Professional Documents
Culture Documents
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.
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 as changes in the production plan or market demand. In such si-
this paper under the four carbon emission regulations mentioned tuations, infinite time horizon models are not quite exact; there-
above. fore, in recent years, many studies have begun to focus on finite
The IRP is to determine simultaneously the optimal inventory time horizon (a single period or multiple periods) IRP. Huang and
control strategy and vehicle schedules to minimize the supply Lin (2010) considered multi-item replenishment with uncertain
chain’s total cost. In the past, inventory and transportation costs demands in a single period. The demands of items only became
have been minimized separately by different departments without known upon arrival at the demand locations. It was necessary to
considering their relationship. However, the interrelationship be- determine shipping polices to optimize the trade-off between
tween the two has recently motivated many researchers to model stockout costs and transportation costs. A modified ant colony
them simultaneously. Studies show that the IRP can indeed reduce optimization algorithm was developed for the problem. Moin et al.
the supply chain’s total cost (Ramkumar et al., 2012; Coelho et al., (2011) considered an inbound logistics system with multi-periods.
2014; Qin et al., 2014). In this paper, we explore the effects of At the beginning of the time horizon, they had to make decisions
carbon emission regulations on a multi-period inventory routing about time of collection, size of shipment and vehicle routes to
problem (MIRP) and examine the relationships involved. minimize the system’s total cost. A genetic algorithm was devel-
The remainder of the paper is organized as follows. In Section oped to solve the mathematical formulation. Mjirda et al. (2014)
2, we review previous research on IRP and green supply chains, proposed a two-phase variable neighborhood search metaheur-
and the contributions of this paper are also noted. We describe our istic to solve the same problem as Moin et al. (2011). Computa-
problem and present the mathematical formulations in Section 3. tional results showed that their algorithm was more efficient than
The hybrid genetic algorithm is described in Section 4. The com- Mion’s. Coelho and Laporte (2013) developed the first branch-and-
putational tests and analyses are presented in Section 5. We draw cut algorithm to solve a multi-product, multi-period and multi-
some conclusions and suggest directions for future research in vehicle IRP. Vidović et al. (2014) addressed a multi-product, multi-
Section 6. 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-
2. Literature review blem was solved by the use of a mixed integer programming
model and a heuristic approach.
IRP can be mainly classified according to the following criteria: In the conventional supply chain system, many researchers
time horizon, supply chain topology, demand pattern, inventory and companies only focus on profits, costs or service levels.
strategy and vehicle fleet (Andersson et al., 2010). In addition, However, with increasing pressure on the environment, many
several other types of IRP can also be found depending on different researchers have begun to consider environmental issues in their
assumptions in the problems. In this paper, we present the studies models. Letmathe and Balakrishnan (2005) proposed two models
in terms of time horizon, which are essential and closely related to for companies to determine their optimal product mix and pro-
the proposed research. Other types of IRP are covered in some duction quantities in the presence of different types of environ-
review papers (Andersson et al., 2010; Coelho et al., 2014). mental constraints. Elhedhli and Merrick (2012) considered a
IRP can be divided into two categories in terms of time horizon: supply chain network design problem that takes CO2 emissions
infinite time horizon IRP and finite time horizon IRP. The goal of into account. They found that the optimal configuration of the
infinite time horizon IRP is to determine the optimal replenish- supply chain would be changed if emission costs were included.
ment intervals of retailers (or suppliers) and plan vehicle routes so Martí et al (2015) studied the supply chain network design pro-
that the supply chain system’s average cost is minimized. With blem under different carbon polices. Benjaafar et al. (2013) pre-
regard to finite time horizon IRP, we need to decide when to visit sented a set of simple and widely used models to illustrate how
retailers (or suppliers) and the resulting products quantities de- carbon emissions can be integrated into operations decisions. Pan
livered (or collected), as well as vehicle routes, to minimize the et al. (2013) explored the environmental impact of the pooling of
supply chain system’s total cost. supply chains and found that supply network pooling was an
For infinite time horizon IRP, some computationally efficient efficient way to reduce CO2 emissions. Hoen et al. (2010) explored
policies have been introduced, such as fixed-partition policies transport mode choices under the carbon cap policy. Bozorgi
(FPP) and power-of-two (POT) policy. Under FPP, retailers are di- (2016) formulated multi-product inventory model for cold pro-
vided into different regions. Each time if one of the retailers in a ducts with emission consideration. A detailed review of the re-
specific region receives a delivery, then other retailers in this re- search in the green logistics can be found in Dekker et al. (2012)
gion also receive deliveries from the same vehicle. Anily and and Seuring (2013).
