Professional Documents
Culture Documents
Discrete Optimization
CIRRELT and Canada Research Chair in Distribution Management, HEC Montral, 3000, chemin de la Cte-Sainte-Catherine, Montral, H3C 3J7, Canada
CIRRELT and Department of Mathematics and Industrial Engineering, cole Polytechnique de Montral, C.P. 6079, succ. Centre-ville, Montral, H3C 3A7, Canada
c
CIRRELT and HEC Montral, 3000, chemin de la Cte-Sainte-Catherine, Montral H3C 3J7, Canada
b
a r t i c l e
i n f o
Article history:
Received 23 September 2014
Accepted 9 July 2015
Available online 29 July 2015
Keywords:
Traveling salesman problem
Time-dependency
Service times
Lower and upper bounds
a b s t r a c t
This paper introduces a version of the classical traveling salesman problem with time-dependent service
times. In our setting, the duration required to provide service to any customer is not xed but dened as
a function of the time at which service starts at that location. The objective is to minimize the total route
duration, which consists of the total travel time plus the total service time. The proposed model can handle
several types of service time functions, e.g., linear and quadratic functions. We describe basic properties for
certain classes of service time functions, followed by the computation of valid lower and upper bounds. We
apply several classes of subtour elimination constraints and measure their effect on the performance of our
model. Numerical results obtained by implementing different linear and quadratic service time functions on
several test instances are presented.
2015 Elsevier B.V. and Association of European Operational Research Societies (EURO) within the
International Federation of Operational Research Societies (IFORS). All rights reserved.
1. Introduction
The purpose of this paper is to introduce, model and solve a version of the classical traveling salesman problem (TSP) with timedependent service times (TSP-TS), an extension of the asymmetric
TSP. ncan, Altnel, and Laporte (2009), and Roberti and Toth (2012)
present comprehensive reviews of the available mathematical formulations for the asymmetric TSP, some of which will be extended to
model the TSP-TS.
In most of the research on the TSP, service times are either ignored, or assumed to be constant and thus accounted for in travel
times. However in practice, it can easily be observed that service
times vary according to several factors which naturally depend on
the time of day (e.g., availability of parking spaces, accessibility to
the customer at its location, and so on). In the TSP-TS, the service
time required at each customer is not xed a priori, but depends
on the start time of service. The TSP-TS aims to minimize the total
route duration including the total travel time and the total service
time. This problem can formally be dened on a connected digraph
G = (N, A). In this graph, N = {0, 1, . . . , n, n + 1} is the set of nodes
and A = {(i, j) | i, j N, i = j} is the set of arcs. Nodes 0 and n + 1
correspond to the starting and ending points of the salesmans tour,
respectively. Each node in N \ {0, n + 1} corresponds to a distinct
Corresponding author. Tel.: +1 514 343 6111; fax: +1 514 343 7121.
E-mail addresses: duygu.tas@cirrelt.ca, d.tas@tue.nl (D. Tas),
customer. With each arc (i, j) in A is associated a travel time tij . Each
customer i has a service time dened as a continuous function si (bi ),
where bi corresponds to the start time of service at that customer
location.
In the TSP literature, time-dependency is usually addressed in
terms of travel times. The interested reader is referred to Gouveia and
Vo (1995) who present a classication of formulations proposed for
the time-dependent TSP. Picard and Queyranne (1978), Vander Wiel
and Sahinidis (1996), and Bigras, Gamache, and Savard (2008) consider a time-dependent TSP in which the travel time between any
two nodes depends on the time period of the day. It is further assumed that when the salesman starts traversing an arc, no transition
from one time period to the next takes place during this travel, in
other words there is no transient zone. More specically, the travel
time from node i to node j depends on the time period in which node
i is visited. This problem with discrete travel times can be viewed
as a single machine scheduling problem with sequence-dependent
setup times. Picard and Queyranne (1978) provide three integer programming formulations for the time-dependent TSP. The authors analyze the relationships between the relaxations of these models by
comparing their lower bounds. It is observed that the shortest path
relaxation (related to the rst model) is very similar to a formulation proposed by Hadley (1964) for the classical TSP. Vander Wiel and
Sahinidis (1996) propose an algorithm for the time-dependent TSP,
based on applying Benders decomposition to a mixed-integer linear
programming formulation. The authors also develop a network-based
algorithm to identify Pareto-optimal dual solutions of the highly
degenerate subproblems. Results indicate that the performance
http://dx.doi.org/10.1016/j.ejor.2015.07.048
0377-2217/ 2015 Elsevier B.V. and Association of European Operational Research Societies (EURO) within the International Federation of Operational Research Societies (IFORS).
All rights reserved.
373
forms of subtour elimination constraints. We also present the computation of a lower bound on the bigM which is employed in our model.
Section 4 reports computational results corresponding to different
subtour elimination constraints and different service time functions.
This section also provides the computation of a valid upper bound
on the total route duration of an optimal solution. Finally, our main
ndings and conclusions are highlighted in Section 5.
2. Service time function
In this section, we present the certain properties of the service
time function si (bi ) at node i.
2.1. First-In-First-Out property
The rst property is related to the First-In-First-Out (FIFO) principle which states that if service at node i starts at a time bi , any service
starting at a later time bi at that node cannot be completed earlier
than bi + si (bi ).
Proposition 2.1. si (bi ) satises the FIFO property if and only if
dsi (bi )
1.
dbi
Proof. ( necessity)
If si (bi ) satises the FIFO property, then
si (bi + ) si (bi ) ,
si (bi + ) si (bi )
lim
1,
si (bi + ) si (bi )
lim 1,
0
dsi (bi )
1.
dbi
( suciency)
It is given that si (bi ) is continuous on [bi , bi ] where bi > bi . From
the mean value theorem, we know that there exists at least one point
bi in (bi , bi ) such that
si (bi ) si (bi )
dsi (bi )
=
.
dbi
bi bi
If
dsi (bi )
dbi
si (bi ) si (bi )
1,
bi bi
si (bi ) + si (bi ) bi bi ,
bi + si (bi ) bi + si (bi ),
which means that the FIFO property is satised.
