Applied Soft Computing 39 (2016) 154–164

Contents lists available at ScienceDirect

Applied Soft Computing

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

The green vehicle routing problem: A heuristic based exact solution

approach
Çağrı Koç a,∗ , Ismail Karaoglan b
a
CIRRELT and Canada Research Chair in Distribution Management, HEC Montréal, Montréal, Canada
b
Department of Industrial Engineering, Selçuk University, Konya, Turkey

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

Article history: This paper develops a simulated annealing heuristic based exact solution approach to solve the green
Received 22 March 2015 vehicle routing problem (G-VRP) which extends the classical vehicle routing problem by considering a
Received in revised form 4 August 2015 limited driving range of vehicles in conjunction with limited refueling infrastructure. The problem par-
Accepted 22 October 2015
ticularly arises for companies and agencies that employ a fleet of alternative energy powered vehicles on
Available online 20 November 2015
transportation systems for urban areas or for goods distribution. Exact algorithm is based on the branch-
and-cut algorithm which combines several valid inequalities derived from the literature to improve lower
Keywords:
bounds and introduces a heuristic algorithm based on simulated annealing to obtain upper bounds. Solu-
The green vehicle routing problem
Mixed integer programming formulation
tion approach is evaluated in terms of the number of test instances solved to optimality, bound quality
Branch-and-cut algorithm and computation time to reach the best solution of the various test problems. Computational results
Simulated annealing show that 22 of 40 instances with 20 customers can be solved optimally within reasonable computation
time.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction The classical VRP assumes that the vehicle fuel tank capacity
is unlimited and the fuel amount in the tank is always sufficient
Today’s competitive economic environment requires strate- to serve all customers in any possible route. However, in real-
gic and operational decisions for companies in order to optimize life, vehicles need to refuel their tanks to continue and complete
and manage their logistic processes more efficiently. One of the their tour. This situation is frequently encountered in the case of
most important operational decision concerns the design of vehi- companies or agencies having alternative energy powered fleets
cle routes since it offers great potential to reduce the costs and to (i.e., natural gas, electricity, ethanol) in which routes have to be
improve the service quality. planned taking additional difficulties associated with the limited
The classical vehicle routing problem (VRP) aims at routing a refueling infrastructure into account. The alternative energy pow-
fleet of vehicles on a given network to serve a set of customers ered fleet operations (or more generally green logistics concept)
under specified supply and demand related constraints. Minimiz- have emerged as one of the latest extensions of the VRP literature
ing the total distance traveled by all vehicles or minimizing the in recent years. For example, many studies suggest that there are
overall travel cost are the typical objectives of the VRP and usually several opportunities for reducing carbon dioxide (CO2 ) emissions
the cost is computed as a linear function of distance. Since its intro- by extending the traditional VRP objectives to account for wider
duction by Dantzig and Ramser [1], the VRP and its variants have environmental and social impacts rather than just economic costs
been studied extensively by researchers. Many heuristics have been [10–12]. These studies are motivated by the activities of the trans-
developed in recent years for several variants of the VRP (see [2–5]). portation industry which has significant negative impacts on the
For a recent coverage of the state-of-the-art models and solution environment, economy and human health. These impacts include
algorithms, the reader is referred to the survey by Cordeau et al. [6] increased resource consumption, toxic effects on ecosystems and
and Laporte [8], and to the books by Golden et al. [7] and Toth and humans, increased air and noise pollution, and also the climatic
Vigo [9]. effects induced by greenhouse gas (GHG) emissions. GHG and in
particular CO2 emissions are the most concerning ones as they
have direct effects on human health, e.g., pollution, and indirect
∗ Corresponding author. Tel.: +1 514 343 6111; fax: +1 514 343 6111. ones, e.g., climate change. Growing concerns about such hazardous
E-mail addresses: [email protected] (Ç. Koç), [email protected] effects of transportation on the environment call for revised plan-
(I. Karaoglan). ning approaches to road transportation by explicitly accounting for

http://dx.doi.org/10.1016/j.asoc.2015.10.064
1568-4946/© 2015 Elsevier B.V. All rights reserved.
 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164 155

such negative impacts. Juan et al. [13] studied the fleet size and capacity. The authors formulated the electric vehicle routing prob-
mix vehicle routing problem with multiple driving ranges in which lem with time windows and recharging stations (E-VRPTW) as a
the total distance that each vehicle type can travel is limited. This MIP and employed a hybrid heuristic solution procedure that com-
problem arises in the routing of electric and hybrid-electric vehicles bines a variable neighborhood search algorithm with a Tabu Search
which can only travel limited distances due to the limited capac- heuristic. The authors have also improved the MIP formulation of
ity of their batteries. A mathematical model is formulated and a Erdogan and Miller-Hooks [18] and used it to solve the set of small
multi-round heuristic is developed. The method is based on a biased G-VRP benchmark instances with CPLEX. The algorithm is tested
randomized algorithm which can be used alone to create alterna- on a benchmark instances derived from the literature and on a
tive fleet choices whenever the feasibility of the prespecified fleet newly generated benchmark instances. The results have shown
configuration is not guaranteed. A set of benchmark instances were that the proposed heuristic performed well and the hybridization
created to analyze how distance-based costs increase when consid- mechanism had positive impacts on the solution quality. Schnei-
ering “greener” fleet configurations. The method performed well on der et al. [28] later studied the vehicle routing problems with
all benchmark instances and many different alternative solutions intermediate stops (VRP-IS) in which stopping requirements at
offer competitive distance-based costs while using fewer long- or intermediate facilities may include replenishment/disposal and
medium-range vehicles than normally required. More recently, Koç refueling/recharging stops. The authors have developed an adap-
et al. [14] studied the fleet size and mix pollution-routing problem tive variable neighborhood search algorithm to solve the VRP-IS
where the objective is a linear combination of vehicle, fixed cost, instances. The algorithm is tested on several related VRP instances
fuel cost and CO2 emissions, and driver cost. The authors formally (G-VRP, E-VRPTW, etc.) in the literature and on the new benchmark
defined the problem, presented a mathematical model and devel- instances. The authors reported that the proposed algorithm shows
oped a hybrid evolutionary metaheuristic. For a further coverage a satisfactory performance compared to the methods from the lit-
of green issues at the operational level the reader is referred to the erature and is able to obtain numerous new best solutions. Felipe
book chapter of Eglese and Bektaş [15] and to the surveys of Demir et al. [29] proposed constructive and local search heuristics within
et al. [16] and Lin et al. [17]. a simulated annealing framework to solve a variant of the G-VRP
Erdogan and Miller-Hooks [18] introduced the green vehicle which considers multiple technologies and partial recharges. The
routing problem (G-VRP). The problem design least-cost delivery authors tested their solution method on a newly generated bench-
routes from a depot to a set of geographically scattered customers mark instances, which indicate the efficiency of the algorithm.
within a pre-specified time limit and without exceeding the vehi- This brief review shows that only several heuristic algorithms
cle’s driving range that depends on fuel tank capacity to minimize have already solved the G-VRP, which does not guarantee the opti-
the total distance traveled and/or total cost. In the G-VRP, vehi- mality. We believe there exists merit for the development of a
cles have limited fuel tank capacity and are allowed to refuel when new solution approach based on simulated annealing heuristic
needed. The vehicles may be refueled at a limited number of fuel and branch-and-cut (B&C) algorithm, which is capable of optimally
stations (FSs) which are available in the service area and at the solving the G-VRP. This is the main motivation of this paper. The
depot node. In practice, the G-VRP is encountered, particularly, contributions of this paper are as follows. We introduce an effi-
when the vehicle fleet includes the alternative fuel vehicles (AFVs). cient and powerful new solution approach to solve the G-VRP. We
The authors proposed a mathematical formulation, and developed develop a new mathematical formulation having fewer variable
two construction heuristics; the Modified Clarke and Wright Sav- and constraints without network augmentation, and adapts a set of
ings heuristic and the density-based clustering algorithm, and a valid inequalities to strengthen the linear programming relaxation
customized improvement technique. of the formulation. We propose a simulated annealing heuristic to
Another problem that is closely related to the G-VRP is the VRP improve initial solution and upper bounds found during the search
with satellite facilities (VRP-SF) in which replenishment of a vehicle process of the solution approach.
is allowed from another facility different from the depot. Bard et al. The rest of this paper is organized as follows. The problem
[19] have formulated the VRP-SF as a mixed integer programming description is given in Section 2. Section 2.1 shows the formulation
(MIP) formulation with capacity and tour duration limitation con- of Erdogan and Miller-Hooks [18]. Section 2.2 presents the pro-
straints. Vehicles with capacity limitations have the option to stop posed mixed integer programming formulation, and subsequently
at satellite facilities to reload in order to serve customer demand Section 3 introduces valid inequalities. The proposed solution
at the nodes. In their formulation, dummy nodes are introduced approach is described in Section 4. Finally, Section 5 presents the
to represent multiple stops at intermediate depots for reloading result of computational experiments, followed by conclusions in
vehicles with goods for delivery process. The authors have also Section 6.
developed an exact solution procedure (B&C algorithm) for solv-
ing VRP-SF to optimality. Regarding the complexity of the problem,
different heuristic/meta-heuristic approaches have been also pro- 2. Problem definition and formulations
posed to solve larger VRP-SF instances [20–22].
Frade et al. [23] studied the location of electric-vehicle charging The G-VRP is defined on a complete directed graph
stations in the city of Lisbon. The authors used a maximal covering G = (N, A), where N = “0” ∪ Nc ∪ F is the set of nodes, “0”
model, and the aim to minimize the level of service and the number corresponds to the depot, Nc = {n1 , n2 , . . ., nc } is the cus-
of charging stations. Chen et al. [24] later studied the electric vehicle tomer nodes, F = {nc+1 , nc+2 , . . ., nc+s } is the fuel stations, and
charging stations location problem where a parking-based assign- A = {(ni , nj ):ni , nj , ∈ N} is the set of arcs that connect nodes in N. An
ment method is presented for the city of Seattle. Nie and Ghamami unlimited number of homogeneous vehicle fleet is available at the
[25] developed a conceptual optimization model to investigate depot to serve customers with fuel tank capacity Q (gallons) and
travel of electric vehicles along a long corridor. Cavadas et al. [26] fuel consumption rate r (gallons per mile). Each vehicle travels
later described a method to locate the electric charging stations on the graph with constant speed sp (miles per hour). Each arc
in the city of Coimbra where the aim is to maximize the number (ni , nj ) ∈ A is associated with a nonnegative distance dij , travel
of electric vehicles under a fixed budget for building the stations. time tij associated with distance (tij = dij /sp) and also the triangular
More recently, Schneider et al. [27] introduced an extended ver- inequality holds (dij + djk ≥ dik ). Note that our formulation is also
sion of the G-VRP by considering an electric vehicle fleet with valid when the triangular inequality does not hold. Each node is
time windows, recharging at stations and limited vehicle load associated with a service time pi , which represents the service
156 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164

