See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221704650

The vehicle routing problem with multiple use of vehicles

Article · January 1990

CITATIONS READS

75 611

1 author:

Bernhard Fleischmann
Universität Augsburg
111 PUBLICATIONS   1,849 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Bernhard Fleischmann on 05 June 2018.

The user has requested enhancement of the downloaded file.

The vehicle routing problem with multiple use of the vehicles

Bernhard Fleischmann

Universität Hamburg, Fachbereich Wirtschaftswissenschaften,

Von-Melle-Park 5, D - 2000 Hamburg 13, F.R. Germany
June 1990

c Bernhard Fleischmann, Universität Augsburg, 1990, 2010

Scope and Purpose

In the vehicle routing problem (VRP) a fleet of vehicles at a central depot is to be
scheduled so as to service a given set of customers. The usual assumption, that every
vehicle can run one tour from and to the depot, is not correct for the daily scheduling
of delivery tours in an urban area, where demands are high and travel times short:
Then a vehicle has to run several tours a day. While this case is considered in most
commercial software packages, there is a lack of relevant theory and methods in
the literature. The paper proposes a procedure for this problem and reports on
applications.

Abstract
We consider the single-depot VRP with a limited heterogeneous fleet of vehicles,
where every vehicle can run several tours within a limited time of use. In addition,
customer time window constraints are included. We extend the parallel savings pro-
cedure to this case and report on computational experiences with large-scale real-life
delivery problems.

1 Introduction
The vehicle routing problem (VRP) consists in designing routes or tours from and
to a central depot which service a set of customers with given demands, where every
tour is restricted by the vehicle capacity and / or a maximal duration, and the
total distance, travel time or cost is to be minimized. One of the most important
situations where the VRP occurs in practice, is the delivery of consumer goods for
daily varying customer sets and demands, requiring daily routing. However, this
problem typically involves additional complications which are not considered in the
VRP and the usual solution methods:

(a) There is a given fleet of vehicles of different types with, of course, a limited
number of vehicles.
1 Introduction 2

(b) A vehicle that returns from a short tour may be used a second or third time.
This is important in urban areas, where the vehicle capacity is the limiting
factor and travel times are rather short.

(c) There are time window constraints with respect to the arrival time at certain
customers.

While the extension (c), the vehicle routing and scheduling problem with time win-
dows (VRSPTW), has been studied intensively in the recent literature [6,8,18,19,20],
there is little work on problem (a) and nearly none on problem (b).

The best way to deal with problem (a) is the method of Fisher and Jaikumer [9],
the only one that guarantees a feasible solution, if any, for a given fleet by solving
a general assignment problem. However, this method does not apply to problem
(b) which involves a two-stage assignment of customers to tours and of tours to
vehicles. The usual heuristics, in particular the Savings procedure (SP) in its
basic form (see e.g. [2,17]) assume an unlimited homogeneous fleet. Extensions to
fleet restrictions have been considered already in the original paper of Clarke and
Wright [5], but rarely in the later literature [7], except for short comments, e.g. in
[4,15 p. 193 f].

The lack of theory on problems (a) and (b) is the more surprising, as most com-
mercial software packages cover these problems using the Savings procedure [3,12],
but with unpublished details. The most common way, as we know from authors of
software, seems to be a two-phase iterative procedure: First, a standard VRP is sol-
ved, and in the second phase the resulting tours are assigned to the vehicles. If there
is no feasible assignment, the VRP is repeated in total or partially with reduced
time and/or capacity limits. This is also proposed in the case report [11]. Another
way is the sequential Savings procedure, which can construct the tours vehicle
by vehicle, but which is known for its poor performance compared to the parallel
procedure [16 p. 16ff].

The main subject of this paper is the problem (b), the vehicle routing problem
with multiple use of the vehicles (VRPMU). It always includes problem (a) and
may also include time windows. We propose a method that integrates all restrictions
relevant to (a), (b) and (c) into a one-phase parallel SP called Savings procedure
for multiple use of vehicles (SPMU). For the practical problems considered (see
Section 6) with up to 360 customers per day, a SP seems to be most appropriate due
to its low computational effort. It works for any definition of the savings function
which reflects the objective of the VRP. When daily scheduling a given fleet, however,
the most important objective is total distance.
2 A Framework for the Savings Procedure 3

The next section summarizes the necessary notations and terms for the VRP and
the basic SP. Section 3 explains a concept for considering a given fleet. Section 4
extends this concept for the multiple use of vehicles and presents the SPMU. Section
5 shows how to include time windows into the SPMU. Section 6 reports on several
applications and numerical results.

2 A Framework for the Savings Procedure

The VRP with customers i = 1, . . ., n, the depot (i = 0) and vehicles v = 1, . . ., m
is characterized by the following data:
qi demand of customer i
Qv capacity of vehicle v
dij distance between depot or customers i, j
tij travel time
si stop time at customer i
tmax maximal duration of a tour.