Note that Proposition 2.1 holds for any service time function. At
this point, it is worth observing that a TSP solution is not always optimal for the TSP-TS, even when we apply a service time function
that satises the FIFO property starting from the rst customer in the
route. To illustrate, suppose that there are three customers (denoted
374
Table 1
Travel time matrix of a small TSP instance.
Nodes
Nodes
AT
DT
TT
ST
0
1
2
3
4
0.00
5.00
4.00
4.00
0.00
5.00
0.00
2.00
1.75
5.00
4.00
2.00
0.00
1.50
4.00
4.00
1.75
1.50
0.00
4.00
0.00
5.00
4.00
4.00
0.00
0
1
2
3
4
(Starting point)
0.50
3.25
5.31
16.79
0.00
2.75
4.81
16.29
(Ending point)
0.00
0.50
1.00
1.50
2.00
0.00
2.25
3.81
14.79
14.79
Table 2
Details of the TSP solution evaluated with respect to time-dependent service times.
Table 5
Details of the route for the case with waiting.
Nodes
AT
DT
TT
ST
Nodes
AT
DT
TT
ST
BS
0
3
1
2
4
(Starting point)
4.00
6.75
22.81
419.35
0.00
5.00
20.81
415.35
(Ending point)
0.00
4.00
5.75
7.75
11.75
0.00
1.00
15.06
407.06
407.06
0
1
2
3
4
(Starting point)
0.50
2.25
2.81
3.97
0.00
1.75
2.31
3.47
(Ending point)
0.00
0.50
1.00
1.50
2.00
0.00
0.25
0.31
0.97
0.97
(No service)
1.50
2.25
2.81
(No service)
Table 3
Details of the optimal TSP-TS solution.
Nodes
AT
DT
TT
ST
0
2
3
1
4
(Starting point)
4.00
6.50
20.50
331.75
0.00
5.00
18.75
326.75
(Ending point)
0.00
4.00
5.50
7.25
12.25
0.00
1.00
13.25
319.50
319.50
by nodes 1, 2 and 3) and one depot (denoted by nodes 0 and 4), with
the travel time matrix of Table 1. The classical TSP has the two optimal solutions (0, 3, 1, 2, 4) and (0, 2, 1, 3, 4), with a total travel time
equal to 11.75.
Now assume that for the TSP-TS represented by the same graph,
the service time function is dened as si (bi ) = b2i 6bi + 9 for all
i N \ {0, n + 1}. From Proposition 2.1, the FIFO property is satised
for all bi 2.50. Note that in the TSP solutions, the salesman arrives
at the rst customer after the earliest time at which the FIFO property starts holding. Moreover, it is obvious that waiting at the depot
or at a customer location does not bring any reduction in the total
route duration. When we evaluate the two TSP solutions with the
time-dependent service times dened above, we observe that the total route durations of the corresponding routes substantially increase
to 419.35 and 501.81. The arrival times (AT) and departure times (DT)
at each node of the TSP solution (0, 3, 1, 2, 4), which has a lower total route duration, are provided in Table 2 where there is no waiting.
This table also gives the total travel time (TT) and the total service
time (ST) spent until the departure of the salesman from each node.
The optimal solution of the TSP-TS is (0, 2, 3, 1, 4), with a total route
duration of 331.75. Table 3 provides the details of the corresponding
route at each node. As in the TSP solutions, waiting is not needed
since it does not bring any reduction in the total route duration.
In a TSP-TS setting, the solution (0, 2, 3, 1, 4) has a lower duration
than that of the TSP solutions (0, 3, 1, 2, 4) and (0, 2, 1, 3, 4). This
observation implies that a TSP solution may not be optimal for the
TSP-TS even when all customers have the same service time function
and waiting does not provide any reduction in the total service time.
Moreover, it shows that our problem may not correspond to the TSP
even when all customers have the same quadratic service time function (with a unique minimum).
2.2. Waiting property
The second property is related to waiting. If si (bi ) does not satisfy the FIFO property when the salesman arrives at customer i, then
it may pay to wait at that customer before starting service. We rst
provide a small example to illustrate this property. The related propositions are then presented.
Suppose that there are three customers (denoted by nodes 1, 2
and 3) and one depot (denoted by nodes 0 and 4). In this setting,
ti j = 0.50 for all (i, j) in A, and si (bi ) = b2i 4bi + 4 for all i N{0,
4}. From Proposition 2.1, the FIFO property is satised for all bi
1.50. Consider a route that visits customers in the order 1, 2 and 3.
Tables 4 and 5 provide the arrival times and the departure times at
each node in the current route if waiting is not allowed and if waiting
is allowed, respectively. These tables also present the total travel time
and the total service time spent until the departure of the salesman
from each node. In Table 5, we also provide in the last column the
start time of service at each customer. If there is no waiting, the total service time is equal to 14.79. When waiting is allowed at the rst
customer, the total service time is reduced to 0.97.
This example shows that when the FIFO property is not always
satised by the service time function, then the total route duration
may be decreased by waiting. The related proposition follows.
Proposition 2.2. If the salesman arrives at a customer i before bi (the
earliest time at which the FIFO property starts holding at that customer),
waiting at that location to begin service at time bi is then benecial.
