Sei sulla pagina 1di 15

Omega 76 (2018) 85–99

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

Multi-depot multi-compartment vehicle routing problem, solved by a


hybrid adaptive large neighborhood search✩
Mahdi Alinaghian∗, Nadia Shokouhi
Department of Industrial and Systems Engineering, Isfahan University of Technology, No. 17, ESteghlal Avenue, Daneshgah Street, Isfahan 84156-83111, Iran

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

Article history: This paper presents a mathematical model for multi-depot multi-compartment vehicle routing problem.
Received 29 June 2016 The objective function of the proposed problem includes the minimization of the number of vehicles
Accepted 5 May 2017
and then minimization of the total traversed routes. In this type of problem, the cargo space of each
Available online 22 May 2017
vehicle has multiple compartments, and each compartment is dedicated to a single type of product. In
Keywords: the proposed model, split delivery for one given product is not allowed, therefore demand of a customer
Multi depot vehicle routing problem for a certain product must be fully delivered by a single vehicle; however, split delivery for a set of
Multi compartment requested products is allowed, so different products can be delivered to a customer by different vehicles.
Adaptive large neighborhood search Considering the NP-Hardness of the proposed problem, a hybrid algorithm composed of adaptive large
neighborhood search and variable neighborhood search is developed to solve the large scale instances.
Performance of the proposed algorithm is evaluated by comparing its results with the results of exact
method, adaptive large neighborhood search algorithm and variable neighborhood search algorithm. The
results demonstrate the good performance of the proposed hybrid algorithm.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction of fuel [4] or different kinds of frozen food [5] is an example of


multi-compartment vehicle routing problem applications.
Vehicle Routing Problem (VRP) is one of the fundamental prob- So, to transport these products simultaneously with a single ve-
lems in operational research and physical distribution. In general, hicle, cargo space of that vehicle should have multiple separate
this problem can be defined with a graph such as G (E, V), where compartments. Naturally, each of these compartments has a lim-
V = {0, 1, …, n} represent the demand nodes or customers, {0} rep- ited capacity. Sometimes the partitions separating the compart-
resents the depot, and E is the set of arcs constituting the trans- ments are fixed, but in some vehicles these partitions are ad-
portation network and therefore represents the routes between the justable, so, for example, capacity of a single compartment can be
demand nodes. Assuming that transportation cost between nodes increased to include total available cargo space. Despite extensive
(i,j) is denoted by Cij , in cases where Cij = Cji VRP is called symmet- research carried out on the subject of VRP, the multi-compartment
ric, and otherwise it should be called asymmetric. The classic VRP variant of this problem has attracted scarce attention.
and its variants including CVRP (Capacitated VRP), VRPTW (VRP In some cases, distribution problem has multiple depots, but
with Time Window), PVRP (Periodic VRP), DVRP (Dynamic VRP), tours of each depot get optimized separately. However, developing
VRPMT (VRP with Multiple Trips), and SDVRP (Split Delivery VRP) a multi-depot multi-compartment VRP can lead to better integra-
are very popular and extensively studied subjects of research. The tion of the routing process and provide a more efficient solution.
definition provided by Renaud et al. [1] for multi-depot VRP di- Fig. 1 can better clarify the issue. This figure illustrates a problem
vides the set V that includes all nodes into two sets: the set of with 10 customers, 2 types of products, and 5 vehicles, in a case
customers VC = {V1 , V2 ,…, VN } and the set of depots VD = {VN + 1 ,…, where number of depots increases from 1 to 4. Fig. 2 shows that
VM }, and defines other parameters like a single depot problem. as the number of depots increases, transportation cost decreases
Transportation of different kinds of products such as different dramatically.
kinds of oil [2], or different kinds of milk [3], or different kinds Several applications of the proposed problem are distribution
of the frozen products, fuel distribution, milk and oil collection,
etc. In addition, some distribution companies have several depots.

This manuscript was processed by Associate Editor Campbell. Integrated and simultaneous optimization of the different depots

Corresponding author. routes can properly decrease the distribution cost instead of sep-
E-mail address: alinaghian@cc.iut.ac.ir (M. Alinaghian). arate optimization of the routes for each depot. Regardless of the

http://dx.doi.org/10.1016/j.omega.2017.05.002
0305-0483/© 2017 Elsevier Ltd. All rights reserved.
86 M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99

Fig. 1. Changing the tours of vehicles by increasing the number of depots.

Fig. 2. The decrease in the total travel distance with the increase in the number of depots.

importance of the multi-depot multi-compartment vehicle routing erogeneous multi-compartment fleet servicing a set of determinis-
problem, to the best of the authors’ knowledge, no research has tic demands over multiple periods. In their model, there was only
been conducted so far. a single central depot, daily travel cost depended on the work time
This paper presents a new mathematical model for multi-depot of each truck per day, and the inventory costs were also incor-
multi-compartment vehicle routing problem. Considering the NP- porated into the formulation. They solved this problem with the
Hardness of the proposed problem, a hybrid algorithm composed branch and price heuristic and the exact method and compared
of adaptive large neighborhood search and variable neighborhood the results. They also proposed a fast heuristic algorithm based on
search is designed to solve the large scale instances. To evalu- the Clarke and Wright algorithm [6] to solve the problem. To eval-
ate the performance of the proposed algorithm, its results are uate the performance of the algorithms, they investigated several
compared with the results of exact method, adaptive large neigh- data sets from distribution companies with 60 customers in which
borhood search algorithm and variable neighborhood search algo- 25 customers were serviced each day. Results demonstrate that so-
rithm. lutions from heuristic and exact methods have 12–15% and 22–25%
The remainder of this paper is organized as follows: Section 2 less total travel cost than the result from the solutions proposed by
reviews the literature on the subject, and Section 3 presents the distribution company. In addition, the number of used vehicles
the problem description and the proposed mathematical model. in the planning horizon in both methods was less than the number
In Section 4 the proposed solution methods are described, and of used vehicles for the distribution company. The running time
in Section 5 the numerical results are presented. Finally, the of heuristic and exact method was almost 1.5 and 4.5 s for sam-
conclusions are presented in Section 6. ple problems. Cornillier et al. [7] proposed a heuristic solution ap-
proach for multi-period multi-compartment fuel delivery problem.
2. Literature review In this problem, each station had a backup reserve and the service
time at each delivery was also incorporated into the formulations.
This section first reviews the previous works on the subject of They generated 100 sample problems to investigate the heuris-
multi-compartment vehicle routing problem and then reviews the tic algorithm. All the sample problems include 200 customers and
literature on multi-depot vehicle routing problem. 28 days. They compared the results from heuristic algorithm with
the result from a heuristic algorithm in the literature of problem.
2.1. Multi-compartment vehicle routing problem The proposed heuristic algorithm generated solution with 35% less
cost in compare with the heuristic algorithm from literature. Re-
Avella et al. [4] studied a fuel distribution company and pro- garding the processing time, the average processing time of the
posed a formulation to minimize the total cost of travel for a het- proposed algorithm was 15 s, while the available heuristic algo-
M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99 87

