Journal of the Operational Research Society (2008) 59, 1220 --1228 © 2008 Operational Research Society Ltd.

y Ltd. All rights reserved. 0160-5682/08

www.palgrave-journals.com/jors

Vehicle routing with soft time windows

and Erlang travel times
RA Russell and TL Urban∗
The University of Tulsa, Tulsa, Oklahoma, USA
This paper investigates vehicle-routing problems in which the travel times are random variables, and deliveries
are made subject to soft time-window constraints. In particular, we model the travel time using a shifted gamma
distribution. Penalties are incurred for deviations from the customers’ time windows—early or late—and are
developed using a fixed cost, a linear cost penalty, and/or a quadratic loss penalty. Alternatively, specifying
a given probability of meeting the time-window constraints is considered. A tabu-search metaheuristic is
developed, and computational results on test problems from the literature are reported.
Journal of the Operational Research Society (2008) 59, 1220 – 1228. doi:10.1057/palgrave.jors.2602465
Published online 25 July 2007

Keywords: stochastic vehicle routing; metaheuristics; time-window constraints

1. Introduction earliness and lateness penalties as part of the objective and/or

by constraining the probability of delivering outside the time
The performance of a vehicle-routing system—or any distri-
window, including the allowable total route duration.
bution and delivery system—is critically dependent on the
Because of the complexity of the VRP, the vast majority
time it takes to travel between destinations. Travel times
of VRP models are deterministic in nature. Gendreau et al
become even more important when the deliveries are to be
(1996) presented a review of the stochastic vehicle-routing
scheduled within specified time windows. Still, most of the lit-
(SVRP) literature, identifying a number of papers that
erature on the vehicle-routing problem (VRP) treats the travel
consider stochastic demands and/or stochastic customer
time to be a known, linear function of the distance travelled,
realization (ie each customer has some probability of realizing
such that minimizing the total distance is equivalent to mini-
demand); however, very few consider stochastic travel times,
mizing the travel time.
despite the practical relevance of nondeterministic travel
Many applications of VRPs include time windows (see, eg
times. Cook and Russell (1978) used a simulation model to
Golden and Assad, 1988). These scheduling constraints can
evaluate the suitability of deterministically generated routes
either be hard (no violation of the time window is allowed) or
on the stochastic problem. However, the first model that
soft (early or late deliveries are allowed, although incurring
explicitly incorporated stochastic travel times into the VRP
some cost). In practice, time windows tend to be soft, as travel
was presented by Laporte et al (1992). They considered two
times are generally stochastic and not precisely predictable.
versions of the problem: (1) using chance-constrained pro-
It is usually acceptable, to some extent, to deliver outside the
gramming, which includes a constraint to ensure some min
time window, although large deviations may be considered
imum probability of completing a route within a given time
to be a breach of service. Thus, to ignore the randomness
frame, and (2) stochastic programming with recourse, which
of travel times can significantly impact the performance of a
incorporates a linear penalty function for a route exceeding
vehicle-routing system (see Noland and Polak (2002) for a
the time limit. Lambert et al (1993) incorporated a constant
review of the theoretical and empirical issues in travel-time
penalty for a route exceeding the time limit, and treated the
variability).
stochastic travel time between two customers with a proba-
In this paper, we generalize the VRP to include soft time
bility that the route will be congested, thus taking more time
windows as well as stochastic travel times. As with the classic
to travel the route. The models developed by Kenyon and
VRP, our objective is to minimize the number of vehicles
Morton (2003) focus on the length of the longest route by
required as well as the distance travelled; however, we also
minimizing the expected completion time of all vehicles or
consider consequences that may arise due to the stochas-
by maximizing the probability of completing the routes by
ticity of the problem. This may be in the form of expected
a given time. They presented a branch-and-cut scheme em-
bedded within a sampling-based procedure to solve the SVRP.
∗ Correspondence: TL Urban, Operations Management, The University of While a great deal of research has been conducted on the
Tulsa, 600 South College Avenue, Tulsa, Oklahoma 74104-3187, USA. VRP with hard time windows (see reviews by Bräysy and
E-mail: [email protected] Gendreau, 2005a, b), far less has addressed the soft time-
 RA Russell and TL Urban—Vehicle routing with soft time windows 1221

