Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F.

Yu

The Application of Metaheuristics to Solving Location Routing Problem

with Time Windows
Meilinda F.N. Maghfiroh1, A.A.N. Perwira Redi2 and Vincent F. Yu3

Department of Industrial Management,

National Taiwan University of Science and Technology, Taiwan
[email protected]
[email protected]
[email protected]
Corresponding Author: [email protected]

Abstract
In most distribution systems, depot location and vehicle routing decisions are
implemented independently. However, in location-routing problems (LRP), depot location
and vehicle routing problems are solved simultaneously. Here, our goal is to combine the
capacitated facility location problem (CFLP) and the vehicle routing problem with time
windows (VRPTW) into LRP with time windows (LRPTW), which considers more realistic
aspects, especially the time aspects, of many real problems. In order to efficiently solve the
LRPTW for small, medium and large-scale instances, variable neighborhood search (VNS)
and simulated annealing (SA) are proposed. The proposed metaheuristic approaches can
solve LRP and LRPTW with nearly optimal solution.

Keywords: Location routing problem, time windows, metaheuristic, variable neighborhood

search, simulated annealing

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

1. Introduction
Nowadays, chain stores and modern retailers have become the solutions to providing
daily goods in urban area. The number of chain stores and modern retailers is increasing to
cover the growing demand in urban areas. Thus, the need of distribution centers is also
increasing. The location of each DC varies. To minimize the opening cost and operational
cost, management tends to locate DCs far from urban area. As a consequence, distribution
activities of goods also grow enormously. In fact, location planning for urban distribution
centers is a vital consideration in lowering distribution costs and minimizing traffic
congestion arising from goods movement in urban area. It is surely a complex decision which
involves multiple considerations, such as minimizing total cost (opening cost, distribution
cost), maximizing costumers’ coverage, and minimizing freight movements in urban area.
The needs of having low logistic costs represent a large portion of the company
expenses (Belenguer et al., 2011). Therefore, designing distribution system becomes one of
the major issues for many industries. Significant productivity of the DC can be achieved
through the design of location routing models as this model can determine true least-cost
solutions to a logistic problem taking into both strategic policy (facility location) and
operational decisions (vehicle routing) (Yu et al., 2010).
Such situation can be modeled as the location routing problem (LRP). Considering that
in this delivery system each chain store has to be serviced within a specified time frame, this
research adds additional time constraint, i.e. time windows, to the LRP. The goal of the
resulting location routing problem with time windows (LRPTW) is to determine the depots to
be opened, the customers and the vehicles to be assigned to each open depot and the routes to
be constructed to fulfill the demand and the time windows of the customers at a minimum
global cost.

2. Literature Review
Location routing problem belongs to the class of NP-hard problems since it combines
two difficult sub-problems: the facility location problem (FLP) and the vehicle routing
problem (VRP), where both of them are shown to be NP-hard (Lenstra & Rinnooy, 1981).
Laporte and Nobert proposed an exact algorithm for the single facility fixed fleet size LRP
without tour length restrictions. Due to the exponential growth in the problem size, exact
approaches for the LRP have been limited to small and medium size instances with 20–50
customers (Laporte & Nobert, 1981). For this reason, heuristics and meta-heuristics are often
used to solve realistic sized LRP instances in recent studies.
Tuzun and Burke (1999) developed a two-phase tabu search (TS), but for the LRP with
capacitated routes and uncapacitated depots so a depot may have as many routes as desired.
The two phases of their TS algorithm are also dedicated to routing and location. A genetic
algorithm was applied to the design of a physical distribution system where both the location

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

of facilities and the routing of vehicles were considered by Su (1999). Barreto (2004)
developed a class of three-phase heuristics based on clustering techniques. In the first phase,
customers were aggregated into clusters fitting vehicle capacity. The second phase solved a
traveling salesman problem (TSP) for each cluster. Finally in the third phase, the depots to be
opened were determined by solving an FLP, where the TSP cycles were combined to form
super nodes. Prins et al. (1999) proposed a hybrid method to solved LRP with capacitated
depots and routes by combining greedy randomized adaptive search procedure (GRASP) with
a learning process and a path relinking mechanism. SALRP heuristic for the LRP based on
the popular simulated annealing heuristic is introduced by Yu et al. (2010). The proposed
SALRP heuristic is tested on three sets of well-known benchmark instances and the results
are compared with other heuristics in the literature. The computational study indicates that
the proposed SALRP heuristic is competitive with other well-known algorithms.