rithm in the literature solved sample problems in almost 1 s. They consumption (demand) of each station was assumed to be stochas-
did not use the mathematical model in the paper and results were tic. They proposed a mathematical model for the problem. They
not compared with the exact method. Cornillier et al. [8] also stud- also proposed a heuristic method to solve the problem. To investi-
ied the fuel delivery problem with time window. Here, the aim of gate the quality of the obtained solutions from heuristic algorithm,
the proposed model was to optimize the delivery of several types they generated 100 small-sized sample problems, and results from
of fuel to fuel stations, so a bi-objective model aimed at maximiza- heuristic method were compared with the optimal results from the
tion of revenue and minimization of routing cost was provided. solved mathematical model. The small-sized problems included 10
This model was based on the use of multi-compartment vehicles customers, 6 days and 2 vehicles and 3 products. The average er-
with a single depot and its planning horizon was a single day. Like ror of the heuristic algorithm was 7% for problems with small size.
the previous work, the service time at each delivery was taken into The processing time was 2.8 s in average. To investigate the ef-
account, and the fuel destined for each fuel station needed to ar- fect of demand uncertainty on the final solution, they generated
rive within a specified time window. They proposed two heuris- 20 sample problems with 20 customers, 70 days, 4 vehicles and 3
tic algorithms to solve the problem. The first one was capable of statistical distributions for customers’ demands (Exponential, Nor-
solving problems with almost 15 customers. However, the second mal and uniform) and studied the safety stock and urgent transfers
heuristic algorithm was able to solve larger problems. To evaluate costs. Popović et al. [16] also proposed a heuristic solution method
the heuristic algorithms, they used two set of randomly-generated for the fuel delivery problem with multi-compartment vehicles of
sample problems with 15 and 50 customers. Finally, they inves- unlimited size. In this work, inventory costs depended on average
tigated a case study with 42 customers. The result from heuris- daily inventory and a model was developed with the aim of min-
tic algorithm had 22% less cost than the solution proposed by the imizing the costs of routing and inventory. In the end, the devel-
company. Fallahi et al. [9] studied a vehicle routing problem where oped model was solved with a variable neighborhood search algo-
vehicles were assumed to have multiple compartments, and then rithm. To investigate the quality of the meta-heuristic algorithm,
developed a solution approach based on memetic and tabu search they generated three sets of problems. The number of customers
meta-heuristics. They solved some instances of this problem but in the problem sets was 10, 15 and 20, respectively. In addition,
since there was no prior solution to be used as the point of com- the number of days in the test problems set was 3, 4 and 5. In
parison, they compared the results of two proposed solution ap- each set of problems, 30 sample problems were generated. The
proaches with each other. They used 20 sample problems with the quality of the result from meta-heuristic algorithm was compared
context of vehicle routing problem while applying some modifica- with quality of the results obtained from exact method and heuris-
tions. The first 14 sample problems were extracted from the prob- tic algorithm. In the small-sized problems, the meta-heuristic al-
lem set of Christofides [10] with the number of customers rang- gorithm was reach to optimal solution in 27 samples and in gen-
ing from 50 to 199. The last 6 sample problems were from Eilon eral, for 87 problems out of 90 sample tests, it reached to the
[11] sample problems that included 76–484 customers. In all the optimal solution. Vidović et al. [17] proposed a heuristic method
sample problems, it was assumed that vehicles have two compart- model for the fuel delivery IRP. The objective of this model was to
ments. The average values of the objective functions for the sample minimize the maximum travel distance per period, or in other
problems from MA and TS were 965.8 and 961.9, respectively. Re- words to minimize the fleet costs. This model was developed for
garding the processing time, values of 209.84 and 367.7 s were ob- the problems where the goal is to maximize the benefit from an
tained for MA and TS, respectively. In the end they reported that already known number of vehicles. They solved several small and
tabu search provides better solutions than memetic algorithm, but large-sized sample problems with the mathematical model and
with a higher computational time. Mendoza et al. [12] studied the the meta-heuristic algorithm. The result from small-sized prob-
multi-compartment VRP with stochastic demand, and then solved lems has shown that the error of the meta-heuristic algorithm
problem instances of up to 200 customers with three heuristic was almost 7% in average. Abdulkader et al. [18] also studied the
methods: an improved saving-based algorithm and two route-first multi-compartment VRP with a single depot and developed a hy-
cluster-second methods all enhanced by a post optimization pro- brid method composed of ant colony algorithm and local search
cedure. They then compared their computational results with the to solve it. They then compared the result of proposed algorithm
results available in the literature and reported that in most cases with the result of basic ant colony algorithm and demonstrated the
their solutions are the best results yet obtained. Derigs et al. improvements.
[5] studied the distribution of food and fuel products with multi- A number of newer articles on the subject of multi-
compartments vehicles. They proposed an integer programming compartment VRP have been published in 2016. For instance, Kaabi
model and then solved 200 different problem instances (with dif- [19] has developed a hybrid meta-heuristic algorithm to solve the
ferent number of customers, different vehicle capacities, different multi-compartment VRP with time windows and a limited number
compartment capacities, different customer distribution, etc.). They of vehicles. To investigate the effect of the number of vehicles on
proposed one algorithm to solve the problem and then compared the profit, they solved several problems and studied the result.
their proposed algorithm with the algorithm proposed by Muyl- They used the test problems that were proposed by Solomon with
dermans and Pang [13]. Their proposed algorithm in 60 min cal- some modifications. They multiplied the number of products by
culated the best solution for all the sample problems except for two and then investigated the effect of increasing the number of
one. Hanczar [14] presented a bi-objective multi-item formulation vehicles on the profit. Result demonstrates that with increasing the
seeking to minimize the cost of inventory and transportation. This number of vehicles from 1 to 3, the profit will increase by 145%.
author set the planning horizon to 1 day and solved the problem On the other hand, the processing time was also increased by
with a cluster-first route-second approach. They estimated the data 161%. Due to the lack of other solution approaches, they performed
for a real case by interviewing the employees of the system. Based no experiments to evaluate the quality of the solutions. Sethanan
on the estimation, they generated 5 test problems that included 36 et al. [3] have provided several solution algorithms for raw milk
customers, three types of products and 7 days. They demonstrated transportation scheduling problem with a multi-routing arrange-
that when using IRP model, expenses will decrease up to 19%. ment and a heterogeneous fleet of multi-compartment vehicles,
Popović et al. [15] used a simulation approach to analyze the appli- where each compartment is dedicated to a certain type of milk de-
cation of deterministic IRP in solving stochastic fuel delivery prob- pending on the destination. To solve the problem, they proposed 5
lems. In their model, vehicles were assumed to have multiple com- improved meta-heuristic algorithms based on differential evolution
partments, stations were assumed to have a safety stock, and fuel algorithm. Then, they generated 14 sample problems to investigate
88 M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99

the performance of the algorithms. Sample problems included for a specific product was not allowed and it was instead allowed
5–40 customers. The first two sample problems that included 5 for all demands of customer. Liu et al. [26] assessed the MDVRP
and 8 customers were solved with the aid of exact method and with heterogeneous fleet and delivery split. To achieve their ob-
results were compared with that of meta-heuristic algorithms. In jectives, they developed a mathematical formulation in which the
both problems, all the algorithms reached the optimal solution. objective function was to minimize the movement of empty vehi-
Then, result from improved meta-heuristics was compared with cles. They also proposed a greedy algorithm to solve this model. To
the result from the basic algorithm. Result has shown that the evaluate their proposed algorithm, they created two lower bounds.
fifth algorithm had the best performance among all the algo- The computational results have shown a reasonable performance
rithms. The total improvement in the costs was calculated from and solution in a timely manner. Gulczynski et al. [27] presented
0.246% to 0.968%, in comparison with the basic algorithm. In ad- an integer programming-based heuristic to solve the combination
dition, the improvement in the total number of used vehicles was of two problems of MDVRP and split delivery VRP, which they
0–48.558%. called MDSDVRP. The objective function of their model was to min-
imize the distance traveled by vehicles by the use of split deliv-
2.2. Multi-depot vehicle routing problem ery concept. They used their heuristic method to solve 30 prob-
lems and showed a decrease in the total traveled distance by using
Jin et al. [20] modeled the multi-depot vehicle routing problem split delivery approach. Lalla-Ruiz et al. [28] presented a novel
as a binary programing problem and proposed 2 solution meth- mathematical formulation for the multi-depot open vehicle rout-
ods. The first method decomposes the problem into two indepen- ing problem by adding some new constraints and improving oth-
dent sub-problems (1) allocation and (2) routing, and then solves ers provided by the literature. The competition results carried out
them by a two-stage approach; but in the second method, allo- over problem instances from the literature indicates that their pro-
cation and routing sub-problems are integrated. After comparing posed model outperform the existing ones. Rahimi-Vahed et al.
the results, this article concluded that solutions of integrated ap- [29] proposed a model for finding the optimal number of vehi-
proach are better than results of two-stage method. Tsirimpas et cles in three VRP variants including multi-depot VRP, periodic VRP,
al. [21] studied multi-product multi-depot VRP with capacitated multi-depot periodic VRP, and then solved all three problems with
vehicles. This model, in which customer sequence was assumed modular heuristic algorithm (MHA).Extensive computational exper-
to be predefined, was solved using a dynamic programing algo- iments shows that MHA performs impressively well, in terms of
rithm suited for determination of optimal routing policies. Ray et solution quality and computational time for the 3 problems. Kun-
al. [22] proposed a new integer linear model for solving multi- napapdeelert et al. [30] presented two differential evolution algo-
depot logistic problem with the assumption of depots already be- rithms for solving multi-depot vehicle routing problem with multi
ing selected. This model has been developed for customer service pickup and delivery requests. They compared the solution of pro-
application but a minor change in one of its constraints can turn it posed algorithm with the results of classic algorithms and con-
into a product delivery problem. This model has a flexible design, cluded that the proposed algorithm provides better solutions. A
which allows it to be transformed into other forms of VRP and summary of the assumptions of reviewed papers is provided in
the process of transformation is explained within that paper. These Table 1.
authors solved their model with a heuristic algorithm and com- Review of past research shows that despite the importance of
pared their solutions with the instances available in the literature. considering multiple depots for multi-compartment VRPs, this as-
It was found that the proposed algorithm provides good solutions pect of problem has remained neglected. Therefore, this paper con-
for small and medium-sized problems and reasonably acceptable tributes to the literature by presenting a new mathematical model
solutions for larger problem instance. They also solved a real prob- for multi-depot multi-compartment vehicle routing problem. Con-
lem with and without the split delivery assumption and showed sidering the NP-Hardness of the proposed problem, a hybrid al-
that the costs (objective function) obtained with the split delivery gorithm composed of adaptive large neighborhood search and
assumption are better. Salhi et al. [23] presented a formulation for variable neighborhood search is designed to solve the large scale
multi-depot VRP with heterogeneous fleet based on the assump- instances of the problem. To evaluate the performance of the pro-
tions that the total number of vehicles and the number of vehicles posed algorithm, its results are compared with the results of exact
in each depot were limited and pre-determined. In their model, the method, adaptive large neighborhood search algorithm and vari-
vehicle load is calculated at each delivery, and vehicle returning to able neighborhood search algorithm.
depot is guaranteed to be empty. They solved their model with a
modified VNS algorithm and used 26 problem instances available 3. Problem description
in the literature as benchmark. In the end they reported that in 23
cases their results were better than what was previously obtained The goal of this study is to develop a new mathematical model
in the literature. Allahyari et al. [24] studied the multi-depot VRP for multi-depot multi-compartment vehicle routing problem. The
with covering tour. They stated that in many practical applications, objective function of the proposed problem includes the minimiza-
the constraints of time, budget and availability of resources (e.g. tion of the number of vehicles and then minimization of the total
the number and capacity of vehicles) and routes may not allow a traversed routes. This problem includes multiple depots; so each
certain group of customers to be visited. To overcome this chal- vehicle starts its tour from one of these depots, visit its desig-
lenge, the demand can be delivered to an acceptably close loca- nated customers, and then returns to the same depot. Cargo space
tion (in walking distance) to the unsuitable customer. CSP (Cover- of each vehicle has multiple compartments with a limited capac-
ing Salesman Problem) is one of the first VRPs in which visiting ity. The distance traveled by each vehicle is also subject to con-
all customers on a tour is not necessarily required. They called this straint. Due to the fuel limits or the maximum working hours of
problem MDCTVRP (Multi Depot Covering Tour VRP) and used a the employees, a maximum travel distance is considered. Each de-
combination of simulated annealing algorithm and iterated local pot has limited capacity and maximum number of allocated ve-
search to solve it. Their model also incorporated the split deliv- hicles to a depot is limited. Split delivery for one product is not
ery concept originally introduced by Dror and Trudeau [25]. Split allowed, which means that customer demand for a certain product
delivery is a VRP assumption according to which a customer can must be fully delivered by a single vehicle, Therefore, demand of
be served through several visits. But the split delivery used for each customer of each product has to be less than or equal to the
this multi-compartment model was different in the sense that split capacity of each compartment. This means that a vehicle cannot
M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99 89

