Algorithms for the Vehicle Routing and Scheduling Problems with Time Window

Constraints

Marius M. Solomon

Operations Research, Vol. 35, No. 2. (Mar. - Apr., 1987), pp. 254-265.

Stable URL:
http://links.jstor.org/sici?sici=0030-364X%28198703%2F04%2935%3A2%3C254%3AAFTVRA%3E2.0.CO%3B2-E

Operations Research is currently published by INFORMS.

Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/about/terms.html. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained
prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in
the JSTOR archive only for your personal, non-commercial use.

Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/journals/informs.html.

Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.

The JSTOR Archive is a trusted digital repository providing for long-term preservation and access to leading academic
journals and scholarly literature from around the world. The Archive is supported by libraries, scholarly societies, publishers,
and foundations. It is an initiative of JSTOR, a not-for-profit organization with a mission to help the scholarly community take
advantage of advances in technology. For more information regarding JSTOR, please contact [email protected].

http://www.jstor.org
Fri Jan 25 09:47:51 2008
 ALGORITHMS FOR
THE VEHICLE ROUTING AND SCHEDULING PROBLEMS
WITH TIME WINDOW CONSTRAINTS
MARIUS M. SOLOMON
Northeastern University, Boston, Massachusetts
(Received February 1984; revisions received October 1984; March, October 1985; accepted December 1985)

This paper considers the design and analysis of algorithms for vehicle routing and scheduling problems with time window
constraints. Given the intrinsic difficulty of this problem class, approximation methods seem to offer the most promise
for practical size problems. After describing a variety of heuristics, we conduct an extensive computational study of their
performance. The problem set includes routing and scheduling environments that differ in terms of the type of data used
to generate the problems, the percentage of customers with time windows, their tightness and positioning, and the
scheduling horizon. We found that several heuristics performed well in different problem environments; in particular an
insertion-type heuristic consistently gave very good results.

A key element of many distribution systems is the

routing and scheduling of vehicles through a set
of customers requiring service.
scheduling. So far, this type of constraint has been
handled in an ad hoc manner, mostly by manual
adjustments to routing-based schedules.
The vehicle routing problem (VRP) involves the While the spatial problem of routing vehicles has
design of a set of minimum-cost vehicle routes, origi- been intensively studied in the literature (Bodin et al.),
nating and terminating at a central depot, for a fleet very little work has been done on the VRSPTW, which
of vehicles that services a set of customers with known encompasses both spatial and temporal aspects of
demands. Each customer is serviced exactly once and, vehicle movements. Almost all approaches to the rout-
furthermore, all the customers must be assigned to ing problem suffer from the limitation that they do
vehicles without exceeding vehicle capacities (see not consider time window constraints. The existing
Bodin et al. 1983 for a comprehensive survey). In the literature on the problem with time windows has dealt
vehicle routing and scheduling problem with time mainly with case studies (Pullen and Webb 1967,
window constraints (VRSPTW), these issues must be Knight and Hofer 1968, and Cook and Russell 1978).
addressed under the added complexity of allowable The latter authors report computational experience
delivery times, or time windows, stemming from the with a k-optimal improvement heuristic, M-TOUR
fact that some customers impose delivery deadlines (Russell 1977) for an actual problem involving 163
and earliest-delivery-time constraints. customers, 4 vehicles and only 15% time-constrained
In the presence of time windows, the total routing customers. The computational requirements of Rus-
and scheduling costs include not only the total travel sell's method are, however, very large (1.27 minutes
distance and time costs considered for routing prob- of IBM 3701168 CPU time).
lems, but also the cost of waiting time incurred when The rest of the literature has been directed primarily
a vehicle amves too early at a customer location or at special structures. Christofides, Mingozzi and Toth
when the vehicle is loaded or unloaded. (1981) discuss state space relaxations for dynamic
The VRSPTW has emerged as an important area programming approaches to the traveling salesman
for progress in handling realistic complications and problem with time windows, while Baker (1983), and
generalizations of the basic routing model (Schrage Baker and Rushinek (1982) present a branch-and-
1981, Bodin et al.). Time windows arise naturally in bound algorithm for a new, time-oriented formulation
problems faced by business organizations that work of the problem. Some progress has been made in
on fixed time schedules. Specific examples include the multiperiod routing problem (Federgruen and
bank deliveries, postal deliveries, industrial refuse col- Lageweg 1980, Fisher et al. 1982). In this problem,
lection, dial-a-ride service, and school bus routing and the time windows are full days, and a service activity

Subject classification: 632 heuristic programming, 83 1 transportation, 483 distance algorithms.

