cost and inventory holding cost to evaluate the system’s total cost,
Keywords: in which fuel consumption cost is determined by fuel consumption rate, distance and fuel price. Then we
Inventory routing problem investigate the impacts of carbon emission regulations on the traditional IRP problem, wherein carbon
Carbon emission regulation
emissions are generated by fuel consumption. A series of mixed integer nonlinear programming models
Green logistics
are constructed and linearization methods are used. A hybrid genetic algorithm based on allocation first
Hybrid genetic algorithm
and routing second is proposed to find near-optimal solutions for these problems. Numerical tests are
performed to show the effectiveness of the proposed algorithm, and several managerial insights are
observed from parameter sensitive analyses which may help both the government and the industry to
adopt appropriate carbon reduction regulations.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction choose to invest in carbon offset projects to obtain extra carbon

It is well known that global warming is among the greatest
challenges in this century, and carbon emission is a dominant cause
of global warming. To respond to the challenge and control carbon
emissions effectively, many governments and researchers have be-
gun to pay close attention to green logistics. Recently, the Chinese
government published an outline about the development planning
of logistics between 2014 and 2020, in which the government ex-
plicitly noted that the whole society should develop green logistics
and reduce carbon emissions (Chinese State Council, 2014). Many
other governments and organizations (e.g., the United States, the
United Nations and the European Union) have also designed me-
chanisms or enacted legislations to control carbon emissions. Cur-
rently, there are four carbon emission regulation policies: carbon
cap, carbon cap and offset, carbon cap and trade, and carbon taxing
(Konur and Schaefer, 2014). Under the carbon cap policy, companies
plan their operations so that the level of carbon emissions, known
as a carbon cap, is not violated. Under the carbon cap and offset
policy, a company is subjected to a carbon cap, but the company can
n
Corresponding author.
E-mail address: qimy@sz.tsinghua.edu.cn (M. Qi).
emission credits. In other words, the company can get extra emis-
sion credits by paying for them. The carbon cap and trade policy
means that a company can buy (or sell) emission credits if its
emission level is higher (or lower) than the cap. Lastly, under the
carbon taxing policy, a company must pay for its carbon emissions
in taxes.
Supply chain activities, such as production, transportation and
inventory, all contribute to carbon emissions. However, transpor-
tation is the most visible sector of the supply chain that emits
most of the carbon (Dekker et al., 2012). The Intergovernmental
Panel on Climate Change (IPCC) reported that transportation ac-
counted for 14% of the greenhouse gas emissions by economic
sectors in 2010 (IPCC, 2014). Therefore, curbing the carbon emis-
sions by transportation will make a good environmental sense. To
achieve this goal, many companies tend to adopt more fuel effi-
cient vehicles, equipment and facilities (Hua et al., 2011a). They
can also adjust operational decisions to reduce carbon emissions,
which may provide a greater reduction in carbon emissions with
less cost than using a low-energy-consumption technology (Ben-
jaafar et al., 2013). Unfortunately, many companies and re-
searchers seem to have overlooked this solution to environmental
preservation, and studies about operations decisions under carbon
emission regulations are scarce (Hua et al., 2011b). To enrich the

http://dx.doi.org/10.1016/j.ijpe.2016.09.001
0925-5273/& 2016 Elsevier B.V. All rights reserved.
264 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275
research in this area, an inventory routing problem is studied in
this paper under the four carbon emission regulations mentioned
above.
The IRP is to determine simultaneously the optimal inventory
control strategy and vehicle schedules to minimize the supply
chain’s total cost. In the past, inventory and transportation costs
have been minimized separately by different departments without
considering their relationship. However, the interrelationship be-
tween the two has recently motivated many researchers to model
them simultaneously. Studies show that the IRP can indeed reduce
the supply chain’s total cost (Ramkumar et al., 2012; Coelho et al.,
2014; Qin et al., 2014). In this paper, we explore the effects of
carbon emission regulations on a multi-period inventory routing
problem (MIRP) and examine the relationships involved.
The remainder of the paper is organized as follows. In Section
2, we review previous research on IRP and green supply chains,
and the contributions of this paper are also noted. We describe our
problem and present the mathematical formulations in Section 3.
The hybrid genetic algorithm is described in Section 4. The com-
putational tests and analyses are presented in Section 5. We draw
some conclusions and suggest directions for future research in
Section 6.
2. Literature review
IRP can be mainly classified according to the following criteria:
time horizon, supply chain topology, demand pattern, inventory
strategy and vehicle fleet (Andersson et al., 2010). In addition,
several other types of IRP can also be found depending on different
assumptions in the problems. In this paper, we present the studies
in terms of time horizon, which are essential and closely related to
the proposed research. Other types of IRP are covered in some
review papers (Andersson et al., 2010; Coelho et al., 2014).
IRP can be divided into two categories in terms of time horizon:
infinite time horizon IRP and finite time horizon IRP. The goal of
infinite time horizon IRP is to determine the optimal replenish-
ment intervals of retailers (or suppliers) and plan vehicle routes so
that the supply chain system’s average cost is minimized. With
regard to finite time horizon IRP, we need to decide when to visit
retailers (or suppliers) and the resulting products quantities de-
livered (or collected), as well as vehicle routes, to minimize the
supply chain system’s total cost.
For infinite time horizon IRP, some computationally efficient
policies have been introduced, such as fixed-partition policies
(FPP) and power-of-two (POT) policy. Under FPP, retailers are di-
vided into different regions. Each time if one of the retailers in a
specific region receives a delivery, then other retailers in this re-
gion also receive deliveries from the same vehicle. Anily and
Federgruen (1990, 1993) and Anily and Bramel (2004) have ana-
lyzed FPP in detail and demonstrated its effectiveness. Under POT,
the replenishment intervals of retailers are restricted to be power-
of-two multiples of a base planning period. Roundy (1985) has
proved that an optimal power-of-two policy has an effectiveness
of at least 98%. Thus, POT is widely used in IRP. Viswanathan and
Mathur (1997) proposed a stationary nested POT policy for a two-
echelon supply chain system. Under a stationary policy, the re-
plenishment intervals of retailers are constant. A nested policy
means that if the replenishment interval Ti of product i is larger
than the replenishment interval Tj of product j , then Ti is a multiple
of Tj . This policy was adopted by Custódio and Oliveira (2006) in a
real-world application concerning the distribution of frozen pro-
ducts in a Portuguese company. Zhao et al. (2008) employed both
FPP and POT policies to solve a three-echelon IRP.
For most companies, logistics planning may require some
changes after running for a period of time for various reasons, such
as changes in the production plan or market demand. In such si-
tuations, infinite time horizon models are not quite exact; there-
fore, in recent years, many studies have begun to focus on finite
time horizon (a single period or multiple periods) IRP. Huang and
Lin (2010) considered multi-item replenishment with uncertain
demands in a single period. The demands of items only became
known upon arrival at the demand locations. It was necessary to
determine shipping polices to optimize the trade-off between
stockout costs and transportation costs. A modified ant colony
optimization algorithm was developed for the problem. Moin et al.
(2011) considered an inbound logistics system with multi-periods.
At the beginning of the time horizon, they had to make decisions
about time of collection, size of shipment and vehicle routes to
minimize the system’s total cost. A genetic algorithm was devel-
oped to solve the mathematical formulation. Mjirda et al. (2014)
proposed a two-phase variable neighborhood search metaheur-
istic to solve the same problem as Moin et al. (2011). Computa-
tional results showed that their algorithm was more efficient than
Mion’s. Coelho and Laporte (2013) developed the first branch-and-
cut algorithm to solve a multi-product, multi-period and multi-
vehicle IRP. Vidović et al. (2014) addressed a multi-product, multi-
period IRP in fuel delivery. An identical fleet of vehicles with
compartments was used for fuel distribution from a depot to many
petrol stations with deterministic fuel consumption. Their pro-
blem was solved by the use of a mixed integer programming
model and a heuristic approach.
In the conventional supply chain system, many researchers
and companies only focus on profits, costs or service levels.
However, with increasing pressure on the environment, many
researchers have begun to consider environmental issues in their
models. Letmathe and Balakrishnan (2005) proposed two models
for companies to determine their optimal product mix and pro-
duction quantities in the presence of different types of environ-
mental constraints. Elhedhli and Merrick (2012) considered a
supply chain network design problem that takes CO2 emissions
into account. They found that the optimal configuration of the
supply chain would be changed if emission costs were included.
Martí et al (2015) studied the supply chain network design pro-
blem under different carbon polices. Benjaafar et al. (2013) pre-
sented a set of simple and widely used models to illustrate how
carbon emissions can be integrated into operations decisions. Pan
et al. (2013) explored the environmental impact of the pooling of
supply chains and found that supply network pooling was an
efficient way to reduce CO2 emissions. Hoen et al. (2010) explored
transport mode choices under the carbon cap policy. Bozorgi
(2016) formulated multi-product inventory model for cold pro-
ducts with emission consideration. A detailed review of the re-
search in the green logistics can be found in Dekker et al. (2012)
and Seuring (2013).
Although most studies in the field of green supply chains focus
on production processes or network design decisions, contribu-
tions that incorporate environmental aspects into IRP are scarce
(Treitl et al., 2014). Treitl et al. constructed an IRP model and ap-
plied it to the petrochemical industry. They noted that detailed
transportation processes analysis was important to reduce costs
and carbon emissions in a vendor managed inventory (VMI) sys-
tem. Mirzapour Al-E-Hashem and Rekik (2014) considered
greenhouse gas emissions in IRP models. They proposed a trans-
shipment option to improve the performance of the supply chain.
Konur (2014) addressed an integrated inventory management and
truck-load transportation with a carbon cap constraint. They for-
mulated an economic order quantity (EOQ) model and developed
a heuristic search method to solve the problem. Konur and
Schaefer (2014) explored the EOQ model with less-than-truckload
and truckload transportation under the four carbon emission
regulations.
 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 265