The capacities and demands may be vectors with several components such as weight,
number of pallets, etc. The distances and travel times are supposed symmetric. Note
that their definition is ambiguous in a road network with speeds varying from arc to
arc, as they cannot be minimal at once. The most realistic definition is to determine
tij as the shortest travel time from i to j and dij as the length of the corresponding
path. In the following, we assume the reader to be familiar with the savings heuristic
(see e.g. [2,4;5] and for implementation techniques [13,17]). We only outline the main
steps and introduce the notations which are required for the extensions in the next
sections.

A tour T = (i1 , . . ., iK ) is a sequence of customers visited by a vehicle, including the

travel from the depot to i1 and from iK to the depot. It induces the quantity shipped
q(T ), the length d(T ) and the duration t(T ) (travel and stop times). T is feasible if
t(T ) ≤ tmax and, for some vehicle v, q(T ) ≤ Qv . The reverse tour, denoted by T 0 ,
is feasible if and only if T is feasible. If i is the first or last customer in tour T , T (i)
denotes the (eventually reversed) tour starting with i, i.e.

T if i = 1
T (i) =
T 0 if i = iK
A shuttle tour is a tour with a single customer. All shuttle tours are supposed
feasible. A schedule S is a set of feasible tours T ∈ S, such that every customer is
contained in exactly one tour. If T1 , T2 ∈ S, T 1 = (i1 , . . ., iK ) and T2 = (j1 , . . ., jL ),
2 A Framework for the Savings Procedure 4

the combined tour is denoted T 1 × T 2 = (i1 , . . ., iK , jl , . . ., jL ).

The SP starts with the schedule S0 consisting of all shuttle tours. The current
schedule S is changed by the following basic operation: Let i and j be the first or
last customers of two different tours T1 , T2 ∈ S. Then T1 and T2 are replaced by the
combined tour which links i and j,

T ∗ = T1 (i)0 × T2 (j)

i.e. S is replaced by S ∪ {T ∗ }\{T1 , T2 }.

This operation yields the savings C(i, j) = d0i +d0j −dij with respect to distance and
analogous savings with respect to travel time. The definition of the savings function
C(i, j), which determines the order of carrying out the basic operations, is the only
part of the procedure, where the overall objective function of the VRP is taken into
account. According to different possible objectives (minimal distance, time, number
of vehicles, cost) there are different definitions of the savings function proposed in
literature [2]. Moreover, perturbations of shape parameters can be used to generate
several solutions, from which to choose the best one [13,16,17]. In the following, we
admit any appropriate savings function C(i,j). But it should be a function of i, j only
and not depend on the current schedule. Otherwise, one has to update the savings
during every iteration, so that the main advantage of the SP gets lost.

The general scheme of the SP is then as follows:

Initial step
S = S0 ; order all pairs (i, j) (i = 2, . . ., n; j = 1, . . ., i − 1) from largest to smallest
values C(i, j).

Iteration

(a) Proceed to the next (i, j) such that i and j are the first or last customer in
different tours T1 , T2 ∈ S.

(b) If the tour T ∗ = T1 (i)0 × T2 (j) is feasible, go to step (c), otherwise to the next
iteration.

(c) S = S ∪ {T ∗ }\{T1 , T2 }.

The algorithm terminates in step (a), if all (i, j) have been processed.

A major deficiency of the basic savings algorithm is that it does not take into account
the available fleet of vehicles. For this purpose, one has to ensure, in step (b), not
only the feasibility of the single tours T , but also the feasibility of the schedule S
3 Considering a given fleet of vehicles 5

as a whole, with respect to the fleet. In the next section, a concept of feasibility is
proposed for the sequence of schedules occuring in the savings algorithm.

3 Considering a given fleet of vehicles

In this section, we still assume that every vehicle v can run a single tour. In the next
section, the multiple use of the vehicles is allowed for. The notation of feasibility of
a schedule S is obvious. S is feasible if there exists an assignment v(T ) ∈ V of
the tours T ∈ S to the vehicles such that

v(T ) 6= v(U ) for T 6= U and q(T ) ≤ Qv(T ) . (1)

However, this definition is not useful in the SP, because the initial schedule S0 which
requires a separate vehicle for every customer, is infeasible in all realistic situations.
Therefore, the feasibility of S cannot be postulated at every iteration of the algo-
rithm. One possible way to overcome this difficulty, proposed in [4,7,15] is to assign
only the combined tours, which arise in step (b), to the vehicles. Then, the parti-
al schedule consisting of all tours with at least two customers can be kept feasible
during the whole algorithm, but all schedules, incl. the final schedule, may contain
unassignable shuttle tours. This procedure has the major disadvantage, that already
in early iterations, as soon as the number of combined tours equals the number of
vehicles, favourable combinations among shuttle tours are blocked. Retrying blocked
combinations at every iteration improves the solution, but increases the effort dra-
stically, compared to the basic SP. And even then, the result may be very poor,
because the order of the combinations deviates from the order of decreasing savings.