Operations Research 0030-364X/87/3502-0254 $01.25

Vol. 35, No. 2, March-April 1987 254 0 1987 Operations Research Society of America
 The Vehicle Routing and Scheduling Problem / 255
must occur on a specified number of days of the The service at a customer, say i, i = 1, . . . , n,
planning horizon. involving pickup and/or delivery of goods or services
Other classes of problems with time windows have for s, units of time, can begin at time b,, within a time
received increased attention lately: the dial-a-ride window defined by the earliest time e, and the latest
problem (Psaraftis 1983, Sexton and Bodin 1985a,b) time 1, that customer i will permit the start of service.
and the school-bus scheduling problem. Orloff (1976) Hence, if a vehicle travels directly from customer i to
presents a matching-based heuristic, while Swersey customer j and anives too early at j, it will wait, that
and Ballard (1982) discuss an optimal approach to is, 4 = max(eJ, b, + s, + t,), where t, is the direct
this problem with the time window discretized. travel time between i and j. Note that the times b, for
Desrosiers, Soumis and Derochers (1983a,b) develop i = I , . . . , n, at which services begin are decision
exact methods for this problem. One algorithm uses a variables.
column generation approach in which the columns We assume that the cost of direct travel from cus-
are generated by using a shortest-path-with-time- tomer i to customer j, is given by cIj = p,d, + Pz(bJ-
windows algorithm (Desrosiers, Pelletier and Soumis b,), p, 3 0, pz z 0, as defined by the direct distance d,
1982). Two other branch-and-bound algorithms in- from customer i to customer j. For example, if p, = 0
volve relaxations of the time-window-related con- and p2 = 1, one seeks to minimize total schedule time.
straints. We also assume that the number of vehicles used is
Problems with time windows are from a computa- free, i.e., the fleet size is determined simultaneously,
tional complexity perspective quite difficult. Since the using the best set of routes and schedules. The vehicles
(VRP) is NP-hard, by restriction, the VRSPTW is NP- leave the depot, denoted by node 0, at time eo, at the
hard. Furthermore, Savelsbergh ( 1984) has recently earliest and must return to the depot by time lo, at the
shown that even finding a feasible solution to the latest. Note that vehicle departure times from the
VRSPTW when the number of vehicles is fixed is itself depot are decision variables.
a NP-complete problem. Therefore, the development Solomon (1983) presents an MIP formulation for
of heuristic algorithms for this problem class is of the VRSPTW. It is based on a model for the basic
primary interest. routing problem given by Fisher and Jaikumar ( I 98 l),
This paper is concerned with the design and analysis and on a generalization of the Miller, Tucker and
of tour-building algorithms for the VRSPTW. All the Zemlin (1960) subtour elimination constraints that is
algorithms presented are extensions of known VRP similar to the generalization used by Desrosiers,
heuristics. The novelty of our approach consists in Soumis and Desrochers (1983a) in the context of the
incorporating not only the distance but also the time traveling salesman problem with time windows.
dimension in the heuristic process. The extensive Of primary importance to the effectiveness and
modifications to the original algorithms have pro- efficiency of heuristics for this problem is the way in
duced more robust and flexible methods that can which the time window constraints are incorporated
accommodate time window constraints. in the solution process. Since we will concentrate on
route-building procedures, let us first examine the
necessary and sufficient conditions for time feasibility
1. Heuristics
when inserting a customer, say u, between the cus-
Tour-building algorithms for the VRSPTW can be tomers i,,, and i,, 1 s p s m, on a partially con-
divided into sequential and parallel methods. Sequen- structed feasible route, (io, i,, i2, . . . , im),io = im = 0,
tial procedures construct one route at a time until all for which the times to begin service, b,, for 0 s r s m,
customers are scheduled. Parallel procedures are char- are known.
acterized by the construction of a number of routes We will assume that initially each vehicle leaves the
simultaneously. The routes are either allowed to form depot at the earliest possible time, eo. After the com-
freely or their number is fixed a priori. In the light of plete vehicle schedules have been created, we can
Savelsbergh's results (1984), we will focus on free- adjust the depot departure time separately for each
routing type of parallel procedures and on sequential vehicle to eliminate any unnecessary waiting time.
methods. Denote by "6: the new time when service at cus-
Before describing the heuristics, let us first state the tomer i,, begins, given the insertion of customer u.
assumptions and introduce the concepts that will be Also, let w,, be the waiting time at i, for p s r 6 m. If
needed throughout the paper. For notational simplic- we assume that the triangle inequality holds both for