As multi-period logistics planning is widely adopted in practice, product type i only

we focus on MIRP with carbon emission regulations, considering O={0} Depot
the carbon dioxides generated by transportation sector in our P ={ N +1} Assembly plant
models. We define the emissions as Hoen et al. (2010), Pan et al. H = { 1, 2, …T } Time horizon
(2013) and Konur (2014) did. It is mainly influenced by load and
travel distance. There are three highlights in our work. First, to the Parameters
best of our knowledge, this paper is the first to employ MIRP to
examine the mutual effects between operational decisions and Dit Demand for product type i in period t (t ∊ H )
carbon regulations. Second, from numerical experiments, we ob- F Fixed vehicle cost per trip
tain several managerial insights that are distinctly different from C The weight capacity of each vehicle
our intuitions. Third, we propose a hybrid genetic algorithm that hi Inventory holding cost at assembly plant for product type i
can be used to solve large-sized problems efficiently, whereas in per unit product per period
Treitl et al.’s and Mirzapour and Rekik’s studies, CPLEX was used to cij Distance between node i and node j
solve models, and it was impossible for CPLEX to solve large-sized u Unit fuel price (e.g., $/l)
problems efficiently. ρ0 The empty-load fuel consumption rate (e.g., l/km)
ρ* The full-load fuel consumption rate (e.g., l/km)
ε Emissions generated by unit fuel consumption (e.g., kg CO2 /l).
3. Problem definition and mathematical formulation Gt Carbon cap in period t
f The price (tax) per unit carbon emission whether it is bought
In this section, we present the mathematical model of a tra- or sold
ditional MIRP and MIRP models with different carbon emission M A sufficiently large number
regulations. In Section 3.1, we describe assumptions and notations.
In Section 3.2, we introduce the method of computing fuel con- Decision variables
sumption and carbon emissions. In Section 3.3, the formulation of
a traditional MIRP is proposed. In Sections 3.4–3.8, we present xijt xijt =1 If arc( i, j ) is traversed by a vehicle in period t , otherwise 0
models under different emission regulations and make a summary Q ijt The total product weight carried by a vehicle through arc
of these models. ( i, j) in period t
ait The product quantity picked up at supplier i in period t
3.1. Problem description and notations Iit The inventory level of product type i at the assembly plant at
the end of period t
As the outbound logistic problem of a one-to-many distribution FCijt Fuel consumption from node i to node j in period t
network is equivalent to the inbound logistic problem of a many-
to-one network (Moin and Salhi, 2007), this paper only focuses on 3.2. Computation of fuel consumption
MIRP in an inbound product collection system consisting of an
assembly plant, a depot and N geographically dispersed suppliers. Vehicles’ fuel consumption is subjected to three major factors:
In each period, vehicles depart the depot to collect products from travel distance, truckload and speed (Elhedhli and Merrick, 2012).
suppliers to meet the assembly plant’s demand and return to the In this study, we assume travel speed to be constant and model
depot again at the end of the trip. Our assumptions are as follows: fuel consumption as Xiao et al. (2012) did:

Each supplier offers one type of distinctive product. At the be- ⎛ ρ* −ρ0 ⎞
FCijt =⎜ρ0 + Q ijt ⎟ cij
ginning of the time horizon, all products are assumed to be ⎝ C ⎠ (1)
ready for collection.

The demand at the assembly plant is product specific and de- For a specific type of vehicles, ρ0 , ρ*and C are constant, then we
terministic, but variable over periods. simplify Eq. (1) as follows:

Stockouts are not permitted at the assembly plant. FCijt =(ρ0 +βQ ijt ) cij (2)

Inventory holding costs are only considered at the assembly
plant. The initial inventory levels of all products are assumed to Wherein β = (ρ* −ρ0 ) /C . Therefore, the fuel consumption cost from
be zero. node i to node j is u (ρ0 +βQ ijt ) cij . According to Zhang et al. (2014)

Vehicles are homogeneous and capacitated. The number of ve- and Cachon (2014), fuel consumption is the direct cause of carbon
hicles is unlimited. This assumption provides flexibility because emissions, then we compute vehicles’ emissions from node i to
it allows the actual number of vehicles to be determined later. node j as ε (ρ0 +βQ ijt ) cij .

Split pick up is not allowed, that is, a supplier is visited by at
most one vehicle in each period. 3.3. Formulation of a traditional MIRP

The carbon emission regulations are imposed on each period
during the time horizon. In the traditional MIRP, companies only pursue economic
profits while ignoring environmental issues. The following is a
The objective of the proposed MIRP is to determine when to mixed integer programming formulation of the traditional MIRP,
visit the suppliers, the corresponding product quantities collected, and we refer to this as model (P).
and the vehicle schedules such that the total cost of the supply Model (P):
chain is minimized.
The proposed MIRP uses the following notations: Min Zp = ∑ ∑ hi Iit + ∑ ∑ Fx 0jt
i ∈ V′ t ∈ H t∈H j∈V′

Sets ⎛ ⎞
+ u ⎜⎜ ∑ ∑ ∑ ( ρ0 + βQ ijt ) cij x ijt + ∑ ∑ ρ0 cN + 1,0 x 0jt ⎟⎟
V ={0, 1, … ,N +1} A set of all nodes ⎝ t∈H i ∈ V ′∪ O j ∈ V ′∪ P t∈H j∈V′ ⎠ (3)

V ′ = {1, 2, …N} A set of suppliers where supplier i (i ∈ V ′) offers Subject to:

266 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275

Iit = Ii ( t − 1) + ait − Dit ∀ i ∈ V′, ∀ t ∈ H (4) wijt ≥ 0 (19)

∑ Q jit + ait = ∑ Q ijt ∀ i ∈ V′, i ≠ j, ∀ t ∈ H wijt ≥ Cxijt + Q ijt − C (20)

j ∈ V′∪ O j ∈ V′∪ P (5)
wijt ≤ Q ijt (21)
∑ xjit = ∑ xijt ∀ i ∈ V′, ∀ t ∈ H
j ∈ V′∪ O j ∈ V′∪ P (6) wijt ≤ Cxijt (22)

∑ xjit ≤ 1 ∀ i ∈ V′, ∀ t ∈ H
j ∈ V′∪ O (7) 3.4. Formulation with a carbon cap policy

We now discuss the first extension to model (P), in which a

∑ x 0it = ∑ xi (N + 1) t ∀t∈H
i ∈ V′ i ∈ V′ (8) carbon cap policy is considered in each period. That is, the carbon
emission level cannot exceed the upper limit (cap) in each period.
Q ijt ≤ Cxijt ∀ i ∈ V′, ∀ j ∈ V′ ∪ P , ∀ t ∈ H (9) To represent such a policy, we add an extra constraint to the tra-
ditional MIRP model. We refer to this as model (P_cc).
x 0 ( N + 1) t = 0 ∀t∈H Model (P_cc):
(10)
Min Zcc = Z p′ (23)
xi0t = 0 ∀ i ∈ V′, ∀ t ∈ H (11)
Subject to (4)–(17), (19)–(22) and

x ( N + 1) jt = 0 ∀ j ∈ V′, ∀ t ∈ H ⎛ ⎞
(12) ε⎜ ∑ ∑ ( ρ0 cij x ijt + βcij wijt ) + ∑ ρ0 cN + 1,0 x 0jt ⎟ ≤ Gt ∀t∈H
⎜ ⎟
⎝ i ∈ V ′∪ O j ∈ V ′∪ P j ∈V ′ ⎠ (24)
xiit = 0 ∀ i ∈ V, ∀ t ∈ H (13)
The objective of model (P_cc) is the same as model (P). Con-
straint (24) is added to limit the total emissions in each period.
Q 0it = 0 ∀ i ∈ V′, ∀ t ∈ H (14)