Federgruen (1990, 1993) and Anily and Bramel (2004) have ana- Although most studies in the field of green supply chains focus
lyzed FPP in detail and demonstrated its effectiveness. Under POT, on production processes or network design decisions, contribu-
the replenishment intervals of retailers are restricted to be power- tions that incorporate environmental aspects into IRP are scarce
of-two multiples of a base planning period. Roundy (1985) has (Treitl et al., 2014). Treitl et al. constructed an IRP model and ap-
proved that an optimal power-of-two policy has an effectiveness plied it to the petrochemical industry. They noted that detailed
of at least 98%. Thus, POT is widely used in IRP. Viswanathan and transportation processes analysis was important to reduce costs
Mathur (1997) proposed a stationary nested POT policy for a two- and carbon emissions in a vendor managed inventory (VMI) sys-
echelon supply chain system. Under a stationary policy, the re- tem. Mirzapour Al-E-Hashem and Rekik (2014) considered
plenishment intervals of retailers are constant. A nested policy greenhouse gas emissions in IRP models. They proposed a trans-
means that if the replenishment interval Ti of product i is larger shipment option to improve the performance of the supply chain.
than the replenishment interval Tj of product j , then Ti is a multiple Konur (2014) addressed an integrated inventory management and
of Tj . This policy was adopted by Custódio and Oliveira (2006) in a truck-load transportation with a carbon cap constraint. They for-
real-world application concerning the distribution of frozen pro- mulated an economic order quantity (EOQ) model and developed
ducts in a Portuguese company. Zhao et al. (2008) employed both a heuristic search method to solve the problem. Konur and
FPP and POT policies to solve a three-echelon IRP. Schaefer (2014) explored the EOQ model with less-than-truckload
For most companies, logistics planning may require some and truckload transportation under the four carbon emission
changes after running for a period of time for various reasons, such regulations.
C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 265
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)
∑ xjit ≤ 1 ∀ i ∈ V′, ∀ t ∈ H
j ∈ V′∪ O (7) 3.4. Formulation with a carbon cap policy
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)
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
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
Table 2 Table 4
Pseudocode of generating initial population. Crossover Operator.
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 bound). A comparison between HGA and CPLEX will be evaluated
and causes unnecessary costs. A post-optimization operator is on computation effectiveness and efficiency. For each instance, the
proposed to improve the solution. This operator is performed on heuristic is run 10 times, and the best objective value is reported.
the best chromosome found in HGA. Suppose that in period t (t >1), Table 5 summarizes the results got from CPLEX and HGA. The
the utilization rate of the last vehicle is rt . If rt +rt ′≤1 ( rt ′ is the first column is data sets. The second and third columns are the
utilization rate of the last vehicle in period t′( t′ < t )), then we best bounds and the best integer solutions found by CPLEX after a
consider deleting the last route in period t and transferring the limited amount of time, respectively. The fourth column (Gap1) is
suppliers in this route to be visited by the last vehicle in period t′. the gap between the best integer and the best bound. The fifth and
We construct the last vehicle’s route again in period t′ and obtain a eighth columns are the CPU time of CPLEX and HGA, respectively.
new route cost. Similarly, inventory cost and carbon emission cost The sixth column is the best objective found in HGA. The solutions
are also renewed. If this post-optimization operator reduces the in bold are as good as the optimal solutions obtained from CPLEX.
total cost, then accept it; otherwise, return to the situation before The seventh column (Gap2) is the gap between the best objective
this operator. Note that if the cost after this operator is equal to the and the best integer.
