Sie sind auf Seite 1von 10

A Cluster Insertion Heuristic for Single and Multiple Depot Vehicle Routing Problems with

Backhauling
Author(s): S. Salhi and G. Nagy
Source: The Journal of the Operational Research Society, Vol. 50, No. 10 (Oct., 1999), pp. 10341042
Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society
Stable URL: http://www.jstor.org/stable/3009928 .
Accessed: 07/04/2014 11:59
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

Palgrave Macmillan Journals and Operational Research Society are collaborating with JSTOR to digitize,
preserve and extend access to The Journal of the Operational Research Society.

http://www.jstor.org

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

journal

of the Operational

Research

Society

(1999) 50, 1034-1042

1999 Operational

Research Society Ltd. All rights reserved. 0160-5682/99

$15.00

http://www.stockton-press.co.uk/jors

A cluster insertionheuristic for single and multiple


depot vehicle routing problems with backhauling
S Salhil* and G Nagy2
1University of Birmingham; and 2University of Greenwich
We investigate an extension to the classical insertion-based heuristic for the vehicle routing problem with backhauling
(VRPB). It is based on the idea of inserting more than one backhaul at a time. This method is tested on data sets with
single and multiple depots with encouraging results at no additional computational burden. This approach can also be
useful in generating good starting solutions for the more computer-intensive meta-heuristics.
Keywords: routing; heuristic; insertion; clustering; backhauling

Introduction
The vehicle routing problem with backhauling (VRPB),
also called the backhaulingproblem, is an extension to the
classical vehicle routing problem (VRP). In this problem,
the vehicles are not only required to deliver goods to
customers,but also to pick up goods at customerlocations,
and bring them back to the depot. Customerswhose goods
are being picked up are also called backhauls.
Several researchers have made the assumption that
vehicles can only serve backhaul locations after they
have finished delivering all their load. One reason for this
is that it may be difficult to re-arrangegoods on the vehicle
and having both delivery and pickup goods may necessitate
this. Furthermore,accepting pickups before finishing all
deliveries results in a fluctuatingload. This may cause the
vehicle to be overloaded during its trip, resulting in an
infeasible vehicle tour. Although such an assumption
makes the implementation issue easier, in our view, it
may be possible to design vehicle routes which can be
practically feasible as well as more cost-effective if some
pickups are allowed to occur before all the deliveries are
completed.
If we do not make this assumption,then we can further
differentiate between two categories: simultaneous and
mixed backhaulingproblems.In the formercase, customers
may simultaneously send and receive goods. In the latter
case, customersare either delivery or pickup locations, but
not both. (The term 'mixed' denotes the fact that deliveries
andbackhaulsmay occur in any sequence on a vehicle tour.)
We note that mixed and simultaneousVRPB problems can

*Correspondence:Dr S Salhi, The Universityof Birmingham,School of


Mathematicsand Statistics, Edgbaston,BirminghamB15 2TT, UK

be modelled in the same framework.Mixed problemscan be


thought of as simultaneousones with either the pickup or
the delivery load being nil; while the customersof simultaneous problems can be divided into pickup and delivery
entities to give a mixed formulation.
The aim of this study is to develop an insertion-based
heuristicusing cluster insertionfor the VRPB, and to adapt
it to the case of multiple depots. Both the simultaneousand
the mixed cases are addressedin this study.
The literatureon the VRPB is very scant compared to
that of vehicle routing, see for example the review of
Laporte1for the VRP and that of Salhi and Sari2 for the
multi-depot routing. The literature on the VRPB can be
classified into three main categories: (i) simultaneous
pickups and deliveries; (ii) mixed pickups and deliveries;
and (iii) the case when deliveries are allowed to occur
before pickups only.
The only paper which, to our knowledge, tackles the
simultaneousproblem is that of Min.3 The authorsolves a
practical problem of transportingbooks between libraries
(one depot, two vehicles and 22 customers).The customers
are first clustered into two groups and then the two travelling salesman problems are solved. Feasibility is achieved
by penalising the overloaded arcs and resolving the infeasible TSP.
There are also very few papers which accept mixed
pickups and deliveries. The approachdescribed in Golden
et a1 is based on insertingbackhaul(pickup)customersinto
the routes formed by linehaul (delivery) customers. Their
insertion formula uses a penalty factor which takes into
account the numberof delivery customersleft on the route
afterthe insertionpoint. Casco et a15 develop a 'load-based
insertion procedure'where the insertion cost for backhaul
customerstakes into accountthe load still to be deliveredon