3. Problem Definition
3.1 Problem Descriptions
A feasible set of DC locations and demand of chain stores are given. Vehicles start from
a DC, deliver goods to a subset of the chain stores and then return to the same DC. The
problem determines the location of the DCs and vehicle routes to minimize total cost
associated with establishing the DC and distribution the goods. Some practical constrains that
also have to be satisfied are:
a. Each vehicle route must start from at DC and ends at the DC.
b. Each customer belongs to exactly one route.
c. The total demand of each route cannot exceed the capacity of the vehicle.
d. Each vehicle route must satisfy the route duration and time windows constraints.
This research focuses on solving the LRP with time windows and considers both
capacitated routes and capacitated depots. The following formal mathematical model for the
problem is modified from the models of Prins et al. (1999) and Tan et al. (2001). The
LRPTW can be stated as follows: there is a graph G = (V, E) where V and E represent
vertices and edges of the graph. V is a set of nodes consist of a subset I of m potential depot
sites and a subset J = V\I of n customers. E is a set of edges connecting each pair of nodes in
V. Associated with each edge (i, j)  E is a traveling cost cij. Each depot site i  I has a
capacity Wi and opening cost Oi. Each customer j  J has a demand dj which must be fulfilled
by a single vehicle. A set K of vehicles with capacity Q is available. Each vehicle used by
depot i incurs a depot dependent fixed cost Fi and performs a single route.
The customers and depots have time windows. The time window of a site, i, is specified
by an interval [ei, li] where ei and li represent the earliest and the latest arrival time,
respectively. All vehicles must arrive at a site before the end of the time windows li. The
notation of e0 represents the time that all vehicles in the routing plan leave the depot, while l0

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

corresponds to the time that all vehicles must return to the depot. In fact, the interval [e0, l0] is
the largest time window for which all customers’ time windows must be within the range. The
travel time between edges (i, j)  E is denoted by tij. aik is the arrival time of vehicle k at
customer i and pik is the departure time at customer i. Each vehicle k has an identical duration
Bk equal to the length of a driver’s working hours. The service time of customer i, δi, is a
fixed number indicating the time needed to service customer i. The waiting time wi for
customer i is incurred if a vehicle arrives before the start of customer i’s service time window.
If the vehicle arrives at customer i within its time windows, wi = 0.
3.2 Mathematical Model
Objective:
min z  Oi yi  cij xijk  Fk xijk (1)
iI iV jV kK kK iI jJ

Subject to:

x
kK iV
ijk  1j  J (2)

x
jJ iV
ijk d j  Qk  K (3)

d f
jJ
i ij  Wi yi i  I (4)

x
jV
ijk  x jik  0i V , k  K
jV
(5)

x
iI jJ
ijk  1k  K (6)

x
iS jS
ijk  S  1S  J , k  K (7)

x
uJ
iuk  x
uV { j }
ujk  1  fij i  I , j  J , k  K (8)

 t
iV jV
ij xijk  ( i  wi )xijk  Bk k  K
iV jV
(9)

a0  w0   0 (10)
 
x
kK iV
 aik  t ij   i  wi   a jk j  J

ijk

(11)

ei  aik  lii V , k  K (12)

Define binary variables yi = 0 iff depot i is opened, fij = 1 iff customer j is assigned to
depot i, and xjlk = 1 iff edge (j, l) is traversed from j to l in the route perform by vehicle k Є K.
Then the problem can be formulated as the following binary integer program.

xijk 0,1 , i  I , j  J , k  K (13)

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

yi 0,1 , i  I (14)

fij 0,1 , i  I , j V (15)

The objective function (1) of this problem is the total cost for both depot opening cost
and distribution cost. Constraint (2) ensures that each customer belongs to exactly one route.
Constraints (3) and (4) are capacity constraints for the route and depot while constraints (5)
and (6) guarantee that the route starts and ends in the same depot. Constraint (7) is sub-tour
elimination constraint. Constraint (8) ensure a customer to be assigned to a depot if there is a
route that connecting them. Next constraints are for explaining time windows constraints.
Constraint (9) is the maximum travel time constraint and (10-12) are the time windows
constraints. Finally, constraints (13-15) specify the binary variables used in the formulation.
The time windows constraint is treated as hard time windows, where it cannot be
violated. But during the computational procedure, it is allowed to violate the constraints.
Every time the solution exceed or violate the time constraint, it will get a penalty. In the end,
the best solution that was chosen must be the feasible one, which the sum of the penalty is
equal zero. The penalty can be calculated using formulation below:
 
P
kK
k  a jk  xijk  aik  t ij   i  wi  j  J
kK iV  
(16)