Table 1
A summary of the reviewed papers.

Year Authors Constraints Solution method

Time- Heterogeneous Stochastic Periodic Pick-up Split Multi- Multi- Exact Heuristic Metaheuristic
window fleet and delivery delivery compartment depot
√ √
1996 Renaud et al. [1]
√ √ √ √ √
2004 Avella et al. [4]
√ √
2004 Jin et al. [20]
√ √ √
2008 Tsirimpas et al. [21]
√ √ √
2008 Cornillier et al. [7]
√ √
2008 Fallahi et al. [9]
√ √ √ √ √
2009 Cornillier et al. [8]
√ √ √ √
2010 Liu et al. [26]
√ √ √
2011 Gulczynski et al. [27]
√ √ √
2011 Mendoza et al. [12]
√ √ √
2011 Derigs et al. [5]
√ √ √
2012 Hanczar, [14]
√ √ √ √ √
2012 Popović et al. [15]
√ √ √ √ √
2013 Vidović et al.
2014 Popović et al. [16]
√ √ √
2014 Ray et al. [22]
√ √
2014 Salhi et al. [23]
2015 Lalla-Ruiz et al. [28]
√ √ √
2015 Rahimi-Vahed et al. [29]
√ √ √
2015 Kunnapapdeelert et al. [30]
√ √
2015 Abdulkader et al. [18]
√ √ √ √
2015 Allahyari et al. [24]
√ √ √
2015 Lahyani et al. [2]
√ √ √
2016 Kaabi [19]
√ √ √ √
2016 Sethanan et al. [3]

partly satisfy the demand of a customer. Customers’ demands for Parameters


different products are deterministic and predetermined. The fleet is dig demand of node i for product g
homogeneous. This means that the capacity of all vehicles and the Cij the distance between nodes i and j
number of each vehicle’s compartment, are equal. However, split Qg the capacity of vehicle compartment dedicated to prod-
delivery for multiple products is allowed, so customer can receive uct g
different products by different vehicles, also each depot can supply DQi the maximum number of vehicles departing from depot i
all kinds of products. MC the maximum distance, each vehicle is allowed to travel
fk fixed cost of using vehicle k
3.1. Problem assumptions M a very large number

Decision variables
The proposed formulation is based on the following assump-
xijk equals 1 if the route between nodes i and j is traveled by
tions:
vehicle k, and is zero otherwise
• Vehicles have multiple compartments. yigk equals 1 if the demand of node i for product g is deliv-
• Each compartment is dedicated to one type of product. ered by vehicle k, and is zero otherwise
• Each customer can be served by multiple vehicles. uk equals 1 if the vehicle k is used, and is zero otherwise
• Each vehicle is assigned to one depot. STik variable used for elimination of sub-tours
• Demands of all customers must be fully met.
• Each vehicle starts its tour from a depot and in the end returns 3.3. Mathematical formulation
to the same depot.
• A customer’s demand for each product must be fully met by a This section presents the mathematical formulation proposed
single vehicle, but customer can receive different products by for multi-depot multi-compartment vehicle routing problem,
different vehicles.    
min Ci j xi jk + f k uk (1)
• Each depot has a limitation on the number of assigned vehicles.
i∈Ntot j∈Ntot k∈K k∈K
• Each depot can supply all kinds of products.
• The distance traveled by each vehicle is subject to constraint. s.t:
• The fleet is homogeneous. 
xi jk ≤ uk , ∀k ∈ K (2)
i∈ND j∈Nv
3.2. Sets, parameters and variables

yigk = 1, ∀i ∈ Nv , g ∈ G (3)
This section defines the sets, indices, parameters, and variables k∈K
used in the proposed formulation.  
Sets yigk ≤ M x jik , ∀i ∈ Nv , k ∈ K (4)
Ntot the set of all nodes; denoted by indices i and j g∈G j∈Ntot
ND the set of depots  
Nv the set of customers x jik ≤ yigk , ∀k ∈ K, i ∈ Nv (5)
G the set of products; denoted by index g j∈Ntot g∈G
K the set of vehicles; denoted by index k
90 M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99


xi jk ≤ 1, ∀ j ∈ Nv , k ∈ K (6)
i∈Ntot
 
xi jk = x jik , ∀ j ∈ Ntot , k ∈ K (7)
i∈Ntot i∈Ntot

yigk × dig ≤ Qg , ∀g ∈ G, k ∈ K (8)
i∈Nv
 
Ci j × xi jk ≤ MC , ∀k ∈ K (9)
i∈Ntot j∈Ntot

xi jk ≤ DQi , ∀i ∈ ND (10)
k∈K j∈Nv

STik = 0 (11) Fig. 3. An example of solution representation.

i∈ND k∈K

cells with the value of zero. So this string has Nv × G + K − 1 cells.