   
time at the customer node and refueling time at fuel station node. t0j ≤ j ≤ Tmax − tj0 + pj ∀j ∈ N  \ 0 (9)
The depot can serve as a fuel station and all fuel stations have   
unlimited capacities. There is no limitation on the number of stops fj ≤ fi − r · dij xij + Q 1 − xij ∀j ∈ Nc ; ∀i ∈ N ; i =
/ j (10)
for refueling and fuel tank of vehicle is assumed to be full after  
leaving a fuel station (FS). There exists a time horizon (Tmax ) which fj = Q ∀j ∈ N  \Nc (11)
establishes the duration of a workday. fj ≥ r · dji xji

∀i ∈ (F ∪ F) ; ∀j ∈ Nc (12)
The problem is to determine the corresponding vehicle routes so
 
as to minimize the total cost subject to the following assumptions: xij ∈ 0, 1 ∀ (i, j) ∈ N  (13)

• Each vehicle is used for at most one route. Objective function (1) minimizes the total distance traveled by
• Each route starts and ends at the depot. the vehicles. Constraint sets (2)–(4) are known as degree and flow
• Each customer is served by exactly one vehicle. conservation constraints. While constraint sets (2) and (3) ensure
• Fuel level at the vehicle’s tank must be greater than or equal to that each customer must be visited exactly once, and each FS (and
the fuel consumption between any two nodes. associated dummy nodes) will have at most one successor node,
• The amount of fuel in a vehicle’s tank is sufficient to be able to constraint set (4) guarantees that entering and leaving arcs to each
visit between any pair of nodes. node are equal. Constraint sets (5) and (6) ensure that at most m
• The duration of the route assigned to a vehicle cannot exceed vehicles are used to serve customers. Value m should be set as
Tmax . the number of available vehicles in case of the number of vehicle
restrictions, otherwise these constraints should be removed. Con-
2.1. The formulation of Erdogan and Miller-Hooks [18] straint sets (7)–(9) both determine the arrival time at each node
and ensure that each vehicle returns to the depot no later than
In this section, we present the formulation for the G-VRP in the Tmax . Constraint set (10) is associated with a vehicle’s fuel level
literature. This formulation is originally developed by Erdogan and based on node sequence and type. Constraint sets (7) and (10) are
Miller-Hooks [18] (abbreviated as FEMH ) and improved by Schnei- also eliminates the illegal routes (such as cycle without depot node).
der et al. [27]. Constraint set (11) replenishes the vehicle fuel tank level to Q when
In FEMH , graph G is augmented as G = (N , A ) with a set of s it arrives at any FS. Constraint set (12) guarantees that there will
dummy nodes, F̂ = nc+s+1 , nc+s+2 , . . ., nc+s+s , one for each poten- be enough remaining fuel to return to the depot by the way of a FS
tial visit to a fuel station (FS) or depot serving as a FS (i.e., N  = from any customer location en route. Finally, constraint set (13) is
N ∪ F̂ and A = {(ni , nj ) : ni , nj ∈ N  }). Based on this augmentation, known as integrality constraints.
the decision variables and FEMH are presented as follows:
Decision variables: 2.2. A new formulation for the G-VRP
xij Binary variable equal to 1 if a vehicle travels from node i to
j, 0 otherwise ∀i, j ∈ N . The G-VRP formulation described in previous section requires
fi Fuel level variable specifying the remaining fuel level upon the augmentation of a network with dummy nodes in order to rep-
arrival to node i and it is reset to Q at each refueling station node i resent multiple visits to an alternative fuel station. In the worst
and at the depot (∀ i ∈ N ). case, the number of dummy nodes for each FS visit can be equal
 i Time variable specifying the arrival time of a vehicle at node to the number of customers i . e . , (|N|) which increase the total
i, initialized to zero upon departure from the depot (∀ j ∈ N ). number of nodes in the augmented network from (|Nc | + |F|+1) to
Model (FEMH ): (|Nc | + |Nc ||F|+1). Although, it is acceptable for small test instances,
 this method requires quite more solution time for larger instances
min dij xij (1) since the number of binary variables in a formulation increases with
(i,j) ∈ N  ,i =
/ j the additional dummy nodes. As it is known, the size of enumer-
ation tree in any MIP solver increases exponentially based on the
subject to number of binary variables in a problem. Furthermore, the mem-
 ory requirement of any commercial software increases with the
s.t. xij = 1 ∀i ∈ Nc (2)
size of the matrix dimensioned with the number of variables and
j ∈ N  ,i =
/ j
constraints. Thus, it is expected that the formulation obtained by
   augmentation of a network has poorer performance than that of one
xij ≤ 1 ∀i ∈ N  \Nc (3)
with no augmentation in terms of linear programming (LP) relax-
j ∈ N  ,i =
/ j
ation, CPU time, etc. In this paper, the G-VRP is formulated without
  network augmentation.
xji − xij = 0 ∀j ∈ N  (4)
In our formulation, Miller–Tucker–Zemlin (MTZ) capacity and
i ∈ N  ,i =
/ j i ∈ N  ,i =
/ j
subtour elimination constraints for the VRP are adapted for the G-
 VRP, which were proposed by Miller et al. [30] for the traveling
xoj ≤ m (5)
salesman problem (TSP). Kulkarni and Bhave [31] adapted them
j∈( N  \ 0 ) to the capacitated VRP. Desrochers and Laporte [32] lifted them
 and introduced new bounding constraints on additional variables.
xi0 ≤ m (6)
Kara et al. [33] corrected the lifted version of these constraints, and
i ∈ (N  \ 0 ) finally, exact meanings to auxiliary variables were given by Kara
[34].
   