window problem (VRPSTW). Sexton and Choi (1986) anal- 2. Model development
ysed the single-vehicle pickup and delivery problem with soft
To describe the VRP with soft time windows and stochastic
time windows using a Bender’s decomposition approach in
travel times, let G = (C, A) be a directed graph with a vertex
conjunction with a route-improvement heuristic. Ferland and
set C = {c0 , c1 , . . . , cn } and an arc set A = {(i, j)|i, j ∈ C,
Fortin (1989) presented a routing problem with ‘sliding’ time
i = j}. Vertex c0 represents the depot at which V identical
windows, such that the time windows can be modified as long
vehicles with known capacity, Q, are based; the remaining
as they remain within a given range. Min (1991) formulated
vertices of C represent customer locations with demand, qi ,
a mixed-integer goal programming model to address the mul-
service time, si , and soft time windows, [ei , li ], where ei (li )
tiobjective nature of the VRPSTW. An optimization-based
is the earliest (latest) that service can begin without penalty.
heuristic was developed by Koskosidis et al (1992), incorpo-
Each arc has an associated nonnegative distance, dij , and a
rating a generalized assignment procedure to assign customers
nonnegative, random travel time, tij .
to vehicles and a travelling-salesman algorithm to schedule
The SVRPSTW involves the assignment of customers to
the final routes. Balakrishnan (1993) developed an approach
routes such that (i) every route starts and ends at the depot,
based on simple heuristics that included a penalty for devi-
(ii) every customer is visited exactly once by exactly one ve-
ations from the time-window constraint, but only within a
hicle, and (iii) the total demand of any vehicle route does not
given maximum allowable violation; the vehicle must wait if
exceed the vehicle capacity. Similar to the VRPSTW with
arriving before that time. Taillard et al (1997) used a tabu-
known travel times, we consider multiple objectives in that
search metaheuristic exploiting an adaptive memory, identi-
we wish to minimize the number of vehicles required, the
fying better results for the soft time-window problem as well
total distance travelled, and the expected penalties incurred
as some new best-known solutions to the hard time-window
(both earliness and tardiness) from violating the time window
problem.
of each customer. Of course, the priorities of each of these
A parallel tabu-search heuristic for the dynamic VRPSTW
objectives will depend on the particular application; in this
was developed by Gendreau et al (1999), in which a number
paper, we assume priorities in order of vehicles required, dis-
of service requests are revealed after the vehicle has started
tance travelled, and time-window penalties incurred and min-
servicing its route. Ioannou et al (2003) presented a method
imize a weighted average of these objectives (in the compu-
of solving Balakrishnan’s formulation of the problem by
tational analyses, though, we will evaluate the effects of al-
applying the nearest-neighbourhood heuristic on example
tering the relative priorities of distance and penalties). The
sets created by a problem generator. Chiang and Russell
distinguishing feature of this problem lies in the stochastic
(2004) developed a tabu-search metaheuristic to solve the
nature of the travel times; even if the routes are assigned such
Balakrishnan formulation; their implementation included,
that the ‘expected’ arrival time falls within the time window
among other aspects, an advanced recovery strategy to iden-
of a customer, the variability in travel times results in some
tify superior results. The VRPSTW with stochastic demand
probability of arriving outside the time window and incurring
and stochastic customer realization was presented by Mak
an associated penalty.
and Guo (2004); they used a genetic algorithm to solve the
Chiang and Roberts (1980) developed an empirical model
problem. Recently, Ibaraki et al (2005) used a piecewise-
to predict the transit time for regular-route, less-than-truckload
linear penalty function when violating the time-window
trucking between two points. They found that a minimum
constraint and assumed a fixed number of vehicles; their
travel time exists (using reasonable speeds) and that the dis-
solution methodology is based on variants of multistart local-
tribution tends to be skewed toward longer travel times. Con-
search techniques with a dynamic-programming algorithm to
sequently, the shifted gamma distribution provides reasonable
schedule the final routes.
estimates of the travel time, with parameters determined by the
Despite the practical relevance of stochastic travel times
length of haul. We will apply their findings to the SVRPSTW
and of customer time windows for the VRP, no research
by assuming the time required to travel one unit distance, u, is
has appeared in the VRP literature that incorporates both
a three-parameter, shifted gamma (, , ) distribution—see,
of these aspects. In this paper, we formulate a model for
for example, Johnson et al (1994)—with probability density
the VRP with soft time windows and stochastic travel times
function, f (u), and cumulative distribution function, F(u):
(SVRPSTW). Closed-form expressions for the stochastic
(Erlang) travel times and for the expected earliness and tardi- (u − )−1 e−(u−)/
f (u) = for u >  0 (1)
ness time-window penalties are presented in the next section.  ( − 1)!
In Section 3, a tabu-search metaheuristic with a postoptimiza-  −1 
tion routine is developed to solve the problem. In Section 4,  (u − )r
F(u) = 1 − e−(u−)/ (2)
test problems from the literature are solved to illustrate the
r =0
r
r !
performance of the metaheuristic and to gain insight into the
problem. Finally, we close the paper with some general com- where  ∈ Z+ (ie a positive integer),  ∈ R+ (ie a positive real
ments, including the benefits and limitations of the proposed number), and  ∈ R+ 0 (ie a nonnegative real number). While
model. we must restrict our analysis to the Erlang distribution—a
1222 Journal of the Operational Research Society Vol. 59, No. 9

special case of the gamma distribution in which the shape Penalty

parameter, , can only take on integer values—in order to
maintain tractability, we lose little flexibility by doing so.
This distribution is particularly relevant for the VRP as
it is restricted to nonnegative values (obviously, travel time
cannot be negative), it is a very ‘rich’ function (it can take on
a variety of shapes, ranging from the exponential distribution Time
for  = 1, to the ‘near-normal’ as  → ∞), and it has a small ei li