Another way of relaxing the feasibility restriction for the intermediate schedules,
which is already proposed by Clark and Wright [5] is to assume an unlimited number
of vehicles of the smallest size. We generalize this idea for the VRP and in the
next section for the VRPMU. At the beginning, we introduce an appropriate set
of fictitious vehicles, which render the initial schedule S0 feasible. The range
of capacities of the fictitious vehicies is that of the real capacities Q1 , . . ., Qm . At
any iteration, we impose the feasibility condition on the schedule S, that for any
v the number of fictitious vehicles of size Qv or greater, required for schedule S
must not increase. This restriction is very weak at the early iterations, when most
combinations improve the feasibility, but strengthens gradually, when the tour sizes
approach the vehicle sizes. Once a feasible schedule is reached, i.e. all fictitious
vehicles are removed, feasibility is maintained up to the end. Of course, like any
savings heuristics, this procedure does not guarantee that feasibility is reached at
3 Considering a given fleet of vehicles 6

all, but it attempts to improve the degree of feasibility at every step.

If the vehicles can be used once only, the feasibility condition can be implemented
easily without constructing an explicit assignment v(T ) using the following lemma.
We order the vehicles according to decreasing capacity:

Q1 ≥ Q2 ≥ . . ., Qm .

(If the capacity is a vector, we assume, that the order is identical for all components;
this is a realistic assumption about the vehicle sizes.) Moreover, we group the vehicles
into classes of identical size, say

V1 = {1, . . ., v1 }, Vc = {vc−1 + 1, . . ., vc } (c = 2, . . ., k),

i.e. Qv = Qvc for all v ∈ Vc . For any tour T ∈ S, the class of the smallest feasible
vehicles is

c(T ) = max{c : Qvc ≥ q(T )}, (2)

and the number of vehicles of size Qvc or greater, needed for S, is

Nc (S) = |{T ∈ S : c(T ) ≥ c}| (c = 1, . . ., k) (3)

Lemma
(a) A schedule S is feasible, if and only if

Nc (S) ≤ vc for c = 1, . . ., k. (4)

(b) The number of fictitious vehicles of size Qvc or greater, required for S, is

Fc (S) = max(max{Ni (S) − vi : i = 1, . . ., c}, 0) (5)

Proof

As vc is the number of available vehicles of size Qvc or greater, (4) is obviously

necessary. If (4) holds, the assignment of the tours T to the vehicles v(T ) = 1, . . ., m
in the order of increasing c(T ) is feasible; because q(T ) ≤ Q1 for all T ∈ S by
definition, and if there was a T0 ∈ S and v = v(T0 ) ∈ Vc (c ≤ 2) such that
q(T0 ) > Qv , then c(T ) < c for the first v tours and hence Nc−1 (S) ≥ v > vc−1 which
contradicts (4). Statement (b) follows immediately from (a).


4 Multiple use of the vehicles 7

The numbers Nc (S) need to be determined only for the initial schedule. They can
be updated at every change of S in step (c) of the SP as follows:



 −1 for c ≥ max(c(T1 ), c(T2 ))

Nc (S) = Nc (S) + 1 for c(T ∗ ) ≤ c < min(c(T1 ), c(T2 )) (6)



0 otherwise.

The numbers Fc (S) can be determined from Nc (S) recursively:

F1 (S) = max(N1 (S) − v1 , 0)

Fc (S) = max(Nc (S) − vc , Fc−1 (S)).

Therefore, the feasibility condition is satisfied, if Nc (S) does not increase beyond
the present value of Fc (S).

In the whole, the consideration of the given fleet requires the following supplements
to the basic SP:
In the initial step:
Determine c(T ) for all T ∈ S0 , Nc (S0 ) and Fc (S0 ) according to (2), (3) and (6),
respectively.

In Step (b) of the iteration:

Determine c(T ∗ ); let c = min(c(T1 ), c(T2 )).
If c(T ∗ ) < c and Nc (S) + 1 − vc > Fc (S) for some c such that c(T ∗ ) ≤ c < c, reject
T ∗ and go to the next iteration.
In step (c) of the iteration:
Update Nc (S) according to (6) and Fc (S) as follows:

Fc (S) = max(Fc (S) − 1, Fc−1 (S)) (c = max(c(T1 ), c(T2 )), . . ., m). (7)

4 Multiple use of the vehicles

A vehicle that runs several tours on the same day has to stay at the depot between
two tours for a certain time for being reloaded. The time of use of the vehicle, from
the beginning of its first tour up to the end of its last tour, is limited, e.g. due to
the drivers work time. Thus, the VRPMU involves the additional data:
t0 minimal time between tours
umax maximal time of use of a vehicle.
In the VRPMU the definition (1) of the feasibility of a schedule S is replaced as
4 Multiple use of the vehicles 8

follows: S is feasible, if there exists an assignment v(T ) ∈ V of the tours T ∈ S to

the vehicles such that

X
(t(T ) + t0 ) ≤ umax + t0 for v ∈ V and (8)
T :v(T )=v

q(T ) ≤ Qv(T ) for T ∈ S.