4. Proposed Methods
4.1 Solution Representation
The solution is represented by the permutation of n customers (1, 2,…, n), m potential
depot (n+1, n+2,…, n+m), and Ndummy zeros. The Ndummy zeros are used to randomly
terminated the route in order to better explore the solution space. Routes can also be
 d 
terminated based on capacity constraints. The number of Ndummy is calculate as   i  ,
 i Q
where di is the demand of customer i and Q as the capacity of the vehicle.
The first number occur in the solution is always a depot, i.e. in {n+1, n+2,…, n+m}. It
will then be followed by a number in {1, 2,…, n} which represents the 1st customer to be
served. Each depot will serve the customer between the depot and the next depot based on
the solution representation. However for both constraints, capacity and time, the violation is
allowed to get a better solution during the computational study. However, the final solution
should be feasible.
4.2 Initial Solution
The initial solution is constructed randomly. The first position of the solution is the first
selected open depot followed by the customers on the first route, customers on the second
route and so forth. After all the customers for the first depot are assigned, then the second

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

depot will be selected and added in to the solution representation. This whole process is
continued until all customers are assigned to a depot. Then, Ndummy zeros are randomly added
to the solution. Finally, the closed depot is appended to the solution. Figure 1 illustrates a
randomly generated initial solution.

14 13 7 3 2 1 6 9 8 10 12 11 5 4

Figure 1. Solution Representation

In the example above, the first depot is 14. Since the next bit is 13 which is also a depot,
depot 14 is closed. Depot 13 will open and serves customers 7, 3, 2 and 1 respectively. Each
route can exceed the time constraint only in the temporary solution. The final solution must
not exceed the capacity and time constraint.
4.3 Simulated Annealing for LRPTW
Simulated annealing (SA) is a random-search technique which exploits an analogy
between the way in which a metal cools and freezes into a minimum energy crystalline
structure (the annealing process) and the search for a minimum in a more general system; it
forms the basis of an optimization technique for combinatorial and other problems. SA's
major advantage over other methods is an ability to avoid becoming trapped in local minima.
The algorithm employs a random search which not only accepts changes that decrease the
objective function f (assuming a minimization problem), but also some changes that increase
it. The latter are accepted with a probability p = exp (-δf / T), where δf is the increase in f and
T is a control parameter, which by analogy with the original application is known as the
system ''temperature" irrespective of the objective function involved.
To improve the solution, the neighborhood operator that performs in SA Algorithm are
swap, insert and reverse. In the beginning, the current temperature T is set to be the same as
T0. Then an initial solution X is generated using the random generator, in which the first
number in the solution representation must be a depot. The current best solution Xbest and the
best objective function value obtained so far are set to be X and obj (X, P), respectively. For
each iteration, the next solution Y is generated from N(X) and its objective function value is
evaluated. Let Δ denote the difference between obj (X, P) and obj (Y, P), that is Δ = obj (Y, P)
– obj (X, P). The probability of replacing X with Y, given that Δ > 0, is exp (-Δ/KT). This is
accomplished by generating a random number r ∈ [0, 1] and replacing the solution X with Y
if r < exp (-Δ/KT). Meanwhile, if Δ≤0, the probability of replacing X with Y is 1. Xbest records
the best solution found so far as the algorithm progresses.
The current temperature T is decreased after running Iiter iterations since the previous
decrease, according to the formula T←αT, where 0<α<1. After each temperature reduction, a
local search procedure which sequentially performs swap and insertion is used to improve the
current best solution. The algorithm is terminated when the current temperature T is lower

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

than TF or the current best solution Xbest is not improved in Nnon-improving consecutive
temperature reductions. Following the termination of the SALRPTW procedure, the facility
locations and vehicle routes can be derived from Xbest.
4.4 Variable Neighborhood Search for LRPTW
The systematic change of the neighborhood within the local search to evade local optima
traps is suggested by Mladenović & Hansen (1997). This leads to a metaheuristic called
variable neighborhood search (VNS), which has been further developed into various
extensions. The modified VNS, variable neighborhood descent (VND), is used by Rousseau
et al. (2002) to take advantage of different neighborhood structures for the vehicle routing
problem. Polacek et al. (2004) use a VNS for the multiple depot vehicle routing problem with
time windows. Several variants of location problems are also solved efficiently with the VNS
method such as uncapacitated p-median problem (Mladenović & Hansen, 1997) and
capacitated p-median problem (Fleszar & Hindi, 2008).
Consider a combinatorial problem to minimize a function f defined on a solution space
X. To each solution x Є X, is associated a subset N(x) ⊆ X called neighborhood of x. Let us
denote with Nk (K=1,2,….,kmax) a finite set of pre-selected neighborhood structures, and with
Nk(x) the set of solutions in the kth neighborhood of x. In general, VNS starts with an
initial solution x and a set of neighborhoods Nk, k = 1, 2, ...,kmax. At each iteration, a random
solution x’ is computed with respect to the kth neighborhood, Nk(x). Then, a local search is
applied to the solution x’ to yield a second solution x”. If solution x” is better than x, the
solution is updated and the process continues with the first neighborhood N1(x), otherwise the
same steps are repeated with the next neighborhood, Nk+1.
The final solution found will be a local optimum with respect to all neighborhood
structures. The VNS combines three basic steps: a stochastic phase, which is the shaking step
that finds a random neighbor of the incumbent solution, a deterministic phase which
represents the application of any local search algorithm, and the evaluation step which
accepts only solutions that improve the objective (sometimes the fitness) function. In fact, the
VNS can be seen as a variable depth search if we make a move to the best neighborhood
among the predefined ones. The different steps of VNS are presented in algorithm 1 below.