cost before this operator, we also accept it, as the change will It is observed that for 17 out of 18 small-sized problems, HGA
decrease the number of vehicles used in period t . 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
5. Numerical tests and analyses give satisfying solutions in a relatively shorter time.
For all small-sized problems, the average gap between HGA and
The aim of this section is threefold: CPLEX is 0.04%; however, the CPU time of HGA is dramatically
short. This advantage becomes more obvious when instance size
To evaluate the effectiveness of the proposed HGA. increases. For medium- and large-sized problems, the average
To indicate the effects of carbon emission regulations on MIRP. gaps are 4.25% and 5.32% respectively. Therefore, HGA is able to
To perform sensitive analyses and give some managerial in- produce good solutions for proposed models in a relatively short
sights into the proposed problem. time.
As to MIRP, we use data sets based on Lee et al. (2003), which To further study the effects of different carbon emission reg-
are also available at http://www.mie.utoronto.ca/labs/ilr/IRP. The ulations on MIRP, a series of sensitivity analyses based on a
coordinates, demands and inventory holding cost coefficients can medium-sized instance, the S12H14 , are performed. This is be-
be found at the website. Note that the depot is located at (0, 0) for cause the medium-sized instance is more representative than a
all the data sets. We define SxHy as an instance with x suppliers small-sized one, and it is also time-efficient compared with large-
and y periods. Based on the S12H14 instance given at the website, sized ones. For each circumstance, HGA is run 10 times and the
we generate different test problems. For example, consider the average values are reported. The carbon emissions shown in the
S 3H 3 instance, which means that we use the information of the following figures are the total emissions during the whole time
first 3 suppliers and the first 3 periods in the S12H14 instance as horizon. Based on the sensitivity analyses, some interesting in-
our data set. For small-sized problems, we generate 18 instances: sights are obtained. We present these insights in the form of a
S 3Hy , S 4Hy and S5Hy ( y = 3,5,7,9,11,14). For medium-sized pro- series of observations.
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- 5.3.1. Comparison with the traditional MIRP model
stances based on the S 20H 21 and S50H 21 instances. S 20Hy and In this section, we compare the traditional model with the
S50Hy mean that the information of all the suppliers and the first (P_cc) model and observe the effects of carbon limits on cost and
y periods is used from the S 20H 21 and S50H 21 instances, re- emissions. For the (P_cc) model, in each scenario the cap in the
spectively. Other parameters are as follows: F =20, C = 10 . first period is fixed to be 500 (large enough to ensure a feasible
We assume that the vehicles used in our study are powered by solution), and the caps in other periods are changed step by step.
diesel and use the data based on Ubeda et al. (2011) and Cachon For instance, for scenario 1: G1=500, Gt =500 ( t = 2,3…T ); for sce-
(2014) to compute fuel consumption cost and carbon emissions. nario 2: G1=500, Gt =400 ( t = 2,3…T ); and for scenario 3: G1=500,
The values of parameters are as follows: ρ0 =0.296, ρ*=0.390, Gt =300 ( t = 2,3…T ). The vehicle capacity is also enlarged to be 50
ε = 2.61, u = 1.05, f = 1. in these two models, to make sure that a feasible solution can be
For small- and medium-sized problems, G1=200, Gt =100 ( t = found even though the cap is much tight. The comparison is re-
2,3…T ) . For large-sized problems, G1=500, Gt =400 ( t = 2,3…T ); ported in Table 6. The penultimate column is the percent gap
Parameters in HGA are as follows: K = 200,Pc =0.9,Pm=0.02, G′=300 between the two models’ total costs. The last column is the
(for small-sized problems), G′=500 (for medium- and large-sized emission gap, and the negative values indicate the percentage of
problems). decreased emissions.