The determination of an assignment v(T ) which satisfies (8) is a difficult problem.

It is a generalized bin packing problem, where the vehicles v are the bins of size
umax + t0 and the tours T are the items to be packed with size t(T ) + t0 . The items
bear the capacity index c(T ), defined by (2), and can only be packed into a bin
v ∈ Vc with c ≤ c(T ). To solve this problem, we modify the First-Fit-Decreasing
(FFD) algorithm which is known as the most favourable bin packing heuristic [10].
The decreasing “item size” t(T ) is only used as a second order criterion, the first
criterion being increasing c(T ), because tours requiring bigger vehicles must have
priority on these vehicles over smaller tours; moreover, the vehicles are scanned in
the order of increasing size. We use the variable uv = time of use of vehicle v.

Assigning the tours of a given schedule to the vehicles:

uv = 0 (v ∈ V );
for every T ∈ S in the order of increasing c(T ) and, for
equal c(T ), decreasing t(T ):
for every v ∈ Vc (c = c(T ), c(T ) − 1, . . ., 1):
if uv + t(T ) ≤ umax : v(T ) = v,
uv = uv + t(T ) + t0 ,
go to the next T ;
if T has not been assigned: stop, no assignment found.

A similar assignment procedure is used by Matthäus [15, p.190f], but only for the
combined tours, as explained in section 3.

In order to enforce a feasible assignment for the initial schedule S0 , we again in-
troduce fictitious vehicles and minimize their number using the above assignment
heuristic. As soon as a fictitious vehicle becomes idle during the savings procedure,
it is withdrawn. The feasibility condition, which is checked before every combina-
tion, is simply that the new schedule admits an assignment to the current set of
vehicles. Thus, like in the procedure of the previous section, the number of fictitious
vehicles never increases, except for the initial step. The fictitious vehicles are num-
bered v = m + 1, . . ., m + m̃ and augment V and the appropriate classes Vc . In the
assignment algorithm, they are scanned after the real vehicles of the same class.
4 Multiple use of the vehicles 9

As the assignment algorithm is rather time-consuming compared to the operations

of the SP, we try to avoid running it before every combination. We first try to update
the current assignment by assigning only the new tour T ∗ either to one of the vehicles
v(T1 ) or v(T2 ), relieved of the tours T1 and T2 , (Test 1) or to any other vehicle (Test
2). Only if these tests fail, a completely new assignment is constructed using the
assignment algorithm (Test 3). This way, the assignment algorithm is performed
only at a very small proportion of the savings iterations (cf. Section 6).

In the whole, the SPMU requires the following supplements to the basic SP:

In the initial step:

Determine c(T ) for T ∈ S0 according to (2); m̃ = 0. Determine an assignment v(T )
for T ∈ S0 using the assignment algorithm. If it fails to find a vehicle for a tour T ,
set

m̃ = m̃ + 1; v = m + m̃; c = c(T );
Qv = Qvc ; Vc = Vc ∪ {v}; V = V ∪ {v};
v(T ) = v; uv = t(T ) + t0

and continue the assignment algorithm.

In Step (b) of the iteration:
Test 1:

Set v 0 (T ) = v(T ) for T 6= T1 , T2 and u0v = uv for v ∈ V ;

u0v(Ti ) = u0v(Ti ) − t(Ti ) − t0 (i = 1, 2);
for i = 1, 2:
if c(Ti ) ≤ c(T ∗ ) and u0v(Ti ) + t(T ∗ ) ≤ umax :
v 0 (T ∗ ) = v(Ti );
u0v(Ti ) = u0v(Ti ) + t(T ∗ ) + t0 ;
go to step (c).

Test 2: Search for a vehicle v 0 (T ∗ ) for T ∗ , using one iteration of the assignment
algorithm. If it is found, go to Step (c).

Test 3: Determine a new assignment v 0 (T ) for T ∈ S\{T1 , T2 } ∪ {T ∗ } using the

assignment algorithm. If it fails, reject T ∗ and go to the next savings iteration.

In step (c) of the iteration:

v(T ) = v 0 (T ) (T ∈ S);
uv = u0v (v ∈ V );

withdraw all unused fictitious vehicles from V and Vc .

5 Time windows 10

The sorting of the tours T of the current schedule according to increasing c(T )
and decreasing t(T ), as required by the assignment algorithm in step (b), can be
accomplished easily, starting from an initial sorting of the tours T ∈ S0 : At every
combination in Step (c), T1 and T2 are removed and T ∗ is inserted into the current
order.