STik + 1 ≤ ST jk + M (1 − xi jk ) , ∀i ∈ Ntot , j ∈ Nv , k ∈ K (12)
Fig. 3 shows a solution representation with 5 customers (Nv ), 3
vehicles (K), 2 depots (ND ) and 2 product (G).
xi jk , yigk ∈ {0, 1}, ∀i, j ∈ Ntot , k ∈ K, g ∈ G (13) In the solution string, the products of each customer are se-
quentially assigned with a number. For example, the first and sec-
ond products of the first customer are shown with 1 and 2. The
STik ≥ 0, ∀i ∈ Ntot , k ∈ K. (14)
customer and the product related to each cell can be obtained from
Eq. (1) is the objective function, which includes two sections, Eqs. (16) and (17),
the first section is related to minimizing the total traveled dis-
customer (ssa ) = (ssa )/G (16)
tance, and the second section is related to the fixed cost for us-
ing the vehicles. Constraint (2) indicates that in case each vehicle
product (ssa ) = ( (ssa − 1 ) mod G ) + 1. (17)
is used, then it should start its tour from a depot. This constraint
also indicates that each vehicle can ultimately leave the depot only In these equations, ssa is the numeric value of the ath cell and
once. Constraint (3) states that each customer’s demand for a given G is the number of products.
product must be delivered by a single vehicle. Constraints (4) and For example, the value of 9 in the solution string demonstrates
(5) state that the demand of customer i for products g can be ful- the first product of customer 5. To assign customers to vehicles, all
filled by vehicle k only when this vehicle visits customer i. Con- the cells from the beginning of the string will be assigned to the
straint (6) ensures that each vehicle visits any given customer at first vehicle until we reach a zero value. Having the same proce-
most once. Constraint (7) states that any vehicle that enters a node dure, other cells will be assigned to other vehicles. As shown in
should also depart from that node. Constraint (8) limits the ca- the figure, both products of the first customer and both products
pacity of vehicles and constraint (9) limits the maximum distance. of the fourth customer are assigned to the first vehicle. When dif-
Constraint (10) is related to the capacity of depots. Constraints (11) ferent products requested by a customer are assigned to one vehi-
and (12) are used for elimination of sub-tours. And constraints (13) cle, solution string will be amended such that all those products
and (14) define the domains of decision variables. will be placed successively behind the product to be delivered at
the first visit. This ensures that each vehicle visits each customer
4. Solution method just once. After the tours are determined, they will be assigned
to depots. Tours are assigned to depots with consideration of the
This section presents the meta-heuristic algorithms proposed maximum number of tours that can be assigned to each depot. To
for the solution of described problem. In this section, first the so- assign tours to depots, the following procedure is performed: first,
lution representation is described, then the procedures of Vari- the cost of assignment of a tour to all the depots is calculated.
able Neighborhood Search (VNS) and Adaptive Large Neighborhood Then, the tour with highest value of the difference between lowest
Search (ALNS) are explained, and finally the proposed hybrid meta- assignment cost and the second lowest assignment cost is assigned
heuristic algorithm is presented. to the depot with lowest assignment cost. This process will go on
until all vehicles are assigned to depots. To calculate the cost of
4.1. Solution representation assignment of a tour to a depot, the depot will be placed among
all the customers in the tour and the related cost is calculated and
The solution procedure starts with estimating the lowest pos- the location with minimum cost if selected.
sible number of vehicles required for solution, and proceeds with
solving the problem based on this estimation; when this estima- 4.2. Variable Neighborhood Search (VNS)
tion cannot yield a feasible solution, the number of vehicle will be
incremented by 1. The above mentioned initial estimation is ob- VNS is a meta-heuristic algorithm commonly used for solving
tained from Eq. (15), combinatorial and global optimization problems. The core idea of
  this algorithm is to use local search for changing the search neigh-
i∈Nv dig
Max arg . (15) borhood. VNS was first proposed by Mladenović and Hansen in
g∈G Qg
1997 [31] as a method of systematic neighborhood change based
The customers allocated to each vehicle and the order of visit on repeating local searches. VNS is a simple algorithm and its ad-
are represented by a solution string. In the solution string, for each vantage is the low number of its parameters, which allows it to
product of each customer, a cell has been considered. In addition, exhibit a very good speed. VNS is composed of two main steps:
the assignment of customers to vehicles is determined with K − 1 shaking and local search. Assume that Nl (where l = 1, 2, ..., lmax )
M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99 91

Fig. 4. VNS pseudo-code.

is a predetermined neighborhood structure and Nl (x) is the set of 4. Two-exchange: algorithm randomly exchanges two customer of
neighbors of point x under structure Nl . In the shaking step, algo- a tour with two other customers of another tour. This exchange
rithm uses the lth neighborhood structure to randomly move from includes all products pertaining to the exchanged customers.
the current (incumbent) solution to a neighbor solution (S’). In the 5. Product split: algorithm selects a random customer on a tour,
next step, VNS runs a local search around S’ to find the optimal and randomly transfers one or several of its products on that
local solution S
∗ . If S
∗ is better than current solution S, it will be tour to another tour.
used as the new incumbent solution; otherwise algorithm moves 6. Product aggregation: algorithm selects a random customer on a
to the next neighborhood (Nl+1 ).The search continues as long as l tour, and randomly transfers one of its products on other tours
≤ lmax or until termination conditions are met. Fig. 4 shows the to the selected tour.
pseudo-code of this algorithm.
Local search
The purpose of this step is to find the local optimums. After
4.2.1. The proposed VNS algorithm
creating a neighborhood, in this step algorithm performs a local
The proposed VNS algorithm is composed of four main steps:
search on the changed solution. In this paper, the swap and the 2-
generation of initial solution, shaking phase, local search, and ter-
opt local searches are used for this purpose. So at each start of this
mination, which are all described in the following.
step, algorithm randomly selects one of these two local searches
to be applied on the solution. The solution selection policy used in
Generation of an initial solution. The initial solution of algorithm
this step is the selection of overall best improvement.
will be generated randomly.
SWAP algorithm: Assume a tour in form of {1, ..., i, i +
1, ..., j, j + 1, ..., n}. The swap algorithm selects two points like i
Shaking phase. The purpose of this phase is to make a great
and j on this tour and swaps them such that the tour changes to
change in the current solution [32]. This phase is based on the
use of 6 neighborhood structures. These neighborhood structures
{1, ..., j, i + 1, ..., i, j + 1, ..., n}. To apply this algorithm, all possible
values for i and j (i ∈ {1, 2, ..., n − 1} and j ∈ {i + 1, ..., n}) will be
are arranged in the order of magnitude of their effect on the so-
assessed and the swap that makes the best improvement in the
lution, from simple neighborhood structures to complex neighbor-
objective function, will be performed.
hood structures, and will be used according to this order. In other
2-Opt algorithm: This algorithm selects two points like i and
words, the simplest structure will be used first and the most com-
j and reverse the string between them {1, ..., j, j − 1..., i + 1, i, j +
plex one will be used in the end. It should be noted that these
1, ..., n}. It then checks all possible values for i and j (where i ∈
6 neighborhood structures have been chosen by assessing 12 can-
didate structures. The neighborhood structure will change after 3
{1, 2, ..., n − 1}, j ∈ {i + 1, ..., n}) and selects the change making the
best improvement in the objective function.
consecutive iterations without any improvement (the number 3
Termination condition
has been determined through trial and error). It is worth mention-
In each neighborhood, failure of algorithm to improve the so-
ing that only the feasible solutions are studied in a shake. If an
lution after number of iterations (Shakeiter ) leads to increment of
infeasible solution is generated by applying a shake (violation of
neighborhood index, and if the last unsuccessfully explored neigh-
the capacity limit for each compartment or the maximum traveled
borhood is lmax , then algorithm restarts the exploration from l1 .
distance), such shake is performed again. This will go on until a
The algorithm stops when the iteration counter reaches predeter-
feasible solution is found.
mined number of iterations (Iterationmax ).
1. One-move: algorithm selects a customer on a tour, and transfers
this customer and all its products on that tour to another tour. 4.3. Adaptive large neighborhood search (ALNS)
2. Two-move: algorithm selects two customers on a tour, and
transfers them and all their products on that tour to another First introduced in 1998 by Shaw [33], ALNS is a developed ver-
tour. sion of large neighborhood search (LNS), with the difference that
3. One-exchange: algorithm randomly exchanges a customer of a it allows user to employ a variety of destroy and repair operators.
tour with another customer in another tour. In this procedure Recently, LNS-based meta-heuristic algorithm has yielded good
all products pertaining to the exchanged customers are moved. results for transportation scheduling problems. This algorithm
92 M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99

Fig. 5. The procedures of destroy and repair operators.

e−(c(x )−c(x ))/T ,if c(xt )<c(xb ). Steps of this algorithm are presented
t
explores an expanded set of neighborhoods and improves an initial
solution by alternating use of destroy and repair operators. Destroy schematically in Fig. 5. This figure shows 2 depots and 9 customers,
operator – as the name suggests – destroys part of the solution, of which 3 are removed by destroy operator, and are then rein-
and then repair operator rebuild that part in another fashion. For serted in other positions by the repair operator.
example, in a VRP, a simple destroy operator can randomly remove
20% of customers from the solution, and then a repair operator can
4.3.1. The proposed hybrid ALNS algorithm
remake the removed part of solution by the use of a greedy heuris-
The proposed algorithm is a combination of VNS and ALNS al-
tic method.
gorithms. The use of VNS allows the algorithm to carry out a more
In the original paper of Shaw, the accept operator only ac-
extended search around the neighborhood of solution. The pro-
cepts the improved solutions, but the accept operator of later
posed hybrid algorithm also uses a banned list in each step of its
articles published by Ropke [34] and Schrimpf [35], is based
procedure. This list prevents repeated changes in the location of
on the concept that is presented on simulated annealing algo-
one node in the solution string and thus allows a broader neigh-
rithm, meaning that temporary solution Xt will always be ac-
borhood search. Pseudo-code of the proposed hybrid ALNS algo-
cepted if: c(xt ) < c(xb ), and will be accepted with a probability of
rithm is presented in Fig. 6.