From Table 6, we can see that when the cap is much loose
5.2. Evaluation of HGA (between 300 and 500), it affects the system’s performance very
slightly. As the cap becomes tight (decreases from 200 to 120), the
As we indicate in Section 3.8 that the proposed five models are (P_cc) model tends to increase inventory and decrease fuel con-
closely related, in this section we solve one of them, the (P_cco) sumption, thereby curbing carbon emissions. It is noted that the
model, to evaluate the effectiveness of our algorithm. The pro- system’s total cost increases dramatically, far surpassing the de-
posed HGA is implemented in the C þ þ language and run on a PC crease of emissions, when the cap becomes tight. This is due to the
with an Intel Core i5 Processor (2.3 GHZ) and 4 GB memory. The significant increase of inventory cost. From instance data, we find
mathematic formulation is solved by CPLEX 12.6 to obtain the best that the unit inventory holding cost for each product is much
integer solution (the upper bound) and the best bound (the lower 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.
Best bound Best integer Gap1 (%) CPU (s) Best objective Gap2 (%) (%) CPU (s)
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.
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. 3. The result comparison between the carbon cap and offset policy and the
carbon cap and trade policy.
Fig. 2. The results when unit inventory holding costs are changed.
unused carbon emission credits, then they have stronger moti-
vation to decrease emissions. As the cap becomes tighter (less
dramatically although fuel consumption decreases. To verify this
than 225), the system under the carbon cap and trade policy has
assumption, we reduce all products’ unit inventory holding cost by
no extra carbon emission credits to sell and they can only
5 and 10 times respectively, and observe corresponding results,
maintain the operations by buying more credits. In this situa-
which are shown in Fig. 2. In the figure, “Cost 1” represents the
tion, the optimal results under the two policies are the same. On
system’s total cost when inventory holding cost coefficients are
the other hand, it is surprising to find that when the cap de-
not changed and “Emissions 1” is the corresponding emissions.
creases from 190 to 180, the carbon emission level increases
“Cost 2” and “Cost 3” are the system’s total cost when inventory
rather than decreases under these two polies. This unexpected
holding cost coefficients are reduced by 5 and 10 times respec-
phenomenon was also observed by Benjaafar et al. (2013) when
tively, and “Emissions 2” and “Emissions 3” are resulting emission
investigating a lot-sizing problem, which indicates that some-
levels.
times a tighter carbon cap does not mean better environmental
As Fig. 2 shows, the system’s total cost does not increase that
effects. This observation is also illustrated by the case that when
dramatically when the inventory holding cost coefficient is re-
the cap changes from 160 to 150. Thus, we come to Observation
duced. It is also noted that the decreases of emissions are almost
2:
the same in these three sceneries when the cap reduces from 190
to 120. However, their costs have significant differences. This Observation 2. A tighter carbon cap can sometimes paradoxically
phenomenon indicates that different industries (or commodities) lead to a higher carbon emission level under the carbon cap and
have different sensitives to carbon caps. For those whose unit in- offset policy and the carbon cap and trade policy.
ventory holding cost is very high, it is difficult for companies to
The reason for this counter-intuitive phenomenon is that im-
control emissions at lower price, and tight caps will put heavy
posing carbon cap on a period by period basis gives companies less
burdens on companies’ cost. Whereas carbon caps are meaningful
flexibility, and they cannot emit more carbon dioxides in one
in reducing emissions for enterprises whose unit inventory hold-
period if this allows them to emit less in the future. To testify this
ing cost is low. Therefore, governments should consider the dif-
presumption, we have performed experiments to compare the
ference of industries when establishing carbon caps. Here we get
situation in which carbon caps are imposed on the whole time
the first observation:
horizon with that on a period by period basis. The comparison
Observation 1. The reaction to carbon caps varies significantly result supported our presumption (To keep this paper concise, the
due to different industries. Rigid emission limits will put heavy detailed information is not given).