The ability of the SPMU to create a feasible solution in case of tight vehicle capacity
can be improved by strengthening the feasibility condition. It may happen that a
fictitious vehicle, charged with a small tour only, is reloaded with additional tours
or customers. This augmented infeasibility can be avoided by controlling not only
the number of fictitious vehicles, but also their load, which can be expressed by the
number FC of customers supplied by the fictitious vehicles. The additional condition
is, that FC must never increase. The implementation of this rule in the SPMU is
easy: In the initial step, the starting value of FC is registered; in the Tests 1 and
2, the tour T ∗ can be assigned to a fictitious vehicle only if both v(T1 ) and v(T2 )
are fictitious. In Test 3, the corresponding number FC’ is registered for the new
assignment v 0 (T ); a tour is assigned to a fictitious vehicle only, if the number of its
customers does not exceed F C − F C 0 . In step (c), FC is updated.

5 Time windows
Time window constraints cause additional difficulties for scheduling the repeated use
of vehicles. However, the SPMU can be well extended to this case. We explain the
main ideas of the algorithm, which has been used for some of the results in Section
6, but suppress straightforward details of the implementation. For simplicity, we
consider only a single time window per customer; the extension to several time
windows is easy. We then need the following additional data:
ai , b i time window of customer i (earliest and latest arrival time at the
customer),
Av , Bv time window of vehicle v (earliest start of the first tour, latest end
of the last tour),
wmax maximal waiting time per tour.
The time window constraints restrict the feasibility of both the single tours and the
whole schedule. The first aspect, the check on the feasibility of the combined tour T ∗
in SP, has been dealt with frequently in literature, see e.g. [6,8,18] for the parallel
and [20] for the sequential SP. Similarly, we record the following variables for all
tours T ∈ S and the reverse tours T 0 :
5 Time windows 11

a(T ), b(T ) time window of tour T (earliest and latest arrival at the first custo-
mer)
w(T ) waiting time of tour T
k(T ) ’kernel time’ (duration of the tour without travel time from and to
the depot and without waiting time).
T 0 is feasible w.r.t. the time windows if a(T ) ≤ b(T ) and w(T ) ≤ wmax . The initia-
lization for a shuttle tour T = (i) is

a(T ) = ai , b(T ) = bi , w(T ) = 0, k(T ) = si ,

the update for the combined tour T ∗ = U1 × U2 , where U1 = T1 (i)0 , U2 = T2 (j) is

w(T ∗ ) = max{a(U2 ) − b(U1 ) − k(T1 ) − tij , w(U1 )} + w(U2 )

b(T ∗ ) = min{b(U1 ), b(U2 ) − k(T1 ) − tij − w(U1 )}

max{a(U ), a(U ) − k(T ) − t } if w(T ∗ ) = 0
∗ 1 2 1 ij
a(T ) =
b(T ∗ ) otherwise
k(T ∗ ) = k(T1 ) + k(T2 ) + tij

and T ∗ is feasible, if T1 (i)0 and T2 (j) are feasible and

w(T ∗ ) ≤ wmax and

a(U1 ) ≤ b(U2 ) − k(T1 ) − tij − w(U1 ).

The feasibility of the reverse tour T ∗ ’ is checked analogously. The combination is

performed, if at least one of the tours T ∗ and T ∗ ’ is feasible.

The feasibility of a schedule S now involves not only an assignment of the tours to
the vehicles, but also a feasible sequence of all tours assigned to the same vehicle.
S is feasible if, for T ∈ S, there exists an assignment v(T ) satisfying (8) and starting
times

ST (T ) starting time of tour T at the depot

such that

a(T ) ≤ ST (T ) + t0i ≤ b(T ), (9)

where i is the first customer of T , and for any v ∈ V and {T1 , . . ., Tz } = {T : v(T ) =
v}
5 Time windows 12

Av ≤ ST (T1 ) (10)
ST (Tp ) + t(Tp ) + t0 ≤ ST (Tp+1 ) (p = 1, . . ., z − 1) (11)
ST (Tz ) + t(Tz ) ≤ Bv (12)
ST (Tz ) + t(Tz ) ≤ ST (T1 ) + umax . (13)

Note that the starting times ST (T ) are related to the assignment v(T ) and change
with it, whereas the time window a(T ), b(T ) and the waiting time w(T ) are fixed
properties of the tour T . The starting time ST (T ) has to be determined in the
assignment algorithm simultaneously with v(T ). An easy way to do that would be
to append every tour as the last tour to those tours already assigned to the same
vehicle. But this works only, if the tours are considered in the order of their time
windows, which may be quite different from the FFD-order, and gives poor results,
to our experience. It is preferable to use the order of the tours as in the assignment
algorithm in Section 3 and to insert the current tour at an appropriate position, as
follows:

Determination of the starting time ST(T) in the assignment

algorithm:
Let T be the tour and v the vehicle under consideration and i the first customer of
T.