Fig. 6. Pseudo-code of the proposed hybrid ALNS algorithm.


M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99 93

current
In the pseudo-code shown in Fig. 6, w j represents the the S number of products exhibiting greatest difference be-
weight of the operator j in the inner loop; Nuj is the current num- tween current position cost and best position cost during pre-
ber of uses of the operator j in the inner loop. The initial temper- vious iterations.
ature T is assumed to be equal to the value of objective function 6. The NR operator (Neighborhood Removal): this operator divides
of the first solution found by algorithm, and to decrease after each the length of each tour by the number of its customers to
iteration by h = 0.999. With the above description, the steps of hy- calculate the average travel distance per customer. It then re-
brid ALNS algorithm are as follows: moves S products from customers with the greatest difference
between their position cost and the average travel distance of
4.3.1.1. Generation of initial solution. Algorithm generates a random their tour.
initial solution. 7. The ZR operator (Zone Removal): this operator divides the area
to smaller zones, selects a random zone and then removes all
nodes within it. In this operator, the process of selection is
4.3.1.2. Assigning weights to destroy and repair operators. Weights
current based on Eq. (21):
of operators, which are denoted by w j , are initially identical.
At this step, there is a probability as high as Eq. (18) for operator j s = { j∗ |x(i1 ) ≤ x( j∗ ) ≤ x(i2 ), y(i1 ) ≤ y( j∗ ) ≤ y(i2 )} (21)
to be accepted as the repair operator and a probability as high as And the process continues until S nodes are removed.
Eq. (19) for it to be accepted as the destroy operator (these proba- 8. The NNR operator (Node Neighborhood Removal): this operator
bilities are denoted by Pr and Pd respectively), first selects a random node, and then delineates a rectangular
wj perimeter around the node and removes all nodes within it. If
Pr j =  (18) the number of removed nodes is less than S, operator expands
j∈r wj
the rectangular perimeter until S nodes are removed.
wj 9. The SR’ operator: this operator is a variant of SR operator; the
Pd j =  . (19)
j∈d wj different weights used in this operator (φ1 = 0.5, φ2 = 0.5, φ3 =
0) increase the chance of a customer and all its related products
In each iteration, destroy and repair operators are selected by being removed from the tours.
the use of a roulette wheel. 10. The ZWDR operator (Zone Worst Distance Removal): this opera-
tor is a combination of ZR and WDR operators. This operator
4.3.1.3. Destroy operation. Each destroy operator removes a certain divides the area to smaller zones, selects S random zones, and
percentage of the total number of customers. Here, the number of then removes the customer with the highest position cost in
removed customers are denoted by S, and the set of destroy opera- each zone.
tors is denoted by D. The proposed ALNS algorithm uses 11 destroy 11. The improved SR operator: this operator first removes a product
operators, of which 8 are derived from [36] and are only slightly of a customer from the tour and adds it to a set of removals.
modified to become fit for the discussed problem, and the remain- Then it removes a point (a product of a customer) closest to
ing two (operators 9,10 and 11) are proposed by the authors. the set of removals, and continues this process until S nodes
1. The RR operator (Random Removal): this operator randomly re- are removed.
moves S products of different customers. This action triggers a 4.3.1.4. Repair operation. In this step, one of the repair operators
split in different products of customers. selected based on calculated probabilities will be applied on the
2. The WDR operator (Worst distance Removal): this operator first solution destroyed in the previous step. Repair operations used for
calculates each customer’s position cost, which is the sum of its this purpose are as follows:
distances from previous and next customers on the tour, and
then selects a number of customers with the highest position 1. The GI operator (Greedy Insertion): this operator inserts the re-
costs and removes them along with their products. The total moved nodes in the best possible position in the tour. To do so,
number of removed products should be equal to S. it assesses all possible edges and selects the one with lowest
3. The SR operator (Shaw Removal): this operator first removes cost (fi1 ).
randomly a product of a customer (one point on the solution 2. The RI operator (Regret Insertion): one problem of GI operator
string), and then removes S − 1 number of products of other is the delay that appears in the insertion of some nodes and
customers who are related to the initially selected customer. may lead to reaching an unfeasible solution. To avoid this prob-
Points that are on the same tour with the initially selected lem, RI operator uses a regret criterion. Like GI operator, RI first
point, have different demands, and are closer to that point have calculates the insertion cost for all edges, but it then finds the
a higher chance of being selected. Eq. (20) will be used to cal- best and second best positions of insertion, denoted by fi1 and
culate a score for each point, and points with lowest scores will fi2 , and places the node with greatest difference between these
be selected. In this equation, φ 1 , φ 2 and φ 3 are the weights of two values (regret = fi2 − fi1 ) at top insertion priority.
criteria; li j = −1 if points i and j are on the same tour and is 3. The GIN operator (Greedy Insertion with Noise function): this op-
zero otherwise; qi j = −1 if demand of two points are not of the erator is an extension of GI operator, with its difference being
same product and is zero otherwise. After examining the re- the use of a freedom degree for selection of best insertion po-
sults, the weights are selected to be φ1 = 0.5, φ2 = 0.25, φ3 = sition. This operator adds the cost of insertion of a customer to
0.25. a given edge with the value of d¯με , where d¯ is the maximum
  distance between nodes,ɛ ∈ [0, 1], and μ is a parameter used
aw j = arg min φ1 .Disi j + φ2 .li j + φ3 .qi j (20) for diversification of solution (μ = 0.1).
4. The PR operator (Proximity-based Removal): this operator is a 4. The RIN operator (Regret Insertion with Noise function): this oper-
variant of SR operator. It first removes a randomly selected cus- ator is an extension of RI operator, and its difference (like GIN)
tomer, and then removes S − 1 customers closest to the initially is the use of a freedom degree for selection of best insertion
removed customer (it means:φ1 = 1, φ1 , φ2 = 0). position.
5. The HR operator (Historical knowledge node Removal): this oper- It is worth mentioning that both the capacity constraint of each
ator calculates the position cost of each product of each cus- compartment and maximum travel distance were included in the
tomer at each iteration. When applied, this operator removes repair operators.
94 M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99

Table 2
Parameter values determined by Taguchi analysis.

VNS ALNS Proposed ALNS

Parameter Value Parameter Value Parameter Value

Iterationmax 110 max iter1 100 max iter1 90


Shakeiter 3 max iter2 50 max iter2 40
S U[0.05Nv × G, 0.15Nv × G] S U[0.05Nv × G, 0.15Nv × G]