Proof. Suppose that in a given route, each customer j N \ {0, n + 1}
has a service time function sj (bj ) and waiting is not considered. The
latter structure implies that the arrival time at any customer i is
equivalent to the start time of service at that customer, which is rep.
resented by b
i
Suppose that si (bi ) satises the FIFO property for all bi where bi
< b . More specically, dsi (bi ) < 1 for all b such that
bi bi and b
i
i
i
dbi
, b ). From the mean value
0 bi < bi , and therefore such that bi [b
i i
, b )
theorem, we know that there exists at least one point b in (b
i
such that
)
si (bi ) si (b
dsi (bi )
i
=
.
dbi
b b
i
i
Since
dsi (bi )
dbi
)
si (bi ) si (b
i
< 1,
b b
i
) < (b b
),
si (bi ) si (b
i
i
i
.
) + b
si (bi ) + bi < si (b
i
i
The last inequality implies that instead of beginning service immediately, waiting at customer i until bi (until si (bi ) starts satisfying the
FIFO property) is benecial.
Proposition 2.2 also holds for any service time function, even if
there may be several points at which the FIFO property starts holding. For example, a cubic function may have a structure in which the
FIFO property does not always hold. In such a case, we should dene
at which the FIFO property starts holdbi as the point closest to b
i
ing, where b < b . More specically, if customer i has a cubic service
i
time function and the salesman arrives at i before bi (the point closest to the arrival time at which the FIFO property starts holding at
that customer), waiting at that location to begin service at time bi is
still benecial. Note that in the following propositions, it is assumed
that the service time function has only one point at which the FIFO
property starts holding (linear or quadratic service time functions).
From Proposition 2.2, we distinguish two categories of service
times: (i) each customer has a distinct service time function, and (ii)
all customers have the same service time function. In Proposition 2.3,
we consider the latter case and prove that all the waiting time can
be shifted to the depot (instead of spending idle times at customer
locations).
Proposition 2.3. If all customers have the same service time function,
then the waiting time required to satisfy the FIFO property can be spent
at the depot.
Proof. Consider a service time function sj (bj ) for all j N \ {0, n + 1},
satisfying the FIFO property for all b j bj 0. In other words, bj is
the earliest time at which the FIFO property starts holding at each
customer j. Let i be the rst customer in a given route. First, the arrival time at customer i is equivalent to the start time of service at that
. From Proposition 2.2, we know that it is benecial to
customer, i.e., b
i
delay service at customer i until bi . Since all customers have the same
service time function, waiting at the rst customer is sucient to satisfy the FIFO property for all other customers (no additional waiting
is needed to satisfy the FIFO along the route). This waiting time can
be shifted to the depot. More specically, waiting is needed only at
the depot to satisfy the FIFO property for all customers.
375
and the depot needs to be considered twice in the list L (the connection from the depot to the customer, and the connection from
the customer to the depot). The list is then sorted in ascending order so that k < m implies L[k] < L[m], where L[k] is the kth element
) where
A
= (N,
of list L. We then construct a new directed graph G
N = {n + 2, n + 3, . . . , 2n + 3} is the set of nodes, and A = {(i, j) | i
\ {2n + 3}, j = i + 1, j N
\ {n + 2}} is the set of arcs. In this graph,
N
nodes n + 2 and 2n + 3 can be viewed as the starting and ending
\ {n + 2, 2n + 3} has a service
points of each route. Each node i N
time dened by the original function si (bi ). Moreover, the travel time
on each arc (i, j) in A,
ti j is equal to L[i n 1]. From the denition
of G, it can be observed that only one route (route p) starts and ends
exactly once. Suppose that the
at the depot and visits each node i N
start time of service at the rst node of this route (node n + 2) is set
to 0. With respect to bn+2 and to the service time spent at node n + 2,
tn+2,n+3 . We
which is equal to 0 by denition, bn+3 is computed as
know that bn+3 b(r,2) for any possible route r generated for the original problem since
tn+2,n+3 = min {ti j } and b(r,1) = 0 for all r. Simi(i, j)A
minimize
xi j = 1,
si (bi )
(1)
iN\{0,n+1}
iN jN
s.t.
ti j xi j +
i N \ {n + 1 },
(2)
jN\{i}
j N \ {0},
xi j = 1,
(3)
iN\{ j}
bi + si (bi ) + ti j M(1 xi j ) b j ,
bi 0,
i N,
xi j {0, 1},
i N, j N.
i N, j N,
(4)
(5)
(6)
376
The objective (1) is to minimize the total route duration which consists of the sum of the total travel time and the total service time.
Constraints (2) and (3) ensure that each customer is visited exactly
once, and impose the degree requirements for the two depot nodes.
Constraints (4) link the departure time from a node and the starting
time of service at its successor, and act as subtour elimination constraints. Constraints (5) ensure that the start time of service at each
customer is non-negative. Constraints (6) indicate that partial service
at customers is not allowed.
In Section 3.1, we present a number of additional subtour elimination constraints which provide valid cuts to the model and will be
compared in Section 4.1. In Section 3.2, we provide a procedure to
compute a lower bound on the bigM.
3.1. Additional subtour elimination constraints
We consider the three additional subtour elimination constraints,
which are reported by ncan et al. (2009) and Roberti and Toth (2012)
as the best when the model is solved directly by CPLEX IBM (2015),
to enhance the performance of our formulation. In other words,
these constraints can be used as valid inequalities in the model (1)
(6) (they are not needed to eliminate subtours since constraints (4)
inherently prevent them).
3.1.1. The Miller, Tucker and Zemlin (MTZ) formulation
Miller, Tucker, and Zemlin (1960) proposed the following MTZ
subtour elimination constraints:
ui u j + nxi j n 1,
i, j = 1, . . . , n,
(7)
ui u j + nxi j + (n 2)x ji n 1,
ui + (n 2)xi,n+1 +
i, j = 1, . . . , n,
x ji 1,
i = 1, . . . , n,
(8)
(9)
jN\{0,n+1}
ui + (n 2)x0i +
xi j n,
i = 1, . . . , n.
(10)
jN\{0,n+1}
Desrochers and Laporte (1990) also proved that constraints (8) are
facet dening.