j ≥ i + tij + pi xij − Tmax 1 − xij For the sake of simplicity, we add the service time (pi ) to
  the succeeding arc (i . e . , tij = pi + (dij /sp)) and introduced two new
∀i ∈ N  ; ∀j ∈ N  \ 0 ; i =
/ j (7) parameters which are indirect distance d̂ijk and indirect travel time
t̂ijk calculated as follows:

0 ≤ 0 ≤ Tmax (8) dijk = dik + dkj ∀i, j ∈ N; ∀k ∈ F; i =
/ j (14)
 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164 157

  
t ijk = pi + pk + dijk /sp ∀i, j ∈ N; ∀k ∈ F; i =
/ j (15) Objective function (16) minimizes the total distance traveled
by the vehicles. Constraint sets (17) and (18) are known as degree
where d̂ijk defines the distance between node i and j through fuel constraints. While constraint set (17) ensures that each customer
station k and t̂ijk determines the total duration which includes the must be visited exactly once, constraint set (18) guarantees that
service time, refueling time and en route time on arcs (i, k) and (k, j). entering and leaving arcs to each node are equal. The arrival time
Based on these definitions, the decision variables of the G-VRP are of a vehicle to each customer is determined by constraint set (19).
presented as follows: Constraint set (19) along with constraint sets (20)–(22) guarantee
Decision variables: that each vehicle must return to the depot before time Tmax . Con-
xij Binary variable equal to 1 if a vehicle travels from node i to straint sets (20) and (21) both bound the decision variables ( i ) and
j directly, 0 otherwise (∀ i, j ∈ (0 ∪ Nc )). specify the arrival time of a vehicle at the first customer in any
yijk Binary variable equal to 1 if a vehicle travels from node i to route. Constraint set (22) ensures that the arrival time of a vehicle
j through fuel station k, 0 otherwise (∀ i, j ∈ (0 ∪ Nc ) ; ∀ k ∈ F). at the last customer must allow the vehicle to return to the depot
fi Fuel level variable specifying the remaining tank fuel level before time Tmax . Constraint set (23) determines the fuel level at
upon arrival to node i (∀ i ∈ Nc ). each customer node based on the distance and the vehicle’s fuel
 i Time variable specifying the arrival time of a vehicle at node consumption rate. Constraint set (24) specifies the vehicle’s fuel
i (∀ i ∈ Nc ). level at the first customer node of any route or first customer node
Model (FKK ): after refueling. Constraint set (25) guarantees that vehicle’s fuel
   
level is enough to return to the depot or to reach a FS. Constraint sets
min dij xij + dijk yijk (16) (26)–(28) are known as non-negativity and integrality constraints,
i,j ∈ (0∪Nc ) i,j ∈ (0∪Nc )k ∈ (0∪F) respectively.
⎛ ⎞ It should be noted that whereas FKK has O(|NC |2 |F|) binary vari-
  ables, O(|NC |) continuous variables and O(|NC |2 |F|) constraints, FEMH
s.t. ⎝xij + yijk ⎠ = 1 ∀j ∈ Nc (17) has O(|NC |2 |F|2 ) binary variables, O(|NC | |F|) continuous variables
i ∈ (0∪Nc ) k ∈ (0∪F) and O(|NC |2 |F|2 ) constraints.
     
xij − xji + yijk − yjik = 0 ∀i ∈ (0 ∪ Nc ) (18)
3. Valid inequalities for the G-VRP
i ∈ (0∪Nc ) k ∈ (0∪F)

      The success of any exact solution procedure mainly depends

i − j + M1ijk − t ijk xij + M1ijk − t ijk − tij − tji xji on the performance of the upper and lower bounding procedures.
      Rapid improvements on these bounds shorten the solution proce-
+ M1ijk − tij yijk + M1ijk − tij − t ijk − t jik yjik
dure. In this section, we describe the valid inequalities derived from

≤ M1ijk − tij − t ijk ∀i, j ∈ Nc ; ∀k (0 ∪ F) (19) the VRP literature to improve the lower bounds.
Valid inequalities are one of the most practical ways to
strengthen the linear relaxations of the formulations. These

j ≥ t0j x0j +

(t 0jk y0jk ) ∀j ∈ Nc (20) inequalities eliminate some fractional solutions from the solution
space in such a way that a stronger lower bound can be obtained
k∈F
for the problem. The first valid inequality which bounds below the
   
j ≤ Tmax − Tmax − t0j x0j − Tmax − t̂0jk y0jk ∀j ∈ Nc (21) number of routes originating from the depot is given as follows:
k∈F
  
 x0j + y0jk ≥ rG-VRP (NC ) (31)
i ≤ Tmax − ti0 xi0 − t̂i0k yi0k ∀i ∈ Nc (22)
j ∈ NC k∈F
k∈F

  where rG-VRP (NC ) is the minimum number of vehicles. Similar

fj − fi + M2ij xij + M2ij − r · dij − r · dji xji
bounding constraint has been used by Achuthan et al. [35] for the
≤ M2ij − r · dij ∀i, j ∈ Nc ; i =
/ j (23) capacitated VRP and [36,37] for the location routing problem with
simultaneous pickup and delivery.
Although, the minimum number of vehicles required can easily
  
fj ≤ Q − r · d0j x0j − r · dkj yijk ∀j ∈ Nc (24) ⎡ having ⎤customer load
be evaluated for vehicle routing problems
i ∈ (0∪Nc ) k ∈ F 
  and vehicle capacity (i.e., rVRP (NC ) = ⎢
⎢ pi /Cap⎥
⎥: where pi is
 ⎢i ∈ NC ⎥
fi ≥ r · di0 xi0 + r · dik yijk ∀i ∈ Nc (25)
j ∈ (0∪Nc ) k ∈ F the demand of customer i, and Cap is the vehicle capacity), it is
not as easy as for routing problems without load restrictions. In
i , fi ≥ 0 ∀i ∈ Nc (26) this paper, we use the following property of any feasible route for
  G-VRP to obtain rG-VRP (NC ).
xij ∈ 0, 1 ∀i, j ∈ (0 ∪ Nc ) (27)
 
yijk ∈ 0, 1 ∀i, j ∈ (0 ∪ Nc ) ; k ∈ F. (28) Property 1
In the given formulation, two constants (M1ij and M2ij ) are used  Let us consider a feasible route P=
for all (i, j) pairs and are evaluated as follows: v0 , v1 , v2 , . . ., vs−1 , vs , vs+1 with arbitrary s customers (as
illustrated in Fig. 1) where v0 and vs+1 represent the depot node
M1ijk = Tmax + tij + t̂ijk − ti0 − t0j ∀i, j ∈ Nc ; k ∈ F (29) "0". The total en route time of P, called as tP , is evaluated as
   
M2ij = Q + r · dij − min r · dkj − min r · dik ∀i, j ∈ Nc (30)
tP = tv0 ,v1 + tv1 ,v2 + · · · + tvs−1 ,vs + tvs ,vs+1 (32)
k∈F k∈F
158 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164

knowledge, this evaluation scheme is the first for the VRP without
customer demands.
Constraint (31) is a special case of the following exponential-
size constraints derived from capacity and subtour elimination
constraints of the VRP [38].

     
xij + yijk ≤ S  − rG-VRP (S) ∀S ⊂ NC , S  > 2 (35)
(i,j) ∈ S k∈F

where rG-VRP (S) is calculated as in constraint (31). The constraint

set (35) guarantees that the number of vehicles visiting a set of
customers is not smaller than the corresponding lower bound.
Another exponential-size inequality for the G-VRP is based on
the generalized large multistar (GLM) inequalities which have been
originally proposed for the VRP [39,40]. We have adapted the GLM
for the G-VRP as follows:
  
xij + yijk
i ∈ S j ∈ (N\S) k∈F
Fig. 1. The total en route time evaluation for any feasible solution. ⎛
1   −
→ ←−
≥ ⎝t̄S + tx ij xij + tx ji xji
or equivalently Tmax
t  t  i ∈ S j ∈ (N\S)
tv0 ,v1 v0 ,v1 + tv1 ,v2 v1 ,v2 + tv2 ,v3
tP = + +
2 2 2  −
→ ←− 
t  + ty ijk yijk + ty jik yjik ∀S ⊂ NC , S =
/ ∅ (36)
vs−1 ,vs + tvs ,vs+1 tvs ,vs+1
+··· + + (33) k∈F
2 2
where t̄S is given as in Eq. (34) for the customer set S and
and more formally