probability of large values (an accident in the roadway or a

Figure 1 Illustration of proposed penalty function.
flat tire, for example, can disrupt the delivery). Furthermore,
the additive property of the gamma distribution allows us to
extend the unit travel time, u, to the time to travel between routes, were already determined). Taguchi’s approach explic-
two customers, tij ; that is, the time required to travel between itly considers the failure to meet customer’s expectations and
these two points is tij ∼ gamma(dij , , dij ). The mean of a assumes that larger deviations cause larger losses (hence, the
gamma (, , ) distribution is  =  + , so the expected quadratic function). While the quadratic functional form is
time required to travel from customer i to customer j is dij  + not necessarily an exact representation of the earliness and
dij  = ( + )dij ; therefore, minimizing the total expected tardiness penalties in all situations, it has been widely used in
travel time is equivalent to minimizing the distance travelled. practice and in research (if desired, a generalized loss func-
The arrival time, i , at any customer will then be the cu- tion can be found in Stevens and Baker, 1994).
mulative travel time to customer i (ie all arcs taken from the These penalties will likely be different for early and tardy
depot to the customer) plus the sum of the service times of all arrivals and may, if desired, be different for each customer; we
preceding customers as well as the waiting time at all previous will use the notation a0i , a1i , and a2i , for the constant, linear,
customers. Furthermore, the travel time can be broken down and quadratic penalty coefficients, respectively, for early ar-
into its stochastic (represented by a two-parameter gamma rivals, and b0i , b1i , and b2i for tardy arrivals. We do not place
distribution) and deterministic (the total amount of the shift) limits on the extent of the time-window violations; however,
components. Let Pik be the set of customers served before the quadratic loss function is effective in restricting the level
customer i on the same route k, including the depot (this al- of violation due to the ‘increasing-at-an-increasing-rate’ prop-
lows a service time associated with the depot before the ve- erty (as illustrated in Figure 1).
hicle’s departure time, such as loading and securing cargo, An earliness penalty, i , is incurred if the arrival time at
inspecting and servicing the vehicle, acquiring and reviewing customer i (after waiting, if justified) precedes the earliest
necessary paperwork, etc), then: time that service can begin without penalty, ei . Since the
service times are deterministic, we evaluate the cumulative
  travel time, i , over the available time for travel, E i = ei −
i = i +  d j, j  + (s j + w j ) (3)
wi − d j, j+1 − (s j + w j ):
j∈Pik j∈Pik
 Ei

where i ∼ gamma( , ),  =  j∈Pik d j, j  , and  ∈ Z+ , i = a0i [F(E i )] + a1i (E i − i ) f (i )di
 ∈ R+ , and  ∈ R+  0
0 ; j is the customer immediately fol-  Ei
lowing j in route k; and w j is the length of time the vehicle + a2i (E i − i )2 f (i )di (4)
waits at customer j before beginning service. 0

A penalty is assessed if the vehicle arrives at the customer After some calculus and algebraic manipulation, we obtain:
outside the service time window. We will allow a variety of  
−1
 
penalty structures: −E i / (E i /)r
i = a0i 1 − e + a1i  [(E i /) −  ]
r =0
r !
(i) a constant penalty, reflecting fixed compensation to the
customer if a delivery guarantee was agreed upon (eg a  
−1
 
r (E i /) −r

 E i /
clause in a supply contract), +  + e
( − r )!
(ii) a penalty per unit time the delivery is early or tardy (a r =1
linear loss function), and ⎡ [(E /) −  ]2 +  ⎤
i 
(iii) a penalty based on the Taguchi loss function (a quadratic ⎢ ⎥
⎢ −  [( + 1) + ( − 1)(E i /)] ⎥
loss function). 2⎢ ⎥
+ a2i  ⎢ ⎥
⎢  ⎥
The quadratic loss function has become quite popular in the ⎣  
−1  ⎦
r (r −1)(E i /) +1−r E i /
quality control and product design fields, but has not yet ap- + ( +1−r )!
e
r =2
peared in the vehicle-routing literature (Dumas et al, 1990,
considered convex inconvenience costs within the time win- Similarly, a tardiness penalty, i , is incurred if the arrival
dows but assumed the routes, and customer sequence on the time exceeds the latest time that service can begin without
 RA Russell and TL Urban—Vehicle routing with soft time windows 1223

penalty, li . In the same way, we evaluate i , over the residual Russell (2004) for the deterministic VRPSTW problem. The
time for travel, L i = li − wi − d j, j+1 − (s j + w j ): mean of the travel time between two points, t¯ij = ( + )dij ,
 ∞ is used; the departure time of each vehicle from the depot
i = b0i [1 − F(L i )] + b1i (i − L i ) f (i )di is set equal to e j − t¯0 j , where j is the first customer on
 ∞ Li the route; and no wait times are allowed before each of the
remaining customers on a route. This method was applied to
+ b2i (i − L i )2 f (i )di
Li the soft time-window problems developed by Balakrishnan
 −1  (1993), which specified a zero tolerance for early or late
 (L i /)r
= b0i e−L i / arrivals. Thus, applying the deterministic VRPSTW method
r =0
r! results in an initial solution for which each arrival is expected
 

−1
  to arrive within the specified time window of a customer.
r (L i /) −r


+ b1i   + 
e−L i / The approach of Chiang and Russell (2004) is a
r =1
( − r )! reactive tabu-search method that performs route construc-
 tion in conjunction with route improvement. It employs a
+ b2i 2  [( + 1) + ( − 1)(L i /)] mixed-neighbourhood search procedure utilizing both node
exchanges and edge exchanges. It also uses search intensifi-


−1
  cation and diversification strategies. The advanced-recovery
r (r − 1)(L i /) +1−r