If Av + t0i > b(T ) or Bv + t0i − t(T ) < a(T ), reject v. If uv = 0, set

ST (T ) = min(Av , a(T ) − ti0 ) (14)

otherwise, a sequence of tours T1 , . . ., Tz is already assigned to v. Try to insert T at

any position p = z, z − 1, . . ., 0, i.e. after Tp or as the first tour on v, as follows: Set

ST (T ) = min(ST (Tp ) + t(Tp ) + t0 , a(T ) − ti0 ), if p > 0

and ST (T ) according to (14), if p = 0. Shift the tours Tp+1 , . . ., Tz to the right as

far as (11) requires; if this is prevented by (9) for some Tq , try the same with Tq0 ,
if feasible. If only (13) is violated, shift the tours T1 , . . ., Tp to the right as far as
possible. If (9) - (13) are satisfied, perform the assignment v(T ) = v and go to
the next tour, otherwise try the next position p. If all positions have been checked
without success, try the same with T 0 , if feasible. If no feasible start times are found,
reject v and go to the next vehicle.
6 Application and Results 13

Note that the assignment algorithm is used in Test 3 to produce preliminary starting
times ST 0 (T ) only, like the assignment v 0 (T ). Only if a complete assignment has been
established successfully, ST 0 (T ) and v 0 (T ) replace the current ST (T ) and v(T ). The
above procedure is also used for determining ST (T ∗ ) in Test 2. In Test 1, T ∗ ’ is
inserted on vehicle v(Ti ) (i = 1, 2) at the position of the (withdrawn) tour Ti only.

Finally, we improve the sequence of customers within all tours of the final schedule
using the OROPT procedure as described in [19]. The criterion is distance, but we
impose the additional restrictions on the tour T to be improved that (i) the travel
time must not increase and (ii) a(T ), b(T ) remain within the limits given by av and
the preceeding tours, bv and the following tours, respectively, the latter being shifted
to the right, if necessary. The tours within a vehicle are processed from the last one
to the first one.

6 Application and Results

In this section we report on numerical experiences with the SPMU and on some
applications. Unfortunately, there are no standard test data for this case in the
literature, so that comparative studies are not possible. However, the data for the
three problem classes presented in the following are available from the author. All
data stem from real-life VRSP problems for daily deliveries. The first case (“Berlin”)
concerns a carrier delivering miscellaneous consumer goods in West-Berlin. These
data from 1985 have also been used in the fleet capacities studies [1,14]. We use
them to illustrate the difference between the standard SP and the SPMU. The two
other cases (“Duisburg” and “Dortmund”) are current applications of our procedure,
concerning the delivery of food from a depot in Duisburg/Rhein to supermarkets
and of beverages from a depot near Dortmund to retailers.

Table 1 summarizes the characteristics of the three cases. In all cases the multiple
use of the vehicles is very important, the average number of tours per vehicle being
at least 1.6, a typical situation for deliveries in an urban area such as Berlin or the
Rhein-Ruhr area. In the Duisburg case, this ratio is even nearly 3, as most tours are
very short (107 minutes in the average); this is due to a small number of customers
per tour and to the use of containers allowing for short unloading times.

Time windows occur only in the Duisburg and Dortmund cases. In the former, about
two thirds of the customers receive fresh products which are to be delivered by 9.00
a.m. within a time window of 2 or 3 hours. In the latter, nearly all customers have
a time window with an average length of 4.5 hours or two time windows, one in
the morning and one in the afternoon. The tightness of the time windows, i.e. the
ratio of the average time window length per customer over the total time horizon,
6 Application and Results 14

as defined in [6,20], is about 50% in both cases. Note that the time horizon (13.5
hours) is longer than the maximal time of use of a vehicle (10 hours). These time
window structures are quite different from those generated at random [18,20]. We
found that they can be well satisfied without waiting times, so we set wmax = 0.
Then there is no necessity for considering the time windows in the Savings function,
as proposed e.g. in [6].

The specified utilizations with respect to capacity and time are:

total weight delivered

Ucapacity = P
v Qv · (no. of tours of vehicle v)

and

total time of use

Utime =
umax · (no. of vehicles used)
The number of vehicles used is in the average somewhat lower than the number
of vehicles available. In fact, in the Dortmund case, one vehicle has been saved
permanently by the use of the VRP system. As an effect of the multiple use of the
vehicles, the utilization of the used vehicles is well balanced between capacity and
time. Only in the Duisburg case, the time utilization is low, as there is a bottleneck
of capacity for the delivery of the fresh products in the early morning and excess
capacity for the rest of the day. However, the all-day schedule saves about 8% of the
total distance, 4% of the time of use and 4% of the tours as compared to the separate
planning of the fresh and non-fresh products performed so far by the company. The
high utilization of the used vehicles is an indicator of a good quality of the schedule
for a given fleet. Note that the values in Table 1 are averages over several days, the
utilization at certain days is still higher.