3.1.3. The Gavish and Graves (GG) formulation
Several single-commodity, two-commodity and multicommodity ow formulations were introduced for the asymmetric
TSP. The interested reader is referred to Langevin, Soumis, and
Desrosiers (1990) for a comprehensive classication on commodity
ow formulations. Gavish and Graves (1978) developed a singlecommodity ow formulation for which the LP relaxation of the
resulting model is stronger than that of the MTZ formulation but
weaker than that of the classical Dantzig, Fulkerson and Johnson
formulation (Dantzig, Fulkerson, & Johnson, 1954) (see Gouveia,
1995). In the GG formulation, subtours are eliminated by using
n(n + 1) non-negative variables gij , where i N \ {0, n + 1} and j
N{0}. The GG subtour elimination constraints of this formulation
are stated as follows:
jN\{0}
gi j
jN\{0,n+1}
g ji = 1,
i = 1, . . . , n,
(11)
0 gi j nxi j ,
i = 1, . . . , n, j = 1, . . . , n + 1.
(12)
In (11) and (12), the gij variables denote the number of arcs on a tour
from the depot to arc (i, j). Note that the starting point of each tour is
represented by node 0. These two sets of constraints yield a network
ow model in which the gij variables naturally take integer values for
xed values of xij variables.
3.2. Computing a lower bound on the bigM
In this section, we provide the computation of a suciently large
value for the bigM to be used in constraints (4) of the formulation (1)
(6), where all customers rst have the same quadratic service time
function (with a unique minimum).
Proposition 3.1. When all customers have the same quadratic service
time function (with a unique minimum), the arrival time of the salesman
at the last node in a route, computed by considering the rst n + 1 largest
travel times, yields a valid value for M in formulation (1)(6).
Proof. Suppose that each customer i N \ {0, n + 1} has the same
2
service time function si (bi ) = bi bi + , where , , > 0 and
si (bi ) 0, and that si (bi ) has a unique minimum value (the minimum
ds (b )
d2 si (bi )
service time) at bi = /2 . Since i i 0 and
> 0 for all
dbi
db2i
bi /2 , si (bi ) is an increasing function with an increasing derivative for bi [ /2 , ). From Proposition 2.1, we know that si (bi )
satises the FIFO property at each customer i N \ {0, n + 1} for all
bi ( 1)/2 . From Proposition 2.3, we know that for any route
r in G (starting and ending at the depot, and visiting each node exactly once), waiting at the depot to arrive at the rst customer at
( 1)/2 is sucient to satisfy the FIFO property (no additional
waiting is needed for the FIFO along the route r). Moreover, the latest possible arrival time at the rst customer in any route r in G is
equal to /2 due to the behavior of si (bi ) for bi [ /2 , ). Let (r,
i) denote the ith node visited by route r. Note that (r, 1) and (r, n + 2)
correspond to the depot in each route r.
We have a symmetric travel time matrix in which nodes 0 and n +
1 both correspond to the depot. In order to compute a lower bound
on the bigM, we rst dene a list L that contains travel times tij for all
j > i and j = n + 1. When the graph comprises only one customer, the
travel time between that customer and the depot needs to be considered twice in the list L, once in each direction. This list is then sorted
in decreasing order so that L[k] > L[m] when k < m, where L[k] is
the kth element of list L. We then construct a new directed graph
G = (N, A), where N = {n + 2, n + 3, . . . , 2n + 3} is the set of nodes
and A = {(i, j) | i N \ {2n + 3}, j = i + 1, j N \ {n + 2}} is the set of
arcs. In this graph, nodes n + 2 and 2n + 3 can be viewed as the starting and ending points of each route. Each node i N \ {n + 2, 2n + 3}
has a service time dened by the original function si (bi ). Moreover,
the travel time on each arc (i, j) in A, t i j is equal to L[i n 1]. From
the denition of G, it can be observed that there exists only one
route (route p) starting and ending at the depot, and visiting each
node i N exactly once. Suppose that the starting time of service at
node n + 2 is set to /2 . More specically, si (bi ) satises the FIFO
property at each node i in route p and waiting along that route does
not yield any reduction in the total service time. With respect to
bn+2 and the service time spent at node n + 2, which is equal to 0
by denition, bn+3 is computed as ( /2) + t n+2,n+3 . We know that
bn+3 b(r,2) for any possible route r generated for the original prob-
lem since t n+2,n+3 = max {ti j } and ( /2 ) b(r, 1) for all r. Similarly,
(i, j)A
Thus, we can conclude that b2n+3 b(r,n+2) for any possible route
r generated for the original problem dened on G, where (r, n + 2)
corresponds to the depot. Note that the departure time of each route
r from the starting point is optimized. In other words, the arrival time
of the salesman at the end of route p provides a valid and reasonable
value for the bigM to be employed in the original problem dened
with a quadratic service time function (with a unique minimum) at
each customer.
If we consider a linear service time function, the route constructed
by using the rst n + 1 largest travel times is also valid to compute
a value for bigM. Suppose that each customer i N \ {0, n + 1} has
the same linear (increasing) service time function si (bi ) = bi + ,
where , > 0 and si (bi ) > 0. We know that si (bi ) satises the FIFO
property at each customer i N \ {0, n + 1} for all bi 0 (see the
proof of Proposition 2.4). In other words, waiting at the depot or at
customer locations does not yield any reduction in the total route duration. It is easy to observe that when we adjust the latest possible
departure time from the depot with respect to the linear function
considered at each customer, all properties provided in the proof of
Proposition 3.1 hold. Thus, one can employ the computation of the
bigM when all customers have the same quadratic service time function (with a unique minimum) or the same linear (increasing) service
time function.