+ e−L i / (5) feature employs a pool of elite solutions that serve as

r =2
( + 1 − r )! the starting point for subsequent search paths. While this
achieves a high-quality solution, the method proposed for the
Each vehicle may also have a latest finish time, l0 , such that
VRPSTW with stochastic travel times can utilize any initial,
the vehicle is expected to return to the depot within a given
feasible solution.
time frame. We can deal with this in one of two different
ways: (1) by including a penalty in the objective function if 3.2. Improvement using tabu search
the time is violated, in the same manner as the customers,
and (2) by including a constraint that ensures the probability 3.2.1. Tabu search neighbourhood moves. The computa-
of returning by time l0 is at least a certain level: tional requirements of the VRPSTW with stochastic travel
 −1  times suggest a more basic and streamlined tabu-search ap-
 Lr proach. The implemented tabu-search procedure is based
P(0  l0 ) = 1 − F(L 0 ) = k
e−L i /  (6) on the reactive tabu-search method that Chiang and Russell
r =0
r
r !
(1997) developed for the VRP with hard time windows.
The -interchange moves (Osman, 1993) consist of three
3. Solution methodology possible node exchanges—insert, delete, and swap:
In order to study the computational performance of the pro-
posed VRPSTW with stochastic travel times, we develop a • The format of the insert movement is insert(r1 , r2 , i, j). It
tabu-search solution procedure. The NP-hard nature of the moves a node from route r1 position i to route r2 position
problem calls for a metaheuristic approach for an effective j. We label the insert movement as a (0, 1) movement.
solution method. Simple or local search heuristics yield • The format of the delete movement is delete(r1 , r2 , i, j). It
inferior solutions and mathematical optimization (with the ex- moves the node from route r2 position j to route r1 position
ception of column generation methods) cannot solve realistic- i. The delete movement is also called a (1, 0) movement.
sized problems. Other researchers have found that tabu search • The format of the exchange (swap) movement is
has been effective in solving VRPs, and the majority of the exchange(r1 , r2 , i, j). It exchanges the node in route r1
reported applications use tabu search. Other metaheuristic position i and route r2 position j. The exchange movement
approaches, such as evolutionary algorithms, have also been is labelled as a (1, 1) movement.
successful. We also compare the results to a previous deter-
ministic travel time approach to assess the potential benefits The current neighbourhood structure was effective in solving
of incorporating stochastic travel times in the solution pro- VRPs with hard time windows as demonstrated in Chiang and
cedure. The proposed approach consists of three phases: (i) Russell (1997).
obtaining an initial solution, (ii) improving the solution via 3.2.2. Tabu search list structure. A node is said to have tabu
tabu search, and (iii) postprocessing to optimize the wait status with respect to a route if it cannot be switched back to
times before each customer. that route. Let us define TABU(u, r ), ∀u ∈ {1, . . . , |C|}, ∀r ∈
{1, . . . , |R|}, where |C| is the cardinality of the set of nodes
3.1. The initial solution
and |R| is the cardinality of the set of routes. TABU(u, r )
The initial solution is obtained using the tabu-search pro- records the tabu status of node u in route r, after being moved
cedure with advanced recovery developed by Chiang and to another route; that is, if the TABU(u, r ) = x, then node
1224 Journal of the Operational Research Society Vol. 59, No. 9

u cannot return to route r until iteration x finishes. Let t be solution space and exploiting promising regions. In our
a tabu list size. If node u switches from route r to r  at the reactive tabu scheme, the size of the list is learned in an
current iteration, then TABU(u, r ) is updated as follows: automated way by reacting to the occurrence of repeated
solutions. Thus, if solutions are found to be repeating too
TABU(u, r ) = iter + t where iter is the often, the parameter t is increased to discourage subsequent
current iteration number repetitions. In addition, if the search appears to be repeating
solutions excessively often, then the search is diversified by
Node u is tabu w.r.t. route r if TABU(u, r ) > iter making a number of random moves proportional to a moving
average of the repetition cycle length.
Hence, for operators (0, 1) or (1, 0), if u is to be switched
from r1 to r2 and if TABU(u, r2 ) > iter, then the move is 3.2.4. Computational issues. The evaluation of earliness
tabu; otherwise, it is not tabu. For operator (1, 1), if node and tardiness penalties, i and i , in Equations (4) and
u in route r1 is exchanged with node u  in route r2 , and if (5) may involve very large integers resulting from factorial
TABU(u, r2 ) > iter or TABU(u  , r1 ) > iter, then the move is and exponential calculations. The resulting numbers can
tabu; otherwise, it is not. When a node is tabu, it cannot exceed the largest integer that can be represented on a 32-bit
return to its original route in the next t iterations if it switches computer. For this reason, all moves and their penalty evalu-
from that route to another unless the aspiration criterion is ations were calculated by using natural log transformations.
satisfied. The aspiration criterion is as follows: Tabu status These transformations increase the computational require-
can be overridden if the switch of a tabu-prohibited node can ments but allow evaluations that otherwise would not be
result in a new solution with an objective value less than the possible.
best solution found so far.
3.3. Postprocessing
3.2.3. Reactive search and diversification. We employ a re-
active scheme to dynamically vary the prohibition parameter, The deterministic starting point solution method assumes
t, in order to obtain a better balance between exploring the that the departure time at the depot for each route is equal to

STEP 1: Generate an initial solution using deterministic tabu search with advanced recovery.
The deterministic travel times are set equal to the expected Erlang-distributed
travel times, tij =(+)dij

STEP 2: Initialize tabu search parameters, iter and maxiter, and reactive parameters,
solrepetition and cyclemax
Calculate mincost = 1000V + 0.5D + 0.5P