4.3.1.5. Improvement phases. The solutions are improved by VNS And in the third category, which is called “nominal-the-best”, de-
algorithm. This means that when the value of objective function sirability increases with the value of objective function approach-
of new solution is better than the incumbent one, VNS algorithm ing a certain target value. The formula of S/N ratio depends on
searches around the new solution to find an improvement. This the type of assessed objective function. In this paper the aim is to
procedure allows the hybrid algorithm to search the neighborhood minimize the value of objective function, so a smaller-the-better
of this solution for improvements before re-destroying it. It should Taguchi approach was used for parameter adjustment. To do so,
be noted that in this combined use of VNS and ALNS, the number first parameters affecting each algorithm were determined by trial
of iterations in each neighborhood (in VNS) is limited to 1, so after and error and three levels were defined for each parameter. Then
searching the last neighborhood, algorithm ends. the Minitab software was used for data analysis, and afterwards,
an appropriate standard orthogonal table was selected based on
4.3.1.6. Acceptance criterion. The acceptance criterion used in this the number of factors and levels selected for analysis. Three prob-
study is based on a simulated annealing algorithm, which means lem instances were then randomly selected, and each problem was
that temporary solution xcurrent will always be accepted if: solved 10 times based on parameter values listed in the developed
c(xnew ) > c(xcurrent ), and will be accepted with a probability of Taguchi table. The mean value of objective function was then re-
e(−(c(Xnew )−c(Xcurrent )/T ) if c(xnew ) < c(xcurrent ). The initial temperature ported to Minitab software. And finally, the level showing the high-
T is assumed to be equal to the value of objective function of the est S/N ratio for each given parameter was determined as the best
first solution found by algorithm, and to decrease at each iteration level for that parameter. The results of parameter adjustment pro-
by h  (0,1). In this paper h is assumed to be 0.999. cedure carried out for the discussed meta-heuristic algorithms are
shown in Table 2.
4.3.1.7. Weight updating. At the end of each iteration, the used de- In this table, Iterationmax and shakeiter are the maximum num-
stroy and repair operators will be given a score. When an operator ber of iterations in total and in each neighborhood structure of
improves the best solution, improves the incumbent solution, or VNS algorithm; max iter1 and max iter2 are the maximum number
leads to a worse solution, its score increases respectively by σ 1 , σ 2 of iterations in the main loop and inner loop in the ALNS algorithm
and σ 3 . These weights are used in Eq. (22) tasked with updating and S is the number of points (a product of a customer) selected
the weight of operators. In this paper, these weights are selected in ALNS algorithm.
to be σ1 = 20, σ2 = 10, σ3 = 2,
5. Numerical experiments
wnew
j = (1 − λ ) × wcurrent
j + λ × π j /N u j . (22)
To evaluate the performance of the proposed algorithm, 18
In this equation, Nuj is the number of uses of operator j, π j is randomly generated small-size problem instances is solved, and
the current score of operator j, wcurrent
j
is its current weight, and its solutions are compared with the results of exact solution ob-
wnew
j
is the new weight of this operator. The parameter λ in this tained by Gams software and the results of ALNS and VNS algo-
equation determines how static the weights are. When λ is close rithms. To generate large-size problem instances, 2 groups of prob-
to zero weights will be more influenced by their previous values, lems including 33 large-size multi-depot VRP instances provided
and when λ is close to 1 weights will be updated according to the by Cordeau1 are modified to fit the problem description discussed
latest score. In this paper, λ is assumed to be 0.1. in this paper. It is worth mentioning that the number of vehicles’
compartments was randomly assigned with the numbers of 2 or 3;
4.3.1.8. Termination criteria. Algorithm stops after iteration counter so demands were randomly divided into 2 or 3 classes and capac-
of the main loop reaches the predetermined value of max iter1 . ity of each vehicle was divided into 2 or 3 equal compartments
respectively. It should be noted that all algorithms were coded
4.4. Parameter adjustment in MATLAB 2013 environment and were run on a computer with
2.2 GHz Core i7 CPU, 6 GB of RAM, and Windows 8 (64-bit) operat-
Parameter values of meta-heuristic algorithms affect their per- ing system. It is worth noting that each algorithm is executed for
formance, so determination of the most suitable values of these five times, and the best results are reported. Also the error per-
parameters can contribute greatly to the improvement of final re- centages are also calculated according to the best answer. More-
sults. Taguchi analysis is a statistical method widely used for pa- over, the reported times in the tables are the mean times to solve
rameter adjustment. The aim of this method is to adjust the pa- the algorithms in 5 different executions.
rameters such that algorithm efficiency would be maximized and
5.1. Algorithms performance for small size problem instances
algorithm stability would be maintained. Instead of comparing the
solutions themselves, Taguchi method evaluates a function of solu-
The results obtained by solving the small size problem in-
tions called signal to noise ratio or S/N. In this ratio, S represents
stances are presented in Table 3. The times shown in the table are
the desirability and N represents the degree of undesirability. So
in seconds and the errors are expressed as a percentage. These er-
the ultimate goal is to maximize the value of this ratio [35]. In
ror percentages were obtained using Eq. (23),
Taguchi method, objective function is divided into three categories.
In the first category, which is called “larger-the-better”, desirability AS − BS
Gap% = × 100. (23)
increases with the increase in the value of objective function. In BS
the second category, which is called “smaller-the-better”, desirabil-
ity increases with the decrease in the value of objective function. 1
http://neo.lcc.uma.es/vrp/vrp- instances/multiple- depot- vrp- instances/.
M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99 95

Table 3
Results in small size instances.

# D/N/G/Q exact method VNS ALNS HALNS

Obj Time K Obj Time Gap% Obj Time Gap% Obj Time Gap%

P01 2/6/2/100 398 0.1 2 398 0.5 0 398 0.6 0 398 1 0


P02 2/7/2/100 349 0.15 2 349 1 0 349 0.6 0 349 1 0
P03 3/7/2/70 448 1 3 448 1 0 448 0.7 0 448 1.2 0
P04 2/8/2/70 449 3 4 449 1 0 449 1 0 449 1 0
P05 2/9/2/100 543 16 3 543 2 0 543 1 0 543 2 0
P06 3/9/2/90 536 53 3 536 2 0 536 1 0 536 1 0
P07 3/10/2/90 620 260 4 620 6 0 620 4 0 620 5 0
P08 3/11/2/90 619 728 5 623 7 0.6 629 6 1.6 619 5.5 0
P09 3/12/2/90 684 1025 5 690 20 0.9 688 14 0.6 684 16 0
P10 2/6/3/100 535 8 4 535 0.9 0 535 0.7 0 535 2 0
P11 2/7/3/100 451 5 4 451 1 0 451 0.7 0 451 3.2 0
P12 2/8/3/70 563 21 4 563 1 0 563 0.7 0 563 1.7 0
P13 3/7/3/70 459 5 4 459 8 0 459 3 0 459 4 0
P14 3/8/3/100 557 212 4 557 24 0 557 10 0 557 11.3 0
P15 3/9/3/90 544 381 5 544 20 0 544 12 0 544 13 0
P16 3/10/3/90 654 1011 5 654 50 0 659 22 0.8 654 20 0
P17 3/11/3/90 625 3601 5 632 51 1.1 625 30 0 625 33 0
P18 3/12/3/100 710 3700 5 720 50 1.4 717 30 1 710 29 0
Average 541.33 612.79 – 542.83 13.69 0.22 542.78 7.67 0.22 541.33 8.38 0

Fig. 7. The processing time in small size problems.

In this equation, AS and BS are the algorithm solution and the In Fig. 7, the vertical axes on the left are related to the pro-
best found solution, respectively. cessing time of the meta-heuristic algorithms and the vertical axes
In Table 3, the first column shows the number of the prob- on right is related to the processing time of the exact method. As
lem and the second column demonstrates the characteristics of the it can be deduced from Fig. 7, the processing times of ALNS and
problems in which, D, N, G and Q are number of depots, number HALNS are close to each other. In addition, the processing time to
of customers, number of product types and each compartment’s solve the problems with exact method is exponentially increasing.
capacity of the vehicles respectively. In this table, Obj, Time, K and Meanwhile, the mean solution time of exact method was almost
Gap% are objective function, processing time, number of vehicles 70 times greater than that of the proposed hybrid algorithm, also
and percent of errors, respectively. Due to the similar value ob- solution time of VNS algorithm was two times greater than the so-
tained by all proposed algorithms, for the number of vehicles for lution time of other two algorithms. Fig. 8 shows the answers ob-
all small-sized problems, this criterion is only reported for exact tained by proposed solution methods for one small size problem
method. As Table 3 shows, these three meta-heuristic algorithms instance (P08). In this example the number of depots, customers
yielded high quality solutions. The average errors of VNS, ALNS, and products are 3, 11, and 2, respectively, also Qg is considered
and hybrid ALNS algorithms were 0.22%, 0.22% and 0%, respec- equal to 90 for both products. Moreover, MC is considered equal
tively, and the maximum error, 1.6%, was resulted by the use of to “400”. Other information of the problem instance is shown in
ALNS algorithm. In addition, VNS and ALNS have reached the op- Table 5 in the appendix.
timal solution in 14 problems out of 18 while HALNS has reached
the optimal solution in all the sample problems. In the small test 5.2. Algorithms performance for large size problem instances
problems HALNS algorithm performs rather better than the other
two algorithms with regard to average standard deviation. Average To evaluate the performance of the proposed meta-heuristic al-
standard deviations (in 5 executions) in this set of test problems gorithms for large-scale problems, 33 large-size problem instances
were 1.79, 1.81, and 0.83 for algorithms VNS, ALNS, and HALNS, are solved using three meta-heuristic algorithms. The results ob-
respectively. All three meta-heuristic algorithms exhibited a good tained by solving these problems are presented in Table 4. In
solution time. Fig. 7 demonstrates the processing time of each al- Table 4, the first column shows the number of the problem and the
gorithm in small size problems. second column demonstrates the characteristics of the problems in
96 M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99

Fig. 8. The answers obtained by proposed solution methods in one test problem.
Table 4
Results in large size instances.

