CRT9513, ORWP 95/03, October 1995. To appear in Journal of heuristics 1, 1995, pp.

147 167
PROBABILISTIC DIVERSIFICATION AND
INTENSIFICATION IN LOCAL SEARCH FOR
VEHICLE ROUTING
Yves Rochat
Dpartement de mathmatiques, cole Polytechnique Fdrale de Lausanne, 1015 Lausanne, Suisse
e-mail : [email protected] WWW: http://dmawww.ep.ch/rose.mosaic/rochat.html
ric D. Taillard
Centre de recherche sur les transports, Universit de Montral, C. P. 6128, succursale Centre-Ville, Montral, Canada H3C 3J7
e-mail : [email protected] WWW: http://www.iro.umontreal.ca/people/taillard
Abstract :
This paper presents a probabilistic technique to diversify, intensify and parallelize a local search adapted for solving vehicle routing problems. This
technique may be applied to a very wide variety of vehicle routing problems and local searches. It is shown that efcient rst level taboo searches
for vehicle routing problems may be signicantly improved with this technique. Moreover, the solutions produced by this technique may often be
improved by a post-optimization technique presented in this paper too. The solutions of nearly 40 problem instances of the literature have been
improved.
Key words : Vehicle routing, local searches, parallel algorithms.
1. INTRODUCTION
More and more, local search methods are used to nd
good solutions to combinatorial optimization problems.
Throughout the paper, we use the term local search as a
synonym of neighbourhood search. Local searches are
sometimes restricted to steepest descent algorithms but
we also include taboo search and simulated annealing in
local search methods. These techniques have two main
weaknesses : first they may sometimes be trapped in a
very poor local optimum and this difculty may be hard
to overcome even by a fastidious tuning of the local
search parameters ; second, these methods require a
large computational effort. Nowadays, the most prom-
ising way of cutting down computational time is to use
parallel computers. Unfortunately, local search is a
process that is intrinsically sequential and therefore not
always easy to parallelize.
In this paper, we describe a technique that overcomes
both weaknesses. We illustrate this technique on two
taboo searches we have developed for vehicle routing
problems (VRPs) : first for the most elementary VRP
(one depot, identical vehicles) and second for the VRP
with time windows (VRPTW). This technique may also
be applied to other local searches or other VRPs. In
section 2, we briefly describe the problems treated and
the local search used to solve them. In section 3, we
present the new technique that allows us rst to diversify
the search by exploring solutions that are very different
from each other, second to intensify the search in order
to identify better local optima in a promising region of
the set of feasible solutions, and third to parallelize the
process. Then, we present a post-optimization technique
that allows frequent improvement in the solutions
produced by the diversification and intensification
procedure. In section 4, we compare the new technique
to the original local searches on which they are based
and nally we conclude in section 5.
2. VRP AND LOCAL SEARCH
2.1. Presentation of the problems
The rst problem we treat is the following : identical
vehicles with a fixed carrying capacity Q must deliver
order quantities q
i
(i = 1, , n) of goods to n customers
from a single depot (i = 0). Knowing the distance d
ij
between customers i and j (i, j = 0, , n), the problem is
to nd tours for the vehicles in such a way that :
The total distance travelled by the vehicles is
minimized.
Only one vehicle handles the deliveries for a given
customer.
The total quantity of goods that a single vehicle
delivers cannot be larger than Q.
The second problem we treat is similar to the
elementary VRP, but distances are interpreted as dura-
tions and each customer i requires a service time s
i
and
must be served during a time window [b
i
, e
i
]. When a
vehicle arrives at customer i before time b
i
, it must wait.
Solomon (1987) has proposed a set of 56 problems
where the first objective is to minimize the number of
vehicles needed to deliver all orders and the second
objective, subject to optimizing the rst, is to minimize
the total distance travelled by the vehicles.
2 Yves Rochat, ric Taillard
2.2. Local search for VRP
Very generally, a local search can be formulated as
follows :
(a) Choose an initial solution s
0
; set k := 0 ;
(b) While a stopping criterion is not satised, repeat :
(b1) Choose a solution s
k + 1
N(s
k
), the set of
neighbour solutions of s
k
.
(b2) Set k := k + 1
The choice of a policy for points (a) to (b2) leads to
various iterative searches. Generally, the initial solution
is chosen in such a way that its generation is fast and
easy. The choice of the stopping criterion is often related
to the type of iterative search chosen : for example, a
descent method stops as soon as there is no better solu-
tion than s
k
in N(s
k
) ; a taboo search stops when the
number k of iterations is greater than a threshold K. In
fact, common terminology equates local search with a
descent method. We use a more general terminology
based on allowing the rules for choosing among its
elements to follow a broader design than customary.
From the perspective of the formulation adopted here,
the way of dening N(s) and the way of selecting a solu-
tion in N(s) are the most difcult parts of the design of a
local search. For the VRP, the reader may refer to the
works of Taillard (1993), Gendreau et al. (1994) or
Gendreau, Laporte and Potvin (1995) for more details
on these choices.
The method of Taillard (1993) is one of the most ef-
cient for the elementary VRP. Its main feature is to
partition large problems into independent subproblems
and to optimize each subproblem independently. A
partition generates sets (subproblems) involving four to
eight tours (or 30 to 60 customers) that are near one
another. Once the subproblems are optimized, all the
tours of the subproblems are grouped together to
construct a solution to the original problem and this
solution is again partitioned, and so on. There is a
random component in the partition process. Therefore,
the algorithm may produce very different solutions from
one run to the next. The taboo search we have used in
this paper has been slightly changed from the original
version of Taillard (1993) : the partition procedure has
been improved and the exact procedure used for the
optimization of the tours has been replaced by a
heuristic approach that is more reliable. Generally, these
modifications improve the behaviour of the method on
non-uniform problem instances.