4. Computational results
The model just described was coded in C++ and solved by using
IBM ILOG CPLEX 12.5 (IBM, 2015). All experiments were conducted on
an Intel(R) Xeon(R) CPU X5675 with 12-Core 3.07 Gigahertz and 96
Gigabyte of RAM (by using a single thread). We have experimented
with data sets presented in the TSPLIB library of Reinelt (1991). We
have focused on the instances with up to 45 nodes dened with
a symmetric travel time matrix (burma14, ulysses16, gr17, gr21,
ulysses22, gr24, fri26, bayg29, bays29, dantzig42 and swiss42). In
addition to these 11 instances, we have generated 11 more instances
by selecting (i) the rst 30 nodes from dantzig42, swiss42, att48,
gr48, hk48 and eil51, (ii) the rst 35 nodes from swiss42, gr48 and
eil51, (iii) the rst 40 nodes from eil51, and (iv) the rst 45 nodes
from eil51. The original travel times were adjusted in such a way that
we have the same average travel time per arc in each instance. Note
that in all these instances, there exists one more node for the ending
depot, e.g., burma14 includes 15 nodes.
4.1. Effects of the additional subtour elimination constraints
Each instance was rst solved directly by CPLEX with the basic
model dened by (1)(6). To assess the effect of the additional constraints on the performance of the model, we have then applied three
types of subtour elimination constraints: (i) MTZ, (ii) DL and (iii) GG.
Each of these four formulations was enhanced by implementing: (i)
an upper bound on the total route duration (B1), (ii) a lower bound on
the total service time (B2), and (iii) a lower and an upper bound on
the start time of service at each customer (B3). A maximum of 7200
377
378
Status
Obj
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Avg
Avg
Optimal
236.44
Optimal
279.57
Optimal
245.40
Optimal
249.32
Optimal
318.06
Optimal
284.93
Feasible
263.01
Optimal
371.22
Optimal
331.90
No integer solution
Feasible
350.65
Feasible
349.16
Feasible
305.23
No integer solution
Optimal
366.78
Feasible
397.42
No integer solution
No integer solution
Feasible
452.89
Feasible
285.12
No integer solution
No integer solution
Feasible
343.36
Optimal
298.18
TT (Percent)
ST (Percent)
Gapf (Percent)
Seconds
93.70
94.51
94.40
90.48
92.91
89.29
89.06
86.74
86.95
6.30
5.49
5.60
9.52
7.09
10.71
10.94
13.26
13.05
0.00
0.00
0.00
0.00
0.00
0.00
15.40
0.00
0.00
86.55
85.92
86.61
13.45
14.08
13.39
18.38
1.80
8.06
87.34
83.79
12.66
16.21
0.00
4.58
82.36
81.84
17.64
18.16
2.60
10.71
85.16
90.70
14.84
9.30
8.79
0.00
1.89
40.30
228.85
11.01
3468.48
4580.27
5248.50
5375.46
534.18
2165.44
Table 7
Solutions with all bounds, MTZ subtour elimination constraints and medium service times, si (bi ) = 102 bi + 6(102 ) for all i N \ {0, n + 1}.
Name
Status
Obj
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Average
Optimal
236.44
Optimal
279.57
Optimal
245.40
Optimal
249.32
Feasible
318.06
Optimal
284.93
Feasible
263.01
Optimal
371.22
Feasible
331.90
Feasible
273.10
Feasible
353.77
Optimal
349.16
Feasible
305.23
No integer solution
Optimal
366.78
Feasible
397.42
No integer solution
Feasible
406.92
Feasible
452.89
Feasible
294.50
No integer solution
Feasible
502.52
Feasible
354.48
Optimal
297.85
TT (Percent)
ST (Percent)
Gapf (Percent)
93.70
94.51
94.40
90.48
92.91
89.29
89.06
86.74
86.95
86.51
86.50
85.92
86.61
6.30
5.49
5.60
9.52
7.09
10.71
10.94
13.26
13.05
13.49
13.50
14.08
13.39
0.00
0.00
0.00
0.00
6.30
0.00
14.30
0.00
3.58
14.56
18.90
0.00
8.39
87.34
83.79
12.66
16.21
0.00
1.35
84.46
82.36
79.80
15.54
17.64
20.20
2.65
2.72
13.53
79.80
85.34
90.30
20.20
14.66
9.70
3.97
8.20
0.00
Seconds
1.07
23.97
62.43
7.24
3628.16
1544.31
147.49
344.64
719.91
379
Table 8
Solutions with all bounds, DL subtour elimination constraints and medium service times, si (bi ) = 102 bi + 6(102 ) for all i N \ {0, n + 1}.
Name
Status
Obj
TT (Percent)
ST (Percent)
Gapf (Percent)
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Average
Optimal
Optimal
Optimal
Optimal
Feasible
Optimal
Feasible
Optimal
Optimal
Feasible
Feasible
Optimal
Feasible
Feasible
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
236.44
279.57
245.40
249.32
318.06
284.93
263.01
371.22
331.90
273.10
349.60
349.16
305.23
347.35
366.78
397.62
306.91
406.92
452.89
285.07
388.64
502.52
353.61
301.64
93.70
94.51
94.40
90.48
92.91
89.29
89.06
86.74
86.95
86.51
86.81
85.92
86.61
87.28
87.34
84.25
84.45
84.46
82.36
81.74
81.89
79.80
85.24
89.92
6.30
5.49
5.60
9.52
7.09
10.71
10.94
13.26
13.05
13.49
13.19
14.08
13.39
12.72
12.66
15.75
15.55
15.54
17.64
18.26
18.11
20.20
14.76
10.08
0.00
0.00
0.00
0.00
7.61
0.00
11.07
0.00
0.00
9.15
3.28
0.00
0.40
3.38
0.00
2.48
6.15
2.30
2.50
11.57
5.84
3.86
5.35
0.00
Seconds
2.64
62.91
38.71
11.54
3207.58
1241.87
4515.15
171.40
355.60
1067.49
Table 9
Solutions with all bounds, GG subtour elimination constraints and medium service times, si (bi ) = 102 bi + 6(102 ) for all i N \ {0, n + 1}.