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

hearistic1035
cluster
insertion
S SalhiandGNagy-A

the delivery route (ratherthan the number of stops). This


latterapproachwas found to be superior.Mosheiov6investigated the TSP with pickups and deliveries (TSPPD), the
single-vehicle version of the backhaulingproblem. It has
been shown that if the TSP solution is infeasible due to
some arcs being overloaded,feasibility can be achieved by
re-insertingthe depot into the arc with the highest load.
There is a much larger body of literatureon the VRPB
with no pickups allowed before deliveries. As we do not
make this assumptionin our study, we refer the interested
readerto the papers of Min et af7 and of Toth and Vigo.8
This paper is organised as follows. The insertion-based
heuristicsfor the single and the multiple depot problemsare
given in sections 2 and 3 respectively. The last section
presents the computationalresults and a summary of our
findings.
Insertion-based heuristics

eabc

Introduction
The insertion-basedmethods of Golden et a1 and of Casco
et a15 selects one backhaulat a time and hence such types of
search can be considered shortsighted.In other words, it
may be possible that because of early insertion of some
backhauls,some otherbackhaulscannotbe insertedcheaply
into a route. One possible way to remedy this problemis to
introduce flexibility in inserting customers in clusters not
necessarily of cardinalityone. We refer to this method as
cluster insertion. In the following, we first describe the
original 1-insertion method, then we provide the steps
needed to achieve cluster insertion.
Our implementationof the 1-insertionheuristic
This method is based on the load-basedinsertionprocedure
of Golden et al.4 Insertion-basedprocedurestreat linehaul
and backhaulcustomersseparately.Linehaulcustomersare
routedusing the heuristicdescribedin Salhi and Sari.2This
basically starts with a giant tour and then improves the
routes by a series of refinementsenhancedby useful reduction tests. The backhaulcustomersare then insertedinto the
linehaul routes one by one. For each backhaul c and each
route arc ab the insertioncost is defined as:
Wabc =dac

decreases.It is importantto note that the degree of arbitrariness for choosing P is much largerthan the one for R. The
choice of the R values can be based on the idea that
backhaulsmay be carriedto the depot by a common carrier.
The cost of doing this is proportionalto the distance of the
backhaul and the depot. This proportionis denoted by R,
and it can easily be observed that the value of R should lie
between 0 and 1.
We propose a more concise implementation for the
choice of P using the following observation.Note that all
terms of (1) are in distance units, while Lb is in units of
mass. Therefore,P should not be chosen to be unit-freebut
should incorporatea term (unit of distance/unitof mass).
One possible way of achieving this is to incorporatethe
ratio Tab/MQ into P, where Tab is the length of the tour
containing arc ab (Tab includes all backhauls already
inserted), and MQ is the maximum capacity constraint.
Therefore,we replace (1) in step 3 with:

+ dcb-dab-(1?

+R)

doc+P

Lb

(1)

where d is the distance function, P and R are penalty


values, 0 is the depot and Lb is the load remaining on the
route containing ab after b.
The method depends to some extent on the parameters P
and R. In Casco et al,5 P is referred to as a penalty
multiplier, its purpose being delaying the insertion of the

backhaulsuntil the end of the route.Golden et att noted that


as P increases, the number of backhauls inserted between
two deliveries ratherthan at the end of the delivery route

= dac

dcb -dab-

a doc +?
.

* Lb(Tab/MQ)

(2)

with ocand /3being penalty multipliers and 0 denoting the


depot. The default values for o and /3 are 1.5 and 1
respectively.
The pseudo-code for this insertion procedure is given
below.
Procedure 1-INS
Step 1. Solve the VRP for the linehaul customersonly.
Step 2. Insert the backhaul with the least insertion cost
using (2).
Step 3. Delete the insertedbackhaulfrom the list of backhauls.
Step 4. Repeat steps 2 and 3 until all backhauls are
inserted.
We note that the method we use for solving the VRP has
complexity 0(14), where 1 is the number of linehauls. In
step 2, at most b remainingbackhaulsmay be insertedinto
less than (I + b) arcs, so the computationalcomplexity of
this step is 0(b(l + b)), where b is the numberof backhauls.
As steps 2 and 3 are repeated b times, the computational
complexity of this procedure is 0(max(14,b2( + b))) =
0(max(14, lb2, b3)) = ((max(14, b3). Denoting the total
number of customers by n (where n = I + b), the computational complexity of 1-INS is 0(n4), the same as that of the
underlying VRP procedure.
While the above procedure has been shown to behave
better than the stop-based insertion procedure of Golden et
al,4 see Casco et al,5 it may still be improved upon. This
greedy method may insert some backhauls in nearby arcs
making the vehicle full on those arcs. This may cause
remaining backhauls to be inserted into a distant arc, therefore increasing routing costs. An illustrative example is
provided in Figure 1. Our greedy method would insert
backhaul c into arc ab before inserting backhaul d. Capacity