For uniform problems (i. e. , with customers
uniformly distributed in the plane around the depot and
ordering quantities much smaller than the vehicle
capacity), this algorithm nds good solutions in a small
amount of computation time. If the problems are non-
uniform, the search may be trapped in a poor solution
with a high probability. Since most real-life problems
are non-uniform (see e. g. the 120-city problem of
Christofides et al. (1979), Taillard (1993), Semet and
Taillard (1993), Fisher (1994), Rochat and Semet
(1994)), it is worth designing a method that works well
on this type of problems.
The taboo search we used for the VRPTW is derived
from the adaptation of Rochat and Semet (1994) for a
real-life problem that is more complex than the
VRPTW; for example this real-life problem incorpo-
rates an heterogeneous fleet, drivers breaks and
accessibility constraints (where each customer can only
be reached by a subset of the vehicles). So our taboo
search for the VRPTW is more nearly a simplication of
this complex taboo search than a new taboo search
speci al l y desi gned for t he Sol omons VRPTW
instances. However, the resulting method is competi-
tive, since it succeeds in improving the quality of 16
previous best known solutions out of the 56 Solomon
instances (see table 6). There is a random component in
our taboo search for the VRPTW. This means that two
runs of the method will generally produce two different
solutions.
3. IMPROVING LOCAL SEARCHES
APPLIED TO THE VRP
3.1. Probabilistic diversication and
intensication technique
A fundamental principle of taboo search is to exploit
the interplay between diversication and intensication
where diversication drives the search to examine new
regions, and intensification focuses more intently on
regions previously found to be good. (Intensification
typically operates by re-starting from high quality solu-
tions, or by modifying choice rules to favour the
inclusion of attributes of these solutions).
Our approach to achieving such an exploitation is
based on two primary perspectives, which we describe
as follows. The first of these comes from probabilistic
taboo search, which is founded on the idea of translating
information generated by the search history, coupled
with current measures of attractiveness, into evaluations
that are monotonically mapped into probabilities of
selection. Operating in a neighbourhood framework, the
approach then successively selects among available
alternatives according to a probability assignment that is
strongly biased to favour the choice of higher evalua-
tions. (In contrast to some terminology, taboo search
refers to neighbourhoods that are constructive and
destructive as well as transitional, since it includes strat-
egies not only of restarting, but also of alternating
between constructive and destructive steps).
Probabilistic diversication and intensication in local search for vehicle routing 3
This approach is motivated by the following
premise : an intelligent use of randomization, which is
not blindly uniform but embedded in probabilities that
account for history and measures of attractiveness,
offers a useful type of diversication that can substitute
for more complex uses of memory. As noted in Glover
(1989), the use of randomization, via assigned prob-
abilities, allows a gain in efficiency by obviating
extensive record keeping and evaluation operations that
a more systematic pursuit of diversity may require. We
take advantage of this means of achieving diversity in a
special way, by rules, we indicate subsequently.
The second main perspective that underlies our
approach derives from one of the most basic (and
earliest) types of intensication strategies. The heart of
this approach lies in generating solutions by reference to
the notions of strongly determined and consistent varia-
bles. A strongly determined variable is one whose value
cannot be changed except by inducing a disruptive effect
on the objective function value or on the values of other
variables. The identification of strongly determined
variables is by reference to best solutions from previous
solution efforts, which motivates us to measure their
strength by the quality of the solutions in which varia-
bles lie (at particular values). This weighting by
objective function values corresponds to the practice of
emphasizing relative attractiveness in probabilistic
taboo search, allowing us to exploit both approaches
together.
A consistent variable is one that is frequently
strongly determined at a particular value (or in a narrow
range). Specically, the more often a variable receives a
particular value in the best solutions (where we weight
these solutions by their objective value), the more highly
it qualies as consistent.
The rationale for isolating such variables, and the
strategy for exploiting them, are embodied in the
following expectations. First, a variable that is highly
consistent is likely to receive its preferred value (or lie
in its preferred range) in optimal and near-optimal solu-
tions. Second, once some variables are assigned specic
values, other variables that previously did not seem
highly consistent will now become a good deal more so.
Third, imposing narrow restrictions on selected varia-
bles will yield increasingly reliable measures of the
relative consistency of remaining variables, given the
imposed restrictions.
The strategy to take advantage of these tendencies
may then be summarized as follows (Glover, 1977).
(a) Select one or more variables with greatest relative
consistencies and constrain these to their preferred
values.
(b) Determine new relative consistencies for the vari-
ables on the basis of the restriction of step (a).
(c) Repeat the process until all variables have been
constrained to specic values.
This process is then joined with a heuristic improve-
ment procedure to transform its value assignments into
a new solution, thus creating an iterative method for
obtaining progressively more rened outcomes.
In our approach for executing the preceding steps in
the vehicle routing setting we select variables in blocks,
composed of specific tours to which customers have
been assigned in the best previous solutions. The new
relative consistency measures produced for step (b) then
result by observing that once particular variables are
assigned values, we cannot reassign them within the
same solution that is, we cannot select a new tour
containing customers already assigned on a previous
step. Our improvement heuristics for obtaining better
solutions, and which are the drivers of the method, are
those of Taillard (1993) and Rochat and Semet (1994).
We now sketch the details of our approach more
precisely. We emphasize that our method can be applied
to most forms of local search (as encompassed by our
rather general denition), and to most types of VRPs. To
play a diversication rle, a local search must be able to
produce solutions that are very different from each
other, but not necessarily among the very best. This
condition is generally satised by iterated local searches
that either start with solutions randomly generated or
that apply a random component in their subsequent deci-
sions, thereby causing different runs generally to
produce different solutions. However, by incorporating
the principle of probabilistic taboo search to guide our
choices, we create a more strategic type of diversifica-
tion than that embodied in simple randomization. The
technique we employ may be described as follows:
In a rst phase (initialization), the search is diversi-
ed by generating, with the local search, I solutions that
are different from one another. For our approach to the
elementary VRP, this generation is done by considering
various initial decompositions of the problem. For the
VRPTW as well as for elementary VRPs of small
size that cannot be decomposed the non-determin-
istic characteristics of our local search guarantee the
diversity of the solutions.