s   ti,0 if node j is depot
tv0 ,v1  tv + tvi ,vi+1 tvs ,vs+1 −
→
tx ij =
i−1 ,vi
tP = + + (34)
2 2 2 ti,j + tj,0 otherwise
i=1
 ∀i, j ∈ Nc (37)
←− t0,i if node j is depot
In the best case, the preceding (vi−1 ) and the succeeding node tx ji =
(vi+1 ) of each customer (vi ) on any route are the first and the second t0,j + tj,i otherwise
nearest nodes to vi in P, respectively. Therefore, the total en route 
−
→ t̂i0 if node j is depot
time evaluation without sequence information in P using the first
ty ijk =
and the second nearest nodes, called t̄P , gives a lower bound on tP .䊐
According to Property 1, a polynomial time procedure is used to  t̂ijk + tj,0 otherwise ∀i, j ∈ Nc ; k ∈ F (38)
←− t̂0i if node j is depot
obtain rG-VRP (NC ) which is illustrated in Table 1. To the best of our ty jik =
t0,j + t̂jik otherwise
Table 1
Procedure to obtain the minimum number of vehicles for a set of nodes.
The constraints set (36) ensure that the number of leaving arcs
(or equivalently the number of routes) from a set of customers (S)
Procedure: Evaluation of rG-VRP (P) must satisfy the required number of routes. These inequalities are
Input: Node set (P) not including the depot node
useful when t̄S /Tmax is close to the next integer.
Output: rG-VRP (P)

Step 1. Initialize parameters

4. Solution approach
t̄P ← 0, NumDepLink ← 0,  ← ∅,
//Route time evaluation without depot node
 (v) in P
Step 2. Perform the following steps to all nodes This section presents a detailed description of our heuristic
Step 2.1. Find the first (v̄cur ) and second vcur nearest node to (vcur ). based exact solution approach for the G-VRP. The method imple-
v̄cur ← argmini ∈ P (ti,vcur ) ments a combination of cutting planes and implicit enumeration
vcur ← argmini ∈ P:i =/ v̄cur (tvcur ,i ) to solve any combinatorial optimization problem. The basic idea is
Step 2.2. Evaluate the effect of vcur to t̄P the identification of the violated inequalities that are valid through-
tv̄ ,v +t ¯
t̄P ← t̄P + cur cur 2 vcur ,v̄cur out the enumeration tree. Therefore, at each step of the solution
//Route time evaluation considering depot node
method, violated inequalities which are identified by solving the
Step 3. Find the best unlinked customer node (i*) in P causing least route
separation problem are added to the formulation and the corre-

time increment because 
of linking to depot
t ¯ sponding linear program is reoptimized.
i,v̄
i∗ ← argmini ∈ P:i/∈ t0,i − 2
i

 the effect
Step 4. Evaluate  of k to t̄P
t∗ ¯
4.1. Simulated annealing
i ,v̄ ∗
t̄P ← t̄P + t0i∗ − 2
i

NumDepLink ← NumDepLink + 1
At each iteration of the simulated annealing (SA) heuristic,
rG−VRP (P) ← t̄P /Tmax neighbors of the current solution are generated by using all of the
←  ∪ i * moving strategies. The best one among them is chosen as a new
Step 5. if NumDepLink < (2 * rG-VRP (P)) then go to Step 3 else report solution (Snew ) for the problem. In recent years, SA has received
rG-VRP (P) and stop
increased attention for solving several routing problems [41]. Banos
 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164 159

et al. [42] developed a SA heuristic which is based on a multi- Table 2

The branch-and-cut algorithm for the G-VRP.
objective procedure to solve the VRP with Time Windows.
The SA algorithm implements a candidate list strategy in gen- Algorithm: Branch-and-cut for the G-VRP
erating the neighbors since searching whole neighborhood of the Input: G-VRP problem data
Output: Optimal solution (Sbest ) for the G-VRP
current solution by a moving strategy is a very time consuming
process. According to this strategy, each moving step randomly Step 1. Construct initial linear relaxation of the MIP model (LP).
generates a subset of neighbors, LS, in order to satisfy the G-VRP Obtain the initial solution (S0 )
Evaluate objective function value of initial solution (f(S0 ))
constraints. Then, this neighbors are gathered in a pool to select
Set Sbest ← S0 , f(Sbest ) ← f(S0 )
the new solution (Snew ). If the new solution is better than the Add the root node to enumeration tree () and set the root node
current solution then it is accepted as the current  solution. Oth- as current node (t)
erwise, it is accepted with probability of exp −s/Titr as the Step 2. if ( = ∅) or (termination criterion is met) then
current solution. s is the relative percent deviation of the quality report Sbest and stop
else
of the new solution from the current solution which is calculated select a node (t) from enumeration tree (t ∈ )
by [(f(Snew ) − f(Scur ))/f(Scur )] * 100. In each iteration of the SA, tem- end if
perature (Titr ) is reduced using a geometric cooling schedule, i.e., Step 3. Solve the corresponding LP and obtain the optimal solution (St∗ ).
Titr ← ˛Titr−1 where ˛ is the cooling rate (0 < ˛ < 1). Step 4. if St∗ is infeasible then
 ← \t,
At the beginning of the SA, the temperature is set to initial
go to Step 2
temperature (i.e., Titr ← T0 ) and the SA stops when the tempera- end if
ture reaches the final temperature (Tf ). SA heuristics use a set of Step 5. if f (St∗ ) > f (Sbest ) then  ← \t and go to Step 2
correlated parameters and configuration decisions. In our imple- Step 6. if St∗ is integer then
mentation, we initially used the parameters suggested by [36,37] if f (St∗ ) < f (Sbest ) then
Sbest ← S ∗t , f (Sbest ) ← f (St∗ )
for the SA, where an extensive meta-calibration procedure was end if
applied to generate effective parameter values. But we have con-  ← \t,
ducted several experiments to further fine-tune these parameters. go to Step 2
After several preliminary experiments, following parameter values end if
if (St∗ is not integer) then
are implemented in the SA. The initial temperature (T0 ) is set 665 in Step 7.
 feas

Obtain feasible solution from current fractional solution St
which an inferior solution (inferior by 70% relative to current solu-
tion) is accepted with a probability of 0.90, the final temperature if f (Stfeas ) < f (Sbest ) then
Sbest ← S feas ,
(Tf ) is chosen 0.15 such that a solution which is inferior by 1% rela- t
f (Sbest ) ← f (Stfeas )
tive to current solution is accepted with a probability of 0.001. The end if
number of neighbors generated by each moving strategy is set |NC | end if
for Swap operator, |F| for Merge, FSAdd and FSDrop operators. Fur- Step 8. if a tailing off exists then
thermore, cooling rate is selected as 0.95. Because of the stochastic go to Step 10
else
nature of the SA, it is run 5 times for each instance with different
solve separation problem to identify violated inequalities and
random seeds. put them cut pool (K)
end if
4.2. B&C algorithm Step 9. if K =/ 0 then
add violated inequalities to the LP,
go to Step 3
The steps of the proposed B&C algorithm for the G-VRP are pre- end if
sented in Table 2, where LP is the LP-relaxation of the MIP model, Step 10. Select a non-integer decision variable according to branching rule,
St∗ is the optimal solution of LP at a particular node t, S0 is the initial create two new nodes and add them to ˚ and go to Step 2
feasible solution, ˚ is the set of unexplored nodes of enumeration
tree.
The B&C is initialized by Step 1 in which the initial LP model figure, routes (0-6-2-5-7-1-0) and (0-3-4-0) are merged into new
is constructed, an initial feasible solution S0 is obtained, and S0 one as (0-6-2-5-7-3-4-1-0).
is accepted as the best feasible solution. The procedure to obtain Swap: Two randomly selected customers, which are in the same
an initial solution in our algorithm can be summarized as follows: route or in two different routes, are exchanged. Swap operator is
first, we implement an extension of the well-known Clarke and illustrated in Fig. 3. In this figure, customers 3 and 6 are swapped
Wright heuristic [43]. In this heuristic, a simple route for each cus- as illustrated in Fig. 3(a) and (b).
tomer from the depot is constructed. Then, all pairs of routes (R1 , R2 ) FSAdd: Two successive customers are selected randomly and
in the solution are examined in terms of cost savings obtained by a FS visit is added between them if there is no previous FS visit.
their merge and the merge is implemented by combining the pair of Although this operator may increase the routing cost, better results
nodes providing the largest cost savings. This strategy is repeated can be obtained at the successive iterations by using other operators
until no feasible combination is found. After an initial feasible solu-
tion (S0 ) is generated, a classical simulated annealing (SA) algorithm
is implemented to obtain an initial solution (S0 ) for the B&C algo-
rithm. SA, which stems from the simulation of the annealing of
solids, is a stochastic search technique that is able to escape from
local optima using a probability function [44]. It starts with an ini-
tial feasible solution (S0 ) and improves it until a stopping criterion
is met. We employ the following moving strategies, two of which
are well-known in the VRP literature and the others are developed
for the G-VRP.
Merge: Two routes (R1 and R2 ) are selected randomly, R2 is
located to each best position in R1 , and the best feasible position is
selected. An example of this operator is illustrated in Fig. 2. In this Fig. 2. A sample for Merge operator.
160 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164

Fig. 5. A sample for FSDrop operator.