Name
Status
Obj
TT (Percent)
ST (Percent)
Gapf (Percent)
Seconds
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Average
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Feasible
Feasible
Optimal
Optimal
Feasible
Optimal
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
236.44
279.57
245.40
249.32
318.06
284.93
263.01
371.22
331.90
273.10
349.60
349.16
305.23
347.35
366.78
397.42
306.91
406.92
452.89
285.07
388.64
502.52
368.11
307.57
93.70
94.51
94.40
90.48
92.91
89.29
89.06
86.74
86.95
86.51
86.81
85.92
86.61
87.28
87.34
83.79
84.45
84.46
82.36
81.74
81.89
79.80
83.92
89.36
6.30
5.49
5.60
9.52
7.09
10.71
10.94
13.26
13.05
13.49
13.19
14.08
13.39
12.72
12.66
16.21
15.55
15.54
17.64
18.26
18.11
20.20
16.08
10.64
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
3.77
3.18
0.00
0.00
2.42
0.00
0.00
2.67
1.55
1.47
5.78
5.46
3.16
3.27
0.00
0.46
7.51
2.33
0.82
521.19
21.27
93.51
129.88
544.31
15.65
412.90
409.23
3342.23
423.18
within less than eight minutes). Table 11 shows that six instances
with up to 31 nodes are solved optimally by the model employing
a large service time function for each customer. The average computation time required to solve these six instances is approximately
22 minutes. Comparing the solutions given in Table 9 (with a medium
service time function) to those in Tables 10 and 11, we observe that
the number of instances solved to optimality decreases as the coefcients in the service time function increase. In addition, more computation time is required to obtain an optimal solution for the same
instance when using a larger service time function for each customer.
4.2.2. Quadratic service times
We now analyze the results obtained by employing a quadratic
service time function for each customer. Recall that the lower bound
on the total service time of the optimal solution is valid for the
case where all customers have the same linear (increasing) service time function (see Section 2.3). We have thus implemented an
380
Status
Obj
TT (Percent)
ST (Percent)
Gapf (Percent)
Seconds
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
228.83
271.74
238.39
237.11
306.44
269.09
247.99
345.49
309.27
253.85
324.21
323.40
283.91
324.20
342.50
363.39
281.82
373.60
410.35
257.37
351.15
448.11
308.74
96.81
97.24
97.18
95.14
96.43
94.54
94.46
93.20
93.31
93.07
92.84
92.76
93.11
93.51
93.53
91.64
91.97
91.99
90.65
90.53
90.63
89.49
93.37
3.19
2.76
2.82
4.86
3.57
5.46
5.54
6.80
6.69
6.93
7.16
7.24
6.89
6.49
6.47
8.36
8.03
8.01
9.35
9.47
9.37
10.51
6.63
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.18
3.84
2.05
0.27
37.72
1.60
4.14
5.38
19.35
150.86
247.35
3.35
13.48
176.20
5.29
56.18
173.19
33.69
367.39
4481.79
1060.69
3125.64
453.17
Table 11
Solutions with all bounds, GG subtour elimination constraints and large service times, si = 2(102 )bi + 1.2(101 ) for all i N \ {0, n + 1}.
Name
Status
Obj
TT (Percent)
ST (Percent)
Gapf (Percent)
Seconds
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Average
Optimal
Optimal
Optimal
Optimal
Feasible
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
252.62
296.28
260.34
275.96
343.58
320.42
297.39
430.35
383.78
316.51
404.54
408.23
353.89
400.88
422.54
474.90
365.75
485.44
556.10
352.36
480.30
638.13
419.15
302.31
87.69
89.18
88.99
81.74
86.01
79.40
78.77
74.82
75.19
74.95
75.52
74.22
74.70
75.78
75.81
70.75
70.87
70.80
67.25
66.22
66.26
63.00
72.92
83.54
12.31
10.82
11.01
18.26
13.99
20.61
21.23
25.18
24.81
25.05
24.48
25.78
25.30
24.22
24.19
29.25
29.13
29.20
32.75
33.78
33.74
37.00
27.08
16.46
0.00
0.00
0.00
0.00
5.19
0.00
4.67
3.28
4.85
10.78
9.99
0.00
7.32
8.69
7.23
3.94
10.66
10.28
5.22
12.72
11.87
6.84
7.72
0.00
4.13
58.41
12.22
10.75
1880.42
5853.90
1303.31
upper bound on the total route duration of the optimal solution (B1),
and a lower and an upper bound on the start time of service at
each customer location (B3) for the model with a quadratic service
time function. Since the quadratic model is inherently more dicult to solve than the linear one, we focus on the 13 instances for
which the optimal solutions can be obtained by the model with a
medium linear service time function. Table 12 presents the solutions
of these 13 instances where si = 4(105 )b2i 4(103 )bi + 101 for all
i N \ {0, n + 1}.
The results presented in Table 12 indicate that nine instances with
up to 31 nodes are solved optimally when a quadratic service time
function is considered. This model cannot provide a feasible solution
to one instance within the 7200 seconds time limit, and the remaining three instances can be solved suboptimally. When we compare
the solutions given in Table 12 to those of Table 9, we observe that
381
Table 12
Solutions with bounds B1 and B3, GG subtour elimination constraints and quadratic service times, with si = 4(105 )b2i 4(103 )bi + 101
for all i N \ {0, n + 1}.