By generating several initial solutions with the local
search, one hopes that all the information necessary to
create solutions of very high quality exists in these solu-
tions, but in a non-apparent way. In the case of the VRP,
this information is included in the tours. So, one hopes
that the initialization creates a set of tours that includes
members not very different from the tours of a good
solution. It is not unrealistic to think that it is easy to
create good tours. The challenge is to nd a set of tours
whose members are simultaneously good for all
customers.
4 Yves Rochat, ric Taillard
The generation of I initial solutions creates a set T of
tours. In gure 1, we give a subset of tours of T after the
initialization phase for a problem of Christofides et al.
(1979) with 199 customers. This figure may be
compared with figure 2 where a very good solution to
the same problem is given. We see that several tours are
similar, two of them being identical. However, T
contains several other tours and, after the initialization
phase, the good tours cannot be directly identified. In
these gures, the area of the customers (empty circles) is
proportional to the quantity ordered ; the lled circle is
the depot and its area is proportional to the capacity of
one vehicle ; the first and last trips of each vehicle are
not drawn.
In a second phase, the goal is to extract these good
tours and improve them. We start from the principle that,
if a solution s includes tours that belong to a good solu-
tion, then the objective function value of s is probably
better than that of a solution that does not contain such
tours. The second phase thus favours the extraction of
tours that belong to the best solutions generated during
the initialization phase. This extraction must not totally
exclude tours belonging to bad solutions.
To implement this phase, each tour is labelled with
the value of the solution to which it belongs. The set T is
sorted by increasing values of the labels and the tours
with only one customer are removed from T, since they
do not contain interesting information. Then, we choose
tours of T probabilistically, by giving preference to tours
with low labels and by ignoring tours that include
customers belonging to tours already extracted. This
choice is repeated until it is not possible to extract new
tours from T.
Figure 1 : Example of tours belonging to T after the initializa-
tion phase.
Let S be the set of tours thus extracted. Since S may
not contain all the customers of the problem, S is a
partial solution. In order to construct a feasible solution
fromS, the set of customers not belonging to the tours of
S may be considered as an independent VRP (of small
size) that can be solved by the local search. The tours of
this independent VRP are added to S to create a feasible
solution to the initial VRP. This feasible solution is
considered as the initial solution of a local search that
will try to improve its quality. Another way to initialize
the local search is to modify the procedure, embedded in
the local search, that produces an initial solution from a
set of customers. This procedure is modified in such a
way that it also accepts the tours of the partial solution
S. In our implementations, we have chosen this way to
initialize the local searches.
Once a new solution is generated by the current
application of local search, the tours of this solution are
labelled with the value of the objective function and are
included in the set T. It is important to note that the same
tour may occur in more than one solution. In order to
represent this fact appropriately (in the sense of identi-
fying the relative consistency of consistent variables), a
tour is included in T multiple times, one for each solu-
tion in the set T. The extraction of tours of T, followed
by the optimization with the local search and the inser-
tion of the new tours in T is repeated until a stopping
criterion is met. More formally, this algorithm may be
formulated as follows :
Figure 2 : Best known solution for the problem of Christodes et
al. (1979) with 199 customers.
Probabilistic diversication and intensication in local search for vehicle routing 5
Diversication and intensication algorithm :
Initialization :
(a) Generate I different initial solutions with the local
search.
(b) Label each tour with the value of the solution to
which it belongs.
(c) Remove the tours having only one customer.
(d) Insert the remaining tours in a set T of tours.
(e) Sort T by increasing values of labels.
Diversication and intensication (to repeat until a
stopping criterion is satised) :
(1) Set T := T, S := .
(2) While T , repeat
(2a) Choose t T, probabilistically, based on its
current relative evaluation.
(2b) Set S := S {t}.
(2c) Remove from T all the tours including one
or several customers belonging to t.
(3) If some customers are not covered by the tours of
S, construct a feasible solution S, including them,
using the partial solution S.
(4) Improve with the local search the solution S.
(5) Label the tours of the improved solutions, remove
tours with only one customer, insert the remaining
tours in T, sort T as in steps (b) to (e) of the
initialization.
To favour the generation of good solutions, the tours
are not uniformly chosen at step (2a), but the ith worst
tour of T has a probability of 2i/(| T| | T + 1|) of being
chosen. For practical reasons, it is necessary to limit the
size T to a value L. After each sorting of T, the last | T|
L tours are removed from T if | T| > L.
We note this procedure also contains an implicit
aspect of combining solutions since we construct new
ones out of components of previous ones. Hence in this
sense our approach embodies some of the spirit of
genetic algorithms (see, e. g., Holland (1976) and Davis
(1987)). This implicit combination by intensication
(by exploiting strongly determined and consistent varia-
bl es ) , whos e or i gi ns ar e appr oxi mat el y
contemporaneously with those of genetic algorithms
provides a useful counterpart to combination by
genetic operators.
The working principles of this algorithm may be
explained as follows : If I and L are large enough, the
initialization phase guarantees that various regions of
the solution space will be explored. The creation of
partial solutions at the beginning of the second phase
(step (2)) allows the search to extend the diversity of the
solutions visited.
As far as the process goes, T grows and the partial
solutions are more and more complete. At the end, they
are often feasible or even, in few cases, better than the
best solution found so far. After having performed the
second phase several times, the process automatically
intensifies the search in promising regions, since the
tours are not uniformly chosen at step (2a) and the worst
solutions of T are removed. Moreover, identical tours
are not removed fromT and, more and more often, there
exist tours that are not modied by the local search. So,
the best tours of T are more and more frequently
extracted during the construction of the partial solution
and the search progressively changes from a diversica-
tion to an intensication process.
This process may easily be parallelized. The steps (a)
and (4) (applying the local search) are those consuming
the most computation time. A master-slave approach is
convenient : The master process executes steps (b) to (e)
of the initialization phase and steps (1), (2), (3) and (5)
of the diversication and intensication phase. In addi-
tion I slave processes independently perform one local
search of step (a), then transmit one initial solution to the
master process and enter a loop in which they wait for a
partial solution from the master process, improve this
solution with the local search (step (4) ) and finally
transmit the improved solution to the master process.