Fig. 3. A sample for Swap operator.

0, . . ., i*, k*, j*, . . ., 0). Note that a simple route is constructed for
customer j who is not assigned to any routes since nodes i with
xij∗ , xji∗ , yijk
∗ , y∗ > 0 are assigned previously to another route. After
jik
generating a feasible solution (St ), an improved solution (Stfeas ) is
obtained by implementing SA algorithm explained in Step 1. To
reduce the computation time spent by this procedure, SA is applied
to all nodes up to depth level ten and all nodes on every tenth depth
level in enumeration tree.
Step 8 determines the valid inequalities, if a tailing off does not
exist. In our B&C algorithm, inequality (31) is imposed directly to
Fig. 4. A sample for FSAdd operator.
the model at the beginning of the algorithm (Step 1) since there
is only one constraint in (31). However, we need a procedure to
(i.e., merge or swap) as it is not possible for the current iteration obtain constraints from the sets (35) and (36). For this purpose,
due to the fuel tank capacity limitations. An illustrative example for we introduce a greedy constructive heuristic procedure to separate
this operator can be seen in Fig. 4. In this figure, a FS visit between these inequalities. In each iteration of this procedure, a customer is
customers 3 and 1 is added. After this insertion it may be possible selected randomly as a seed node and customer set S is initialized
to merge routes (0-2-3-6-1-0) and (0-4-5-0) into new one as (0-2- with this seed. Then, a new customer node (m*) that minimizing
3-6-1-4-5-0) for the next iteration. the slack of the constraint (i.e. m∗ ← argmin {slackm 35 } for (35) and
m ∈ (NC \S)
FSDrop: A randomly selected FS visit is removed if possible.
m∗ ← argmin {slackm
36 } for (36)) is selected and S is expanded with
This operator is illustrated in Fig. 5. In this figure, FS visit between m ∈ (NC \S)
customers 2 and 3 are removed and a new arc is added to these m* (i.e. S ← S ∪ m*). The definitions of the slack values are given as
customers. follows for (35) and (36), respectively.
Steps 2–6 are the classical steps of the B&C algorithm. Step
2 specifies the termination criterion and selects an unexplored  
 
node having the smallest objective function value from enumera- 35
slackm = S ∪ m − rG-VRP (S ∪ m) − xij∗ + ∗
yijk (39)
tion tree for additional processing. Then, the next step solves the
(i,j) ∈ (S∪m) k∈F

⎛ ⎞
⎜ ⎟
   ⎜   − ⎟
⎜ −
→ ∗ → ←− ∗  ⎟
36
slackm = xij∗ + ∗
yijk − 1/Tmax ⎜t̄(S∪m) + tx ij xij + ← txji xji∗ + ∗
ty ijk yijk + ty jik yjik ⎟ (40)
⎜ ⎟
i ∈ (S∪m)j ∈ (N\(S∪m)) k∈F ⎝ i ∈ (S ∪ m) k∈F ⎠
j ∈ (N\ (S ∪ m))

corresponding LP model and obtains the (probably fractional) opti- where rG-VRP (S ∪ m) and t̄(S∪m) is defined as previously. When the
mal solution (St∗ ). Steps 4 and 5 prune the current node t, if St∗ is customer m* is selected, the validity of the current constraint is
infeasible or its objective function value is worse than that of the checked. If the violation is occurred for the set S is added to the cut
best feasible solution found so far (Sbest ), respectively. Step 6 prunes pool. This procedure is repeated until all violated cuts have been
the current node if St∗ is integer, and then Sbest is updated if a better identified. These cuts are then imposed to the LP model and reopti-
integer solution is found. mized in Step 9. Finally, Step 10 creates two new nodes by applying
In B&C literature, it has been observed that optimal fractional branching rule, if a tailing off exists or new violated inequalities are
solutions may include some information about optimal integer not identified.
solutions, and it is highly possible to obtain good quality integer
solutions from fractional solutions by using a heuristic and/or exact
5. Computational results
solution procedure. Based on this idea, we obtain a feasible solu-
tion from St∗ by using the following heuristic procedure in Step 7.
This section presents the computational results of numeri-
Let xij∗ and yijk
∗ be an optimal fractional solution of any node on enu-
cal experiments. Firstly, we give brief information about the test
meration tree. Firstly, a route R is started by selecting customer i* problems. Then, we investigate the effectiveness of our formu-
with largest x0i ∗ (or y∗ ) and then appending a new customer j* with
0ik lation (FKK ) and Erdogan and Miller-Hooks’ formulation (FEMH )
∗
largest xij (or yijk∗ ) to the last customer i* in the route R continues
with respect to linear programming relaxations. Finally, the perfor-
as long as addition is feasible. This step is repeated until all cus- mance of the proposed B&C algorithm is analyzed. All experiments
tomers are assigned to routes. If yijk ∗ is selected as the largest value,
were conducted on Intel Xeon 3.16 GHz equipped with 8 GB RAM
then a FS (k*) visit is included between the customers i* and j* (i.e., computer (the operating system is Windows 7-x64).
 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164 161

Table 3
Linear programming relaxation of formulations on the test instances.

Scenario 1 Scenario 2 Scenario 3 Scenario 4

Instance EMH KK Instance EMH KK Instance EMH KK Instance EMH KK

20c3sU1 61.38 38.05 20c3sC1 62.53 32.60 S1 2i6s 61.66 26.37 S1 4i2s 55.01 33.78
20c3sU2 56.48 26.22 20c3sC2 67.32 52.08 S1 4i6s 54.95 25.02 S1 4i4s 51.36 28.24
20c3sU3 57.49 40.33 20c3sC3 60.39 31.31 S1 6i6s 57.40 28.81 S1 4i6s 54.95 25.02
20c3sU4 51.00 29.30 20c3sC4 61.14 32.44 S1 8i6s 55.24 35.23 S1 4i8s 56.69 25.02
20c3sU5 53.90 30.70 20c3sC5 81.71 48.67 S1 10i6s 41.63 13.32 S1 4i10s 56.85 24.95
20c3sU6 51.91 31.37 20c3sC6 75.85 60.89 S2 2i6s 62.94 49.48 S2 4i2s 59.47 32.44
20c3sU7 61.87 35.54 20c3sC7 88.87 47.10 S2 4i6s 69.45 52.06 S2 4i4s 57.07 35.93
20c3sU8 52.86 35.77 20c3sC8 76.45 51.70 S2 6i6s 70.21 41.70 S2 4i6s 68.19 34.90
20c3sU9 66.51 40.59 20c3sC9 70.75 60.24 S2 8i6s 57.95 38.44 S2 4i8s 68.60 34.90
20c3sU10 41.23 13.90 20c3sC10 76.62 66.98 S2 10i6s 56.79 29.98 S2 4i10s 68.60 34.90