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

Algorithm1. VNS for LRPTW

Input: The set of neighborhood structures Nk, for k = 1, 2, . . . , kmax
1. Initialization: Find an initial solution x;
2. Repeat
3. k1
4. while k≤kmax do
5. x’ Shaking in Nk(x)
6. x”Local Search using VND (x’)
7. if f(x”) < f (x) then
8. xx”; k1;
9. else
10. kk+1
11. until termination condition is met
12. return x

The VNS algorithm incorporates five neighborhood structures to explore different

possibilities of depot locations and to improve customer assignments to each depot. Here, we
also applied the VND in the local search phase to explore another possibility of the solution.

4.5 Neighborhood Operators

We proposed to use five different neighborhood operators as the improvement phase.
The detail of the neighborhood is shown in table 1. In SA, the neighborhood being used
might different based on the generated random number, while in VNS we perform the entire
neighborhood based on the k neighborhood structure.

Table1. Neighborhood operators

Neighborhood
N1 Random insertion
N2 Random swap
N3 2-opt
N4 CROSS-exchange
N5 iCROSS-exchange

The two neighborhoods N1 and N2 correspond to insertion and swap moves respectively.
The intra-insertion neighborhood operator N1 consists of removing a customer from a
position i and inserting it after a position j. The inter-insertion operator consists in inserting a
customer i in a different route, after the customer j. The neighborhood N2 consists of
swapping the positions of two customers inside a route or between two different routes. The
classical 2-opt operator used in standard routing problem is a special case of our

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

neighborhood N3. This 2-opt is used both for improving the route and also the depot used in
the solution.
To deal with time windows, we applied another 2 neighborhoods based on study of
Taillard et al. (1997) whose proposed a neighborhood called CROSS-exchange. The main
idea of this exchange is to take two segments of different routes and exchange them. The
extention of CROSS-exchange called iCROSS-exchange is introduced by Braysy (2003). In
this neighborhood, the sequences of the segment get inverted. Both, CROSS-exchange (N4)
and iCROSS-exchange (N5) operators are used to define a set of neighborhood structures for
the VNS proposed in this paper to deal with time windows constraints.

5. Result and Discussion

The proposed method was implemented in C++ language. In order to evaluate
performance of both proposed approach in solving LRPTW, a well-known LRP problem set
by Barreto (2004) is selected and modified as test problems. All routes in this data set are
capacitated, and the depots are also capacitated. There are no variable costs associated with
depots and the traveling costs are not rounded.
We generate the time windows and defined the service time using the best solution from
Yu et al. (2010) both for the customers and depots. For this research, we decided to generate
data with tight time windows and fixed the service time. Performance of SA and VNS is
investigated in four terms namely best solution, average, standard deviation. Best solution
LRP is present to be compared with the result of LRPTW.

Table2. Computational results of LRPTW

Best
Best Average Standard Average Standard
Solution Best SA
VNS VNS Deviation SA Deviation
LRP
Christofides69-50x5 565.5 565.5 582.25 17.35 672.046 695.27 258.087
Christofides69-75x10 844.4 863.3 876.2 15.74 1042.3 1102.412 47.81
Christofides69-100x10 833.4 875.372 900.01 15.45 1051.83 1163.73 65.65
Gaskell67-21x5 424.9 424.9 424.9 0 424.9 468.24 25.59
Gaskell67-22x5 585.1 585.1 585.1 0 587.398 641.4 66.69
Gaskell67-29x5 512.1 512.1 513.38 1.75 518.344 567.96 61.13
Gaskell67-32x5 562.2 562.2 567.63 7.91 579.257 656.18 74.9
Gaskell67-32x5 504.3 504.3 504.3 0 511.086 609.38 65.58
Gaskell67-36x5 460.4 460.4 460.4 0 503.631 529.96 31.1