3.2. A post-optimization technique
When examining the empirical behaviour of this
procedure, it turns out that during the diversication and
intensication phase, the process sometimes succeeds in
improving the best solution directly from the solution
constructed at step (2) before applying the local search
of step (4). This means that, with the tours contained in
the set T, it may be possible to build solutions better than
those already produced.
Let c
j
be the length of the j
th
tour of T (j = 1,, | T|)
and, for i = 1, , n and j = 1,, | T| :
Then, the best solution that can be built using tours
of T may be found by solving the following set parti-
tioning problem :
The assignment x
j
= 1 indicates that the j
th
tour of T
is chosen. By extension, the solution space may be
a
ij
1 if customer i j
th
tour of T
0 Otherwise

=
min c
j
x
j
j 1 =
T

s. t. a
ij
x
j
j 1 =
T

1 = i 1 n , , =
x
j
0 1 { , } j 1 T , , =
6 Yves Rochat, ric Taillard
enriched by first adding tours containing only one
customer (c
j
= d
0 j | T|
+ d
j | T| 0
, j = | T| + 1,, | T| + n)
and second by replacing the equalities in the constraints
by inequalities (allowing customers to be visited more
than once when distances satisfy the triangle inequality,
since then customers visited more than once can simply
be removed from the inappropriate tours).
However, these extensions make the problem harder
to solve and, practically, do not improve the quality of
the solutions built. We propose instead to include in the
model only the subset of single customer tours that have
been produced by the local search. There is no merit
embedding this post-optimization technique in the
diversification and intensification method, since this
would not enrich the set T with new tours. Thus, the
appropriate application of this partitioning approach is
at the conclusion of the iterated search efforts, which is
why we call it a post-optimization technique.
4. COMPUTATIONAL RESULTS
4.1. Diversication and intensication
technique
In this section, we analyse the behaviour of the diver-
sication and intensication technique when used with
our previous taboo searches (Taillard (1993) and Rochat
and Semet (1994)). First, we identify the parameter
values of our numerical experiments.
The number of initial solutions, I, is set to 20.
The number L of tours kept in the set T is set to
260.
The number of iterations performed by the local
search at steps (a) and (4) has to be xed : for the
elementary VRP, we have chosen to perform six
decompositions of the problem, followed by opti-
mizations, this means a total number of 14n
iterations, where n is the number of customers ;
for the VRPTW, we perform 2000 iterations.
These parameters were chosen in a relatively arbi-
trary fashion and we have not undertaken to fine tune
their values ; therefore, it is likely the results that follow
can be improved. All tests were performed on a Silicon
Graphics Indigo (100 Mhz) and all the computation
times are given in seconds.
Elementary vehicle routing problem
In table 1, we give the computation times for the
elementary VRP required by the previous local search
approach of Taillard (1993) and by our new method to
produce solution at three different average levels of
quality. Specically, these levels are selected to be 5%,
2% and 1% above the best known solution of certain
instances proposed by Christofides et al. (1979),
(without tour length limit), Fisher (1994), (with 71 and
134 customers) and Taillard (1993), (with 385
customers). We have run the initial algorithm and our
new method ve times. A time is printed in bold charac-
ters if it is significantly (more than 1.5 times) smaller
than the time required by the alternative method shown.
We refer to the local search approach of Taillard (1993),
as adapted to the current application, a first level local
search, because we also embed it (with smaller iteration
limits) in our diversication and intensication method.
We see that, for most instances of Christodes et al.
(1979), our new technique is generally not competitive
with the rst level local search (at least for the parame-
ters we have chosen). However, the instances with 120,
71, 134 and 385 customers are generally better solved
with our new technique.
Taillard (1994) has observed that real life quadratic
assignment problems may sometimes be very poorly
solved by good local search methods. Real-life prob-
lems are not necessarily the most difcult to solve (the
local searches sometimes find the best known solution
extremely rapidly, though not at each run), but the local
searches may often be trapped in very poor local optima.
Therefore, the average solution values they produce are
not very good.
We think that similar behaviour occurs for the VRP,
as suggested by the preceding results for the instances of
Christofides et al. (1979) with 120 customers, and the
selected instances of Fisher (1994) and Taillard (1993).
We have tried to generate problems with these chal-
lenging characteristics by using a procedure similar to
the one described in Taillard (1994) to generate non-
uniform quadratic assignment problems : the customers
are located on the plane (we have used Euclidean
distances) and are spread in several clusters. The
number of such clusters and their compactness can be
quite variable. The quantities ordered by the customers
are exponentially distributed (so, one customer may use
almost the entire capacity of one vehicle). We have
generated four problem of each size : 75 customers (9 or
Problem
size
First level local search
Diversication and
intensication
5% 2% 1% 5% 2% 1%
50 2.8 7.3 11 2.8 7.3 11
75 1.5 9.4 27 1.5 11 68
100 15 94 420 15 57 900
100b 33 140 180 51 280 350
120 >5000 690 2100 2700
150 30 250 >3900 25 670 1800
199 110 1100 >14000 89 >3700
71 58 >1400 190 1080 1130
134 32000 >35000 520 2300 3100
385 370 8800 >23000 1300 5400 18000
Table 1 : Computation times needed to find solutions that are,
on average, a specied per cent above the best known
value ; problems of Christofides et al. (1979), Fisher
(1994) and Taillard (1993).
Probabilistic diversication and intensication in local search for vehicle routing 7
10 vehicles), 100 customers (11 or 12 vehicles) and 150
customers (14 or 15 vehicles).
In gure 3, we give the best solution we have found
for problem 150a. This figure is to be compared with
gure 2 in which we give the best known solution for the
problem of Christofides et al. (1979) with 199
customers. We see that the structures of these problems
and their solutions are very different. The solution struc-
ture of problem 150a may also be compared to the real-
life problems proposed by Fisher (1994), and we see
that the structures are similar.