Average 55.46 32.17 72.16 48.40 58.82 34.04 59.68 31.01

5.1. Test problems these gap values report how close the corresponding LP bounds to
the optimal/best known solution. It is well-known that the qual-
The algorithm is tested using the benchmark set generated by ity of the bounds is one of the critical issues in reducing execution
Erdogan and Miller-Hooks [18]. Each instance involves 20 cus- time of exact solution procedure. A tighter bound can help the exact
tomers and FSs varying from 2 to 10 (with increments of two). Each solution procedure to get optimal solutions in a shorter computa-
customer and FS node are located in a grid of 330 by 300 miles and tional time. Table 3 presents the computational results for each test
the depot is located near the center of the grid. To investigate the problem.
effect of the location configuration of both customers and stations In Table 3, each triple of columns reports the name of the
and the number of FSs, four different scenarios have been consid- instance, the percentage gap values for the Erdogan and Miller-
ered. In Scenario-1 (2) customers are located uniformly (clustered) Hooks’ formulation (EMH ) and our formulation (KK ) for each
and three FSs are fixed between the depot and the grid boundaries scenario, respectively. As shown, our formulation provides better
in westerly, northerly and southeasterly directions. In Scenario-3, results than [18] formulation for all instances. Average percentage
while half of the test instances are selected from Scenario-1, others gap values on all scenarios are obtained as 36.41% for our formu-
are selected from Scenario-2, and for each of them 3 additional FSs lation while this value is 61.53% for [18] formulation. These results
are located randomly. In Scenario-4, while half of the test instances indicate that it is plausible to use the proposed formulation in B&C
are created from one instance of Scenario-1, others are created from algorithm since tight bounds are one of the most crucial compo-
one instance of Scenario-2, and the number of FSs is increased from nents for exact solution procedures.
2 to 10 by increments of two for each instance.
In the original test set, some instances are infeasible because of
time and refueling restrictions, such as some customers cannot be 5.3. Results for the heuristic algorithm
served in a given maximum route duration or cannot be reached
without more than one refueling stop. Therefore, these customers In this section, we investigate the effects of heuristic solu-
are identified and removed from the test instances. tion procedures (i.e., extended Clarke and Wright heuristic (ECW)
and simulated annealing (SA) algorithm). Table 4 presents the
computational results for each scenario, respectively. In this
5.2. Comparison of MIP formulations table, successive three columns give instance name, average
percentage gap values for ECW and SA heuristics at the begin-
In this section, we compare two MIP formulations with respect ning of B&C, respectively. The gap values are evaluated as
to LP relaxation quality. The LP/MIP solver IBM ILOG CPLEX 12.5 100 * ((Zheur − UB)/Zheur ), where Zheur is the feasible solution value
is used as LP solver. To analyze the computational results, we use for corresponding heuristic solution value and UB is optimal/best
lower bound gap which is the percentage gap between the objective found solution value. Please note that, the computation times of
function value of the LP relaxation (ZLP ) corresponding to a partic- these heuristics are less than 1 s.
ular formulation and the optimal or the best found integer feasible Table 4 shows that average percentage gap values are 14.35%
solution (Z*) obtained by proposed B&C algorithm. The percent- and 5.98% for ECW and SA heuristic, respectively. For ECW, the aver-
age gap values are calculated as  = 100 * ((Z* − ZLP )/Z*). Note that, age gap values are 11.18%, 13.15%, 16.20% and 16.87% for scenario 1,

Table 4
Computational results of the heuristic solution procedures.

Scenario 1 Scenario 2 Scenario 3 Scenario 4

Instance GapECW GapSA Instance GapECW GapSA Instance GapECW GapSA Instance GapECW GapSA

20c3sU1 17.17 0.87 20c3sC1 14.81 12.01 S1 2i6s 6.34 0.65 S1 4i2s 7.98 0.00
20c3sU2 6.56 0.00 20c3sC2 2.67 0.16 S1 4i6s 22.27 13.79 S1 4i4s 17.01 1.61
20c3sU3 15.56 0.30 20c3sC3 23.05 2.98 S1 6i6s 6.76 3.56 S1 4i6s 22.27 13.79
20c3sU4 15.28 7.95 20c3sC4 4.76 4.76 S1 8i6s 23.77 4.91 S1 4i8s 22.27 0.03
20c3sU5 14.76 0.00 20c3sC5 9.04 1.61 S1 10i6s 11.59 11.59 S1 4i10s 22.38 13.34
20c3sU6 5.85 1.90 20c3sC6 10.61 7.68 S2 2i6s 1.18 0.75 S2 4i2s 4.76 4.76
20c3sU7 2.51 0.98 20c3sC7 43.86 43.86 S2 4i6s 46.59 46.59 S2 4i4s 3.49 1.45
20c3sU8 22.74 3.06 20c3sC8 9.49 0.00 S2 6i6s 25.62 0.20 S2 4i6s 22.87 2.26
20c3sU9 0.56 0.00 20c3sC9 3.26 0.00 S2 8i6s 0.00 0.00 S2 4i8s 22.87 7.11
20c3sU10 10.85 10.85 20c3sC10 9.97 2.19 S2 10i6s 17.85 4.61 S2 4i10s 22.80 7.05

Average 11.18 2.59 13.15 7.52 16.20 8.67 16.87 5.14

162 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164

Table 5
Computational results for the test problems of Scenario-1.

Instance #Cust #FS UB LP Root B&C

Gap CPU Gap CPU #(35) #(36) Gap CPU #(35) #(36)

20c3sU1 20 3 1797.49 38.0 0.0 15.6 0.6 70 257 0.0 172.3 3918 12,445
20c3sU2 20 3 1574.78 26.2 0.0 11.6 0.4 76 114 4.7 3600.0 136,128 351,763
20c3sU3 20 3 1704.48 40.3 0.0 28.0 0.1 38 40 0.0 1789.0 126,733 192,425
20c3sU4 20 3 1482.00 29.3 0.0 18.3 0.1 24 32 9.5 3600.0 543,111 915,252
20c3sU5 20 3 1689.37 30.7 0.1 13.7 0.6 57 247 0.0 2165.5 76,617 199,557
20c3sU6 20 3 1618.65 31.4 0.0 14.9 0.2 48 77 2.7 3600.0 206,853 593,611
20c3sU7 20 3 1713.66 35.5 0.0 15.0 0.3 61 134 9.8 3600.0 144,715 411,704
20c3sU8 20 3 1706.50 35.8 0.0 9.0 0.4 63 199 0.0 1601.3 42,187 182,116
20c3sU9 20 3 1708.82 40.6 0.0 20.3 0.5 64 182 8.3 3600.0 157,942 349,572
20c3sU10 20 3 1181.31 13.9 0.0 0.8 0.1 27 25 0.0 2.3 27 25

Average 1617.71 32.2 0.0 14.7 0.3 53 131 3.5 2373.0 143,823 320,847

Table 6
Computational results for the test problems of Scenario-2.

Instance #Cust #FS UB LP Root B&C

Gap CPU Gap CPU #(35) #(36) Gap CPU #(35) #(36)

20c3sC1 20 3 1173.57 32.6 0.0 21.2 0.1 18 20 7.6 3600.0 388,839 396,159
20c3sC2 19 3 1539.97 52.1 0.0 26.2 0.3 51 95 0.0 1164.5 21,531 44,110
20c3sC3 12 3 880.20 31.3 0.0 9.3 0.1 36 60 0.0 25.4 1527 1793
20c3sC4 18 3 1059.35 32.4 0.0 15.3 0.1 30 24 11.0 3600.0 237,796 485,153
20c3sC5 19 3 2156.01 48.7 0.0 37.9 0.2 39 117 0.0 2246.4 79,900 324,194
20c3sC6 17 3 2758.17 60.9 0.0 42.5 0.1 15 31 0.0 61.6 2160 8332
20c3sC7 6 3 1393.99 47.1 0.0 26.2 0.0 3 1 0.0 0.1 3 1
20c3sC8 18 3 3139.72 51.7 0.0 25.6 0.2 29 42 0.0 53.7 860 748
20c3sC9 19 3 1799.94 60.2 0.0 39.7 0.3 52 149 0.0 113.9 2075 5153
20c3sC10 15 3 2583.42 67.0 0.0 48.0 0.2 58 101 0.0 2067.5 53,172 222,441

Average 1848.43 48.4 0.0 29.2 0.2 33 64 1.9 1293.3 78,786 148,808

2, 3 and 4, respectively. For SA heuristic, the average gap values are Tables 5–8 summarize the computational results for each sce-
2.59%, 7.52%, 8.67% and 5.14% for scenario 1, 2, 3 and 4, respectively. nario, respectively. In these tables, the first four columns display
According to these results, although the solutions obtained by ECW the name of instance, the number of customers, the number of FSs
heuristic are far from the optimal/best found solution proposed SA and upper bounds obtained by the proposed B&C algorithm. Bold
heuristic improves this solution within short computation time. and italic entries in UB columns indicate that the corresponding
problem solved to optimality in one hour. Successive columns give
5.4. Results for the solution approach information about LP relaxation, root node of the enumeration tree
and overall B&C algorithm. While the columns labeled Gap report
In this section, we analyze the performance of our B&C algo- the percentage gap, the columns labeled CPU show the solution
rithm in terms of the number of test instances solved to optimality, time. The columns labeled #(35) and #(36) display the number of
computation time, and bound qualities of several components added valid inequalities (35) and (36), respectively.
(i.e., LP relaxation, root node of enumeration tree, and overall In these tables, it can be seen that the average lower bound per-
B&C algorithm). Bound quality is defined as the gap between cor- centage gap values of LP relaxation from FKK are obtained as 32.2%,
responding lower bound (LB) and the optimal/best found solution 48.4%, 34.0% and 31.0% for each scenario, sequentially. However, at
(UB) and is evaluated as 100 * ((UB − LB)/UB). the root node, these values are reduced to 14.7%, 29.2%, 10.9% and
We implement our B&C algorithm using the SCIP v1.2.0 frame- 14.7% without significant increase in solution time (i.e., less than
work [45] and [46] as underlying LP solver, the default parameters 1 s). This result indicates that valid inequalities (35) and (36) are
were used, and a time limit of 1 h is set for each instance. very efficient in improving the solution quality.