burdens on companies whose products are high-valued, perish-
able or quickly updated. 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
5.3.2. Parameter sensitivity analyses analyses under various policies, which are shown in Fig. 4. As the
To indicate the impacts of different parameter values on cost carbon cap policy has the least flexibility compared to other cap-
and emissions, we have performed many numerical experiments. related policies, in order to get feasible solutions we set the carbon
To keep this paper reasonably concise, we only present the results cap in the first period to be 500 and caps in other periods to be 100
which provide interesting managerial insights. 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
5.3.2.1. Carbon cap. In model (P_cco) and (P_cct), we change the in Section 5.1.
carbon cap step by step (from 150 to 200 with step 10, from 200 to From Fig. 4(a)–(d), we can find that when u increases, the total
350 with step 25) and observe the changes in costs and emissions. cost increases due to the increase of fuel cost, and the carbon
Other parameters’ values, except for the carbon cap, are the same emissions decrease in general. This forces decision makers to
as in Section 5.1. In Fig. 3, if we say “carbon cap ¼100”, we mean control the fuel consumption, which will benefit the environment.
that the cap imposed on each period is equal to 100. However, we also find that sometimes as u increases the
As expected, the tighter the cap, the higher the cost. From emission level keeps almost unchanged. For instance, in Fig. 4(d),
Fig. 3, it is also observed that when the cap is equal to or more when u varies between 4 and 7, the total fuel consumption (also
than 225 the emissions are less under the carbon cap and trade the emission) is almost the same under each u value. This can be
policy, this is because the system can make a profit by selling explained that with the constraint of vehicle capacity and that split
C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 273
Emissions
10000
Emissions
4340
7000
Cost
1520
Cost
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
Emissions
12000 4320
12000
Emissions
Cost
4340
10000 10000 4310
Cost
4330
8000 8000
4300
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 In this paper, a multi-period two-echelon supply chain is ad-
increase in cost and the decrease in emissions; however, some- dressed. In contrast to previous studies, the carbon emission reg-
times the emissions may keep stable with increased u value due to ulation policies are considered in our study. A binary chromosome
the constraints like vehicle capacity, non-split delivery and so on. representation using HGA is proposed to solve the problems. A
post-optimization operator is performed on the best chromosome
5.3.2.3. Unit carbon price (tax). Fig. 5 shows the influences of dif- to enhance the solution. A series of numerical experiments have
ferent f (unit carbon price) on cost and emissions under the three demonstrated the effectiveness and efficiency of the proposed
carbon price-related polices. In the three models, parameters’ va- HGA. And some interesting phenomena are observed from the
lues are the same as in Section 5.1. parameter sensitivity analyses under different polices. These in-
From Fig. 5(a)–(c), it is observed that with the increase of sights can not only offer advices for companies to balance costs
unit carbon price, the system’s cost increases and the emissions and environmental effects, but also suggest ideas for governments
decrease as a whole. However, from the curves of emissions, we to employ suitable policies and establish reasonable carbon caps
can find with some parts the emissions don’t exhibit a various and prices.
decreasing trend as the price becomes higher. For example, in Avenues for future study are numerous. First, we can consider
Fig. 5(a), when the unit carbon price increases from 3.5 to 4.5, imposing the carbon cap on the whole planning horizon, not on
the emissions almost keep constant. Also in Fig. 5(b), the car- each period. Second, as different commodities (for example, high/
bon emissions only reduce by 0.25% when the price varies low value) need different inventory and transportation tactics; we
from 2.5 to 4.5. The same phenomenon can be found in may investigate the carbon reduction policies on different in-
Fig. 5(c) when f varies from 1.5 to 3. Therefore, we have dustries. Future studies can also include the carbon emissions of
Observation 4: other supply chain activities, such as warehouse operations. In
274 C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275
12000
12000 4300
4300
Emissions
10000
Emissions
10000
Cost
4280
Cost
4280
8000
8000
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 Anily, S., Bramel, J., 2004. An asymptotic 98.5%-effective lower bound on fixed
more complex supply chain structures, for example, three-echelon 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.
supply chain systems. 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
Acknowledgments
supply chains: Insights from simple models. IEEE Trans. Autom. Sci. Eng. 10 (1),
99–116. http://dx.doi.org/10.1109/TASE.2012.2203304.