Name
Status
Obj
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
eil30
gr30
swiss30
eil35
Average
Average
Optimal
224.83
Optimal
268.14
Optimal
234.82
Optimal
232.77
Optimal
301.58
Optimal
263.04
Optimal
239.08
Feasible
345.11
Feasible
305.46
Feasible
320.74
Optimal
279.94
Optimal
340.42
No integer solution
Feasible
323.77
Optimal
264.96
TT (Percent)
ST (Percent)
Gapf (Percent)
Seconds
98.53
98.54
98.66
96.91
97.98
96.72
97.98
93.30
94.47
93.53
94.43
94.10
1.47
1.46
1.34
3.09
2.02
3.28
2.02
6.70
5.53
6.47
5.57
5.90
0.00
0.00
0.00
0.00
0.00
0.00
0.00
4.34
2.69
3.26
0.00
0.00
93.77
97.09
6.23
2.91
3.43
0.00
0.77
9.17
3.94
2.77
639.74
47.61
30.85
2247.43
1462.91
493.91
Table 13
Solutions with GG subtour elimination constraints and medium service times, with si = 102 bi + 6(102 ) for all i N \ {0, n + 1}.
Name
Status
Obj
TT (Percent)
ST (Percent)
Gapf (Percent)
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Average
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
Feasible
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
236.44
279.57
245.40
249.32
318.06
284.93
263.01
371.22
331.90
273.10
349.60
349.16
305.23
347.35
366.78
397.42
306.91
406.92
452.89
285.07
388.64
502.52
366.36
283.19
93.70
94.51
94.40
90.48
92.91
89.29
89.06
86.74
86.95
86.51
86.81
85.92
86.61
87.28
87.34
83.79
84.45
84.46
82.36
81.74
81.89
79.80
84.52
90.92
6.30
5.49
5.60
9.52
7.09
10.71
10.94
13.26
13.05
13.49
13.19
14.08
13.39
12.72
12.66
16.21
15.55
15.54
17.64
18.26
18.11
20.20
15.48
9.08
0.00
0.00
0.00
0.00
0.00
0.00
0.00
6.40
7.09
6.08
5.53
6.55
0.00
7.12
0.00
11.68
9.00
7.59
13.71
13.72
14.05
17.52
9.70
0.00
upper bound on the total route duration (B1), (ii) a lower bound on
the total service time (B2), and (iii) a lower and an upper bound on
the start time of service at each customer (B3). Table 13 presents the
solutions obtained within the 7200 seconds time limit by employing
the model without any additional lower and upper bounds, where we
consider the GG subtour elimination constraints and medium service
times, si = 102 bi + 6(102 ) for all i N \ {0, n + 1}. Comparing the
results of Table 13 to those of Table 9 (the model with a medium service time function and all valid bounds), we observe that the number
of instances solved to optimality decreases from 13 (with up to 36
nodes) to nine (with up to 31 nodes). These nine instances, where
both models can obtain corresponding optimal solutions, are solved
by the model with all bounds in smaller computation times (on average by 73.89 percent). The enhanced model also provides smaller
nal gaps for the instances which are suboptimally solved by the
model without any additional lower and upper bounds (on average
76.64 percent).
In addition to the solutions obtained by the model including all
bounds (B1, B2 and B3) and by the model without bounds, Table 14
presents the results for each case where the model includes (i) only
B1, (ii) only B2, (iii) only B3, (iv) B2 and B1, and (v) B2 and B3. Note
that the last two cases are considered since we observe that the
Seconds
0.80
6.66
4.51
1.91
709.47
122.47
749.30
2353.52
1678.08
625.19
Table 14
Average results obtained by employing different bounds with GG subtour
elimination constraints and medium service times, si = 102 bi + 6(102 )
for all i N \ {0, n + 1}.
Bounds
Opt
Seconds
Gapf (Percent)
No bounds
B1
B2
B3
B2, B1
B2, B3
B1, B2, B3
9
9
13
7
12
13
13
625.19
1015.91
752.61
282.30
227.69
615.79
423.18
9.70
9.55
3.25
9.45
3.09
3.38
3.27
performance of the model for case (ii) is better than that for cases
(i) and (iii). In other words, when each bound is incorporated into
the model separately, the lower bound on the total service time is
the most effective one. For each formulation, we report the number
of instances solved to optimality, the average corresponding computation times, and the average optimality gap over the instances
solved suboptimally. The results presented in Table 14 indicate that
the enhanced model including all bounds provides the best performance in terms of the number of instances solved to optimality and
382
Status
Obj
TT (Percent)
ST (Percent)
Gapf (Percent)
Seconds
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Average
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Optimal
Feasible
Optimal
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
228.83
271.74
238.39
237.11
306.44
269.09
247.99
345.49
309.27
253.85
324.21
323.40
283.91
324.20
342.50
363.39
281.82
373.60
410.35
257.37
351.15
448.11
366.07
291.87
96.81
97.24
97.18
95.14
96.43
94.54
94.46
93.20
93.31
93.07
92.84
92.76
93.11
93.51
93.53
91.64
91.97
91.99
90.65
90.53
90.63
89.49
90.59
94.18
3.19
2.76
2.82
4.86
3.57
5.46
5.54
6.80
6.69
6.93
7.16
7.24
6.89
6.49
6.47
8.36
8.03
8.01
9.35
9.47
9.37
10.51
9.41
5.82
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
2.78
0.00
0.00
5.95
3.31
5.06
8.00
5.02
0.00
0.38
3.38
2.36
0.42
59.44
7.04
17.68
417.16
467.83
525.07
1062.66
1138.93
62.28
1743.14
21.92
1109.01
270.03
406.40
Table 16
Solutions with GG subtour elimination constraints and large service times, with si = 2(102 )bi + 1.2(101 ) for all i N \ {0, n + 1}.