Table 7
Computational results for the test problems of Scenario-3.

Instance #Cust #FS UB LP Root B&C

Gap CPU Gap CPU #(35) #(36) Gap CPU #(35) #(36)

S1 2i6s 20 6 1578.12 26.4 0.0 11.5 0.5 64 150 0.0 1626.6 51,655 112,355
S1 4i6s 20 6 1397.27 25.0 0.0 13.6 0.1 21 33 4.3 3600.0 249,452 484,415
S1 6i6s 20 6 1560.49 28.8 0.1 7.9 0.5 66 168 0.0 523.5 10,447 25,838
S1 8i6s 20 6 1692.32 35.2 0.0 8.1 0.6 58 202 0.0 817.8 12,168 52,278
S1 10i6s 20 6 1173.48 13.3 0.0 0.1 0.1 26 28 0.0 1.9 29 30
S2 2i6s 20 6 1633.10 49.5 0.0 11.8 0.3 47 85 0.0 66.0 853 2824
S2 4i6s 19 6 1505.07 38.4 0.0 22.8 0.3 55 76 9.5 3600.0 98,062 157,808
S2 6i6s 20 6 2431.33 52.1 0.0 8.8 0.2 35 88 0.0 1801.0 44,320 121,929
S2 8i6s 16 6 2158.35 41.7 0.0 16.7 0.1 30 59 0.0 3.2 41 140
S2 10i6s 16 6 1585.46 30.0 0.0 7.8 1.5 69 174 0.0 1.3 69 174

Average 1671.50 34.0 0.0 10.9 0.4 47 106 1.4 1204.1 46,710 95,779
 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164 163

Table 8
Computational results for the test problems of Scenario-41.

Instance #Cust #FS UB LP Root B&C

Gap CPU Gap CPU #(35) #(36) Gap CPU #(35) #(36)

S1 4i2s 20 2 1582.21 33.8 0.0 23.7 0.1 18 24 6.2 3600.0 652,305 1,020,534
S1 4i4s 20 4 1460.09 28.2 0.1 17.3 0.2 28 32 3.8 3600.0 398,654 632,741
S1 4i6s 20 6 1397.27 25.0 0.1 13.6 0.1 21 33 4.3 3600.0 249,514 484,537
S1 4i8s 20 8 1397.27 25.0 0.1 13.6 0.2 20 28 0.0 2133.6 144,411 191,305
S1 4i10s 20 10 1396.02 24.9 0.1 13.3 0.2 32 43 5.0 3600.0 146,158 265,565
S2 4i2s 18 2 1059.35 32.4 0.0 14.5 0.1 37 31 8.2 3600.0 182,052 381,171
S2 4i4s 19 4 1446.08 35.9 0.0 19.7 0.3 67 65 6.2 3600.0 93,266 146,004
S2 4i6s 20 6 1434.14 34.9 0.0 10.5 0.3 54 79 6.8 3600.0 104,626 152,519
S2 4i8s 20 8 1434.14 34.9 0.1 10.6 0.5 92 89 6.6 3600.0 61,921 87,563
S2 4i10s 20 10 1434.13 34.9 0.1 10.2 0.5 72 109 6.7 3600.0 49,624 66,906

Average 1404.07 31.0 0.0 14.7 0.2 44 53 5.4 3453.4 208,253 342,885

Tables 5–8 also show that 22 out of 40 instances are solved to bounds. The results also show that the SA heuristic improves this
optimality using the proposed B&C algorithm. For the other test solution within short computation time, however the solutions
instances that is not solved in optimally, the same results found obtained by ECW heuristic are far from the optimal/best found solu-
as in [27] which reported the best G-VRP results in the literature. tion. We realized that the G-VRP is easy to solve when the number
Average percentage gap values are obtained as 3.5%, 1.9%, 1.4% and of fuel stations increases. Our results show that optimal solutions
5.4% with the average solution time of 2373.0, 1293.3, 1204.1 and for 22 out of 40 test instances with 20 customers are obtained
3453.4 s for each scenario, respectively. within reasonable computation time. Furthermore, the computa-
Tables 5 and 6 indicate that test problems where customers tional results indicate that the simulated annealing heuristic is able
are uniformly located (i.e., Scenario-2) are more difficult to solve to improve the performance of the B&C algorithm in finding high
than clustered test problems (i.e., Scenario-1). Average percentage quality solutions.
gap values are obtained as 3.5% and 1.9% with the average solution Concerning the future research directions, the proposed MIP
time of 2373.0 and 1293.3 s for uniformly and clustered located test formulation and B&C algorithm can be modified by considering
problems, respectively. more realistic aspects of the G-VRP including customer demands,
Table 7 reports the results obtained from the test problems in time windows, multiple depots, etc. The proposed B&C algorithm
Scenario-3 which are obtained by increasing the number of FSs implements a straightforward branching scheme (i.e., branching
from three to six on the selected test problems from Scenarios most fractional variable), and three set of valid inequalities. Fur-
1 and 2. As is seen in this table, the average gap value and CPU ther researches may also address how to affect different branching
time are found as 1.4% and 1204.1 s while these values are 2.8% and schemes and valid inequalities to the G-VRP. CO2 emissions or other
1935.1 when the original problems for each row are considered (see greenhouse gas emission can be embedded into the objective func-
Tables 5 and 6), meaning that the problem is easy to solve when the tion.
number of FSs increases.
In another experiment, we analyze the effect of the number Acknowledgements
of FSs for the same customer configuration. Table 8 shows the
results obtained from the test problems in Scenario-4. These results The authors are thankful to the SCIP team for providing us the
are obtained by increasing the number of FSs from two to ten source code of branch-and-cut-and-price framework. The authors
by increments of two on the selected test problems from Sce- are also grateful to “IBM Academic Initiative” program to provide
narios 1 and 2 (20c3sU4 and 20c3sC4, respectively). The results us with CPLEX callable library. Thanks are due to referees for their
clearly show that the objective function value for the test problems valuable comments.
derived from 20c3sU4 is improved (i.e., decreased from 1582.21 to
1396.02) when the number of FSs increases. However, there is some
fluctuation on the objective function value for the test problems References
derived from 20c3sC4 since the some customers become feasible
[1] G.B. Dantzig, J. Ramser, The truck dispatching problem, Manag. Sci. 6 (1959)
as the newly added fuel stations and the number of customers are 80–88.
increased. Therefore, the objective function value increases when [2] A.A. Juan, J. Faulin, R. Ruiz, B. Barrios, S. Caballé, The SR-GCWS hybrid algo-
the number of customers increases. rithm for solving the capacitated vehicle routing problem, Appl. Soft Comput.
10 (2010) 215–224.
[3] M.R. Khouadjia, B. Sarasola, E. Alba, L. Jourdan, E.G. Talbi, A comparative study
6. Conclusions between dynamic adapted PSO and VNS for the vehicle routing problem with
dynamic requests, Appl. Soft Comput. 12 (2012) 1426–1439.
[4] Y. Marinakis, G.R. Iordanidou, M. Marinaki, Particle swarm optimization for the
We have developed a heuristic based solution approach to solve vehicle routing problem with stochastic demands, Appl. Soft Comput. 13 (2013)
the green vehicle routing problem (G-VRP). We have presented 1693–1704.
a mixed integer programming formulation and developed a B&C [5] Z. Ursani, D. Essam, D. Cornforth, R. Stocker, Localized genetic algorithm for
vehicle routing problem with time windows, Appl. Soft Comput. 11 (2011)
algorithm for the exact solution of the G-VRP. We have adapted sev- 5375–5390.
eral valid inequalities from literature to solve the G-VRP. We have [6] J.-F. Cordeau, G. Laporte, M.W.P. Savelsbergh, D. Vigo, Vehicle routing, in: C.
also developed several separation algorithms for these inequalities Barnhart, G. Laporte (Eds.), Transportation, Handbooks in Operations Research
and Management Science, 1st ed., Elsevier Science Publishers, Amsterdam,
and a simulated annealing heuristic to improve the quality of solu- 2007, pp. 367–428.
tions. The solution approach has evaluated in terms of optimality [7] B.L. Golden, S. Raghavan, E.A. Wasil, The Vehicle Routing Problem: Latest
gap and computation time to reach the best solution on the various Advances and New Challenges, Springer Science + Business Media, LLC, New
York, 2008.
test problems derived from the literature. [8] G. Laporte, Fifty years of vehicle routing, Transp. Sci. 43 (2009) 408–416.
Our results clearly indicated that the proposed mathematical [9] P. Toth, D. Vigo, Vehicle Routing: Problems, Methods, and Applications, MOS-
formulation in the B&C algorithm is effectively able to tighten the SIAM Series on Optimization, Philadelphia, 2014.
164 Ç. Koç, I. Karaoglan / Applied Soft Computing 39 (2016) 154–164