This work is supported by the National Natural Science Foun- Berger, J., Barkaoui, M., 2004. A parallel hybrid genetic algorithm for the vehicle
dation of China (71272030), National Key Technology Support Pro- routing problem with time windows. Comput. Oper. Res. 31 (12), 2037–2053.
http://dx.doi.org/10.1016/S0305-0548(03)00163-1.
gram of China (2014BAH23F04), as well as the Science & Technology Borade, A.B., Sweeney, E., 2014. Decision support system for vendor managed in-
Foundation of Shenzhen City (CXZZ20130321145336439). We ventory supply chain: a case study. Int. J. Prod. Res. 53 (16), 4789–4818. http:
would like to thank the anonymous referees for their helpful //dx.doi.org/10.1080/00207543.2014.993047.
Bozorgi, A., 2016. Multi-product inventory model for cold items with cost and
comments. 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.
References Chinese State Council, 2014. Long-Term Logistics Industry Development Planning
(2014–2020). 〈http://www.gov.cn/zhengce/content/2014-10/04/content_9120.
Abdelmaguid, T.F., Dessouky, M.M., 2006. A genetic algorithm approach to the in- htm〉, (accessed on 04.10.14).
tegrated inventory-distribution problem. Int. J. Prod. Res. 44 (21), 4445–4464. Cho, D.W., Lee, Y.H., Lee, T.Y., Gen, M., 2014. An adaptive genetic algorithm for the
http://dx.doi.org/10.1080/00207540600597138. time dependent inventory routing problem. J. Intell. Manuf. 25 (5), 1025–1042.
Andersson, H., Hoff, A., Christiansen, M., et al., 2010. Industrial aspects and litera- http://dx.doi.org/10.1007/s10845-012-0727-5.
ture survey: Combined inventory management and routing. Comput. Oper. Res. Coelho, L.C., Laporte, G., 2013. A branch-and-cut algorithm for the multi-product
37 (9), 1515–1536. http://dx.doi.org/10.1016/j.cor.2009.11.009. multi-vehicle inventory-routing problem. Int. J. Prod. Res. 51 (23–24),
Anily, S., Federgruen, A., 1990. One warehouse multiple retailer systems with ve- 7156–7169. http://dx.doi.org/10.1080/00207543.2012.757668.
hicle routing costs. Manag. Sci. 36 (1), 92–114. http://dx.doi.org/10.1287/ Coelho, L.C., Cordeau, J.F., Laporte, G., 2014. Thirty years of inventory routing.
mnsc.36.1.92. Transp. Sci. 48 (1), 1–19. http://dx.doi.org/10.1287/trsc.2013.0472.
Anily, S., Federgruen, A., 1993. Two-echelon distribution systems with vehicle Custódio, A.L., Oliveira, R.C., 2006. Redesigning distribution operations: a case study
routing costs and central inventory. Oper. Res. 41 (1), 37–47. http://dx.doi.org/ on integrating inventory management and vehicle routes design. Int. J. Logist.:
10.1287/opre.41.1.37. Res. Appl. 9 (2), 169–187. http://dx.doi.org/10.1080/13675560600649982.
C. Cheng et al. / Int. J. Production Economics 182 (2016) 263–275 275
Dekker, R., Bloemhof, J., Mallidis, I., 2012. Operations research for green logistics–an jors.2602264.
overview of aspects, issues, contributions and challenges. Eur. J. Oper. Res. 219 Moin, N.H., Salhi, S., Aziz, N., 2011. An efficient hybrid genetic algorithm for the
(3), 671–679. http://dx.doi.org/10.1016/j.ejor.2011.11.010. multi-product multi-period inventory routing problem. Int. J. Prod. Econ. 133
Dye, C.Y., Yang, C.T., 2015. Sustainable trade credit and replenishment decisions (1), 334–343. http://dx.doi.org/10.1016/j.ijpe.2010.06.012.
with credit-linked demand under carbon emission constraints. Eur. J. Oper. Res. Palak, G., Ekşioğlu, S.D., Geunes, J., 2014. Analyzing the impacts of carbon regulatory
244 (1), 187–200. http://dx.doi.org/10.1016/j.ejor.2015.01.026. mechanisms on supplier and mode selection decisions: An application to a
Elhedhli, S., Merrick, R., 2012. Green supply chain network design to reduce carbon biofuel supply chain. Int. J. Prod. Econ. 154, 198–216. http://dx.doi.org/10.1016/j.
emissions. Transp. Res. Part D: Transp. Environ. 17 (5), 370–379. http://dx.doi. ijpe.2014.04.019.
org/10.1016/j.trd.2012.02.002. Pan, S., Ballot, E., Fontane, F., 2013. The reduction of greenhouse gas emissions from
Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization and Machine freight transport by pooling supply chains. Int. J. Prod. Econ. 143 (1), 86–94.