In table 2, we provide the same statistics for non-
uniform problems as in table 1. We see that our new
technique may vastly improve the performance of the
local search on these problems, especially when very
Figure 3 : Best known solution for a non-uniform problem with
150 customers.
good solutions are required. In addition, we show in this
table that the non-uniform problems we have generated
are not only ill-conditioned for our first level local
search (Taillard, 1993) but also for the taboo search of
Gendreau et al. (1994) : we give the computation times
needed by this method and the solution quality it has
found (measured in per cent above the best known solu-
tions). We see that the algorithm is very often trapped by
local optima of poor quality, even for the problems with
75 customers. It is shown in Taillard (1993) that the
algorithm of Gendreau et al. (1994) is superior for a few
problem instances of Christodes et al. (1979).
All the computational results given here are for
sequential computers ; our new technique may easily be
parallelized and an implementation on distributed
computers would reduce the computation times by one
order of magnitude or more, depending on the number of
processors.
Vehicle routing with time windows
We now present computational results obtained on
Solomons VRPTW instances. Each of these 56
instances has 100 customers and the travel time between
customers are equal to the corresponding Euclidean
distance. As these problems vary in fleet size, vehicle
capacity, travel time of vehicles, spatial distribution of
customers, time window density and width they are
divided into three categories : R-type (uniformly distrib-
uted customers), C-type (clustered customers) and RC-
type (a mix of R and C types). Two sets of problems are
proposed for each of these three categories. Problem sets
R1, C1 and RC1 have narrow scheduling horizon while
problem sets R2, C2 and RC2 have large scheduling
horizon. Narrow scheduling horizon problems have
vehicles with small capacities and short route times
hence only a few customers can be served by the same
vehicle. Conversely, large scheduling horizon problems
Problem
size
First level local search
Diversication and
intensication
Gendreau et al.
(1994)
5% 2% 1% 5% 2% 1% Time %
75a 3.3 30 170 6.5 18 48 1900 1.2
75b 2.8 14 170 6.4 13 17 1700 0.7
75c 15 > 2100 14 120 170 1600 9.6
75d 4.4 29 52 7.5 13 17 1700 1.1
100a 13 53 > 2000 10 220 720 3600 4.6
100b 21 110 460 15 38 150 2900 2.3
100c 35 190 > 6800 55 140 770 2900 5.1
100d 25 120 > 2100 28 180 680 3300 4.7
150a > 3700 480 1600 2800 7000 4.3
150b 78 > 3100 25 390 1100 8600 2.8
150c 33 > 2800 84 420 820 6400 8.1
150d 100 2000 > 3800 96 980 1700 5800 5.0
Table 2 : Computation time needed to find solutions that are, on average, a specified per cent above the
best known value and performances of the algorithm of Gendreau et al. (1994) ; non-uniform
problems.
8 Yves Rochat, ric Taillard
use vehicles with large capacities and long travel times
so more customers can be served by the same vehicle. In
table 3, we compare the solutions values produced by
five runs of the initial taboo search for the VRPTW to
those produced by one run of the diversification and
intensification method. For each type of problem, we
give : the average CPU time needed to perform 50000,
150000 and 300000 iterations of taboo search (for the
diversification and intensification technique, this CPU
time corresponds to 25, 75 and 150 calls to a rst level
taboo search performing 2000 iterations), the average
solution value of the best out of five runs, the average
value of all runs and the average solution value of the
worst run. The first number is the average number of
vehicles and the second is the average distance travelled.
Again, we see that the diversication and intensication
technique vastly improves the local search. The best
over five of the first level taboo search runs is often
worse than the average solution produced by the diver-
sication and intensication technique, this means that
the new method is more efficient and more robust, at
least for the longest runs. We have also observed that,
when running the method several times, the gap between
the best and the worst run is much smaller for the diver-
sication and intensication than for the rst level taboo
search. Therefore, the new method is more robust and it
was not necessary to have average solutions on several
runs in table 3.
4.2. Post-optimization procedure
The post-optimization procedure has allowed us to
improve the quality of many best known solutions of
problem instances of the literature. The improvements
that can be obtained with the post-optimization proce-
dure are not very large (generally less than one per cent),
but these improvements may often be obtained with a
modest increase in computation time. We illustrate this
fact on the new elementary VRP instances we propose.
We consider 5 runs of our diversication and intensica-
tion technique with I = 20 initial solutions and 50
(respectively 70) diversication and intensication steps
for the problem instances with 100 (respectively 150)
customers. All solutions produced by the rst level local
searches (i. e. 350 or 450 for each problem) have been
stored. Then, for each problem instance, we consider a
set T that contains 250 different tours chosen from the
best solutions produced by the diversification and
intensification procedure. Finally, we have solved the
set partitioning problem induced by T in an exact way,
using the cplexmip software (Cplex inc., 1993).
Problem
size
Diversication
and intensication
Post-optimization
Best
known
solution Value Time Value Time
100a 2062.27
1000
2047.90 5.5 2047.90
100b 1940.61 1940.61 3.9 1940.61
100c 1421.59 1407.44 71 1407.44
100d 1581.25 1581.25 1.2 1581.25
150a 3077.77
2100
3070.91 0.43 3055.23
150b 2735.16 2733.60 1.9 2727.99
150c 2367.65 2364.31 18 2362.79
150d 2668.34 2663.20 22 2655.67
Table 4 : Comparison of the solutions produced with the
diversification and intensification technique and
with the post-optimization procedure.