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

Vol.50,No.10
Research
oftheOperational
1036 Journal
Society

e 'e

(a)

method

Greedy

(b)

A possible

improvement

Figure 1 A drawbackof the 1-insertionmethod. * depot; * linehaul;0 backhaul.

restrictionsmay cause backhauld to be insertedinto arc ef


furtheron. The total insertion cost for c and d is therefore
We can observe that
dac +dcb -dab +ded + ddf -def
considering backhaulsc and d together would insert them
into arc ef since if it could be inserted into any nearerarc
then d would have been inserted there by the greedy 1insertion procedure. The insertion cost for this case is
< dac +
- def. Therefore if dec + dcd-ded
then it is better to insert both backhauls into arc
ef. If backhauls c and d are near each other, that is, dcd is
relatively small then the above case is likely to happen. The
greedy procedure cannot consider this possibility thus we
decided to make a small improvement to it by allowing the
simultaneous insertion of two or more backhauls.

dec + dCd + ddf


dcb - dab

The 2-insertion heuristic


The insertion of pairs of backhauls is a simple extension to
the 1-insertion but this serves as a basis for cluster insertion. The idea is to consider each pair of backhauls for
simultaneous insertion into a delivery route arc. This is
obtained by modifying the insertion cost given in (2) to:
leabcd = min(dad + dab; dac + ddb)
(dOc + dod)/N2

ft

dab + dcd -O

Lb(TablMQ)

(3)

where c and d are backhauls. Then we define the 2-insertion


for all arcs ab. The 2-insertion
cost Wcd as min(6abcd)
procedure inserts either the backhaul with the least insertion cost or the pair of backhauls with the best 2-insertion
cost, whichever of the two is cheaper.

This procedure, referred to as the (2-INS), can be


presented in pseudo-code as:
Procedure 2-INS
Step 1. Solve the VRP for the linehaul customersonly.
Step 2. Find the backhaul with the least insertion cost
using (2).
Step 3. Find the pair of backhaulswith the least 2-insertion cost using (3).
Step 4. Insertthe best individualbackhaulor the best pair
of backhauls based on the lowest of the two
insertioncosts found in steps 2 and 3.
Step 5. Delete the inserted backhaul(s) from the list of
backhauls.
Step 6. Repeat steps 2 and 5 until all backhauls are
inserted.
We note that in step 3 less than b2 pairs of backhauls
may be inserted into less than (1+ b) arcs, therefore the
is
of
2-INS
complexity
computational
0(max(14, lb3, b4)) = O(max(14, b4)), or O(n4).

The cluster insertionheuristic


In this section, we extend the above procedurefor the case
of insertingwhole clusters of backhauls.These clusters are
determined using data generated during the 2-insertion
process. By calculating the 1-insertionand the 2-insertion
values for all backhauls,informationof whether the backhaul is best inserted on its own or together with another
backhaulcan be recorded.If a backhaulis best insertedon
its own then there is no need in considering it for cluster

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

insertion
cluster
hearistic1037
S SalhiandGNagy-A

insertion. On the other hand, if we have, say, three backhauls, c, d and e and we know that the 2-insertioncosts for
each pair of them are better than the corresponding 1insertion costs then it may be worthwhile that the three of
them could be inserted together.
While it is possible to use specific clusteringmethods to
do this we have decided instead to build the clusters on the
basis of the results of the 2-insertion. Taking all backhauls
to be vertices in a graph we draw an edge between two
backhaulsif their simultaneousinsertionis betterthan their
separate insertion. This graph gives rise to two possible
methods of clustering.One way is to consider all groups of
backhaulswhich form a completegraphwhereasthe otheris
to allow all sets which form a connected graph.As completeness is a strongercriterionthan connectedness it can be
observed that the first case generates smaller clusters than
the second. In this study both scenarios are analysed.
To insert these clusters, we determinethe shortestpath
from vertex a in a delivery route to the neighbouringvertex
b such that the path traverses all points in the backhaul
cluster ft. The insertion cost for cluster ?ft into arc ab is
defined to be:
(Pt,

ab) = T(f,, ab) -

dab

*EiaI + /

Lb(Tab/MQ)