3.5. Formulation with a carbon cap and offset policy

Ii0 = 0 ∀ i ∈ V′
xijt = {0, 1} ∀ i ∈ V, ∀ j ∈ V, ∀ t ∈ H
ait , Iit , Q ijt ≥ 0 ∀ i ∈ V, ∀ j ∈ V, ∀ t ∈ H
The objective function of model (P) aims to minimize the total cost
of the supply chain, including inventory holding costs, fixed trans-
portation costs and fuel consumption costs. In order to express the fuel
consumption cost concisely and apply linearization methods, we in-
clude the arc from the depot to the assembly plant in the third term of
the objective, and limit it to be zero in constraint (10). Constraint (4) is
the inventory balance equation for each product at the assembly plant.
Constraint (5) is the product flow balance equation, which guarantees
the flow balance at each supplier and eliminates all subtours. Con-
straint (6) guarantees that the number of vehicles leaving a supplier is
equal to the number of arrival vehicles. Constraint (7) ensures that split
(15)
A carbon cap is imposed on the supply chain system in each
(16) period under a carbon cap and offset policy. However, the decision
makers can choose to purchase extra carbon emission credits in
(17) the carbon offset market under this policy. This mechanism is
often employed by individuals or originations who voluntarily
decrease their carbon emissions by funding a carbon offset project
or activity (Palak et al. 2014; Dye and Yang, 2015). The major
motivations for buying offsets are corporate responsibilities and
public relations (US Government Accountability Office, 2008).
Let et+ denotes the quantity of carbon emission credits pur-
chased in period t , which is a decision variable. The resulting
formulation, denote by (P_cco), is as follows:
Model (P_cco):
Min Zcco = ∑ ∑ hi Iit + ∑ ∑ Fx 0jt
i ∈ V′ t ∈ H t∈H j∈V′

pick up is not allowed. Constraint (8) guarantees that the number of ⎛

vehicles leaving the depot is equal to the number of vehicles arriving at + u ⎜⎜ ∑ ∑ ∑ ( ρ0 cij xijt + βcij wijt )
the assembly plant. Constraint (9) ensures that the vehicle capacity is ⎝ t∈H i ∈ V ′∪ O j ∈ V ′∪ P

respected. Constraints (10)–(13) determine the impossible arcs. Con- ⎞

straint (14) guarantees that the product quantity carried by any vehicle + ∑ ∑ ρ0 cN + 1,0 x 0jt ⎟⎟ + ∑ fet+
from the depot to suppliers is equal to zero. Constraint (15) ensures t∈H j∈V′ ⎠ t∈H (25)
that the initial inventory levels of all products are zero. Constraints Subject to (4)–(17), (19)–(22) and
(16)–(17) define the variable types.
⎛ ⎞
Note that the third term of the objective is nonlinear. We use ⎜
ε⎜ ∑ ∑ ρ 0 cij xijt + βcij wijt + ∑ ρ 0 cN + 1,0 x0jt ⎟⎟ ≤ Gt + et+
( ) ∀t∈H
the method proposed by McCormick (1976) to linearize it. A new ⎝ i ∈ V ′∪ O j ∈ V ′∪ P j ∈V ′ ⎠ (26)
variable wijt is introduced to replace Q ijt xijt . A reformulated ob-
jective and four new constraints are introduced as follows: et+ ≥ 0 ∀t∈H (27)
Min Z′p = ∑ ∑ hi Iit + ∑ ∑ Fx 0jt The objective of model (P_cco) is to minimize the total cost,
i ∈ V′ t ∈ H t∈H j∈V′
including the cost of purchasing extra carbon emission credits.
⎛ Constraint (26) assures that the total emissions in each period do
+ u ⎜⎜ ∑ ∑ ∑ ( ρ0 cij xijt + βcij wijt ) not exceed the sum of the cap and the purchased credits.
⎝ t∈H i ∈ V ′∪ O j ∈ V ′∪ P

⎞ 3.6. Formulation with a carbon cap and trade policy

+ ∑ ∑ ρ0 cN + 1,0 x 0jt ⎟⎟
t∈H j∈V′ ⎠ (18) Unlike the carbon cap and offset policy, companies can not only
 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 267

purchase carbon emission credits if they need, they can also sell * ≤Zcc
the carbon cap model, then we can get Zcco * , where Z * re-
unused credits for profits under a carbon cap and trade policy. presents the optimal objective value.
Currently, many Emissions Trading Systems have been well de-
Proposition 2. The carbon cap and offset model is a special case of
veloped around the world for companies to manage their emis-
* ≤Zcco
the carbon cap and trade model, and Zcct * .
sions, such as USA, EU, Japan and China.
Let et+ (et−) denote the quantity of carbon credits purchased Proof. For the carbon cap and trade model, when the sold carbon
(sold) in period t . The resulting formulation is called model (P_cct). emission credits ( et−) equal to zero, the model (P_cct) is the same
Model (Pcct): as the model (P_cco). This means that the carbon cap and trade
policy has more flexibility than the carbon cap and offset policy,
Min Zcct = ∑ ∑ hi Iit + ∑ ∑ Fx 0jt
* ≤Zcco
* .
i ∈ V′ t ∈ H t∈H j∈V′
then we can get Zcct
⎛ Proposition 3. The carbon cap model is a special case of the
+ u ⎜⎜ ∑ ∑ ∑ ( ρ0 cij xijt + βcij wijt ) * ≤Zcco
carbon cap and trade model, and Zcct * ≤Zcc
*.
⎝ t∈H i ∈ V ′∪ O j ∈ V ′∪ P

⎞ Proof. For the carbon cap and trade model, when both the sold
+ ∑ ∑ ρ0 cN + 1,0 x 0jt ⎟⎟ + ∑ f ( et+ − et−) carbon emission credits ( et−) and the purchased credits (et+) equal
t∈H j∈V′ ⎠ t∈H (28) to zero, the model (P_cct) degrades to the carbon cap model (P_cc).
From Proposition 1 and 2, we can easily get Zcct* ≤Zcco
* ≤Zcc
*.
Subject to (4)–(17), (19)–(22) and
Proposition 4. Zp≤Zct .
⎛ ⎞
ε ⎜⎜ ∑ ∑ ( ρ0 cij xijt + βcij wijt ) + ∑ ρ0 cN + 1,0 x 0jt ⎟⎟ Proof. The feasible regions of the traditional MIRP model and the
⎝ i ∈ V ′∪ O j ∈ V ′∪ P j∈V′ ⎠ carbon taxing model are the same, while the objective of the
+ et− ≤ Gt + et+ ∀t∈H (29) carbon taxing model includes an extra term—the cost of carbon
emissions. Therefore, if f ≥ 0, then Zp≤Zct .
et+ , et− ≥ 0 ∀t∈H (30)

The objective of model (P_cct) is to minimize the total cost, 4. Solution procedure
including the cost of purchasing extra carbon emission credits and
the revenues of selling unused credits. Constraint (29) assures that The MIRP is NP-hard because it subsumes a vehicle routing
the sum of total emissions caused by transportation and the sold problem which is NP-hard. Therefore, to obtain high quality so-
credits is less than or equal to the sum of the cap and the pur- lutions in reasonable computation time, it is necessary to focus on
chased credits. heuristic or metaheuristic approaches. The genetic algorithm (GA),
introduced by John Holland in the 1960 s, has been extensively
3.7. Formulation with a carbon taxing policy used to solve various vehicle routing problems (Baker and Aye-
chew, 2003; Prins, 2004; Berger and Barkaoui, 2004; Ho et al.,
Under a carbon taxing policy, a company is charged a fee for 2008). In recent years, genetic algorithms have been developed to
every unit carbon dioxide it emits. In this paper, we let the tax for generate good solutions for IRP. Abdelmaguid and Dessouky
unit emission equal to the unit carbon price as Benjaafar et al. (2006) solved a MIRP with a GA scheme. Their numerical experi-
(2013) did. The resulting model is called model (P_ct). ments showed that GA was able to generate near-optimal solu-
Model (P_ct): tions for small problems and find good solutions for larger ones.
Cho et al. (2014) proposed an adaptive GA for a time dependent
Min Zct = ∑ ∑ hi Iit + ∑ ∑ Fx 0jt IRP and proved its good potential for solving inventory routing
i ∈ V′ t ∈ H t∈H j∈V′
problems (IRPs). Borade and Sweeney (2014) also adopted a GA-
⎛ based approach to develop a decision support system in a VMI
+ (u + fε) ⎜⎜ ∑ ∑ ∑ ( ρ0 cij xijt + βcij wijt ) supply chain, and their numerical experiments showed that the
⎝ t∈H i ∈ V ′∪ O j ∈ V ′∪ P
GA-based approach could help to serve more customers and to
⎞
improve vehicle utilization and service levels.
+ ∑ ∑ ρ0 cN + 1,0 x 0jt ⎟⎟
Regarding the proposed MIRP, we need to decide whether to
t∈H j∈V′ ⎠ (31)
visit a supplier or not in each period during the time horizon. This
Subject to (4)–(17) and (19)–(22). specific characteristic of MIRP’ solutions can be easily represented
The objective function of model (P_ct) now involves emission by a row of a binary matrix which is exactly a chromosome. Be-
taxes which are proportional to the emissions. sides, using binary matrices to represent MIRP’s solutions makes it
easy to generate new solutions by performing bitwise AND and OR
3.8. Summary of model formulations operations on the existing solutions. Therefore, in this paper, a
hybrid genetic algorithm (HGA) is developed to solve the MIRP,
We proposed four extensions to the traditional MIRP model, and a post-optimization method is further conducted to improve
which capture the effects of carbon emission regulation policies on the best solution found in HGA. The flowchart of the solution
operational decisions. Comparing the five models above, we could procedure is shown in Fig. 1.
get the following propositions. The genetic algorithm is a stochastic search technique based on
ideas from the natural selection process (Goldberg, 1989). It starts
Proposition 1. The carbon cap model is a special case of the with an initial set of randomized solutions called a population.
carbon cap and offset model, and Zcco * ≤Zcc
*. Each individual in the population is called a chromosome and
Proof. If the unit carbon price f is a sufficiently large number, represents a solution to the problem. The chromosomes evolve
then the decision makers will be inclined to plan their operations through successive generations. During each generation, chromo-
without purchasing extra emission credits. In this situation, somes are evaluated by some measures of fitness. Neighborhood
et+=0 ( ∀ t ∈ H ) and the carbon cap and offset model degrades to search operators (crossover and mutation) are applied to create
268 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275

Start Table 1
An example of a binary chromosome representation.