ity, we will assume a homogeneous vehicle fleet. travel distances and times, this insertion defines a
push forward in the schedule at i,: on one route as opposed to servicing them individ-
ually, directly from the depot.
PF, = byw - biP 2 0 Due to the existence of time windows, we now must
Furthermore, account for route orientation. Two partial routes with
end customers i and j, respectively, have compatible
PF,,+,= max(0, PF,, - w,+,J, p s r s m - 1. orientations if i is first (last), and j is last (first), i.e.,
Similarly, one can also define a push backward in the admissible links are from the last customer (I) on
the schedule at i, for r = 0, . . . , p - 1. However, this one route to the first customer ( f ) on another. Fur-
concept is not appealing, since the current schedule thermore, in addition to taking into account vehicle
can be pushed backward only if there is no waiting capacity constraints, we must check time window
time in the schedule up to and the vehicle leaves constraints for violation at every step in the heuristic
the depot later than eo. Note that by initially assuming process. Lemma 1.1 can be easily adapted to the case
that each vehicle leaves the depot at eo, we can take of links of the form (I,f ). As in Lemma 1.1, we simply
advantage of the maximum possible push forward. use the push forward generated at f to eficiently test
If PF, > 0, some of the customers i,, p s r s m, time feasibility.
could become infeasible. It is easy to see that we should We implemented the parallel savings method using
examine these customers sequentially for time feasi- list processing and heapsort structures, as proposed by
bility until we find some customer, say i, with r < m, Golden, Magnanti and Nguyen (1977).
for which PF, = 0, or i, is time infeasible, or, in the As presented, the savings heuristic could find it
worst case all the customers i,, p s r s m are exam- profitable to join two customers very close in distance
ined. We have just proved: but far apart in time. Such links introduce extended
periods of waiting time, which can have a high oppor-
Lemma 1.1. The necessary and sufficient conditions tunity cost since the vehicle could be servicing other
for time feasibility when inserting a customer, say u, customers instead of, for example, waiting for a cus-
between and i,, 1 s p s m, on a partially con- tomer to open. To account for both the spatial and
structed feasible route (io, i , , i2, . . . , im),iO = im = 0 temporal closeness of customers, we propose limiting
are the waiting time when joining two customers: letting
b u s / , , and b i , + P F , , s l l , , p ~ r ~ m . wj'" be the resulting waiting time at f if 1 and f were
to be joined, and W be a parameter, i.e., if wj"* > W,
Note that if we assume non-euclidean travel dis- then we do not use the link (1,f).
tances and times, then it is possible that PF, < 0,
which leaves all the customers time feasible. 1.2. A Time-Oriented, Nearest-Neighbor Heuristic
Note also that since i , = 0, the use of Lemma 1.1
The second heuristic we consider belongs to the class
will ensure that any customer that does not permit
of sequential, tour-building algorithms.
the vehicle to return to the depot within the scheduled
time will not be added to the partial route. The nearest-neighbor heuristic starts every route by
Let us now describe several heuristic methods for finding the unrouted customer "closest" (in terms of
the solution of VRSPTW. a measure to be described later) to the depot. At every
subsequent iteration, the heuristic searches for the
customer "closest" to the last customer added to the
1.1. Savings Heuristics
route. This search is performed among all the cus-
Our initial approach to solving VRSPTW is to extend tomers who can feasibly (with respect to time win-
the savings heuristic originally proposed for the rout- dows, vehicle arrival time at the depot, and capacity
ing problem by Clarke and Wright (1964). This pro- constraints) be added to the end of the emerging route.
cedure begins with n distinct routes in which each A new route is started any time the search fails, unless
customer is served by a dedicated vehicle. The parallel there are no more customers to schedule.
version of this tour-building heuristic is characterized The metric used in this approach tries to account
by the addition at every iteration of a link of distinct, for both geographical and temporal closeness of cus-
partially formed routes between two end customers, tomers. Let the last customer on the current partial
guided by a measure of cost savings given by route be customer i, and let j denote any unrouted
customer that could be visited next. Then the metric
used, c,,, measures the direct distance between the two
For example, when p = 1, save is the savings in customers, d,, the time difference between the com-
distance that results from servicing customers i and j pletion of service at i and the beginning of service at
 The Vehicle Routing and Scheduling Problem / 257
j, T,,, and the urgency of delivery to customer j, u,, as customer j, given that u is on the route;
expressed by the time remaining until the vehicle's
last possible service start.
Formally,