# D/N/G/Q VNS ALNS HALNS

Obj Time Gap% K Obj Time Gap% K Obj Time Gap% K

P01 4/50/3/30 716 80 0 12 716 55 0 12 716 60 0 12


P02 4/50/3/60 530 112 0 6 535 90 0.94 6 530 98 0 6
P03 5/75/2/70 666 250 0 12 666 68 0 12 666 73 0 12
P04 2/100/2/50 1053 245 1.96 15 1048 76 1.47 15 1033 70 0 15
P05 2/100/2/100 786 236 0 8 786 71 0 8 786 81 0 8
P06 3/100/2/50 907 209 0 16 907 53 0 16 919 66 1.34 16
P07 4/100/2/50 931 312 1.32 26 933 79 1.55 26 919 63 0 26
P08 2/249/3/170 4708 550 0 27 4783 356 1.59 27 4756 380 1.02 27
P09 3/249/3/170 4423 422 2.57 27 4360 320 1.11 27 4312 392 0 27
P10 4/249/3/170 3998 635 2.51 27 3909 423 0.23 27 3900 450 0 27
P11 5/249/2/250 3708 496 3.18 27 3612 229 0.5 27 3594 239 0 27
P12 2/80/2/30 1385 115 2.49 9 1351 37 0 9 1362 49 0.82 9
P13 2/80/2/30 1411 121 4.33 10 1420 67 5 10 1352 77 0 10
P14 2/80/2/30 1474 123 5.95 10 1426 57 2.54 10 1391 64 0 10
P15 4/160/2/80 2627 189 3.49 18 2571 110 1.31 17 2538 120 0 17
P16 4/160/3/70 2914 377 3.59 18 2906 333 3.31 18 2813 360 0 18
P17 4/160/3/60 2936 217 1.59 18 2963 187 2.53 18 2890 203 0 18
P18 6/240/2/30 3868 368 3.27 24 3854 203 2.89 24 3746 221 0 24
P19 6/240/2/100 4009 356 3.89 25 3933 158 1.92 25 3859 190 0 25
P20 6/240/2/90 4155 491 1.59 25 4245 215 3.8 26 4090 223 0 25
P21 9/360/2/30 5702 479 3.61 35 5643 253 2.55 35 5503 288 0 35
P22 9/360/2/75 5954 501 3.97 38 5934 276 3.62 37 5727 291 0 37
P23 9/360/2/90 6312 516 3.11 37 6397 263 4.5 38 6122 319 0 37
Pr01 4/48/3/120 915 241 0 4 915 220 0 4 915 225 0 4
Pr02 4/96/3/150 1558 80 3.59 8 1504 85 0 8 1504 78 0 8
Pr03 4/144/2/220 1902 199 3.74 12 1898 75 3.52 12 1833 89 0 12
Pr04 4/192/2/230 2229 225 5.93 17 2126 83 1.05 16 2104 91 0 16
Pr05 4/240/2/230 2476 352 2.63 20 2515 130 4.21 20 2413 114 0 20
Pr06 4/288/2/200 2760 478 0 24 2812 152 1.89 24 2839 140 2.85 24
Pr07 6/72/2/100 1153 115 3.35 6 1171 83 4.9 6 1116 95 0 6
Pr08 6/144/2/235 1720 286 1.31 12 1709 120 0.65 12 1698 130 0 12
Pr09 6/216/2/250 2236 360 2.16 18 2307 180 5.38 18 2189 273 0 18
Pr10 6/249/3/170 3240 736 4.18 24 3221 106 3.57 24 3110 120 0 24
Average 2586.7 317.3 2.4 – 2578.1 157.9 2.02 – 2522.6 173.7 0.18 –

which, D, N, G and Q are number of depots, number of customers, instances respectively and only in a few number of instances they
number of product types and capacity of each compartment of the have reached in different number of vehicles, the instances with
vehicles, respectively. In this table, Obj, Time, Gap% and K are ob- different number of required vehicles are bolded in Table 4. Note
jective function, processing time, percent of errors, and number of that: Instances P01-P07 and P12, P15, P18, P21 do not have any
vehicles, respectively. route length restriction. Fig. 9 shows the error of algorithms in the
As seen in Table 4, for large-scale problems, the hybrid meta- large-scale problem instances.
heuristic algorithm outperforms the other assessed algorithms. In The hybrid meta-heuristic algorithm obtained the best solution
terms of accuracy, the hybrid meta-heuristic algorithm has a 0.18% for 29 of the 33 large-scale problem instances, while VNS and
error, while errors of VNS and ALNS are 2.4% and 2.02%, respec- ALNS algorithms did that for 9 and 8 problems, respectively. The
tively. In all of the large-scale problem instances, the hybrid meta- maximum error of hybrid algorithm was 2.85% while the maxi-
heuristic algorithm obtained the lowest number of required vehi- mum error of VNS and ALNS algorithms were 5.93% and 5.38%,
cles, while VNS and ALNS algorithms did so in 30 and 31 problem respectively. ALNS has reached a better solution in compare with
M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99 97

Fig. 9. The error in solutions of algorithms for large-scale problem instances.

Fig. 10. The processing time of each algorithm when solving problems with large size.

the hybrid algorithm for 2 sample problems out of 32. The rea- the customers’ employees, when the number of customer’s visits
son behind the decreasing performance of the hybrid algorithm is increases, the customer satisfaction decreases. For instance, sup-
the random generation of the initial solution and solution improve- pose there are 10 different products; then if a customer receives
ment. Average standard deviation of HALNS and VNS algorithms in different products in 10 different visits, his satisfaction will be de-
the large size problems were almost similar and equal to 8.3 and creased to a great extent. On the other hand, if splitting various
8.04, respectively. However, the standard deviation mean of ALNS products of a customer is not possible, then each customer is vis-
algorithm was slightly greater than the other two algorithms, being ited only once. However, this causes the total traveled distance to
equal to 12.7. The obtained results from standard deviation of the be increased as compared to the previous situation, or the num-
algorithms indicate stability of the proposed algorithms in finding ber of required vehicles increases. Fig. 11a indicates the possibility
the answers. Fig. 10 demonstrates the processing time of each al- for a customer to be visited only once. Fig. 11b shows the pos-
gorithm when solving problems with large size. sibility of splitting the products. In this example, the number of
As it can be easily deduced from Fig. 10, the processing time depots, customers, products,Qg and MC is 3, 9, 3, 10 0 and 40 0,
of VNS is more than ALNS and HALNS for all the sample prob- respectively.
lems. The processing time for VNS, ALNS and HALNS in the av- As can be seen in Fig. 11, in situation “a” where there is no
erage is 317,158 and 174 s, respectively. Solution time of VNS algo- possibility for splitting the products, the total traveled distance is
rithm is about twice that of hybrid meta-heuristic algorithm, but equal to 407. In situation “b” where there is the possibility for
hybrid meta-heuristic algorithm exhibits slightly shorter solution splitting the products, the total traveled distance has become 395.
time than the ALNS algorithm. Result demonstrates that the pro- In this case, customers 4 and 7 are visited twice, and the other
posed algorithm is competitive in compare with VNS and ALNS. customers are visited only once. Considering customers’ satisfac-
tion as one objective function along with other objective functions,
5.3. Effect of the number of customers’ visits on the customers’ can be an interesting subject for future studies.
satisfaction
6. Conclusion
It has been assumed for the considered case in this study that
splitting the demand of a customer for a product is not possible, This paper presented a mathematical formulation for multi-
but it is possible that various products required by a customer can depot multi-compartment vehicle routing problem. The objective
be supplied by different vehicles. Since receiving the products is function of the proposed problem includes the minimization of the
costly for the customers, requiring relevant coordination between number of the vehicles and then minimization of the total traveled
98 M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99

Fig. 11. Effect of the number of customers’ visits on the customers’ satisfaction.