Supplier Period
Generate initial population
1 2 3 4 5

Compute product collection

Demand matrix
matrixes (Algorithm 1)
1 2 3 2 1 1
2 4 4 2 2 1
Compute individuals total cost 3 3 1 1 2 3
and fitness value 4 2 3 3 1 1
5 4 4 1 2 4

Chromosome representation
Get the best Yes 1 1 0 0 1 0
chromosome Termination condition
2 1 1 1 0 0
3 1 1 0 0 1
NO 4 1 1 0 1 0
5 1 0 0 1 0
Post-Optimization
Chromosomes selection Product collection matrix
1 7 0 0 2 0
2 4 4 5 0 0
Stop
Crossover operator 3 3 4 0 0 3
4 2 6 0 2 0
5 9 0 0 6 0

Mutation operator

The next visit to supplier 5 occurs in period 4; therefore, the

Adjust the binary presentation corresponding collection quantity of product type 5 in period 1 is
matrixes
9.
Renew product collection
matrixes 4.2. Initial population

The initial population comprises K choromosomes. Each

Generate new population
Fig. 1. Flowchart of the solution procedure.
the next generation. After several generations, it is anticipated that
a near-optimal solution will be found.
The following definitions are required in our HGA:
K Population size
G′ The number of generations
Pc Crossover rate
Pm Mutation rate
l Chromosome
4.1. Chromosome representation
A binary matrix of size N × T is used to define a chromosome,
which is also adopted by Moin et al. (2011). N is the number of
suppliers, and T is the number of periods. Each gene of the
chromosome indicates whether the supplier is visited or not in
current period. That is,
⎧ 1 If chromosome l has supplier i visited in period t
Chitl = ⎨
⎩ 0 Otherwise
Thus, the quantity of product type i collected in period t de-
pends on whether there will be a visit to supplier i again in sub-
sequent periods. If we suppose that supplier i is visited again in
period t′(t′>t ), then the quantity of product type i collected in
period t is the sum of the demands in period t , t + 1, …t′−1. Ta-
ble 1 gives an example of chromosome representation for a supply
chain system consisting of 5 suppliers and 5 periods. We find that
all the suppliers are visited in period 1 because the initial in-
ventories of all products are 0, and stockouts are not permitted.
For example, Ch51=1 means that supplier 5 is visited in period 1.
chromosome has N ×T genes. The operator generating the initial
population is illustrated in Table 2.
Take chromosome l for example. If the demand for product i in
period 1 is nonzero, to prevent stockout, we set Chil1=1; otherwise,
Chil1=0 (lines 3–7). For any supplier and any period t (t ≥2), if sup-
plier i has been visited in period t′(t′ < t ), then randomly generate
0 or 1 for Chitl (lines 10–11); if supplier i has not been visited in any
former period, then Chitl =1 (lines 12–16).
After the initial population has been produced, the corre-
sponding product collection matrix for each chromosome can be
computed. However, as we generate some of the genes randomly,
it may produce infeasible solutions, and the collected production
quantity at a supplier may exceed a vehicle’s capacity. Therefore, a
modified algorithm for computing replenishment quantity is de-
signed, which is named Algorithm 1 and shown in Table 3.
If supplier i is not visited in period t , then the collected quantity
of product i is 0 (lines 4–5). If supplier i is visited in period t (t ≠ T )
and will be visited again in later period, then the method in Sec-
tion 4.1 is used to compute the collected quantity ait (lines 6–11). If
supplier i is visited in period t (t ≠ T ) and will not be visited again
T
after period t , then ait = ∑t Dit (lines 12–16). If supplier i is visited
in period T , then the collected quantity equals the demand for
product type i in period T (lines 19–21).
.
If the collected quantities of product i in period t exceed the
vehicle’s capacity, then define a variable q = ait and generate a
number q′ randomly between Dit and C . Set ait =q′. If Chitl + 1=0, then
set Chitl + 1=1. Correspondingly, ait + 1=ait + 1+(q − ait ) (lines 23–31).
This process means that a part of product i is transferred to be
collected in the next period in order to respect vehicles’ capacity.
4.3. Chromosome selection
The fitness value is the criterion for evaluating individuals in
the population. It guides the search process towards more pro-
mising directions in the search space. In this paper, we use roulette
 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 269

Table 2 Table 4
Pseudocode of generating initial population. Crossover Operator.

The operator for generating the initial population Supplier Period

1. for ( l=1; l≤K ; l+ + ) 1 2 3 4 5

2. for i=1; i≤N; i+ + )
3. if ( Di1 >0) Parent 1
1 1 0 0 1 0
4. Chil1=1
2 1 1 1 0 0
5. else 3 1 1 0 0 1
6. Chil1=0 4 1 1 0 1 0
7. end if 5 1 0 0 1 0
8. for ( t =2;t ≤ T;t + + ) Parent 2
9. for ( t′=t −1; t′≥1; t′− − ) 1 1 1 0 0 0
10. if ( Chitl′≠0) 2 1 0 1 0 0
3 1 0 0 0 1
11. Chitl =rand ( ) %2 4 1 1 0 1 0
12. else 5 1 1 1 1 0
13. if ( t′=1)
Child 1
14. Chitl =1
1 1 0 0 0 0
15. end if 2 1 0 1 0 0
16. end if 3 1 0 0 0 1
17. end for 4 1 1 0 1 0
18. end for 5 1 0 0 1 0
19. end for
20. end for Child 2
1 1 1 0 1 0
2 1 1 1 0 0
Table 3 3 1 1 0 0 1
Pseudocode of Algorithm 1. 4 1 1 0 1 0
5 1 1 1 1 0
Algorithm 1

1. for ( l=1; l≤K ; l+ + ) individual in each generation is directly kept in the next genera-
2. for ( i=1; i≤N; i+ + ) tion. With regard to routing cost, the cheapest insertion heuristic
3. for ( t =1;t ≤ T;t + + ) (Rosenkrantz et al., 1977) is used to construct the vehicle routes.
4. if ( Chitl =0)
5. ait =0 4.4. Crossover operator
6. else
7. if (t ≠ T )
The crossover operator is used to mate pairs of chromosomes to
8. for ( t′=t +1; t′≤T ; t′+ + )
produce offspring. First, we select a pair of chromosomes from
9. if ( Chitl ′≠0)
current generation. Second, we perform a bitwise AND on the
10. use the method given in 4.1 to compute collected quantity
11. end if
selected parents to obtain child 1 and a bitwise OR to obtain child
12. if ( t′=T &&Chitl ′=0) 2. Table 4 is an example of this operator.
13. for ( t′ ′=t; t′ ′≤T ; t′ ′+ + )
14. ait =ait +Dit ′′ 4.5. Mutation operator
15. end for
16. end if To increase the diversity of the population, a mutation operator
17. end for is performed in HGA. First, a chromosome is randomly selected
18. end if
19. if ( t = T )
from current population. Second, a gene of this chromosome is
20. ait =Dit randomly selected. If the gene is “1”, then we transform it to 0;
21. end if otherwise, if it is “0”, then we transform it to “1”.
22. end if Note that the binary chromosome representation matrixes
23. if ( ait >C ) should be checked to avoid stockout after the mutation operator. If
24. Define q = ait
the demand for product type i in period 1 is not equal to 0, we
25. Randomly generate q′ ∈ (Dit ,C )
26. Set ait =q′
should make sure that Chi1=1. If the demand for product type i in
27. if (Chitl + 1=0)
period t (t > 1) is not equal to 0, and supplier i has not been visited
in a previous period, then we set Chit =1.
28. set (Chitl + 1=1)
After finishing the operators above, renew the product collec-
29. end if
30. ait +1=ait +1+(q−ait )
tion matrixes using Algorithm 1.
31. end if The last step in HGA is to check whether the current optimal
32. end for solution is good enough to meet our expectations. When the
33. end for stopping criterion is satisfied, the algorithm ends, and the current
34. end for
near-optimal solution is the final result for the problem. In this
study, HGA stops after G’ generations. Note that the value of G’
depends on the problem size.
wheel selection, in which the fitness value is required to be non-
negative, and a larger fitness value indicates a better individual. 4.6. Post-optimization
Thus, we employ a sufficiently large number M to minus the ob-
jective function as the fitness function. An elitism strategy is It is found that the last vehicle in each period usually only
adopted to improve the convergence speed; that is, the best utilizes a small proportion of the vehicle’s capacity, that is, less
270 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275
than 50%. This situation increases the number of vehicles needed
and causes unnecessary costs. A post-optimization operator is
proposed to improve the solution. This operator is performed on
the best chromosome found in HGA. Suppose that in period t (t >1),
the utilization rate of the last vehicle is rt . If rt +rt ′≤1 ( rt ′ is the
utilization rate of the last vehicle in period t′( t′ < t )), then we
consider deleting the last route in period t and transferring the
suppliers in this route to be visited by the last vehicle in period t′.
We construct the last vehicle’s route again in period t′ and obtain a
new route cost. Similarly, inventory cost and carbon emission cost
are also renewed. If this post-optimization operator reduces the
total cost, then accept it; otherwise, return to the situation before
this operator. Note that if the cost after this operator is equal to the
cost before this operator, we also accept it, as the change will
decrease the number of vehicles used in period t .
5. Numerical tests and analyses
The aim of this section is threefold:

To evaluate the effectiveness of the proposed HGA.

To indicate the effects of carbon emission regulations on MIRP.

To perform sensitive analyses and give some managerial in-
sights into the proposed problem.
5.1. Instance data
As to MIRP, we use data sets based on Lee et al. (2003), which
are also available at http://www.mie.utoronto.ca/labs/ilr/IRP. The
coordinates, demands and inventory holding cost coefficients can
be found at the website. Note that the depot is located at (0, 0) for
all the data sets. We define SxHy as an instance with x suppliers
and y periods. Based on the S12H14 instance given at the website,
we generate different test problems. For example, consider the
S 3H 3 instance, which means that we use the information of the
first 3 suppliers and the first 3 periods in the S12H14 instance as
our data set. For small-sized problems, we generate 18 instances:
S 3Hy , S 4Hy and S5Hy ( y = 3,5,7,9,11,14). For medium-sized pro-
blems, the S7Hy and S 9Hy and S12Hy ( y = 3,5,7,9,11,14) instances
are generated. For large-sized problems, we generate several in-
stances based on the S 20H 21 and S50H 21 instances. S 20Hy and
S50Hy mean that the information of all the suppliers and the first
y periods is used from the S 20H 21 and S50H 21 instances, re-
spectively. Other parameters are as follows: F =20, C = 10 .
We assume that the vehicles used in our study are powered by
diesel and use the data based on Ubeda et al. (2011) and Cachon
(2014) to compute fuel consumption cost and carbon emissions.
The values of parameters are as follows: ρ0 =0.296, ρ*=0.390,
ε = 2.61, u = 1.05, f = 1.
For small- and medium-sized problems, G1=200, Gt =100 ( t =
2,3…T ) . For large-sized problems, G1=500, Gt =400 ( t = 2,3…T );
Parameters in HGA are as follows: K = 200,Pc =0.9,Pm=0.02, G′=300
(for small-sized problems), G′=500 (for medium- and large-sized
problems).
5.2. Evaluation of HGA
As we indicate in Section 3.8 that the proposed five models are
closely related, in this section we solve one of them, the (P_cco)
model, to evaluate the effectiveness of our algorithm. The pro-
posed HGA is implemented in the C þ þ language and run on a PC
with an Intel Core i5 Processor (2.3 GHZ) and 4 GB memory. The
mathematic formulation is solved by CPLEX 12.6 to obtain the best
integer solution (the upper bound) and the best bound (the lower
bound). A comparison between HGA and CPLEX will be evaluated
on computation effectiveness and efficiency. For each instance, the
heuristic is run 10 times, and the best objective value is reported.
Table 5 summarizes the results got from CPLEX and HGA. The
first column is data sets. The second and third columns are the
best bounds and the best integer solutions found by CPLEX after a
limited amount of time, respectively. The fourth column (Gap1) is
the gap between the best integer and the best bound. The fifth and
eighth columns are the CPU time of CPLEX and HGA, respectively.
The sixth column is the best objective found in HGA. The solutions
in bold are as good as the optimal solutions obtained from CPLEX.
The seventh column (Gap2) is the gap between the best objective
and the best integer.
It is observed that for 17 out of 18 small-sized problems, HGA
can find optimal solutions with less time. It is difficult for CPLEX to
produce optimal solutions for most medium- and large-sized
problems within a 3600 s time limit. For these problems, HGA can
give satisfying solutions in a relatively shorter time.
For all small-sized problems, the average gap between HGA and
CPLEX is 0.04%; however, the CPU time of HGA is dramatically
short. This advantage becomes more obvious when instance size
increases. For medium- and large-sized problems, the average
gaps are 4.25% and 5.32% respectively. Therefore, HGA is able to
produce good solutions for proposed models in a relatively short
time.
5.3. Observations from carbon regulations
To further study the effects of different carbon emission reg-
ulations on MIRP, a series of sensitivity analyses based on a
medium-sized instance, the S12H14 , are performed. This is be-
cause the medium-sized instance is more representative than a
small-sized one, and it is also time-efficient compared with large-
sized ones. For each circumstance, HGA is run 10 times and the
average values are reported. The carbon emissions shown in the
following figures are the total emissions during the whole time
horizon. Based on the sensitivity analyses, some interesting in-
sights are obtained. We present these insights in the form of a
series of observations.
5.3.1. Comparison with the traditional MIRP model
In this section, we compare the traditional model with the
(P_cc) model and observe the effects of carbon limits on cost and
emissions. For the (P_cc) model, in each scenario the cap in the
first period is fixed to be 500 (large enough to ensure a feasible
solution), and the caps in other periods are changed step by step.
For instance, for scenario 1: G1=500, Gt =500 ( t = 2,3…T ); for sce-
nario 2: G1=500, Gt =400 ( t = 2,3…T ); and for scenario 3: G1=500,
Gt =300 ( t = 2,3…T ). The vehicle capacity is also enlarged to be 50
in these two models, to make sure that a feasible solution can be
found even though the cap is much tight. The comparison is re-
ported in Table 6. The penultimate column is the percent gap
between the two models’ total costs. The last column is the
emission gap, and the negative values indicate the percentage of
decreased emissions.
From Table 6, we can see that when the cap is much loose
(between 300 and 500), it affects the system’s performance very
slightly. As the cap becomes tight (decreases from 200 to 120), the
(P_cc) model tends to increase inventory and decrease fuel con-
sumption, thereby curbing carbon emissions. It is noted that the
system’s total cost increases dramatically, far surpassing the de-
crease of emissions, when the cap becomes tight. This is due to the
significant increase of inventory cost. From instance data, we find
that the unit inventory holding cost for each product is much
higher than unit fuel cost. Therefore, the system’s cost increases
 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 271

Table 5
Comparison of results between HGA and CPLEX.

Data Set CPLEX HGA

Best bound Best integer Gap1 (%) CPU (s) Best objective Gap2 (%) (%) CPU (s)