c2(i, U,j ) = Adou - c,(i, u, j), X 3 0.

and This type of insertion heuristics tries to maximize

the benefit derived from servicing a customer on the
partial route being constructed rather than on a direct
is defined by weights satisfying + 62 + 63 = l , a l 2 0, route. For example, when p = a , = X = 1 and a2= 0,
c2(i, u, j ) is the savings in distance from servicing
6 2 L 0, 83 2 0.
customer u on the same route with customers i and j,
as opposed to individual, direct service. The best
1.3. Insertion Heuristics feasible insertion place for an unrouted customer is
This class of sequential, tour-building heuristics ini- the one that minimizes the weighted combination of
tializes every route using one of several criteria to be its distance and time insertion, i.e., the one that min-
described later (see Section 3.1). After initializing the imizes a measure of the extra distance and extra time
current route, the method uses two criteria, cl(i, u, j ) required to visit the customer.
and c2(i, U, j ) , at every iteration to insert a new Clearly, different values of p and X lead to different
customer u into the current partial route, between two possible criteria for selecting the customer for insertion
adjacent customers i and j on the route. and its best insertion spot. For example, if a 2 = 0 and
Let (io, il, i2, . . . , i,) be the current route, with CO,,= do,,, we obtain the Mole and Jameson (1976)
io = i,,, = 0. For each unrouted customer, we first approach introduced for the routing problem. There-
compute its best feasible insertion place in the emerg- fore, the class of heuristics described in this discussion,
ing route as which considers insertion costs on both distance and
time dimensions, is a generalization of Mole and
Jameson's approach.
The second type of insertion heuristics aims to select
= m i n [ ~ ~ ( i , -u,
~ , i,)], p = 1, . . . , m. customers whose insertion costs minimize a measure
As noted previously, inserting u between i,-I and i, of total route distance and time.
could potentially alter all the times to begin service at ii) cl(i, u, j ) is defined as before:
customers (i,, . . . , i,). In our computational tests. c2(i, 11, j ) = P I R ~ ( u+) PzR~(u),
PI + P2 = 1, PI 3 0,
we have found our time feasibility conditions, i.e., P 2 > 0,
Lemma 1. I, to be much faster than the explicit testing
of time feasibility at each customer. Next, the best where Rd(u) and R,(u) are the total route distance and
customer u to be inserted in the route is selected as time of the current partial route, if u is inserted.
the one for which In our third approach, the temporal aspect of the
criterion used for insertion also accounts for the ur-
c2(i(u*), u*, j(u*)) = optimum[cdi(u), u, j(u))l, gency of servicing a customer.
u unrouted and feasible. iii) cll(i,u, j) and c12(i,u, j ) are defined as before;
c d i , u , j ) = lu - h,,;
Customer u* is then inserted in the route between c~(i,u, j ) = alcll(i, u, j ) + azcldi, u, j )
i(u*) andj(u*). When no more customer with feasible + a3~13(i,U , j),
insertions can be found, the method starts a new route,
where
unless it has already routed all customers.
We now consider three more specific approaches n l + a z + a 3 = 1 , cu120, a 2 2 0 , a 3 3 0 ;
based on the general insertion criteria just described. cdi, u, j ) = cl(i, U,j).
It is easy to see that, in fact, this class of heuristics is
a generalization of the time-oriented, nearest-neighbor
heuristic, in that we allow insertion of an unrouted
where bJuis the new time for service to begin at customer in any feasible location between a pair of
customers on the route, rather than only at the end of behavior of routing and scheduling heuristics. These
the route. factors include: geographical data; the number of
In all the approaches presented, the insertion of customers serviced by a vehicle; and time window
unrounted customers is guided by both geographical characteristics such as percentage of time-constrained
and temporal criteria. As a consequence, we expect customers, and tightness and positioning of the time
that the waiting time in the schedules produced by windows.
these heuristics will be significantly lower than that The data used for the customer coordinates and
produced by distance-driven criteria. demands are based on the data for some of the prob-
lems from the standard set of routing test problems
1.4. A Time-Oriented Sweep Heuristic given in Christofides, Mingozzi and Toth (1979). The
geographical data are randomly generated by a ran-
This heuristic can be viewed as a member of a broad dom uniform distribution (denote the corresponding
class of approximation methods that decompose the problem sets by R1 and R2), clustered (denote the
VRSPTW into a clustering stage and a scheduling corresponding problem sets by C 1 and C2), and semi-
stage. In the first phase, we assign customers to vehi- clustered (denote the corresponding problem sets by
cles as in the original sweep heuristic (Gillett and RCI and RC2). By a semiclustered problem, we mean
Miller 1974). In the second phase, we create a one- one that contains a mix of randomly generated data
vehicle schedule for the customers in this sector, using and clusters. Problem sets R 1, C1, and RC1 have a
a tour-building heuristic. The insertion heuristic of short scheduling horizon. The length of route-time
type (i) was used in our computer implementation. constraint acts as a capacity constraint which, together
Due to the time window constraints, some customers with the vehicle capacity constraints, allows only a
in this cluster could remain unscheduled. few customers to be serviced by the same vehicle. In
After eliminating scheduled customers from further contrast, the sets R2, C2 and RC2 have a long sched-
consideration, we repeat the clustering-scheduling uling horizon; this characteristic, coupled with large
process. To preserve geographical cohesiveness, one vehicle capacities, permits many customers to be
might consider different seed selection criteria for the serviced by the same vehicle.
next cluster. We use a simple rule that bisects the Given certain geographical and demand data, we
sector just considered and, assuming a counterclock- created the VRSPTW test problems by generating
wise sweep in the more counterclockwise half-sector time windows of various widths for different percent-
lets the seed for a new cluster be the unscheduled ages of customers. In terms of time window density,
customer with the smallest angle formed by the that is, the percentage of customers with time win-
ray from the depot through that customer and the dows, we created problems with 25, 50, 75 and 100%
bisector. time windows.
The intuition for partitioning the unscheduled cus- We now present a method we designed for the
tomers in the sector into two subsets is that the cus- random generation of time window constraints. This
tomers in the more clockwise half-sector will be rela- method was used for the development of the problem
tively far away from the new cluster. By inserting these sets R 1, R2, RC 1, RC2. First, we select the percentage
customers at a later stage, we hope to generate a better of customers to receive time windows, say f ; 0 <
schedule. f s 1. We then generate n random numbers from
We then repeat the process until all customers have the random uniform distribution over the interval
been scheduled. (0, 1) and sort them. This approach creates a random
In order to evaluate the computational capabilities permutation of customers, (il, . . . , in), with i, being
of the heuristics presented, we have developed a set of the position after the sort of the j t h random number
test problems. This approach is necessary given that generated. Finally, we assign time windows to the
no benchmark problem set is available in the literature customers i l , . . . , in,,with n , chosen to be the integer