STEP 3: While (iter ≤ maxiter ), do STEP 4 thru STEP 5

STEP 4: iter = iter + 1 // Search pairs of routes for exchange

Reset curcost = 9999999
For nonempty routes r1 and r2
choose customer i in route r1 and customer j in route r2 that can be switched
// Feasible after switching; i.e., vehicle capacities for r1 and r2 are satisfied
calculate the expected time window penalty value, P =Σi (
i+i)
calculate the number of vehicles required, V, and the total distance traveled, D
evaluate the expected cost function, tempcost = 1000V + 0.5D + 0.5P
if ((either i or j is in tabu w.r.t. r1 or r2) and (tempcost < mincost ))
//(aspiration rules)
or ((i and j are not tabu) and (tempcost < curcost ))
then curcost = tempcost

STEP 5: Switch customers i and j

If curcost < mincost then
update mincost and best solution
update tabu status of node i in route r1 and node j in route r2
Check for repeated solutions, update solrepetition, and adjust tabu list length if
necessary

STEP 6: Perform postprocessing on the final tabu search solution using a GRG2 nonlinear
solver to optimize the waiting times before each customer

Figure 2 Pseudo code for the tabu-search algorithm.

 RA Russell and TL Urban—Vehicle routing with soft time windows 1225

e j − t̄0 j , where customer j is the first customer on the short resulting in route lengths of typically no more than 12
route. Thus, there is no wait time before the first customer. customers. For each of the base test problems, additional test
The proposed tabu-search implementation for the stochastic problems were generated using varying parameters for the
problem makes the same assumption. To facilitate computa- shifted gamma distribution of travel times such that (i) the
tional tractability, the wait times before each of the remaining expected travel time between two points is equal to Solomon’s
customers on a route are also assumed to be zero. There- Euclidean travel times, and (ii) different coefficients of vari-
fore, a postprocessing procedure is proposed to determine ation are generated (CV = / = 0.25, 0.50, 0.75, and 1.00)
optimal wait times before each customer—including the de- to provide different levels of ‘stochasticity.’ The , , and 
parture from the depot—for each route provided in the final parameter values associated with the above four coefficients of
tabu-search solution. variation are (1, 0.25, 0.75), (1, 0.50, 0.50), (1, 0.75, 0.25),
The penalty calculations in Equations (4) and (5) involve and (1, 1.00, 1.00), respectively. Equal penalty weights of 1.0
the variables, E i and L i , respectively. Given zero wait times, and 1.0 are assigned for early and tardy deliveries, respec-
E i and L i are based on constants in the calculation of penalties tively, and the quadratic loss function is employed to restrict
in the tabu search. However, explicitly considering wait times excessive time-window violations (a2i = b2i = 1, a0i = b0i =
in both a1i = b1i = 0).
We used two different objective functions in the experi-
E i = ei − wi − d j, j  − (s j + w j ) ments. If we let V = number of vehicles required, D = total
and distance travelled (D = (i, j)∈N dij , where N = the set of
arcs utilized in the solution), and P = time-window penalties
L i = li − wi − d j, j  − (s j + w j ) incurred (P = i (i + i )), then V ? D and V ? P in our
implementation. With V being the highest priority objective,
creates a set of variables, w j , that must be evaluated in min-
we solve the test problems with different weights for D and P.
imizing the earliness and tardiness penalties in Equations (4)
In the first experiment, we minimize 0.5D + 0.5P reflecting
and (5). Given a specific route sequence from the tabu-search
an equal weighting of travel distance and time-window penal-
procedure, final values of the wait times before each cus-
ties. In the second experiment, we used 1.0D +0.2P to reflect
tomer are identified using the generalized reduced gradient
an emphasis on minimizing travel distance at the expense of
method (GRG2) in Excel Solver. The post optimization of the
time-window violations. The weights could be normalized in
route times has been found to generate significant reductions
a real-world application by taking into account the cost of
in penalties (an average of over 20% in all the test problems
vehicles driven per mile, such as $2.00, and the approximate
considered in the next section).
cost of violating a time window. Evans and Lindsay (2002,
The pseudo code presented in Figure 2 summarizes the
p. 377) provide an example of how such a cost could be
metaheuristic approach for solving the VRP with soft time
estimated using a Taguchi cost function. These normalizing
windows and stochastic travel times. The reader interested in
weights would clearly vary by application.
more information concerning the tabu-search metaheuristic
Table 1 presents the solution results in terms of the number
with advanced recovery used for the deterministic starting
of vehicles, the total distance travelled, and the total earli-
solution can find more details in the paper by Chiang and
ness and tardiness penalties for the deterministic starting point
Russell (2004).
(see Section 3.1) and for the three-step, stochastic travel time
approach. The CPU time (in minutes) is also reported for each
4. Computational results
using a 3.2 GHz Pentium IV processor. The starting point so-
To study the performance of the proposed metaheuristic lution times ranged from 2.37 to 2.91 min. The incorporation
approach, we tested the methodology on test problems from of stochastic travel times is computationally intensive and re-
the time-window literature. Four base test problems from quired from 22.2 to 493.0 min to solve the test problems. Be-
the well-known Solomon (1987) hard time-window prob- cause of the computational effort, only 500 iterations were
lems were modified by Balakrishnan (1993) to allow various executed in the tabu-search metaheuristic. As is apparent from
amounts of time-window violation. We chose the data set the table, the CPU times increase considerably as the coeffi-
allowing zero wait times and zero tardiness. Chiang and cient of variation of the travel time increases.
Russell (2004) also experimented with the R101, R102, R103, Table 1a provides the solutions for the instances in which
and R109 test problems in their tabu-search approach to the the travel distance and time-window penalties are equally
VRPSTW. Our objective is to explore solution quality and weighted. The stochastic travel time approach was able to
computational requirements of the proposed stochastic travel reduce the number of vehicles required in 10 of the 16 test
time approach as well as how the solutions compare to a problems; no change was realized in the other six instances.
deterministic travel time approach using expected travel times. The total distance travelled was reduced in 11 test problems,
The R1 data set from the Solomon test problems consists with minimal increases in the other five (at most, a 2.3%
of customers whose location is randomly generated according increase). The time-window penalties were reduced in 12 of
to a uniform distribution. The scheduling horizon is relatively the 16 instances; the increases were primarily in the R101
1226 Journal of the Operational Research Society Vol. 59, No. 9