S 3H 3 194.13 194.13 0.00 0.40 194.13a 0.00 0.30

S 3H 5 332.78 332.78 0.00 0.76 332.78 0.00 0.56
S 3H 7 471.42 471.42 0.00 0.72 471.42 0.00 0.67
S 3H 9 610.06 610.06 0.00 1.00 610.06 0.00 0.89
S 3H11 748.71 748.71 0.00 1.04 748.70 0.00 0.95
S 3H14 955.03 955.03 0.00 1.20 955.03 0.00 1.08
S 4H 3 201.96 201.96 0.00 0.44 201.96 0.00 0.38
S 4H 5 350.49 350.49 0.00 1.03 350.49 0.00 0.80
S 4H 7 499.02 499.02 0.00 1.13 499.02 0.00 1.02
S 4H 9 647.55 647.55 0.00 1.14 647.55 0.00 1.10
S 4H11 796.08 796.08 0.00 1.50 796.08 0.00 1.23
S 4H14 1018.87 1018.87 0.00 1.86 1018.87 0.00 1.56
S5H 3 239.21 239.21 0.00 0.79 239.21 0.00 0.77
S5H 5 402.36 402.36 0.00 1.43 402.36 0.00 0.96
S5H 7 579.20 579.20 0.00 3.40 579.20 0.00 1.73
S5H 9 772.74 772.74 0.00 19.38 772.74 0.00 1.69
S5H11 960.71 960.71 0.00 80.44 960.71 0.00 2.08
S5H14 1234.37 1234.37 0.00 773.49 1242.95 0.70 2.52
Average 49.51 0.04 1.13
S7H 3 456.42 456.42 0.00 2.26 456.42 0.00 1.17
S7H 5 832.21 832.21 0.00 4.68 841.76 1.15 1.68
S7H 7 1265.53 1270.43 0.39 3600.00b 1281.92 0.90 2.29
S7H 9 1642.48 1654.75 0.74 3600.00 1730.65 4.59 2.62
S7H11 2015.78 2096.81 3.86 3600.00 2179.38 3.94 3.03
S7H14 2592.52 2696.96 3.87 3600.00 2864.45 6.21 3.83
S 9H 3 633.50 633.50 0.00 4.18 661.10 4.36 2.43
S 9H 5 1175.70 1175.70 0.00 900.59 1212.64 3.14 2.27
S 9H 7 1692.38 1745.92 3.07 3600.00 1801.48 3.18 3.34
S 9H 9 2184.86 2284.02 4.34 3600.00 2402.56 5.19 6.49
S 9H11 2697.90 2833.46 4.78 3600.00 3041.97 7.36 7.68
S 9T 14 3471.06 3682.88 5.75 3600.00 3984.77 8.20 9.49
S12H 3 958.60 958.60 0.00 1416.02 989.60 3.23 4.20
S12H 5 1679.69 1733.64 0.00 3600.00 1811.64 4.50 5.25
S12H 7 2408.09 2500.49 3.80 3600.00 2585.44 3.40 7.71
S12T 9 3148.76 3275.60 3.87 3600.00 3435.25 4.87 10.26
S12H11 3898.97 4043.30 3.62 3600.00 4264.12 5.46 12.64
S12H14 5002.14 5212.27 4.03 3600.00 5563.24 6.73 16.18
Average 2729.32 4.25 5.70
S 20H10 2757.87 3133.62 11.99 3600.00 3316.15 5.82 21.63
S 20H15 4174.19 5121.19 18.49 3600.00 5425.46 5.94 30.07
S 20H 21 5872.78 7280.94 19.34 3600.00 7827.06 7.50 43.60
S50H10 7732.16 8824.77 12.38 3600.00 8854.03 0.33 96.65
S50H15 12,055.94 15,042.42 19.85 3600.00 16,028.29 6.55 166.40
S50H 21 17,484.65 22,071.99 20.78 3600.00 23,344.89 5.77 245.50
Average 3600.00 5.32 100.64

a
The solutions given by HGA are as good as that obtained from CPLEX.
b
CPLEX is run for a 3600 s time limit.

Table 6
Comparison between the carbon cap model and the traditional MIRP model.

Scenario P_cc model P model Gap

Cap Inventory cost Fixed transport cost Fuel cost Total cost Emissions Inventory cost:322.50 Cost Gap (%) Emission Gap (%)

1 500 313.50 280.00 1149.78 1743.28 2858.01 Fixed transport cost: 280.00  0.44 0.11
2 400 311.10 280.00 1151.20 1742.30 2861.54  0.50 0.24
3 300 310.80 280.00 1148.74 1739.54 2855.43  0.65 0.02
4 200 1247.70 280.00 1054.99 2582.69 2622.41 47.50  8.14
5 190 1912.20 280.00 992.00 3184.20 2465.82 Fuel cost: 1148.47 81.85  13.62
6 180 2548.80 280.00 946.17 3774.97 2351.90 115.59  17.62
7 170 3452.70 286.00 907.10 4645.80 2254.78 Total cost: 1750.97 165.33  21.02
8 160 4386.00 288.00 863.53 5537.53 2146.50 216.26  24.81
9 150 5089.80 294.00 806.90 6190.70 2005.71 253.56  29.74
10 140 5958.60 298.00 769.78 7026.38 1913.44 Emissions: 2854.77 301.28  32.97
11 130 7363.80 300.00 712.66 8376.46 1771.47 378.39  37.95
12 120 8269.80 306.00 677.15 9252.95 1683.21 428.45  41.04
272 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275
Fig. 2. The results when unit inventory holding costs are changed.
dramatically although fuel consumption decreases. To verify this
assumption, we reduce all products’ unit inventory holding cost by
5 and 10 times respectively, and observe corresponding results,
which are shown in Fig. 2. In the figure, “Cost 1” represents the
system’s total cost when inventory holding cost coefficients are
not changed and “Emissions 1” is the corresponding emissions.
“Cost 2” and “Cost 3” are the system’s total cost when inventory
holding cost coefficients are reduced by 5 and 10 times respec-
tively, and “Emissions 2” and “Emissions 3” are resulting emission
levels.
As Fig. 2 shows, the system’s total cost does not increase that
dramatically when the inventory holding cost coefficient is re-
duced. It is also noted that the decreases of emissions are almost
the same in these three sceneries when the cap reduces from 190
to 120. However, their costs have significant differences. This
phenomenon indicates that different industries (or commodities)
have different sensitives to carbon caps. For those whose unit in-
ventory holding cost is very high, it is difficult for companies to
control emissions at lower price, and tight caps will put heavy
burdens on companies’ cost. Whereas carbon caps are meaningful
in reducing emissions for enterprises whose unit inventory hold-
ing cost is low. Therefore, governments should consider the dif-
ference of industries when establishing carbon caps. Here we get
the first observation:
Observation 1. The reaction to carbon caps varies significantly
due to different industries. Rigid emission limits will put heavy
burdens on companies whose products are high-valued, perish-
able or quickly updated.
5.3.2. Parameter sensitivity analyses
To indicate the impacts of different parameter values on cost
and emissions, we have performed many numerical experiments.
To keep this paper reasonably concise, we only present the results
which provide interesting managerial insights.
5.3.2.1. Carbon cap. In model (P_cco) and (P_cct), we change the
carbon cap step by step (from 150 to 200 with step 10, from 200 to
350 with step 25) and observe the changes in costs and emissions.
Other parameters’ values, except for the carbon cap, are the same
as in Section 5.1. In Fig. 3, if we say “carbon cap ¼100”, we mean
that the cap imposed on each period is equal to 100.
As expected, the tighter the cap, the higher the cost. From
Fig. 3, it is also observed that when the cap is equal to or more
than 225 the emissions are less under the carbon cap and trade
policy, this is because the system can make a profit by selling
Fig. 3. The result comparison between the carbon cap and offset policy and the
carbon cap and trade policy.
unused carbon emission credits, then they have stronger moti-
vation to decrease emissions. As the cap becomes tighter (less
than 225), the system under the carbon cap and trade policy has
no extra carbon emission credits to sell and they can only
maintain the operations by buying more credits. In this situa-
tion, the optimal results under the two policies are the same. On
the other hand, it is surprising to find that when the cap de-
creases from 190 to 180, the carbon emission level increases
rather than decreases under these two polies. This unexpected
phenomenon was also observed by Benjaafar et al. (2013) when
investigating a lot-sizing problem, which indicates that some-
times a tighter carbon cap does not mean better environmental
effects. This observation is also illustrated by the case that when
the cap changes from 160 to 150. Thus, we come to Observation
2:
Observation 2. A tighter carbon cap can sometimes paradoxically
lead to a higher carbon emission level under the carbon cap and
offset policy and the carbon cap and trade policy.
The reason for this counter-intuitive phenomenon is that im-
posing carbon cap on a period by period basis gives companies less
flexibility, and they cannot emit more carbon dioxides in one
period if this allows them to emit less in the future. To testify this
presumption, we have performed experiments to compare the
situation in which carbon caps are imposed on the whole time
horizon with that on a period by period basis. The comparison
result supported our presumption (To keep this paper concise, the
detailed information is not given).
5.3.2.2. Unit fuel price. To indicate the impacts of u (unit fuel price)
on cost and emissions, we performed four groups of sensitive
analyses under various policies, which are shown in Fig. 4. As the
carbon cap policy has the least flexibility compared to other cap-
related policies, in order to get feasible solutions we set the carbon
cap in the first period to be 500 and caps in other periods to be 100
in the (P_cc) model. Meanwhile, the vehicle’s capacity is enlarged
to be 50. Parameters’ values in other three models are the same as
in Section 5.1.
From Fig. 4(a)–(d), we can find that when u increases, the total
cost increases due to the increase of fuel cost, and the carbon
emissions decrease in general. This forces decision makers to
control the fuel consumption, which will benefit the environment.
However, we also find that sometimes as u increases the
emission level keeps almost unchanged. For instance, in Fig. 4(d),
when u varies between 4 and 7, the total fuel consumption (also
the emission) is almost the same under each u value. This can be
explained that with the constraint of vehicle capacity and that split
 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 273

Inventory cost Inventory cost

Fuel cost Emissions Fuel cost
1560 Emissions
12000 Total cost 16000 Total cost 4380

11000 1550 4370

14000
10000 4360
1540 12000
9000
4350
8000 1530

Emissions
10000

Emissions
4340
7000

Cost
1520
Cost

6000 8000 4330

5000 4320
1510 6000
4000
4310
3000 1500 4000
4300
2000
1490 2000
1000 4290

0 1480 0 4280
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
u u
(a) Carbon cap policy (b) Carbon cap and offset policy