for vehicle routing and scheduling problems with time that most closely approximates fn. The time windows
windows. have a randomly generated center and width. We
choose the center of the time window for customer i,
for j = 1, . . . , n, as being a uniformly distributed,
2. Development of the Problem Sets
randomly generated number in the interval (eo + to,!,,
We generated six sets of problems; the actual data are lo - tl,,o- sG). To create the time window's width for
available from the author. The design of these test i,, we generate half the width as a normally distributed
problems highlights several factors that can affect the random number.
 The Vehicle Routing and Scheduling Problem / 259
We used a somewhat different method for the clus- vehicles and a higher total schedule time will be better
tered problems, C1 and C2. We first run a 3-opt than one utilizing more vehicles but having a lower
routine (Lin 1965) on each cluster to create routes total schedule time.
and then produce schedules by selecting an orientation In the tables that follow, for each heuristic, the
for each cluster. The time window constraints are numbers on the left-hand side of the column headed
generated by choosing the center as the arrival time at "Average solution values and CPU time" are the total
each customer; the width and density are derived as schedule time, total distance and total waiting time,
before. respectively, of the best of several runs using different
Given the design method, problem sets C1 and C2 parameter values and initialization criteria, averaged
are composed of structured problems in the sense that over the respective problem set. The total schedule
the customers appear in clusters and the time windows time is the sum of the total travel time, total service
are positioned around the amval times at customers. time, and total waiting time, respectively. The num-
This approach permits the identification of a very bers on the right-hand side of this column are the
good, possibly optimal, cluster-by-cluster solution number of routes of the best solution and the total
which, in turn, provides an additional means of eval- CPU time for the different runs, averaged over the
uating heuristic performance. respective problem set. The solution times are in
It is worth mentioning that the best solution we seconds on the DEC- 10, and include the time to read
found for C 1 requires 10 vehicles, for a total schedule the data and compute intercustomer distances and
of 9,829 units, a distance of 829 units and no waiting times. The column headed "Percent deviation from
time. For C2, the best schedule we found requires a the best average solution value" presents the deviation,
fleet of 3 vehicles, a total of 9,59 1 units of time, 59 1 in percentages, of each heuristic's average solution
units of distance and no waiting time. value from the best average solution value, on each
All the test problems are 100-customer euclidean solution dimension. The specific computational
problems. This problem size is not limiting for the results are available from the author.
methods presented, since much larger problems could In obtaining the computational results reported, we
be solved. Travel times between customers are taken used the following parameters:
to equal the corresponding distances. Furthermore, a Savings. The results are the best of two runs, with
homogeneous fleet is assumed. p = 1,0.2.
Savings with Waiting-Time Limit. On R 1, the com-
3. Computational Study putational results are the best of four runs: p = 1,0.2,
and the waiting time limited to 30 and 60 units,
3.1. Computational Results respectively. On C 1, two additional runs were made
Earlier computational experiments for the VRP with the waiting time limited to 120 units.
(Christofides, Mingozzi and Toth 1979) indicate that Insertion-Criterion (i). The results are the best of
the sweep algorithm performed much better than eight runs. The parameters used, (p, A, a,, a2), are:
either the savings or the insertion heuristics on a (1, 1, 1, 01, (1, 2, 1, 01, (1, 1, 0, 11, and (1, 2, 0, 1).
number of randomly generated problems. For several Two initialization criteria were tested:
structured problems, the reverse was true. Also, the (a) the farthest unrouted customer, and
insertion heuristic had a slightly better overall per- (b) the unrounted customer with the earliest dead-
formance than the savings method. All the algorithms line.
used 2-optimal refining procedures to improve the Insertion-Criterion (ii). Eight runs were performed
routes. for this heuristic. The parameters used, (p, a , , a2, PI,
To analyze the behavior of the VRSPTW heuristics P2),are(1,0.5,0.5,0.5,0.5),(1,170,0.5,0.5),and
described in Section 1, we programmed them in (1, 0, 1, 1, 0). Three initialization criteria were used
FORTRAN and performed computational tests on the in conjunction with the first two sets of parameters:
problem classes described in the last section. In ob- criteria (a) and (b), and
taining the computational results, we did not use any (c) the unrouted customer with the minimum
k-optimal improvement procedures. equally weighted combination ofdirect route-time and
Solution quality is measured in terms of the mini- distance.
mum number of vehicles, minimum schedule time, Only criteria (a) and (b) were used with the third
minimum distance, and minimum waiting time in set of parameters.
that order, i.e., we use a lexicographic ordering of the Insertion-Criterion (iii). The results here are the best
solutions. Hence, a schedule with, for example, fewer of ten runs with (p, a , , a2, a3) = (1, 0.5, 0.5, O),
(1, 0.4, 0.4, 0.2), and (I, 0, I, 0). The initialization results is provided by further comparing, on C1 and
criteria (a), (b), and (c) were used with the first set of on C2, this heuristic's values with the best known
parameters. In addition to the previous three criteria, solution (see Section 2). I1 was consistently close to
we used the actual heuristic criterion for initialization the best known total schedule time on both problem
in conjunction with the second set of parameters. For sets; on C1, its smallest deviation was 0.3%, its largest
the third set of parameters, criteria (a), (b) and the deviation was 8.3%, with an average of 2.8%, while
actual heuristic criterion were used for initialization. on C2, its deviation was within 5% in seven out of the
Sweep. The results for this heuristic are the best of eight problems tested.
eight runs with the parameters used for the insertion The insertion-based sweep strategy was found to be
heuristic with criterion (i). The first seed was always very good for problems with many customers per
chosen as the first customer encountered in the first vehicle. This heuristic obtained the best solution in 13
quadrant in a counterclockwise sweep. out of the 27 cases. This method performed very well
Nearest Neighbor. The parameters used for this on the clustered problem sets C1 and C2, where it was
heuristic, (6,, 62, &), are: (0.4, 0.4, 0.2), (0, 1, O), superior to the other heuristics in 4 and 7 cases,
(0.5, 0.5, O), and (0.3, 0.3, 0.4). respectively. When compared to the best known so-
Tables I-VI compare the methods on each problem lution for these problem sets, its largest total schedule
set, for solution quality and running time. time deviation was only 9.7% on C1, and 3.8% on
With respect to solution quality, heuristic I1 per- C2. While it did not behave well on R 1 and RCI, we
formed the best. Its behavior was very stable across all believe its behavior on these problems could be im-
problem environments, obtaining the best solution for proved by the consideration of each of the customers
27 out of the 56 problems tested; even when it did not as initial seeds.
beat the other methods, its deviation from the best The insertion heuristics I2 and I3 dominated the
solution was quite small. Additional support for these waiting time dimension for problems with a short