The performance of proposed VNS and SA are investigated in three terms namely best
solution, average, and standard deviation. According to the best solution and average
mentioned in these tables, VNS outperforms SA. For the chosen dataset, the gap between
VNSLRPTW solutions and optimal solution of LRP ranges from 0.81% to 9.3%. VNS has
smaller standard deviations compared to SA, showing that the results of VNS are more stable
than those of SA. The computational results confirm that the proposed VNS and SA can solve
LRPTW. VNS outperforms SA in that it not only can obtain better solutions, but also more

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

stable solutions and faster convergence. Since the results obtained by VNS are proven to be
the best results, we proposed to use VNS for solving LRPTW.
In summary, the decision maker can use the metaheuristic approach to solve the real
problem represented by LRPTW. The solutions obtained by the proposed VNS are near
optimal solutions. The proposed method can be further employed to solve both determine the
distribution center location and the vehicle routing in the form of LRPTW. In further work,
the decision maker can also escalate the real problem and consider other relevant aspects.

6. Conclusions
In this study we have studied a problem that has not been previously dealt, i.e. a location
routing problem with time windows (LRPTW) based on practical considerations. To solve
this problem, we proposed two metaheuristic approaches, VNS and SA. From the
computational study, VNS outperforms SA. This VNS method can be used as a tool for
solving logistics and transportation problems and can be extended to a Decision Support
System to help the decision makers.

References
Barreto, S. (2004). Análise e modelização de problemas de localização-distribuição [analysis
and modelling of location-routing problems] PhD thesis. Portugal: University of Aveiro,
campus universitario de Santiago.
Belenguer, J.-M., Benavent, E., Prins, C., Prodhon, C., & Wolfler Calvo, R. (2011). A
branch-and-cut method for the capacitated location-routing problem. Computer and
Operation Research, 38(6), 931-941.
Braysy, O. (2003). A Reactive Variable Neighborhood Search for the Vehicle-Routing
Problem with Time Windows. INFORMS Journal on Computing 15(4), 347-368.
Fleszar, K., & Hindi, K. (2008). An effective VNS for the capacitated p-median problem.
European Journal of Operational Research, 191(3), 612-622.
Hansen, P., & Mladenović, N. (1997). Variable neighborhood search for the p-median.
Location Science, 5(4), 207-226.
Laporte, G., & Nobert, Y. (1981). An exact algorithm for minimizing routing and operating
costs in depot location. European Journal of Operational Research, 6(2), 224-226.
Lenstra, J. K., & Rinnooy Kan, A. H. G. (1981). Complexity of vehicle routing and
scheduling problems. Networks, 11(2), 221-227.
Polacek, M., Hartl, R., Doerner, K., & Reimann, M. (2004). A variable neighborhood search
for the multi depot vehicle routing problem with time windows. Journal of Heuristics,
10(6), 613-627.

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.
Meilinda F.N. Maghfiroh, Perwira Redi and Vincent F. Yu

Prins, C., Prodhon, C., & Wolfer-Calvo, R. (2006). Solving the capacitated location-routing
problem by a grasp complemented by a learning process and a path relinking. 4OR—A
Quarterly Journal of Operations Research, 4(3), 221-238.
Rousseau, L.-M., Gendreau, M., & Pesant, G. (2002). Using constraint-based operators to
solve the vehicle routing problem with time windows. Journal of Heuristics, 8(1), 43-58.
Su, C.T. (1999). Dynamic vehicle control and scheduling of a multi-depot physical
distribution system. Integrated Manufacturing Systems, 10(1), 56-65.
Taillard, E. D., Badeau, P., Gendreau, M., Guertin, F., & Potvin, J. Y. (1997). A Tabu Search
Heuristic for the Vehicle Routing Problem with Soft Time Windows. Transportation
Science, 31(2), 170-186.
Tan, K.C., Lee, L.H., & Ou, K. (2001) Artificial Intelligence heuristics in solving vehicle
routing problems with time window constraints. Engineering Application of Artificial
Intelligence, 14(6), 825-837.
Tuzun, D., Burke, L.I. 1999. A two-phase tabu search approach to the location routing
problem. European Journal of Operational Research, 116(1), 87-99.
Yu, V.F., Lin, S.W., Lee, W., & Ting, C.J. (2010). A simulated annealing heuristic for the
capacitated location routing problem. Computers and Industrial Engineering, 58(2),
288-299.

International Conference on Innovation and Management, Sapporo, Japan, Jan. 28-31, 2013.