Inventory cost Inventory cost

Fuel cost Emissions 20000 Fuel cost Emissions 4350
4380
18000
Total cost Total cost
18000
4370 4340
16000 16000
4360 4330
14000 14000
4350

Emissions
12000 4320
12000
Emissions
Cost

4340
10000 10000 4310
Cost

4330
8000 8000
4300

6000 4320 6000

4290
4000 4310 4000
4280
2000 4300 2000

0 4290 0 4270
0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
u u
(c) Carbon cap and trade policy (d) Carbon taxing policy

Fig. 4. Trends of costs and carbon emissions as u changes under different carbon polices.

pick up is not allowed, a vehicle cannot carry too many products Observation 4. A higher unit carbon price does not always result
each time. Thus, in the following periods, vehicles still need to visit in a better environmental benefit; therefore, an appropriate taxing
some suppliers and the travel distance will not decrease. Zhang level should be carefully determined by the government.
et al. (2014) have indicated that travel distance is much more
crucial than load in fuel consumption. Therefore, sometimes al-
though u increases the system’s emissions will not decrease as 6. Conclusion and future work
expected. Here we get Observation 3:

Observation 3. Generally, increasing unit fuel cost will lead to the
increase in cost and the decrease in emissions; however, some-
times the emissions may keep stable with increased u value due to
the constraints like vehicle capacity, non-split delivery and so on.
5.3.2.3. Unit carbon price (tax). Fig. 5 shows the influences of dif-
ferent f (unit carbon price) on cost and emissions under the three
carbon price-related polices. In the three models, parameters’ va-
lues are the same as in Section 5.1.
From Fig. 5(a)–(c), it is observed that with the increase of
unit carbon price, the system’s cost increases and the emissions
decrease as a whole. However, from the curves of emissions, we
can find with some parts the emissions don’t exhibit a various
decreasing trend as the price becomes higher. For example, in
Fig. 5(a), when the unit carbon price increases from 3.5 to 4.5,
the emissions almost keep constant. Also in Fig. 5(b), the car-
bon emissions only reduce by 0.25% when the price varies
from 2.5 to 4.5. The same phenomenon can be found in
Fig. 5(c) when f varies from 1.5 to 3. Therefore, we have
Observation 4:
In this paper, a multi-period two-echelon supply chain is ad-
dressed. In contrast to previous studies, the carbon emission reg-
ulation policies are considered in our study. A binary chromosome
representation using HGA is proposed to solve the problems. A
post-optimization operator is performed on the best chromosome
to enhance the solution. A series of numerical experiments have
demonstrated the effectiveness and efficiency of the proposed
HGA. And some interesting phenomena are observed from the
parameter sensitivity analyses under different polices. These in-
sights can not only offer advices for companies to balance costs
and environmental effects, but also suggest ideas for governments
to employ suitable policies and establish reasonable carbon caps
and prices.
Avenues for future study are numerous. First, we can consider
imposing the carbon cap on the whole planning horizon, not on
each period. Second, as different commodities (for example, high/
low value) need different inventory and transportation tactics; we
may investigate the carbon reduction policies on different in-
dustries. Future studies can also include the carbon emissions of
other supply chain activities, such as warehouse operations. In
274 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275

Inventory cost Inventory cost

Fuel cost Emissions Fuel cost Emissions
18000 Cost of purchased emissions Cost of purchased emissions
Total cost 16000 Total cost 4340
4340
16000
14000
14000 4320 4320

12000
12000 4300
4300

Emissions
10000

Emissions
10000

Cost
4280
Cost

4280
8000
8000

4260 6000 4260

6000

4000 4000
4240 4240

2000 2000
4220 4220
0 0
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5
f f
(a) Carbon cap and offset policy (b) Carbon cap and trade policy

Inventory cost
Fuel cost Emissions
Cost of purchased emissions 4340
26000
Total cost
24000
22000 4320
20000
18000
4300
16000

Emissions
14000
Cost

12000 4280

10000
8000
4260
6000
4000
2000 4240

0
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5
f
(c) Carbon taxing policy

Fig. 5. Trends of costs and carbon emissions as f changes under different carbon polices.