Table 1 Solutions to test problems

Deterministic solution Stochastic solution
Problem Coeff. of var. Vehicles Distance Penalty CPU time Vehicles Distance Penalty CPU time
(a) Objective function = 1000V + 0.5D + 0.5P
R101 0.25 22 2439.13 56.79 2.37 18 2104.32 212.90 32.03
R101 0.50 22 2439.13 400.46 2.37 18 2088.24 646.84 98.17
R101 0.75 22 2439.13 1250.42 2.37 18 2164.59 1910.68 245.24
R101 1.00 22 2439.13 2740.87 2.37 20 2126.15 2715.17 234.79
R102 0.25 19 1957.72 53.80 2.59 17 1792.34 161.54 40.19
R102 0.50 19 1957.72 363.36 2.59 17 1826.80 308.07 127.70
R102 0.75 19 1957.72 1096.07 2.59 17 1871.34 920.66 257.76
R102 1.00 19 1957.72 2356.59 2.59 17 1872.55 2111.35 318.90
R103 0.25 14 1465.72 22.43 2.91 14 1451.60 13.10 48.43
R103 0.50 14 1465.72 188.44 2.91 14 1466.74 142.02 173.30
R103 0.75 14 1465.72 627.25 2.91 13 1462.45 556.55 339.96
R103 1.00 14 1465.72 1416.89 2.91 13 1596.29 1407.90 392.86
R109 0.25 12 1218.74 8.78 2.82 12 1216.62 6.25 23.13
R109 0.50 12 1218.74 54.29 2.82 12 1224.31 28.36 124.06
R109 0.75 12 1218.74 177.02 2.82 12 1230.82 110.43 357.32
R109 1.00 12 1218.74 422.91 2.82 12 1246.88 228.74 493.02
(b) Objective function = 1000V + 1.0D + 0.2P
R101 0.25 22 2439.13 56.79 2.37 17 1806.24 667.44 22.16
R101 0.50 22 2439.13 400.46 2.37 17 1876.64 513.63 87.71
R101 0.75 22 2439.13 1250.42 2.37 16 1820.28 2255.89 226.78
R101 1.00 22 2439.13 2740.87 2.37 17 1766.88 3173.95 312.11
R102 0.25 19 1957.72 53.80 2.59 16 1585.93 225.82 34.05
R102 0.50 19 1957.72 363.36 2.59 15 1695.51 977.96 114.19
R102 0.75 19 1957.72 1096.07 2.59 16 1661.68 1138.40 261.21
R102 1.00 19 1957.72 2356.59 2.59 16 1685.30 2175.82 351.45
R103 0.25 14 1465.72 22.43 2.91 12 1465.64 375.41 22.43
R103 0.50 14 1465.72 188.44 2.91 12 1398.88 671.46 150.33
R103 0.75 14 1465.72 627.25 2.91 13 1372.24 689.17 317.17
R103 1.00 14 1465.72 1416.89 2.91 12 1406.57 2003.59 398.44
R109 0.25 12 1218.74 8.78 2.82 12 1215.22 11.54 22.30
R109 0.50 12 1218.74 54.29 2.82 12 1216.82 42.95 105.00
R109 0.75 12 1218.74 177.02 2.82 12 1219.31 138.23 341.65
R109 1.00 12 1218.74 422.91 2.82 12 1225.11 305.16 484.00

data set which contains the tightest time windows. Examining weights were used. Figure 3 shows the tabu-search trajectory
the effects of stochasticity in terms of the coefficient of vari- for the R101 (CV = 0.25) test problem in Table 1b. The step
ation, it appears that there is a tendency for the distance trav- function aspect of the graph illustrates the effect on the ob-
elled to increase slightly with an increase in the variability jective function of reducing the number of vehicles required
of travel times, and for the time-window penalties to realize as well as the effect of the postprocessing procedure.
substantial increases. One advantage of a stochastic travel time approach is the
In Table 1b, the solutions of the same test problems ability to explicitly assess the probability of events. In Table
are reported and compared to the starting solution with an 2, we report results pertaining to a tardy return to the depot
objective function that emphasizes vehicles and distance (the mean and maximum probability of exceeding l0 for all
relative to time-window penalties. In this case, the solu- routes in each test problem). The probability of route comple-
tion quality is improved (fewer vehicles and less distance tion was not a criterion in the objective function, nor was it
travelled) while time-window penalties are increased. The included as a constraint (it could, however, be incorporated in
stochastic travel-time approach reduced the number of ve- the stochastic travel time approach); still the majority of the
hicles in 12 of the test problems, decreased the distance routes for both the starting point and the stochastic solution
travelled in 14 instances, and increased the time-window had probabilities less than 0.0001. It is interesting—as pre-
penalties in 12 instances. As would be expected, the effect on sented in Table 2a—that the stochastic travel time approach
the solution is more pronounced in this case than when equal achieves a lower mean probability of being tardy (0.037
 RA Russell and TL Urban—Vehicle routing with soft time windows 1227