The OROPT procedure yielded a reduction or 2.4% in distance and 1.1% in travel
time in the average in the Berlin case. In the other case no remarkable improvements
occurred because of the low number of customers per tour.

Table 2 compares the results of the SPMU with the standard parallel SP. The latter
was used iteratively to consider the multiple use of the vehicles, in two different
ways: First, we used a commercial VRP package which assigns the resulting tours
to the existing vehicles, one by one as far as possible, but allows for easy interactive
correction of the schedule, in particular switching customers between tours and fixing
or dissolving tours for a further iteration. This work, which was carried out by a
well-skilled student, took about half an hour for the final schedule of one day; it
is documented in [1]. As this procedure is difficult to follow, we automated it in
the following way: A schedule for the single use of the vehicles is determined as in
section 3, the tours are assigned to the vehicles using the assignment algorithm of
6 Application and Results 15

section 4, the tours in the fictitious vehicles are dissolved and rescheduled in the next
iteration where tmax is reduced to the maximum free time of any (real) vehicle. The
results are worse in any aspect than those of the interactive procedure, indicating
that the latter has been well performed. However, the extended savings procedure
still reduces the total distance by about 7%, with nearly no change in the other
criteria. The reason is, that it selects the best savings that still admit a feasible
schedule, as opposed to the iterative fixing and dissolving of tours.

As to the savings function, we implemented the parametric function of Paessens [17]

for both distances and travel times. However, all results shown in this paper are
based on the original function with distances. We found that parameter variations
yield only slight improvements, if any, and are not worth the additional effort. The
time-based savings function may be advantageous in case of tight time constraints,
but as Table 3 shows, it effected only a small reduction of travel time in all problems
considered and of the number of tours in the Berlin case; on the other hand the total
distance, which is the main cost factor for a given fleet, increases remarkably, and
in the Duisburg and Dortmund Case also the number of vehicles used.

Table 4 shows the computational effort (without OROPT) for selected problems
needed by the implementation in MS-Fortran on a Siemens PC with a 386/387 16
MHz processor. The computation times not only depend on the size of the problem
but also on the occurence of time windows and on the number of times a new
asignment of tours to vehicles is constructed; this equals to the number of successful
Tests 3 plus the number of rejected combinations due to unassignable tours. The
number of rejections also indicates the difficulty of the problem with respect to the
vehicle assignment and the importance of the SPMU compared to the SP. In this
sense, the Berlin and the Dortmund case are rather difficult, and the Duisburg case
is easy because its short tours admit many assignments.

Conclusions
We have developed an algorithm for the VRPMU which is based on the SP. It is dis-
tinguished by the integration of all necessary feasibility checks and the assignment
of tours to vehicles into a parallel procedure, which outperforms a multi-phase or
sequential SP. It yields satisfactory results for large-scale practical delivery problems
at low computation time. It also works well in case of time window constraints, at
least for the typical clustered time windows of delivery problems, although some
authors report on poor performance of the SP for the VRSPTW with randomly ge-
nerated time windows [18]. However, a definitive evaluation of the algorithm requires
further computational studies.
7 REFERENCES 16

7 REFERENCES
1. J. Binsfeld, Erprobung eines Verfahrens zur Bestimmung des optimalen Fuhr-
parks anhand eines Praxis-Falles. Diplomarbeit, Institut für Unternehmens-
forschung, Universität Hamburg (1989).

2. L. Bodin, B. Golden, A. Assad and M. Ball, Routing and scheduling of vehicles

and crews: the state of the art. Computers and Operations Research 10, No.2,
pp. 63-201 (1983).

3. M.A.G. Bocxe and C.B. Tilanus, Testing vehicle scheduling programs for milk
collection (case study). European Journal of Operational Research 20, pp. 25-
33 (1985).

4. N. Christofides, A. Mingozzi and P. Toth, The vehicle routing problem. In

Combinatorial Optimizations, N. Christofides, R. Mingozzi, P. Toth and C.
Sandi (eds.), pp. 315-338. John Wiley & Sons, New York (1979).

5. G. Clarke and J.W. Wright, Scheduling of vehicles from a central depot to a

number of delivery points. Operations Research 12, pp. 568-581 (1964).

6. M. Desrochers, J.K. Lenstra, M.W.P. Savelsbergh and F. Soumis, Vehicle rou-

ting with time windows: optimization and approximation. In: Vehicle routing:
methods and studies, B.L. Golden and A.A. Assad (eds.), pp. 65-84. North-
Holland, Amsterdam (1988).