distance. In this problem, the cargo space of each vehicle has mul- Table 5
Characteristics of small test problem P08.
tiple compartments, and each compartment is dedicated to a sin-
gle type of product. In the proposed formulation, split delivery for # x y di1 di2
one product is not allowed so demand of each customer for a cer- 1 90 52 0 0
tain product must be fully delivered by a single vehicle; however, 2 10 78 0 0
split delivery for multiple products is allowed and customer can 3 61 91 0 0
receive different products by different vehicles. The application of 4 60 13 10 30
5 39 65 30 30
the problem is the distribution of the frozen products, fuel distri-
6 26 84 30 35
bution, milk and oil and etc., while the distribution company has 7 35 15 45 65
more than one depot. Integrated and simultaneous optimization of 8 88 92 35 25
the different depots’ routes can properly decrease the distribution 9 10 20 30 45
10 60 50 50 30
cost instead of separate optimization of the routes for each depot.
11 80 25 45 25
Considering the NP-Hardness of the proposed problem, a hybrid al- 12 40 40 30 20
gorithm composed of adaptive large neighborhood search and vari- 13 95 5 40 50
able neighborhood search was developed. The proposed hybrid al- 14 70 70 20 20
gorithm reached to optimal solution in all small size problem in-
stances. In sample problems with small size, the average process-
ing time for exact method, VNS, ALNS and HALNS is equal to 613, References
14, 8 and 8, respectively. Solving large size problem instances, the
hybrid algorithm achieved to 0.18% error in average which was [1] Renaud J, Laporte G, Boctor FF. A tabu search heuristic for the multi-depot
vehicle routing problem. Comput Oper Res 1996;23:229–35.
significantly lower than errors of VNS and ALNS algorithms. For
[2] Lahyani R, Coelho LC, Khemakhem M, Laporte G, Semet F. A multi-compart-
sample problems with large size, the processing time for VNS, ment vehicle routing problem arising in the collection of olive oil in Tunisia.
ALNS and HALNS is 317, 158 and 174 in average. The assumption Omega 2015;51:1–10.
[3] Sethanan K, Pitakaso R. Differential evolution algorithms for scheduling raw
that each depot can only provide a certain number of products
milk transportation. Comput Electron Agric 2016;121:245–59.
and solve the problem with exact methods or meta-heuristic al- [4] Avella P, Boccia M, Sforza A. Solving a fuel delivery problem by heuristic and
gorithms is an interesting suggestion for future research. In addi- exact approaches. Eur J Oper Res 2004;152:170–9.
tion, investigation of the problem when the transportation fleets [5] Derigs U, Gottlieb J, Kalkoff J, Piesche M, Rothlauf F, Vogel U. Vehicle rout-
ing with compartments: applications, modelling and heuristics. OR Spectr
are heterogeneous can also be an interesting topic. Minimizing in- 2011;33:885–914.
ventory cost while having multiple periods is also another inter- [6] Clarke G, Wright JW. Scheduling of vehicles from a central depot to a number
esting suggestion for future research. of delivery points. Oper Res 1964;12:568–81.
[7] Cornillier F, Boctor FF, Laporte G, Renaud J. A heuristic for the multi-period
petrol station replenishment problem. Eur J Oper Res 2008;191:295–305.
[8] Cornillier F, Laporte G, Boctor FF, Renaud J. The petrol station replenishment
Appendix
problem with time windows. Comput Oper Res 2009;36:919–35.
[9] Fallahi AE, Prins C, Calvo RW. A memetic algorithm and a tabu search
Table 5 shows characteristics of small test problem P08 where for the multi-compartment vehicle routing problem. Comput Oper Res
the first column indicates the points’ number, column 2 and col- 2008;35:1725–41.
[10] Christofides N, Mingozzi A, Toth P. Exact algorithms for the vehicle routing
umn 3 show the coordinates of the points and the last 2 columns problem, based on spanning tree and shortest path relaxations. Math Program
are related to the demands in different points. 1981;20:255–82.
M. Alinaghian, N. Shokouhi / Omega 76 (2018) 85–99 99

[11] Eilon S. Distribution management: mathematical modelling and practical anal- [23] Salhi S, Imran A, Wassan NA. The multi-depot vehicle routing problem with
ysis. Griffin; 1970. heterogeneous vehicle fleet: formulation and a variable neighborhood search
[12] Mendoza JE, Castanier B, Guéret C, Medaglia AL, Velasco N. Constructive implementation. Comput Oper Res 2014;52(Part B):315–25.
heuristics for the multicompartment vehicle routing problem with stochastic [24] Allahyari S, Salari M, Vigo D. A hybrid metaheuristic algorithm for the multi-
demands. Transp. Sci. 2011;45:346–63. -depot covering tour vehicle routing problem. Eur J Oper Res 2015;242:756–68.
[13] Muyldermans L, Pang G. On the benefits of co-collection: Experiments [25] Dror M, Trudeau P. Savings by split delivery routing. Transp Sci
with a multi-compartment vehicle routing algorithm. Eur J Oper Res 1989;23(2):141–5.
2010;206:93–103. [26] Liu R, Jiang Z, Fung RYK, Chen F, Liu X. Two-phase heuristic algorithms for full
[14] Hanczar P. A fuel distribution problem – application of new multi-item inven- truckloads multi-depot capacitated vehicle routing problem in carrier collabo-
tory routing formulation. Proc - Soc Behav Sci 2012;54:726–35. ration. Comput Oper Res 2010;37:950–9.
[15] Popović D, Bjelić N, Radivojević G. Simulation approach to analyse determin- [27] Gulczynski D, Golden B, Wasil E. The multi-depot split delivery vehicle rout-
istic IRP solution of the stochastic fuel delivery problem. Proc - Soc Behav Sci ing problem: an integer programming-based heuristic, new test problems, and
2011;20:273–82. computational results. Comput Ind Eng 2011;61:794–804.
[16] Popović D, Vidović M, Radivojević G. Variable neighborhood search heuris- [28] Lalla-Ruiz E, Expósito-Izquierdo C, Taheripour S, Voß S. An improved for-
tic for the inventory routing problem in fuel delivery. Expert Syst Appl mulation for the multi-depot open vehicle routing problem. OR Spectr
2012;39:13390–8. 2016;38:175–87.
[17] Vidović M, Popović D, Ratković B. Mixed integer and heuristics model for [29] Rahimi-Vahed A, Gabriel Crainic T, Gendreau M, Rei W. Fleet-sizing for multi-
the inventory routing problem in fuel delivery. Int J Prod Econ 2014;147(Part -depot and periodic vehicle routing problems using a modular heuristic algo-
C):593–604. rithm. Comput Oper Res 2015;53:9–23.
[18] Abdulkader MMS, Gajpal Y, ElMekkawy TY. Hybridized ant colony algo- [30] Kunnapapdeelert S, Kachitvichyanukul V. Modified DE algorithms for solv-
rithm for the multi compartment vehicle routing problem. Appl Soft Comput ing multi-depot vehicle routing problem with multiple pickup and delivery
2015;37:196–203. requests. In: Kachitvichyanukul V, Sethanan K, Golinska- Dawson P, editors.
[19] Kaabi H. Hybrid metaheuristic to solve the selective multi-compartment vehi- Toward sustainable operations of supply chain and logistics systems. Cham:
cle routing problem with time windows. In: Abraham A, Wegrzyn-Wolska K, Springer International Publishing; 2015. p. 361–73.
Hassanien EA, Snasel V, Alimi MA, editors. Proceedings of the second inter- [31] Mladenovi N, #263, Hansen P. Variable neighborhood search. Comput Oper Res
national Afro-European conference for industrial advancement, AECIA 2015. 1997;24:1097–100.
Cham: Springer International Publishing; 2016. p. 185–94. [32] Fleszar K, Osman IH, Hindi KS. A variable neighbourhood search algorithm for
[20] Jin T, Guo S, Wang F, Lim A. One-stage search for multi-depot vehicle routing the open vehicle routing problem. Eur J Oper Res 2009;195:803–9.
problem. In: Proceedings of the information systems and control conference; [33] Shaw P. Using constraint programming and local search methods to solve ve-
2004 446-129. hicle routing problems. In: Proceedings of the 4th international conference on
[21] Tsirimpas P, Tatarakis A, Minis I, Kyriakidis EG. Single vehicle routing with principles and practice of constraint programming; 1998.
a predefined customer sequence and multiple depot returns. Eur J Oper Res [34] Ropke S, Pisinger D. An adaptive large neighborhood search heuristic for the
2008;187:483–95. pickup and delivery problem with time windows. Transp Sci 2006;40:455–72.
[22] Ray S, Soeanu A, Berger J, Debbabi M. The multi-depot split-delivery ve- [35] Schrimpf G, Schneider J, Stamm-Wilbrandt H, Dueck G. Record breaking
hicle routing problem: model and solution algorithm. Knowl-Based Syst optimization results using the ruin and recreate principle. J Comput Phys
2014;71:238–65. 20 0 0;159:139–71.
[36] Demir E, Bektaş T, Laporte G. An adaptive large neighborhood search heuristic
for the pollution-routing problem. Eur J Oper Res 2012;223:346–59.

Potrebbero piacerti anche