Table I
Comparison of the Algorithms on R1
Percent Deviation No. of Problems on
Average Solution Valuesn
Algorithm from Best Average Which Method Found
and CPU Timeb
Solution Value Best Solution
Savings (SAV)

Savings, waiting time limit

(SWT)

Insertion criterion i (11)

Insertion criterion ii (12)

Insertion criterion iii (13)

Nearest Neighbor (NN)

Sweep (S)

" Schedule time Number of routes

Distance
Waiting time CPU time.
CPU time in seconds on the DEC-10.
 The Vehicle Routing and Scheduling Problem / 26 1
Table I1
Comvarison of the Algorithms
- on C 1
Percent Deviation No. of Problems on
Average Solution Valuesa from Best Average Which Method Found
Algorithm and CPU timeb Solution Value Best Solution
SAV

SWT

I1

I2

13

NN

" Schedule time Number of routes

Distance
Waiting time CPU time.
CPU time in seconds on the DEC-10.

Table I11
Comparison of the Algorithms on RC1
Average Solution Valuesa Percent Deviation No. of Problems on
Algorithm and CPU Timeb from Best Average Which Method Found
Solution Value the Best Solution

" Schedule time Number of routes

Distance
Waiting time CPU time.
CPU time in seconds on the DEC-10.
 Table IV
Comparison of the Algorithms on R2
Percent Deviation No. of Problems on
Average Solution Valuesa
Algorithm and CPU Timeb from Best Average Which Method Found
Solution Value Best Solution
4

" Schedule time Number of routes

Distance
Waiting time CPU time
CPU time in seconds on the DEC-10.

Table V
Comparison of the Algorithms on C2
Percent Deviation No. of Problems on
Average Solution Valuesa
Algorithm and CPU Timeb from Best Average Which Method Found
Solution Value Best Solution

- - - -

" Schedule time Number of routes

Distance
Waiting time CPU time
CPU time in seconds on the DEC-10.

scheduling horizon. Otherwise, they were rarely able The time-oriented, nearest-neighbor method also
to obtain the best solution. It seems that these heuris- had limited success; this occurred in the long sched-
tics will be successful for problems with idle time as uling horizon problems that permit many customers
the primary objective. to be serviced by the same vehicle.
 The Vehicle Routing and Scheduling Problem / 263
Table VI
Comparison of the Algorithms on RC2
Percent Deviation No. of Problems on
Average Solution Valuesa from Best Average Which Method Found
Algorithm and CPU Timeh
Solution Value Best Solution

" Schedule time Number of routes

Distance
Waiting time CPU time
CPU time in seconds on the DEC-10.