7. G. Diruf, Ein heuristisches Verfahren zur Berücksichtigung heterogener Fahr-

zeugeinsatzcharakteristiken in der computergestützten Tourenplanung.
In: Operations Research Proceedings 1983, H. Steckhan et al. (eds.), pp. 241-
248. Springer-Verlag, Berlin - Heidelberg (1984).

8. G. Diruf, Kundenzeitschranken in der computergestützten Tourenplanung.

Zeitschrift für Operations Research. 24, B207-B220 (1980).

9. M.L. Fisher and R. Jaikumar, A generalized assignment heuristic for vehicle

routing. Networks 11, pp. 109-124 (1981).

10. M.R. Garey and D.S. Johnson, Approximation algorithms for bin packing pro-
blems: a survey. In: Analysis and design of algorithms in combinatorial opti-
mization, G. Ausiello and M. Lucertini (eds.), pp. 147-165. Springer-Verlag,
Wien New York (1981).

11. B. Golden, A. Assad and R. Dahl, Analysis of a large-scale vehicle routing pro-
blem with an inventory component. Working paper MS/S 84-015, University
of Maryland, College Park (1984).
7 REFERENCES 17

12. B. Golden, L. Bodin and T. Goodwin, Microcomputer-based vehicle routing

and scheduling software. Computers and Operations Research 13, pp. 277-285
(1986).

13. B. Golden, T. Magnanti and H. Nguyen, Implementing vehicle routing algo-

rithms. Networks 7, pp. 113-148 (1977).

14. H. Koriath, Zur Optimierung von Fuhrparkstrukturen bei Berliner Unterneh-

men des gewerblichen Straßengüter-Nahverkehrs. Doctoral thesis, Fachbereich
Informatik, Technische Universitat Berlin (1988).

15. F. Matthäus, Tourenplanung. Toeche-Mittler Verlag, Darmstadt (1978).

16. H. Paessens, Tourenplanung bei der regionalen Hausmullentsorgung. Schrif-

ten des Instituts für Siedlungswasserwirtschaft Nr. 26, Universitat Karlsruhe
(1981).

17. H. Paessens, The savings algorithm for the vehicle routing problem. European
Journal of Operational Research 34, pp. 336-344 (1988).

18. M.M. Solomon, Algorithms for the vehicle routing and scheduling problem
with time window constraints. Operations Research 35, pp. 254-265 (1987).

19. M.M. Solomon, E.K. Baker and J.R. Schaffer, Vehicle routing and scheduling
problems with time window constraints: efficient implementations of solution
improvement procedures. In: Vehicle routing: methods and studies. B.L. Gol-
den and A.A. Assad (eds.), pp. 85-105. North-Holland, Amsterdam (1988).

20. H.R.G. van Landeghem, A bi-criteria heuristic for the vehicle routing problem
with time windows. European Journal of Operational Research 36, pp. 217-226
(1988).
A Table 1. Characteristics of the three cases
Case Berlin Duisburg Dortmund
Problems 20 10 5
Customers
average per day 241 314 95
maximum per day 303 361 109
average per tour 19 2.8 3.8
maximum per tour 52 9 11
Time Windows
% customers 0 64 97
% tightness 0 46 50
Fleet
vehicles 9 42 18
vehicle classes 6 3 14
min.-max. capacity 1030-9020 kg 27-45 Contain. 2940-22500 kg
Tours
average per day 13 114 24
maximum per day 19 131 29
average per vehicle 1.8 2.9 1.6
maximum per vehicle 4 5 3
Utilization
% of capacity 88.3 94.4 83.5
% of time 92.1 64.3 77.8
% of vehicles 84.4 91.9 81.5

B Table 2. Comparison of different procedures for

the Berlin case (All values are averages per day
from 20 problems)
SF + iterative adaptation SPMU
automatic interactive
total distance in km 694 668 624
total duration in minutes 3404 3262 3253
tours 14.6 12.6 12.8
vehicles used 7.6 7.1 7.1
 C Table 3. Impact of the Savings function based on
travel times instead of distances (All values are
changes in %)
Case Berlin Duisburg Dortmund
total distance + 2.5 + 2.9 + 1.0
total duration - 0.6 - 1.0 - 1.0
tours - 1.9 + 2.7 + 1.8
vehicles used 0.0 + 1.3 + 5.6

D Table 4. Computation time and effort for vehicle

assignment
Case Berlin Duisburg Dortmund
Time windows no yes yes
Customers 194 240 270 303 290 344 361 68 109
Computation time 12 26 26 32 74 115 63 7 12
(seconds)
No. of checks on vehicle as- 276 835 417 345 173 243 230 198 302
signment
- Test 1 successful 129 144 153 172 19 42 66 32 32
- Test 2 successful 36 42 59 55 74 92 152 5 23
- Test 3 successful 18 38 47 65 74 96 12 14 29
- rejections 93 611 158 53 6 13 0 147 218

View publication stats