Learning. Addison-Wesley Publishing Company Inc. http://dx.doi.org/10.1016/j.ijpe.2010.10.023.
Ho, W., Ho, G.T., Ji, P., Lau, H.C., 2008. A hybrid genetic algorithm for the multi-depot Prins, C., 2004. A simple and effective evolutionary algorithm for the vehicle
vehicle routing problem. Eng. Appl. Artif. Intell. 21 (4), 548–557. http://dx.doi. routing problem. Comput. Oper. Res. 31 (12), 1985–2002. http://dx.doi.org/
org/10.1016/j.engappai.2007.06.001. 10.1016/S0305-0548(03)00158-8.
Hoen, K.M.R., Tan, T., Fransoo, J.C., et al., 2010. Effect of carbon emission regulations Qin, L., Miao, L., Ruan, Q., et al., 2014. A local search method for periodic inventory
on transport mode selection in supply chains (Working paper). Eindhoven routing problem. Expert Syst. Appl. 41 (2), 765–778. http://dx.doi.org/10.1016/j.
University of Technology, Netherlands (Working paper) 〈http://alexandria.tue. eswa.2013.07.100.
nl/repository/books/672727.pdf〉. Ramkumar, N., Subramanian, P., Narendran, T.T., et al., 2012. Mixed integer linear
Hua, G., Cheng, T.C.E., Wang, S., 2011a. Managing carbon footprints in inventory programming model for multi-commodity multi-depot inventory routing
management. Int. J. Prod. Econ. 132 (2), 178–185. http://dx.doi.org/10.1016/j. problem. OPSEARCH 49 (4), 413–429. http://dx.doi.org/10.1007/
ijpe.2011.03.024. s12597-012-0087-0.
Hua, G., Qiao, H., Jian, L., 2011b. Optimal Order Lot Sizing and Pricing with Carbon Rosenkrantz, D.H., Stearns, R.E., Lewis II, P.M., 1977. An analysis of several heuristics
Trade. 〈http://papers.ssrn.com/sol3/papers.cfm?abstract_id ¼ 1796507〉, (ac- for the traveling salesman problem. SIAM J. Comput. 6 (3), 563–581. http://dx.
cessed on 22.02.11).
doi.org/10.1137/0206041.
Huang, S., Lin, P., 2010. A modified ant colony optimization algorithm for multi-item
Roundy, R., 1985. 98%-effective integer-ratio lot-sizing for one-warehouse multi-
inventory routing problems with demand uncertainty. Transp. Res. Part E: Lo-
retailer systems. Manag. Sci. 31 (11), 1416–1430. http://dx.doi.org/10.1287/
gist. Transp. Rev. 46 (5), 598–611. http://dx.doi.org/10.1016/j.tre.2010.01.006.
mnsc.31.11.1416.
IPCC. Climate Change 2014 Synthesis Report–Longer Report 2014. 〈http://www.ipcc.
Seuring, S., 2013. A review of modeling approaches for sustainable supply chain
ch/pdf/assessment-report/ar5/syr/SYR_AR5_LONGERREPORT_Corr2.pdf〉, (ac-
management. Decis. Support Syst. 54 (4), 1513–1520, 10.1016/j.dss.2012.05.053.
cessed on 01.11.14).
Treitl, S., Nolz, P.C., Jammernegg, W., 2014. Incorporating environmental aspects in
Konur, D., 2014. Carbon constrained integrated inventory control and truckload
an inventory routing problem. A case study from the petrochemical industry.
transportation with heterogeneous freight trucks. Int. J. Prod. Econ. 153,
Flex. Serv. Manuf. J. 26 (1–2), 143–169. http://dx.doi.org/10.1007/
268–279. http://dx.doi.org/10.1016/j.ijpe.2014.03.009.
s10696-012-9158-z.