The savings heuristic with waiting-time limit, while We begin by examining the effect of distance inser-
a considerable improvement over the original savings tion ( a , = 1, a2 = 0) versus time insertion ( a , = 0,
method, did not perform well in general. After we a2= I). For the problem sets involving long scheduling
examined its behavior on R1 and C1, preliminary horizons, time insertion proved clearly superior to
computational experiments on the rest of the problem distance insertion. It was used in obtaining the best
sets indicated that this heuristic will, in general, re- solution to 2 1 out of 27 problems. The few exceptions,
quire more vehicles than the number utilized by the two for each problem set, where distance insertion
other heuristics. Therefore, we did not examine its performed best, were problems with less than 100%
behavior any further. Nevertheless, given that it found time window density, or the least tight problems.
the best solutions to 5 problems of C1, it might For the problem sets with a short scheduling hori-
conceivably be used with success in a sweep-type zon, time insertion proved useful for most of the tight
heuristic. and high density problems of R l and RC1. It was
In terms of computation time, all the heuristics used for 5 problems of RC1, and 4 problems of R1.
seem to be very efficient. Algorithms SWT and SAV Not surprisingly, time insertion was used for the
were the fastest. Our results indicate that the efficiency less constrained problems of C 1, 4 problems, since it
of the insertion methods increased with the increase was able to provide more intelligence than distance
in the percentage of time window constraints and their insertion in directing the heuristic toward the optimal
tightness. This result is due to a lower number of solution.
feasible insertions possible in such problems. In conclusion, time insertion proved superior to
distance insertion, especially in problems involving
3.2. Parametric Analysis many customers per vehicle, and/or high density and
tight time windows. Distance insertion, by emphasiz-
Most of the heuristics presented are parameterized. It ing the geographical component, can lead to higher
is thus of interest to know whether there are any total schedule time from accumulated waiting time
relationships between the parameter values and the and, possibly, additional vehicles. We should note that
initialization criteria that produced the best solution we observed the same pattern for the insertion-based-
values, and the corresponding problem structures. sweep strategy.
Given our computational results, we will focus on the Examining now the choice of parameters A, we find
heuristic 11. that X = 2 was used in obtaining the best solution to
35 problems, while h = 1 was used in only 21 cases. tiveness of the various tour-building VRSPTW heu-
On R1, X = 2 was utilized for 9 out of 12 problems, ristics suggested. Furthermore, given the wide variety
while on RC2, it was used for 7 out of 8. Five out of of routing and scheduling environments used in our
9 C1 problems used X = 2. This parameter value was study, we consider the relative performance of these
also used for 5 out of 8 RCI problems, 5 out of 11 heuristics on the test problems indicative of their
R2 problems, and 4 out of 8 C2 problems. relative performance in general.
The initialization criterion used for a heuristic can Based on our study, we recommend the use of the
have a significant impact on its behavior. We used insertion heuristic 11, possibly embedded in a hybrid
two initialization criteria for I1 : the farthest unrouted sweep-insertion approach, to obtain excellent initial
customer and the unrouted customer with the earliest VRSPTW solutions in a reasonable amount of com-
deadline. The former criterion proved successful for puting time. Given its very stable behavior, we believe
the problem sets with short scheduling horizons. It that this heuristic will perform very well on practical
was used in obtaining 23 out of the 29 best solutions problems. Further support for this conclusion is pro-
to these problems: all R l problems except one, all vided by the recent work of Jaw et al. (1984) that
RCl problems, and 4 C l problems. In contrast, the showed the effectiveness of an insertion-based proce-
latter criterion was better on the long scheduling ho- dure in a time-window constrained real dial-a-ride
rizon problems in which, given that only a few vehicles environment.
were used, it was of primary importance to service
customers with early due dates at the beginning of the
scheduling horizon so no additional vehicles would be
Acknowledgment
necessary. In these experiments, the earliest deadline
criterion was used in 16 out of the 27 best solutions:
6 problems of R2, 6 problems of C2, and 4 problems The author has benefited from fruitful discussions
of RC2. with Marshall Fisher, Ramchandran Jaikumar,
Note that most of these problems are characterized Monique Guignard-Spielberg, Pradeep Kedia and
by 100% density and/or tight time windows. This Anand Desai. He also wishes to thank the three anon-
initialization method was quite successful on the struc- ymous referees for their useful comments. The
tured problems since it was able to guide the heuristic computer time for this project was provided by the
search toward the optimal solution. Wharton Computer Center of the University of
Pennsylvania.