Type
CPU
time
Initial taboo search Diversication and
intensication Minimum Average Maximum
R1
450 12.75 1204.81 13.00 1225.62 13.33 1255.71 12.83 1208.43
1300 12.67 1206.54 12.92 1223.74 13.17 1259.10 12.58 1202.31
2700 12.67 1204.76 12.92 1222.36 13.17 1257.72 12.58 1197.42
C1
540 10.00 831.24 10.00 838.93 10.00 863.10 10.00 832.59
1600 10.00 831.01 10.00 838.00 10.00 861.41 10.00 829.01
3200 10.00 830.32 10.00 836.87 10.00 858.75 10.00 828.45
RC1
430 12.63 1383.06 13.00 1418.58 13.25 1475.49 12.75 1381.33
1300 12.25 1396.48 12.88 1417.92 13.25 1470.99 12.50 1368.03
2600 12.25 1385.28 12.77 1418.36 13.25 1470.99 12.38 1369.48
R2
1600 3.36 991.20 3.62 996.03 3.91 1015.97 3.18 999.63
4900 3.36 988.75 3.62 992.48 3.91 1010.86 3.09 969.29
9800 3.36 988.43 3.62 990.09 3.91 1004.07 3.09 954.36
C2
1200 3.00 594.64 3.10 616.44 3.50 654.99 3.00 595.38
3600 3.00 591.43 3.00 611.25 3.00 654.91 3.00 590.32
7200 3.00 591.43 3.00 610.28 3.00 653.38 3.00 590.32
RC2
1300 4.00 1171.18 4.18 1249.80 4.38 1422.50 3.62 1207.37
3900 4.00 1166.78 4.18 1245.06 4.38 1417.02 3.62 1155.47
7800 4.00 1166.78 4.18 1244.77 4.38 1417.02 3.62 1139.79
Table 3 : Average solution values after given computation times for the initial taboo search and for the diver-
sication and intensication method. Problems of Solomon (1987).
Probabilistic diversication and intensication in local search for vehicle routing 9
In table 4, we give, for the non-uniform problems :
the number of customers, the value of the best solution
produced by the diversification and intensification
procedure over 5 runs, the average computation time of
one run of the diversication and intensication proce-
dure, the value of the solution produced by the post-
optimization procedure, the computation time of the
post-optimization procedure and nally we give the best
known solution value. In this table, we see that the diver-
sification and intensification procedure has found the
best known solution on half of the problems instances
with 100 customers. The post-optimization procedure
has succeeded in improving the other solutions and in
finding the best known solutions. For the problem
instances with 150 customers, we see that the average
gap between the best solution produced by the diversi-
cation and intensication procedure and the best known
solution has been reduced by nearly 40%. The running
times of the post-optimization procedure are often very
low but the standard deviation, for a given problem size,
is very high. So, this post-optimization procedure makes
it possible to rapidly improve the solutions produced by
a local search but the set partitioning problems gener-
ated varies enormously.
4.3. Best known solutions
Our diversification and intensification technique,
accompanied by the post-optimization procedure, has
allowed us to improve the quality of several of the best
solutions reported in the literature. Table 5 gives, for
elementary VRP instances : the reference where the
problem is described, the number of the problem in this
reference (when available), the size of the problem, the
value of the best solution published (with the reference
reporting the best solution published) and finally the
value of the best solution obtained during the elabora-
tion of this paper.
Tabl e 6 gi ves t he same i nformat i on for t he
Solomons VRPTW instances, but the origin of the
problem (Solomon, 1987) and the size of the problem
(100) are not reported and the value of the solutions is a
pair (number of vehicles and distance travelled). We see
in this table (bold characters) that the taboo search we
a. Fisher (1994)
b. Taillard (1993)
Problem source
Problem
number
Size
Best
solution
published
New best
Fisher (1994) 12 134
1163.60
a
1162.96
Christodes et al. (1979) 5 199
1298.79
b
1291.45
Christodes et al. (1979) 10 199
1397.94
b
1395.85
Taillard (1993) 385
24599.6
b
24435.5
Table 5 : Best solutions published and new best solutions
found for elementary VRP instances.
a. Desrochers et al. (1992)
b. Thangiah et al. (1994)
c. Potvin & Bengio (1994)
d. Chiang & Russel (1994)
Number
Best solution
published
Best solution
taboo search
Best solution we
found
R101 18 1607.70
a
19 1656.20 (19 1650.80)
R102 17 1434.00
a
18 1477.41 (17 1486.12)
R103 13 1207
b
14 1222.90 (14 1213.62)
R104 10 1048
b
10 1013.26 10 982.01
R105 14 1420.94
c
14 1404.75 14 1377.11
R106 12 1350
b
12 1293.92 12 1252.03
R107 11 1146
b
11 1085.77 10 1159.86
R108 10 989
b
10 965.28 9 980.95
R109 12 1205
c
12 1186.41 11 1235.68
R110 11 1105
b
11 1107.90 11 1080.36
R111 10 1151
b
11 1070.90 10 1129.88
R112 10 992
b
10 965.66 10 953.63
C101 10 827.30
a
10 828.94 10 828.94
C102 10 827.30
a
10 828.94 10 828.94
C103 10 835
b
10 828.06 10 828.06
C104 10 840
b
10 841.59 10 824.78
C105 10 828.94
c
10 828.94 10 828.94
C106 10 827
a
10 828.94 10 828.94
C107 10 827.30
a
10 828.94 10 828.94
C108 10 827
a
10 828.94 10 828.94
C109 10 828.94
c
10 828.94 10 828.94
RC101 14 1669
b
15 1737.03 (15 1623.58)
RC102 13 1557
b
13 1480.66 13 1477.54
RC103 11 1110
b
11 1264.30 (11 1262.02)
RC104 10 1204.07
c
10 1157.23 10 1135.83
RC105 14 1602
b
15 1543.16 13 1733.56
RC106 12 1485.67
c
12 1415.62 12 1384.92
RC107 11 1274.71
c
11 1262.43 11 1230.95
RC108 10 1281
b
11 1149.64 10 1170.70
R201 4 1354
b
4 1485.36 4 1281.58
R202 3 1530.49
c
4 1101.49 (4 1088.07)
R203 3 1126
b
4 912.98 3 948.74
R204 2 914
d
3 824.62 2 869.29
R205 3 1128
b
3 1205.58 3 1063.24
R206 3 833
b
3 956.05 (3 912.97)
R207 3 904
b
3 814.84 3 814.78
R208 2 759.21
c
3 708.78 2 738.60
R209 2 855
b
4 901.88 (3 944.64)
R210 3 1052
b
3 1087.32 3 967.50
R211 3 816
b
3 794.46 2 949.49
C201 3 591.56
c
3 591.56 3 591.56
C202 3 591.56
c
3 591.56 3 591.56
C203 3 591.55
c
3 591.17 3 591.17
C204 3 590.60
c
3 597.76 3 590.60
C205 3 588.88
c
3 588.88 3 588.88
C206 3 588.49
c
3 588.49 3 588.49
C207 3 588.32
c
3 588.49 3 588.29
C208 3 588.49
c
3 588.49 3 588.32
RC201 4 1249
b
5 1469.73 (4 1438.89)
RC202 4 1221
b
4 1443.66 4 1165.57
RC203 3 1203
b
4 1013.99 3 1079.57
RC204 3 897
b
3 843.12 3 806.75
RC205 4 1389
b
5 1286.70 4 1333.71
RC206 3 1213
b
4 1207.76 3 1212.64
RC207 3 1181
b
4 1079.07 3 1085.61
RC208 3 919
b
3 919.83 3 833.97
Table 6 : Best solutions published, best solutions found with
taboo search and new best solutions we found for the
Solomons problems.