25000 Table 2 Probability of tardy return to depot

Deterministic solution Stochastic solution
24000 Coeff.
Problem of var. Mean Maximum Mean Maximum
Objective Function Value

23000
(a) Objective function = 1000V + 0.5D + 0.5P
22000
R101 0.25 0.23 4.35 0.53 6.12
R101 0.50 1.81 18.94 3.04 18.41
R101 0.75 3.86 27.63 5.71 24.43
21000
R101 1.00 5.82 32.67 5.87 24.29
20000 R102 0.25 2.38 21.69 2.69 26.63
R102 0.50 6.26 23.58 4.20 25.59
19000 R102 0.75 8.94 38.23 4.81 24.87
R102 1.00 10.75 40.59 7.54 28.67
18000 R103 0.25 1.39 14.66 0.72 4.22
0 100 200 300 400 500
R103 0.50 5.26 29.12 4.57 17.69
Iterations
R103 0.75 8.93 35.22 6.70 20.62
Figure 3 The tabu-search trajectory of the objective function R103 1.00 11.71 38.45 10.16 30.68
value. R109 0.25 0.00 0.02 0.00 0.03
R109 0.50 0.28 3.00 0.36 3.43
R109 0.75 1.22 9.86 0.95 10.19
versus 0.045) as well as a lower average maximum prob- R109 1.00 2.65 16.21 1.97 15.83
ability (0.176 versus 0.221). In Table 2b, as the emphasis Mean 4.47 22.14 3.74 17.61
shifts from the time-window penalty to the distance travelled,
(b) Objective function = 1000V + 1.0D + 0.2P
the stochastic travel time approach achieved a slightly worse R101 0.25 0.23 4.35 3.08 40.50
mean (0.056 versus 0.045) and maximum (0.286 versus R101 0.50 1.81 18.94 4.44 37.62
0.221) probability. In both cases, the likelihood of a tardy R101 0.75 3.86 27.63 5.49 38.22
return increases as the variability of travel times increases. R101 1.00 5.82 32.67 8.34 38.29
The appropriate trade-off between vehicle and distance on R102 0.25 2.38 21.69 3.75 37.08
the one hand, and tardy route completion on the other hand, R102 0.50 6.26 23.58 2.84 12.86
would depend on the particular application. R102 0.75 8.94 38.23 6.20 23.03
R102 1.00 10.75 40.59 7.98 28.77

5. Conclusions R103 0.25 1.39 14.66 4.70 28.28

R103 0.50 5.26 29.12 18.53 56.80
Delivery reliability—the on-time delivery of products and R103 0.75 8.93 35.22 6.48 29.75
services—is one of the major competitive dimensions of many R103 1.00 11.71 38.45 12.47 30.50
firms. Hill (1994) notes that delivery reliability is now an R109 0.25 0.00 0.02 2.26 27.10
order qualifier (a criterion that a firm must meet just to be R109 0.50 0.28 3.00 0.36 3.43
considered as a possible supplier) rather than an order winner. R109 0.75 1.22 9.86 1.06 10.19
R109 1.00 2.65 16.21 1.97 15.83
Initiatives such as quick-response distribution and just-in-time
delivery have become commonplace as part of the supply- Mean 4.47 22.14 5.62 28.64
chain strategy of many firms. The stochastic nature of the pro-
posed model allows a firm to incorporate delivery reliability
as part of their vehicle-routing and scheduling process. vehicle routes and total travel distance compared to a de-
In this paper, a vehicle-routing model is developed that terministic travel time approach that enforced the expected
incorporates soft time windows as well as stochastic travel arrival to occur during the allowed time window. With an
times. Closed-form expressions of the penalties for early and equal weighting of total distance and time-window penalty
tardy deliveries are included, and can reflect a constant, linear, in the objective function, the stochastic travel time approach
and/or quadratic cost penalty. Also, probabilities of returning also achieved a smaller quadratic time-window penalty in 12
to the depot—or to any particular customer—can be evalu- of the 16 test problems.
ated. A three-phase, metaheuristic is developed to assign and Despite their practical relevance, this is the first time both
sequence customers to routes and to determine the depar- random travel times and soft time windows—or even hard
ture time from the depot as well as any appropriate waiting windows—have been incorporated into a vehicle-routing
times at each customer. Computational results are provided model. In fact, very little VRP research has been conducted
for benchmark problems from the literature, and the effects of involving random travel times, most of the stochastic vehicle-
stochasticity on the problem are identified. Generally, the use routing literature has focused on stochastic demand (see,
of stochastic travel times yielded solutions requiring fewer eg, the Gendreau et al (1996) review). While this increases
1228 Journal of the Operational Research Society Vol. 59, No. 9