4. Conclusions
We have presented the development of heuristics and References
test problem sets and have reported our computational
experiments for VRSPTW. Our results indicate that BAKER,E. 1983. An Exact Algorithm for the Time-
the insertion heuristic I1 proved to be very successful. Constrained Traveling Salesman Problem. Opns.
Res. 31, 938-945.
A parametric analysis of this heuristic revealed the BAKER, E., A N D S. RUSHINEK. 1982. Large Scale Imple-
importance of time driven insertions for heavily time- mentation of a Time Oriented Vehicle Scheduling
constrained problems, particularly when many cus- Model. U.S. Department of Transportation, Urban
tomers are to be scheduled for each vehicle. In the Mass Transit Administration.
latter case, it seems that a very effective strategy is to BODIN,L., B. GOLDEN, A. ASSADAND M. BALL.1983.
embed this heuristic in a sweep-type approach. Routing and Scheduling of Vehicles and Crews: The
The excellent performance of the insertion heuristic State of the Art. Comput. Opns. Res. 10, 62-212.
I1 can be explained if we realize that, while routing CHRISTOFIDES, N., A. MINGOZZI AND P. TOTH.1979. The
problems seems to be driven by the assignment-of- Vehicle Routing Problem. In Combinatorial Opti-
customers-to-vehicles component-as indicated by mizations, N . Christofides, R. Mingozzi, P. Toth,
the success of the Fisher and Jaikumar generalized and C. Sandi (eds.). John Wiley & Sons, New York.
CHRISTOFIDES, N., A. MINGOZZI A N D P. TOTH. 1981.
assignment heuristic-the sequencing aspect of the State Space Relaxation Procedures for the Compu-
problem seems to drive routing problems dominated tation of Bounds to Routing Problems. Networks 11,
by time windows. It is this aspect of the problem that 145-164.
the insertion heuristic I1 captures so well. CLARKE, G., A N D W. WRIGHT.1964. Scheduling of
We believe that our computational study provides Vehicles from a Central Depot to A Number of
some definite guidelines concerning the relative effec- Delivery Points. Opns. Res. 12, 568-58 1.
 The Vehicle Routing and Scheduling Problem / 265
COOK T., A N D R. RUSSELL.1978. A Simulation and LIN,S. 1965. Computer Solutions of the Traveling Sales-
Statistical Analysis of Stochastic Vehicle Routing man Problem. Bell System Tech. 5.44, 2245-2269.
with Timing Constraints. Decision Sci. 9, 673-687. MILLER,C., A. TUCKER AND R. ZEMLIN. 1960. Integer
DESROSIERS, J., P. PELLETIER, AND F. SOUMIS.1982. Programming Formulations and Traveling Sales-
Shortest Path with Time Windows. Working Paper man Problems. J. Assoc. Comput. Mach. 7 ,
8 1-21, Ecole des Hautes Etudes Commerciales, 326-329.
Montreal. MOLE,R., AND S. JAMESON. 1976. A Sequential Route-
DESROSIERS, J., F. SOUMIS A N D M. DESROCHERS. 1983a. Building Algorithm Employing a Generalized Sav-
Routing With Time Windows by Column Genera- ings Criterion. Opnl. Res. Quart. 27, 503-5 1 1.
tion. Working Paper 277, Centre de Reserche sur les ORLOFF,C. 1976. Route Constrained Fleet Scheduling.
Transports, University of Montreal. Trans. Sci. 10, 149- 168.
DESROSIERS, J., F. SOUMIS A N D M. DESROCHERS. 1983b. PSARAFTIS, H. 1983. An Exact Algorithm for the Single
Routing with Time Windows: Synthesis. Working Vehicle Many-to-Many Dial-A-Ride Problem with
Paper G-83-05, Ecole des Hautes Etudes Commer- Time Windows. Trans. Sci. 17,35 1-358.
ciales, Montreal. PULLEN,H., AND M. WEBB. 1967. A Computer Ap-
FEDERGRUEN, A., AND B. J. LAGEWEG. 1980. Hierarchi- plication to a Transport Scheduling Problem.
cal Distribution Modelling with Routing Costs. Comput. J. 10, 10-1 3.
Mathematical Centre, Amsterdam. RUSSELL, R. 1977. An Effective Heuristic for the M-Tour
FISHER,M., A N D R. JAIKUMAR. 1981. A Generalized Traveling Salesman Problem with Some Side Con-
Assignment Heuristic for Vehicle Routing. Networks ditions. Opns. Res. 25, 5 17-524.
11, 109-124. SAVELSBERGH, M. 1984. Private Communication by
FISHER, M., A. GREENFIELD, R. JAIKUMAR AND P. KEDIA. Alexander Rinnooy Kan.
1982. Real-Time Scheduling of a Bulk Delivery SCHRAGE, L. 198 1. Formulation and Structure of More
Fleet: Practical Application of Lagrangian Relaxa- Complex/Realistic Routing and Scheduling Prob-
tion. Working Paper 82- 10-1 1, Department of lems. Networks 11, 229-232.
Decision Sciences, University of Pennsylvania. SEXTON,T., AND L. BODIN.1985a. Optimizing Single
GILLET, B., AND L. MILLER.1974. A Heuristic Algorithm Vehicle Many-to-Many Operations with Desired
For the Vehicle Dispatching Problem. Opns. Res. Delivery Times: I. Scheduling. Trans. Sci. 19,
22, 340-349. 378-410.
GOLDEN,B., T. MAGNANTI A N D H. NGUYEN.1977. SEXTON, T., A N D L. BODIN.1985b. Optimizing Single
Implementing Vehicle Routing Algorithms. Net- Vehicle Many-to-Many Operations with Desired De-
works7, 113-148. livery Times: 11. Routing. Trans. Sci. 19, 4 1 1-435.
JAW,J., A. ODONI,H. PSARAFTIS AND N. WILSON.1984. SOLOMON, M. 1983. Vehicle Routing and Scheduling
A Heuristic Algorithm for the Multi-Vehicle with Time Window Constraints: Models and Algo-
Advance Request Dial-A-Ride Problem with Time rithms. Ph.D. Dissertation, Dept. of Decision Sci-
Windows. Working Paper MIT-UMTA-83-L. ences, University of Pennsylvania.
KNIGHT,K., AND J. HOFER.1968. Vehicle Scheduling SWERSEY, A., AND W. BALLARD. 1982. Scheduling School
with Timed and Connected Calls: A Case Study. Buses. Working Paper, Yale School of Organization
Opnl. Res. Quart. 19, 299-3 10. and Management.