10 Yves Rochat, ric Taillard
used (5 initial solution, 400000 iterations) has
improved or reached the quality of about 27 out of 56
best solutions previously published. Due to rounding
and truncating, the length quoted in the publications
may be inaccurate ; we report our results on real,
double-precision distances. The last digit has been
rounded. The last column reports the best solution we
have found during our computational experiments (in
most cases with the new method). In this column, when
a solution is not the best known, we have put it in paren-
theses. We see that we did not reach or improve a best
solution only 9 times.
Finally, we report the best solution values we found
for the problem instances D417 and E417 (417
customers, Russel, 1995) : For D417 we have found
54 / 6264.80 and 55 / 3467.83 ; for E417, we have found
54 / 7211.83 and 55 / 3693.24. We were not able to
substantially reduce the total length for the solutions
with 54 vehicles, therefore we give also the solution
length when using 55 vehicles. Thangiah et al. (1994)
report the solution values 54 / 4866 (D417) and
55 / 4149 (E417).
5. CONCLUSIONS
We have presented a probabilistic technique that
allows us to diversify, intensify and parallelize almost
any local search for almost any VRP. This technique
makes the local search more robust since it converges
more often toward solutions whose quality is close to
that of the best known solution. This technique has
several advantages :
First, it is relatively easy to design a local search that
locally nds good tours, but it is hard to design a search
that nds good tours for all customers simultaneously ;
our technique makes it possible to overcome this diffi-
culty and to design a fairly robust method more easily.
Second, this technique may be applied to several
types of VRPs, for example those including the
following constraints :
Time windows for the customer deliveries.
Differentiated vehicles (cost of use per kilometre,
volume capacity, carrying capacity).
Constraints on the tours (maximum length, driver
breaks, customers that cannot be reached by any
vehicle).
Backhauls.
Multiple depots.
Third, this technique may easily be parallelized with
an arbitrary number of processors (not depending on
problem size). But we have shown that it can be inter-
esting to use it even on a sequential computer.
Finally, the post-optimization procedure we propose
allows the solutions to be improved with a modest
increase in computation time.
Acknowledgements
The authors would like to thank Gilbert Laporte
whose valuable suggestions have improved the presen-
tation of this paper. This work was partly supported by
an NSERC International Post-Doctoral Fellowship and
by an NSERC Strategic Grant on Parallel Software for
Intelligent Vehicle-Highway System.
REFERENCES
W.-C. Chiang and R. Russell, (1993), Simulated
Annealing Metaheuristics for the Vehicle Routing Prob-
lem with Time Windows, working paper, Department of
Quantitative Methods, University of Tulsa, Tulsa, OK
74104.
N. Christofides, A. Mingozzi and P. Toth, (1979), The
vehicle routing problem, in : Combinatorial optimiza-
tion (N. Christofides, A. Mingozzi, P. Toth and
C. Sandi, eds), Wiley, Chichester.
Cplex Optimization inc. (1993), Using the CPLEX Cal-
lable Library and CPLEX Mixed Integer Library,
Incline Village, NV.
L. Davis (editor), (1987), Genetic Algorithms and Simu-
lated Annealing, Pitman, London.
M. Desrochers, J. Desrosiers and M. Solomon (1992), A
new Optimization Algorithm for the Vehicle Routing
Problem with Time Windows, Operations Research 40,
342 354.
M. L. Fisher (1994), Optimal solution of vehicle routing
probl ems usi ng mi ni mum K-t rees, Operat i ons
Research 42, 626 642.
M. Gendreau, A. Hertz, G. Laporte, (1994), A tabu
search heuristic for the vehicle routing problem, Man-
agement Science 40, 1276 1290.
M. Gendreau, G. Laporte and J.-Y.Potvin (1995), Local
search algorithms for the vehicle routing problem, in :
Local Search Algorithms, (J. K. Lenstra and E. H. L.
Aarts, eds), Wiley, Chichester.
B. E. Gillett and J. Johnson, Sweep algorithm for the
multiple depot vehicle-dispatch problem, ORSA
National Meeting, San Juan, 1974.
F. Glover, (1975), Surrogate constraint duality in math-
ematical programming, Operations Research 23, 434
451.
F. Glover, (1977), Heuristics for Integer Programming
Using Surrogate Constraints, Decision Sciences, Vol. 8,
No. 1, 156 166.
Probabilistic diversication and intensication in local search for vehicle routing 11
F. Glover, (1990), Tabu Search Part II, ORSA Jour-
nal on Computing 2, 4 32.
J. H. Holland, (1976), Adaptations in Natural and
Articial Systems, University of Michigan press, Ann
Arbor.
J.-Y. Potvin, S. Bengio, (1994), A Genetic Approach to
the Vehicle Routing Problem with Time Windows, Publi-
cation CRT-953, Centre de recherche sur les transports,
Universit de Montral.