the complexity of the model, it is important to consider the Cook TM and Russell RA (1978). A simulation and statistical analysis
stochastic nature of the travel times since, even if the expected of stochastic vehicle routing with timing constraints. Decis Sci
arrival time falls within a customer’s time window, the vari- 9: 673–687.
Dumas Y, Soumis F and Desrosiers J (1990). Optimizing the schedule
ability in travel times results in some probability of arriving for a fixed vehicle path with convex inconvenience costs. Transport
outside the time window. Additionally, the quadratic loss Sci 24: 145–152.
function has become quite popular in the quality-control and Evans JR and Lindsay WM (2002). The Management and Control of
product-design fields, but has not yet appeared in the vehicle- Quality, 5th edn. South-Western: Cincinnati.
Ferland JA and Fortin L (1989). Vehicles scheduling with sliding time
routing literature. Taguchi’s approach explicitly considers
windows. Eur J Opl Res 38: 213–226.
the failure to meet customer’s expectations and assumes that Gendreau M, Guertin F, Potvin JY and Taillard É(1999). Parallel tabu
larger deviations from the target cause increasingly larger search for real-time vehicle routing and dispatching. Transport Sci
losses (hence, the quadratic function). 33: 381–390.
One of the limitations of this research is that we are fo- Gendreau M, Laporte G and Séguin R (1996). Stochastic vehicle
routing. Eur J Opl Res 88: 3–12.
cusing solely on the gamma distribution. While there is em-
Golden BL and Assad AA (1988). Vehicle Routing: Methods and
pirical evidence for its application in the transportation in- Studies. North-Holland: Amsterdam.
dustry, and it has a number of advantages for this particular Hill T (1994). Manufacturing Strategy: Text and Cases, 2nd edn.
problem (richness, nonnegativity, additive property, etc), it Irwin: Burr Ridge, IL.
also has some shortcomings: Ibaraki T, Imahori S, Kubo M, Masuda T, Uno T and Yagiura M
(2005). Effective local search algorithms for routing and scheduling
problems with general time-window constraints. Transport Sci
(i) We are assuming that the time to travel one unit dis- 39: 206–232.
tance is independent of the time to travel the next (this is Ioannou G, Kritikos M and Prastacos G (2003). A problem generator-
also implicit in the deterministic vehicle-routing models solver heuristic for vehicle routing with soft time windows. Omega
when it is assumed the travel time is proportional to the 31(1): 41–53.
Johnson NL, Kotz S and Balakrishnan N (1994). Continuous
distance). Univariate Distributions, Volume 1, 2nd edn. John Wiley & Sons:
(ii) We are also assuming that the distribution of the time New York.
to travel any one unit distance is the same as all others Kenyon AA and Morton DP (2003). Stochastic vehicle routing with
(however, this could be handled by multiplying the ap- random travel times. Transport Sci 37: 69–82.
propriate distance by a constant when it differs; eg if the Koskosidis YA, Powell WB and Solomon MM (1992). An
optimization-based heuristic for vehicle routing and scheduling
speed limit is 40 mph instead of 50 mph, simply multiply with soft time window constraints. Transport Sci 26: 69–85.
by 0.8). Lambert V, Laporte G and Louveaux F (1993). Designing collection
routes through bank branches. Comput Opl Res 20: 783–791.
Also, as we found with the test problems, the required com- Laporte G, Louveaux F and Mercure H (1992). The vehicle routing
putational effort for the stochastic problem is certainly greater problem with stochastic travel times. Transport Sci 26: 161–170.
than that for the deterministic problem (and increases as the Mak KL and Guo ZG (2004). A genetic algorithm for vehicle routing
problems with stochastic demand and soft time windows. In:
stochasticity of the problem increases); still, we are able to Jones MH, Patek SD and Tawney BE (eds). Proceedings of the
solve 100-customer problems using the proposed method- 2004 Systems and Information Engineering Design Symposium.
ology (other VRP models that incorporate stochastic travel Department of Systems and Information Engineering, University
time consider from 20 to 44 customers). of Virginia: pp 183–190.
Min H (1991). A multiobjective vehicle routing problem with soft
time windows: The case of a public library distribution system.
Socio-Econ Plann Sci 25: 179–188.
Noland RB and Polak JW (2002). Travel time variability: A review
References of theoretical and empirical issues. Transport Rev 22: 39–54.
Osman IH (1993). Metastrategy simulated annealing and tabu search
Balakrishnan N (1993). Simple heuristics for the vehicle routeing algorithms for the vehicle routing problem. Ann Opns Res
problem with soft time windows. J Opl Res Soc 44: 279–287. 41: 421–451.
Bräysy O and Gendreau M (2005a). Vehicle routing problem with time Sexton TR and Choi YM (1986). Pickup and delivery of partial loads
windows, Part I: Route construction and local search algorithms. with ‘soft’ time windows. Am J Math Mngt Sci 6(3–4): 369–398.
Transport Sci 39: 104–118. Solomon MM (1987). Algorithms for the vehicle routing and
Bräysy O and Gendreau M (2005b). Vehicle routing problem with scheduling problems with time window constraints. Opns Res
time windows, Part II: Metaheuristics. Transport Sci 39: 119–139. 35: 254–265.
Chiang WC and Russell RA (1997). A reactive tabu search Stevens DP and Baker RC (1994). A generalized loss function for
metaheuristic for the vehicle routeing problem with time windows. process optimization. Decis Sci 25: 41–56.
INFORMS J Comput 9: 417–430. Taillard É, Badeau P, Gendreau M, Guertin F and Potvin JY (1997).
Chiang WC and Russell RA (2004). A metaheuristic for the vehicle A tabu search heuristic for the vehicle routing problem with soft
routeing problem with soft time windows. J Opl Res Soc 55: time windows. Transport Sci 31: 170–186.
1298–1310.
Chiang YS and Roberts PO (1980). A note on transit time and
reliability for regular-route trucking. Transport Res 14B(1–2): Received January 2006;
59–65. accepted April 2007 after three revisions
Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.