[10] A.C. McKinnon, M.I. Piecyk, Measurement of CO2 emissions from road [28] M. Schneider, A. Stenger, J. Hof, An Adaptive VNS Algorithm for Vehicle Routing
freight transport: a review of UK experience, Energy Policy 37 (2009) Problems with Intermediate Stops, Technical Report LPIS-01/2014, 2014.
3733–3742. [29] Á. Felipe, M.T. Ortuño, G. Righini, G. Tirado, A heuristic approach for the green
[11] A. Sbihi, R.W. Eglese, Combinatorial optimization and green logistics, Ann Oper vehicle routing problem with multiple technologies and partial recharges,
Res 5 (2007) 99–116. Transp. Res. E: Logist. Transp. Rev. 71 (2014) 111–128.
[12] T. Bektaş, G. Laporte, The pollution-routing problem, Transp. Res. B: Methodol. [30] C. Miller, A. Tucker, R. Zemlin, Integer programming formulations and traveling
45 (2011) 1232–1250. salesman problems, J. ACM 7 (1960) 326–329.
[13] A.A. Juan, J. Goentzel, T. Bektaş, Routing fleets with multiple driving ranges: is [31] R.V. Kulkarni, P.R. Bhave, Integer programming formulations of vehicle routing
it possible to use greener fleet configurations? Appl. Soft Comput. 21 (2014) problems, Eur. J. Oper. Res. 20 (1985) 58–67.
84–94. [32] M. Desrochers, G. Laporte, Improvements and extensions to the
[14] Ç. Koç, T. Bektaş, O. Jabali, G. Laporte, The fleet size and mix pollution-routing Miller–Tucker–Zemlin subtour elimination constraints, Oper. Res. Lett.
problem, Transp. Res. B: Methodol. 70 (2014) 239–254. 10 (1991) 27–36.
[15] R. Eglese, T. Bektaş, Green vehicle routing, in: P. Toth, D. Vigo (Eds.), Vehicle [33] I. Kara, G. Laporte, T. Bektas, A note on the lifted Miller–Tucker–Zemlin subtour
Routing: Problems, Methods, and Applications, MOS-SIAM Series on Optimiza- elimination constraints for the capacitated vehicle routing problem, Eur. J. Oper.
tion, Philadelphia, 2014, pp. 437–458. Res. 158 (2004) 793–795.
[16] E. Demir, T. Bektaş, G. Laporte, A review of recent research on green road freight [34] I. Kara, Two Indexed Polynomial Size Formulations for Vehicle Routing Prob-
transportation, Eur. J. Oper. Res. 237 (2014) 775–793. lems, Başkent University, Ankara/Turkey, 2008.
[17] C. Lin, K.L. Choy, G.T.S. Ho, S.H. Chung, H.Y. Lam, Survey of green vehicle routing [35] N.R. Achuthan, L. Caccetta, S.P. Hill, An improved branch-and-cut algorithm for
problem: past and future trends, Expert Syst. Appl. 41 (2014) 1118–1138. the capacitated vehicle routing problem, Transp. Sci. 37 (2003) 153–169.
[18] S. Erdogan, E. Miller-Hooks, A green vehicle routing problem, Transp. Res. E: [36] I. Karaoglan, F. Altiparmak, I. Kara, B. Dengiz, A branch and cut algorithm for the
Logist. Transp. Rev. 48 (2012) 100–114. location-routing problem with simultaneous pickup and delivery, Eur. J. Oper.
[19] J.F. Bard, L. Huang, M. Dror, P. Jaillet, A branch and cut algorithm for the VRP Res. 211 (2011) 318–332.
with satellite facilities, IIE Trans. 30 (1998) 821–834. [37] I. Karaoglan, F. Altiparmak, I. Kara, B. Dengiz, The location-routing problem
[20] Y. Chan, S.F. Baker, The multiple depot, multiple traveling salesmen with simultaneous pickup and delivery: formulations and a heuristic approach,
facility-location problem: vehicle range, service frequency, and heuristic Omega 40 (2012) 465–477.
implementations, Math. Comput. Model. 41 (2005) 1035–1053. [38] G. Laporte, Y. Nobert, P. Pelletier, Hamiltonian location problems, Eur. J. Oper.
[21] B. Crevier, J. Cordeau, G. Laporte, The multi-depot vehicle routing problem with Res. 12 (1983) 82–89.
inter-depot routes, Eur. J. Oper. Res. 176 (2007) 756–773. [39] L. Gouveia, A result on projection for the vehicle routing problem, Eur. J. Oper.
[22] C.D. Tarantilis, E.E. Zachariadis, C.T. Kiranoudis, A hybrid guided local search for Res. 85 (1995) 610–624.
the vehicle-routing problem with intermediate replenishment facilities, Inf. J. [40] A.N. Letchford, J.-J. Salazar-González, Projection results for vehicle routing,
Comput. 20 (2008) 154–168. Math. Program. 105 (2006) 251–274.
[23] I. Frade, A. Ribeiro, G. Gonçalves, A.P. Antunes, Optimal location of charging [41] K.A. Dowsland, J.M. Thompson, Simulated annealing, in: Handbook of Natural
stations for electric vehicles in a neighborhood in Lisbon, Portugal, Transp. Res. Computing, Springer, Berlin, 2012, pp. 1623–1655.
Rec. J. Transp. Res. Board 2252 (1) (2011) 91–98. [42] R. Banos, J. Ortega, C. Gil, A. Fernandez, F. De Toro, A simulated annealing-
[24] T.D. Chen, K.M. Kockelman, M. Khan, Electric vehicle charging station location based parallel multi-objective approach to vehicle routing problems with time
problem: a parking-based assignment method for Seattle, in: Transportation windows, Expert Syst. Appl. 40 (2013) 1696–1707.
Research Board 92nd Annual Meeting (No. 13-1254), 2013. [43] G. Clarke, J.W. Wright, Scheduling of vehicles from a central depot to a number
[25] Y.M. Nie, M.A. Ghamami, Corridor-centric approach to planning electric vehicle of delivery points, Oper. Res. 12 (1964) 568–581.
charging infrastructure, Transp. Res. B: Methodol. 57 (2013) 172–190. [44] B. Suman, P. Kumar, A survey of simulated annealing as a tool for single and
[26] J. Cavadas, G. Correia, J. Gouveia, Electric vehicles charging network planning, multiobjective optimization, J. Oper. Res. Soc. 57 (2005) 1143–1160.
in: Computer-based Modelling and Optimization in Transportation, Springer [45] T. Achterberg, SCIP – A Framework to Integrate Constraint and Mixed Integer
International Publishing, 2014, pp. 85–100. Programming, Zuse Institute, Berlin, 2004.
[27] M. Schneider, A. Stenger, D. Goeke, The electric vehicle-routing problem with [46] IBM-ILOG, CPLEX 12.2 http://www-01.ibm.com/software/integration/
time windows and recharging stations, Transp. Sci. 48 (2014) 500–520. optimization/cplex-optimizer/.