J.-Y. Potvin, T. Kervahut, B.-L. Garcia, J.-M. Rousseau,
(1993), A Tabu Search Heuristic for the Vehicle Routing
Problem with Time Windows, Publication CRT-855,
Centre de recherche sur les transports, Universit de
Montral.
Y. Rochat, F. Semet, (1994), A tabu search approach for
delivering pet food and our in Switzerland, Journal of
the Operational Research Society 45, 1233 1246.
R. A. Russel (1995), Hybrid Heuristics for The Vehicle
Routing Problem with Time Windows, Transportation
Science 29, 156 166.
F. Semet, . D. Taillard, (1993), Solving real-life vehi-
cle routing problems efficiently using taboo search,
Annals of Operations Research 41, 469 488.
M. M. Solomon (1987), Algorithms for the Vehicle
Routing and Scheduling Problem with Time Window
Constraints, Operations Research 35, 254 265.
. D. Taillard, (1993), Parallel Iterative Search Meth-
ods for Vehicle Routing Problems, Networks 23, 661
676.
. D. Taillard, (1994), Comparison of heuristic methods
for the quadratic assignment problem, Publication CRT-
989, Centre de recherche sur les transports, Universit
de Montral.
S. R. Thangiah, I. H. Osman and T. Sun, (1994), Hybrid
Genetic Algorithm, Simulated Annealing and Tabu
Search Methods for Vehicle Routing Problems with Time
Windows, working paper UKC/OR94/4, Institute of
Mathematics and Statistics, University of Kent,
Canterbury.
P. M. Thompson and H. Psaraftis, (1989), Cyclic Trans-
f er Al gori t hms f or Mul t i -Vehi cl e Rout i ng and
Scheduling Problems, Operations Research 41, 935
946.
12 Yves Rochat, ric Taillard
Appendix : few best known solutions.
Number
of cus-
tomers
Tours Length
Total
capacity
15
89166608417113
8614038144315
57144137
117.07 194
14
59151923798100
19391859310499
966
57.16 197
14
82461241684736
143496411175107
19123
130.76 200
14
101122201886671
651363513516434
7850
122.72 200
14
1951341632429121
169129791851583
77184
83.23 195
13
21781151454122
13375741717321
105
70.37 197
13
541301655525170
67231865619772
40
99.52 200
13
189101089012663
1813213116012830
70
97.09 200
13
51736116141191
4411919214242172
87
83.14 200
12
1468814815962182
48719410615352
73.43 200
11
261491791554139
18739110198180
71.21 194
11
761961166880150
17710912138154
49.69 199
11
2716712719031162
6913217611128
47.26 198
10
1471188319912545
174811418
66.13 195
10
1511031619120
8133157102
77.09 199
10
112183949597117
135815253
41.08 199
1 156 4.47 19
Table 7 : Best known solution to the problem instance of
Christofides et al. (1979) with 199 customers, tour
length not limited. Total length : 1291.45.
Number
of cus-
tomers
Tours Length
Total
capacity
7
120109108107106114
115
335.34 2209
11
4611818132116131
1171191306519
205.33 2029
41
912125262728
922993944544
4340341422
456789
101211148815
131690898786
8584832082
187.06 2145
27
727516253102
104101353699100
983795393896
97105575610355
546160
64.89 2066
13
7374134767764
637967803371
66
89.42 1864
15
473234484950
515258303159
232422
55.63 2140
20
78133687069110
111125112126124123
122121127128129113
8117
225.29 2167
Table 8 : Best known solution to the problem instance of
Fisher (1994) with 134 customers. Total length :
1162.96.
Number
of cus-
tomers
Tours Length
Total
capacity
11
52711628831
1063903270
116.33 138
11
988591441438
86161003797
103.56 197
11
424315412256
474727358
113.57 128
11
60834546884
517619396
112.17 121
10
94959259996
8757213
75.71 152
10
282769764053
26682480
121.70 136
9
217539236755
255412
117.84 159
9
1333020966
713581
132.76 141
9
473664491948
821889
126.64 167
9
516579783429
37750
139.58 119
Table 9 : Best known solution to the problem instance r107 of
Solomon (1987). Number of vehicles : 10, total
length : 1159.86.
Probabilistic diversication and intensication in local search for vehicle routing 13
Number
of cus-
tomers
Tours Length
Total
capacity
13
959759969993
58417458360
89
86.43 182
12
27169507628
53264013946
96.27 167
12
929891441438
8616618510037
106.08 200
12
25715434287
412274732158
107.37 129
11
80242979819
513336812
108.04 172
10
727556236739
5542554
125.95 179
10
5278284636
47194818
108.10 137
10
703020666571
35347877
119.65 134
10
318862114964
63903210
123.05 158
Table 10 : Best known solution to the problem instance r108 of
Solomon (1987). Number of vehicles : 9, total length :
980.95
Number
of cus-
tomers
Tours Length
Total
capacity
11
959298421557
8797431358
116.22 138
10
591444388684
933710091
123.29 172
10
127633817879
3506880
94.64 163
10
27693130519
6620701
106.27 145
10
27273214053
26545525
115.88 118
9
5288718846
176089
105.52 65
9
396723752241
74564
120.25 159
9
83561168599
94696
80.08 160
8
282971653534
2477
143.44 99
7
458219473649
48
124.37 147
7
621164639032
10
105.71 92
Table 11 : Best known solution to the problem instance r109 of
Solomon (1987). Number of vehicles : 11, total
length : 1235.68
Number
of cus-
tomers
Tours Length
Total
capacity
51
95221727523
673912762979
33819657151
309063644936
47484684584
5694965997
13582674564
255554248068
77170
476.06 800
49
282769528219
11623188718
836116863814
448598999287
424315574122
73405337834
356620321050
3793100911760
89
473.44 658
Table 12 : Best known solution to the problem instance r211 of
Solomon (1987). Number of vehicles : 2, total length :
949.49