Name
Status
Obj
TT (Percent)
ST (Percent)
Gapf (Percent)
Seconds
burma14
ulysses16
gr17
gr21
ulysses22
gr24
fri26
bayg29
bays29
att30
dantzig30
eil30
gr30
hk30
swiss30
eil35
gr35
swiss35
eil40
dantzig42
swiss42
eil45
Average
Average
Optimal
Optimal
Optimal
Optimal
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Feasible
Optimal
252.62
296.28
260.34
275.96
343.58
320.42
297.39
430.35
383.78
316.51
404.54
408.23
353.89
400.88
422.54
474.90
365.75
485.44
556.10
352.36
480.30
638.13
413.06
271.30
87.69
89.18
88.99
81.74
86.01
79.40
78.77
74.82
75.19
74.95
75.52
74.22
74.70
75.78
75.81
70.75
70.87
70.80
67.25
66.22
66.26
63.00
73.35
86.90
12.31
10.82
11.01
18.26
13.99
20.61
21.23
25.18
24.81
25.05
24.48
25.78
25.30
24.22
24.19
29.25
29.13
29.20
32.75
33.78
33.74
37.00
26.65
13.10
0.00
0.00
0.00
0.00
5.72
8.03
12.27
18.76
17.91
17.76
18.64
18.80
15.66
18.12
14.33
23.84
22.57
21.45
27.74
28.04
28.90
33.93
19.58
0.00
4.78
56.14
16.11
51.75
32.20
provides smaller nal gaps for the instances which are suboptimally
solved by the model without any additional lower and upper bounds
(on average 64.95 percent).
From the results given in Tables 1316, we conclude that the
model proposed in this paper performs much better with all bounds
explained in Section 4.1.
5. Conclusions
We have considered a traveling salesman problem with timedependent service times. Several analytical properties of the service
time function were derived and a lower bound on the total service
time of the optimal solution was computed. Examples were presented to illustrate properties. We then described a mathematical
formulation capable of solving linear and quadratic service time functions. We have developed a procedure to compute a suciently large
value for the bigM, which was then employed in the proposed mathematical model. In numerical experiments, a lower and an upper
bound on the start time of service at each customer were also implemented. Computational results showed that our model (with the
all valid bounds) can successfully solve instances with up to 46 nodes
with a linear service time function, and with up to 31 nodes with a
quadratic service time function.
Acknowledgments
This research was partly supported by the Canadian Natural Sciences and Engineering Research Council under Grants 338816-10,
436014-2013 and 39682-10. This support is gratefully acknowledged.
Thanks are due to Emanuela Guerriero and to the referees for their
valuable comments.
References
Bigras, L.-P., Gamache, M., & Savard, G. (2008). The time-dependent traveling salesman
problem and single machine scheduling problems with sequence dependent setup
times. Discrete Optimization, 5, 685699.
383
Cordeau, J.-F., Ghiani, G., & Guerriero, E. (2014). Analysis and branch-and-cut algorithm
for the time-dependent travelling salesman problem. Transportation Science, 48(1),
4658.
Dantzig, G. B., Fulkerson, D. R., & Johnson, S. M. (1954). Solutions of a large-scale
traveling-salesman problem. Operations Research, 2, 363410.
Desrochers, M., & Laporte, G. (1990). Improvements and extensions to the Miller
TuckerZemlin subtour elimination constraints. Operations Research Letters, 10(1),
2736.
Gavish, B., & Graves, S. C. (1978). The travelling salesman problem and related problems. Technical Report GR-078-78. Operations Research Center, Massachusetts Institute of Technology.
Gouveia, L. (1995). A result on projection for the vehicle routing problem. European
Journal of Operational Research, 85(3), 610624.
Gouveia, L., & Vo, S. (1995). A classication of formulations for the (time-dependent)
traveling salesman problem. European Journal of Operational Research, 83, 6982.
Hadley, G. (1964). Nonlinear and Dynamic Programming. Massachusetts: AddisonWesley, Reading.
IBM (2015). ILOG CPLEX Optimizer 12.5. http://www-01.ibm.com/software/
integration/optimization/cplex-optimizer.
Ichoua, S., Gendreau, M., & Potvin, J.-Y. (2003). Vehicle dispatching with timedependent travel times. European Journal of Operational Research, 144, 379396.
Langevin, A., Soumis, F., & Desrosiers, J. (1990). Classication of travelling salesman
problem formulations. Operations Research Letters, 9(2), 127132.
Miller, C. E., Tucker, A. W., & Zemlin, R. A. (1960). Integer programming formulations
and travelling salesman problems. Journal of the Association for Computing Machinery, 7, 326329.
ncan, T., Altnel, I. K., & Laporte, G. (2009). A comparative analysis of several asymmetric traveling salesman problem formulations. Computers & Operations Research, 36,
637654.
Picard, J.-C., & Queyranne, M. (1978). The time-dependent traveling salesman problem
and its application to the tardiness problem in one-machine scheduling. Operations
Research, 26, 86110.
Reinelt, G. (1991). TSPLIB a traveling salesman problem library. ORSA Journal on Computing, 3(4), 376384.
Roberti, R., & Toth, P. (2012). Models and algorithms for the asymmetric traveling salesman problem: an experimental comparison. EURO Journal on Transportation and
Logistics, 1, 113133.
Tagmouti, M., Gendreau, M., & Potvin, J.-Y. (2007). Arc routing problems with timedependent service costs. European Journal of Operational Research, 181, 3039.
Tagmouti, M., Gendreau, M., & Potvin, J.-Y. (2010). A variable neighborhood descent
heuristic for arc routing problems with time-dependent service costs. Computers
& Industrial Engineering, 59(4), 954963.
Tagmouti, M., Gendreau, M., & Potvin, J.-Y. (2011). A dynamic capacitated arc routing
problem with time-dependent service costs. Transportation Research Part C, 19(1),
2028.
Vander Wiel, R. J., & Sahinidis, N. V. (1996). An exact solution approach for the timedependent traveling salesman problem. Naval Research Logistics, 43, 797820.