addition, it would make sense to extend the proposed model to
more complex supply chain structures, for example, three-echelon
supply chain systems.
Acknowledgments
This work is supported by the National Natural Science Foun-
dation of China (71272030), National Key Technology Support Pro-
gram of China (2014BAH23F04), as well as the Science & Technology
Foundation of Shenzhen City (CXZZ20130321145336439). We
would like to thank the anonymous referees for their helpful
comments.
References
Abdelmaguid, T.F., Dessouky, M.M., 2006. A genetic algorithm approach to the in-
tegrated inventory-distribution problem. Int. J. Prod. Res. 44 (21), 4445–4464.
http://dx.doi.org/10.1080/00207540600597138.
Andersson, H., Hoff, A., Christiansen, M., et al., 2010. Industrial aspects and litera-
ture survey: Combined inventory management and routing. Comput. Oper. Res.
37 (9), 1515–1536. http://dx.doi.org/10.1016/j.cor.2009.11.009.
Anily, S., Federgruen, A., 1990. One warehouse multiple retailer systems with ve-
hicle routing costs. Manag. Sci. 36 (1), 92–114. http://dx.doi.org/10.1287/
mnsc.36.1.92.
Anily, S., Federgruen, A., 1993. Two-echelon distribution systems with vehicle
routing costs and central inventory. Oper. Res. 41 (1), 37–47. http://dx.doi.org/
10.1287/opre.41.1.37.
Anily, S., Bramel, J., 2004. An asymptotic 98.5%-effective lower bound on fixed
partition policies for the inventory routing problem. Discret. Appl. Math. 145
(1), 22–39. http://dx.doi.org/10.1016/j.dam.2003.09.005.
Baker, B.M., Ayechew, M.A., 2003. A genetic algorithm for the vehicle routing
problem. Comput. Oper. Res. 30 (5), 787–800. http://dx.doi.org/10.1016/
S0305-0548(02)00051-5.
Benjaafar, S., Li, Y., Daskin, M., 2013. Carbon footprint and the management of
supply chains: Insights from simple models. IEEE Trans. Autom. Sci. Eng. 10 (1),
99–116. http://dx.doi.org/10.1109/TASE.2012.2203304.
Berger, J., Barkaoui, M., 2004. A parallel hybrid genetic algorithm for the vehicle
routing problem with time windows. Comput. Oper. Res. 31 (12), 2037–2053.
http://dx.doi.org/10.1016/S0305-0548(03)00163-1.
Borade, A.B., Sweeney, E., 2014. Decision support system for vendor managed in-
ventory supply chain: a case study. Int. J. Prod. Res. 53 (16), 4789–4818. http:
//dx.doi.org/10.1080/00207543.2014.993047.
Bozorgi, A., 2016. Multi-product inventory model for cold items with cost and
emission consideration. Int. J. Prod. Econ. 176, 123–142. http://dx.doi.org/
10.1016/j.ijpe.2016.03.011.
Cachon, G.P., 2014. Retail store density and the cost of greenhouse gas emissions.
Manag. Sci. 60 (8), 1907–1925. http://dx.doi.org/10.1287/mnsc.2013.1819.
Chinese State Council, 2014. Long-Term Logistics Industry Development Planning
(2014–2020). 〈http://www.gov.cn/zhengce/content/2014-10/04/content_9120.
htm〉, (accessed on 04.10.14).
Cho, D.W., Lee, Y.H., Lee, T.Y., Gen, M., 2014. An adaptive genetic algorithm for the
time dependent inventory routing problem. J. Intell. Manuf. 25 (5), 1025–1042.
http://dx.doi.org/10.1007/s10845-012-0727-5.
Coelho, L.C., Laporte, G., 2013. A branch-and-cut algorithm for the multi-product
multi-vehicle inventory-routing problem. Int. J. Prod. Res. 51 (23–24),
7156–7169. http://dx.doi.org/10.1080/00207543.2012.757668.
Coelho, L.C., Cordeau, J.F., Laporte, G., 2014. Thirty years of inventory routing.
Transp. Sci. 48 (1), 1–19. http://dx.doi.org/10.1287/trsc.2013.0472.
Custódio, A.L., Oliveira, R.C., 2006. Redesigning distribution operations: a case study
on integrating inventory management and vehicle routes design. Int. J. Logist.:
Res. Appl. 9 (2), 169–187. http://dx.doi.org/10.1080/13675560600649982.
 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275
Dekker, R., Bloemhof, J., Mallidis, I., 2012. Operations research for green logistics–an
overview of aspects, issues, contributions and challenges. Eur. J. Oper. Res. 219
(3), 671–679. http://dx.doi.org/10.1016/j.ejor.2011.11.010.
Dye, C.Y., Yang, C.T., 2015. Sustainable trade credit and replenishment decisions
with credit-linked demand under carbon emission constraints. Eur. J. Oper. Res.
244 (1), 187–200. http://dx.doi.org/10.1016/j.ejor.2015.01.026.
Elhedhli, S., Merrick, R., 2012. Green supply chain network design to reduce carbon
emissions. Transp. Res. Part D: Transp. Environ. 17 (5), 370–379. http://dx.doi.
org/10.1016/j.trd.2012.02.002.
Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization and Machine
Learning. Addison-Wesley Publishing Company Inc.
Ho, W., Ho, G.T., Ji, P., Lau, H.C., 2008. A hybrid genetic algorithm for the multi-depot
vehicle routing problem. Eng. Appl. Artif. Intell. 21 (4), 548–557. http://dx.doi.
org/10.1016/j.engappai.2007.06.001.
Hoen, K.M.R., Tan, T., Fransoo, J.C., et al., 2010. Effect of carbon emission regulations
on transport mode selection in supply chains (Working paper). Eindhoven
University of Technology, Netherlands (Working paper) 〈http://alexandria.tue.
nl/repository/books/672727.pdf〉.
Hua, G., Cheng, T.C.E., Wang, S., 2011a. Managing carbon footprints in inventory
management. Int. J. Prod. Econ. 132 (2), 178–185. http://dx.doi.org/10.1016/j.
ijpe.2011.03.024.
Hua, G., Qiao, H., Jian, L., 2011b. Optimal Order Lot Sizing and Pricing with Carbon
Trade. 〈http://papers.ssrn.com/sol3/papers.cfm?abstract_id ¼ 1796507〉, (ac-
cessed on 22.02.11).
Huang, S., Lin, P., 2010. A modified ant colony optimization algorithm for multi-item
inventory routing problems with demand uncertainty. Transp. Res. Part E: Lo-
gist. Transp. Rev. 46 (5), 598–611. http://dx.doi.org/10.1016/j.tre.2010.01.006.
IPCC. Climate Change 2014 Synthesis Report–Longer Report 2014. 〈http://www.ipcc.
ch/pdf/assessment-report/ar5/syr/SYR_AR5_LONGERREPORT_Corr2.pdf〉, (ac-
cessed on 01.11.14).
Konur, D., 2014. Carbon constrained integrated inventory control and truckload
transportation with heterogeneous freight trucks. Int. J. Prod. Econ. 153,
268–279. http://dx.doi.org/10.1016/j.ijpe.2014.03.009.
Konur, D., Schaefer, B., 2014. Integrated inventory control and transportation de-
cisions under carbon emissions regulations: LTL vs. TL carriers. Transp. Res. Part
E: Logist. Transp. Rev. 68, 14–38. http://dx.doi.org/10.1016/j.tre.2014.04.012.
Lee, C.G., Bozer, Y.A., White Ⅲ, C.C., 2003. A Heuristic Approach and Properties of
Optimal Solutions to the Dynamic Inventory Routing Problem. University of
Toronto, Toronto, Canada (Working paper) 〈http://www.mie.utoronto.ca/labs/
ilr/papers/IRP(05062003).pdf〉.
Letmathe, P., Balakrishnan, N., 2005. Environmental considerations on the optimal
product mix. Eur. J. Oper. Res. 167 (2), 398–412. http://dx.doi.org/10.1016/j.
ejor.2004.04.025.
Martí, J.M.C., Tancrez, J., Seifert, R.W., 2015. Carbon footprint and responsiveness
trade-offs in supply chain network design. Int. J. Prod. Econ. 166, 129–142. http:
//dx.doi.org/10.1016/j.ijpe.2015.04.016.
McCormick, G.P., 1976. Computability of global solutions to factorable nonconvex
programs: Part I—Convex underestimating problems. Math. Program. 10 (1),
147–175. http://dx.doi.org/10.1007/BF01580665.
Mirzapour Al-E-Hashem, S.M.J., Rekik, Y., 2014. Multi-product multi-period in-
ventory routing problem with a transshipment option: a green approach. Int. J.
Prod. Econ. 157, 80–88. http://dx.doi.org/10.1016/j.ijpe.2013.09.005.
Mjirda, A., Jarboui, B., Macedo, B., et al., 2014. A two phase variable neighborhood
search for the multi-product inventory routing problem. Comput. Oper. Res. 52,
291–299. http://dx.doi.org/10.1016/j.cor.2013.06.006.
Moin, N.H., Salhi, S., 2007. Inventory routing problems: a logistical overview. J.
Oper. Res. Soc. 58 (9), 1185–1194. http://dx.doi.org/10.1057/palgrave.
275
jors.2602264.
Moin, N.H., Salhi, S., Aziz, N., 2011. An efficient hybrid genetic algorithm for the
multi-product multi-period inventory routing problem. Int. J. Prod. Econ. 133
(1), 334–343. http://dx.doi.org/10.1016/j.ijpe.2010.06.012.
Palak, G., Ekşioğlu, S.D., Geunes, J., 2014. Analyzing the impacts of carbon regulatory
mechanisms on supplier and mode selection decisions: An application to a
biofuel supply chain. Int. J. Prod. Econ. 154, 198–216. http://dx.doi.org/10.1016/j.
ijpe.2014.04.019.
Pan, S., Ballot, E., Fontane, F., 2013. The reduction of greenhouse gas emissions from
freight transport by pooling supply chains. Int. J. Prod. Econ. 143 (1), 86–94.
http://dx.doi.org/10.1016/j.ijpe.2010.10.023.
Prins, C., 2004. A simple and effective evolutionary algorithm for the vehicle
routing problem. Comput. Oper. Res. 31 (12), 1985–2002. http://dx.doi.org/
10.1016/S0305-0548(03)00158-8.
Qin, L., Miao, L., Ruan, Q., et al., 2014. A local search method for periodic inventory
routing problem. Expert Syst. Appl. 41 (2), 765–778. http://dx.doi.org/10.1016/j.
eswa.2013.07.100.
Ramkumar, N., Subramanian, P., Narendran, T.T., et al., 2012. Mixed integer linear
programming model for multi-commodity multi-depot inventory routing
problem. OPSEARCH 49 (4), 413–429. http://dx.doi.org/10.1007/
s12597-012-0087-0.
Rosenkrantz, D.H., Stearns, R.E., Lewis II, P.M., 1977. An analysis of several heuristics
for the traveling salesman problem. SIAM J. Comput. 6 (3), 563–581. http://dx.
doi.org/10.1137/0206041.
Roundy, R., 1985. 98%-effective integer-ratio lot-sizing for one-warehouse multi-
retailer systems. Manag. Sci. 31 (11), 1416–1430. http://dx.doi.org/10.1287/
mnsc.31.11.1416.
Seuring, S., 2013. A review of modeling approaches for sustainable supply chain
management. Decis. Support Syst. 54 (4), 1513–1520, 10.1016/j.dss.2012.05.053.
Treitl, S., Nolz, P.C., Jammernegg, W., 2014. Incorporating environmental aspects in
an inventory routing problem. A case study from the petrochemical industry.
Flex. Serv. Manuf. J. 26 (1–2), 143–169. http://dx.doi.org/10.1007/
s10696-012-9158-z.
Ubeda, S., Arcelus, F.J., Faulin, J., 2011. Green logistics at Eroski: A case study. In-
ternational Journal of Production Economics 131 (1), 44–51, http://dx.doi.org/
10.1016/j.ijpe.2010.04.041.
US Government Accountability Office, 2008. Carbon offsets: the U.S. Voluntary
Market Is Growing, but Quality Assurance Poses Challenges for Market-
Participants.〈http://www.gao.gov/new.items/d081048.pdf 〉, (accessed on
08.08).
Vidović, M., Popović, D., Ratković, B., 2014. Mixed integer and heuristics model for
the inventory routing problem in fuel delivery. Int. J. Prod. Econ. 147 (C),
593–604. http://dx.doi.org/10.1016/j.ijpe.2013.04.034.
Viswanathan, S., Mathur, K., 1997. Integrating routing and inventory decisions in
one-warehouse multiretailer multiproduct distribution systems. Manag. Sci. 43
(3), 294–312. http://dx.doi.org/10.1287/mnsc.43.3.294.
Xiao, Y., Zhao, Q., Kaku, I., Xu, Y., 2012. Development of a fuel consumption opti-
mization model for the capacitated vehicle routing problem. Comput. Oper. Res.
39 (7), 1419–1431. http://dx.doi.org/10.1016/j.cor.2011.08.013.
Zhang, S., Lee, C.K.M., Choy, K.L., Ho, W., Ip, W.H., 2014. Design and development of
a hybrid artificial bee colony algorithm for the environmental vehicle routing
problem. Transp. Res. Part D: Transp. Environ. 31, 85–99. http://dx.doi.org/
10.1016/j.trd.2014.05.015.
Zhao, Q., Chen, S., Zang, C., 2008. Model and algorithm for inventory/routing de-
cision in a three-echelon logistics system. Eur. J. Oper. Res. 191 (3), 623–635.
http://dx.doi.org/10.1016/j.ejor.2006.12.056.