Konur, D., Schaefer, B., 2014. Integrated inventory control and transportation de-
Ubeda, S., Arcelus, F.J., Faulin, J., 2011. Green logistics at Eroski: A case study. In-
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. ternational Journal of Production Economics 131 (1), 44–51, http://dx.doi.org/
Lee, C.G., Bozer, Y.A., White Ⅲ, C.C., 2003. A Heuristic Approach and Properties of 10.1016/j.ijpe.2010.04.041.
Optimal Solutions to the Dynamic Inventory Routing Problem. University of US Government Accountability Office, 2008. Carbon offsets: the U.S. Voluntary
Toronto, Toronto, Canada (Working paper) 〈http://www.mie.utoronto.ca/labs/ Market Is Growing, but Quality Assurance Poses Challenges for Market-
ilr/papers/IRP(05062003).pdf〉. Participants.〈http://www.gao.gov/new.items/d081048.pdf 〉, (accessed on
Letmathe, P., Balakrishnan, N., 2005. Environmental considerations on the optimal 08.08).
product mix. Eur. J. Oper. Res. 167 (2), 398–412. http://dx.doi.org/10.1016/j. Vidović, M., Popović, D., Ratković, B., 2014. Mixed integer and heuristics model for
ejor.2004.04.025. the inventory routing problem in fuel delivery. Int. J. Prod. Econ. 147 (C),
Martí, J.M.C., Tancrez, J., Seifert, R.W., 2015. Carbon footprint and responsiveness 593–604. http://dx.doi.org/10.1016/j.ijpe.2013.04.034.
trade-offs in supply chain network design. Int. J. Prod. Econ. 166, 129–142. http: Viswanathan, S., Mathur, K., 1997. Integrating routing and inventory decisions in
//dx.doi.org/10.1016/j.ijpe.2015.04.016. one-warehouse multiretailer multiproduct distribution systems. Manag. Sci. 43
McCormick, G.P., 1976. Computability of global solutions to factorable nonconvex (3), 294–312. http://dx.doi.org/10.1287/mnsc.43.3.294.
programs: Part I—Convex underestimating problems. Math. Program. 10 (1), Xiao, Y., Zhao, Q., Kaku, I., Xu, Y., 2012. Development of a fuel consumption opti-
147–175. http://dx.doi.org/10.1007/BF01580665. mization model for the capacitated vehicle routing problem. Comput. Oper. Res.
Mirzapour Al-E-Hashem, S.M.J., Rekik, Y., 2014. Multi-product multi-period in- 39 (7), 1419–1431. http://dx.doi.org/10.1016/j.cor.2011.08.013.
ventory routing problem with a transshipment option: a green approach. Int. J. Zhang, S., Lee, C.K.M., Choy, K.L., Ho, W., Ip, W.H., 2014. Design and development of
Prod. Econ. 157, 80–88. http://dx.doi.org/10.1016/j.ijpe.2013.09.005. a hybrid artificial bee colony algorithm for the environmental vehicle routing
Mjirda, A., Jarboui, B., Macedo, B., et al., 2014. A two phase variable neighborhood problem. Transp. Res. Part D: Transp. Environ. 31, 85–99. http://dx.doi.org/
search for the multi-product inventory routing problem. Comput. Oper. Res. 52, 10.1016/j.trd.2014.05.015.
291–299. http://dx.doi.org/10.1016/j.cor.2013.06.006. Zhao, Q., Chen, S., Zang, C., 2008. Model and algorithm for inventory/routing de-
Moin, N.H., Salhi, S., 2007. Inventory routing problems: a logistical overview. J. cision in a three-echelon logistics system. Eur. J. Oper. Res. 191 (3), 623–635.
Oper. Res. Soc. 58 (9), 1185–1194. http://dx.doi.org/10.1057/palgrave. http://dx.doi.org/10.1016/j.ejor.2006.12.056.