(4)
where E is the sum of pairwise distances between all
backhauls in the cluster and the depot, n is the numberof
backhaulsin Yt, and T(ft, ab) the length of the path from
a to b passing throughall points within this cluster.
The pseudo-code for this method, called LC-INS or SCINS depending on whether the clusters are formed by
connectedor complete graphsof backhauls,is given below:
Procedure LC-INS/SC-INS
Step 1. Solve the VRP for the linehaul customersonly.
Step 2. Find the backhaul with the least insertion cost
using (2).
Step 3. Find the pair of backhaulswith the least 2-insertion cost using (3).
Step 4. Create, using the outcomes of steps 2 and 3, all
possible clusters of backhauls.
Step 5. Find the backhaulcluster with the least insertion
cost using (4).
Step 6. Insertthe best individualbackhaulor the best pair
or the best cluster based on the lowest insertion
cost found in steps 2, 3 and 5.
Step 7. Delete the inserted backhaul(s) from the list of
backhauls.
Step 8. Repeat steps 2 to 7 until all backhaulsare inserted.
We note that the clusteringprocedurewe use in step 4 is
of complexity 0(b3) for both LC-INS and SC-INS. In step
5, less than b clusters may be inserted into less than (I + b)

arcs. However, to calculate the insertion cost, the shortest

path through the cluster must be found, which requires a


routine of complexity O(b2), therefore the computational
complexity of step 5 is O(b(l + b)b2) = O(max(lb3,b4)).
Steps 2 to 7 may be repeated up to b times, therefore the
computational complexity of LC-INS and SC-INS is
O(max(14,ib', b5)). In terms of the total number of customers n, this is a complexity of O(n5), which is a good
result, especially when comparedto the complexity of the
underlyingVRP algorithm.
The performanceof these insertionheuristics is given in
the computationalresults section.
Adaptation to multi-depot problems
To our knowledge, this is the first time the simultaneous
and the mixed backhauling problems are studied for the
case of multiple depots. We note that even for the case of
'deliveries before pickups', there is only one multi-depot
paper.7
The multi-depot extension is based on the idea of
borderlinecustomers as used by Salhi and Sari.2Roughly
speaking,borderlinecustomersare those customerssituated
approximatelyhalf-waybetween two depots. This flexibility
is importantas it takes into account the local information
about a borderline customer. For instance, a borderline
customer may be situated far from the other customers
that are served by its nearest depot but can be closer to
the other customersthat happen to be served by its second
nearestdepot. In such circumstancessuch a customeris not
necessarily better served from a route originatingfrom its
nearest depot as other routes originating from its second
nearestdepot happento pass close to it and hence requiring
a relatively small extra insertioncost.
For the multi-depot VRPB, step 1 of the procedures 1INS, 2-INS, SC-INS and LC-INS is replaced by the
following four steps:
Step 1: Divide the set of linehaul customers into two
subsets namely borderline and non-borderline
customers.
Step 2: Assign the non-borderlinecustomersto theirnearest depots.
Step 3: For each depot, solve the resultingVRP.
Step 4. Insert the borderline customers into the vehicle
routes one at a time.
The remaining steps of the insertion procedures are
unchanged, however there is one point that needs further
attention.During the calculation of the distances between
backhauls and the depot, the appropriatedepot, that is the
one the route originates from, must be used. As the worstcase complexity of steps 1 to 4 above is 0(14) (althoughthe
average complexity will decrease as the number of depots
increases), the multi-depot modifications have the same
computational complexity as their single-depot counterparts.

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

Vol.50,No.10
oftheOperational
Research
Society
1038 Journal

Explanations of the above steps


In step 1, customerj is considered as a borderline customer
if djp/djq > p where p E [0.5, 1] (usually set to 0.90), p and
q are the nearest and the second nearest depots serving
customer j. In step 3, in each depot the non-borderline
customers are routed using the same VRP method as in the
case of single depots. In step 4, the borderline customers
are inserted one at a time into the routes originating from
the nearest and the second nearest depots serving such an
unassigned customer. Some reduction tests are also incorporated into this search to speed up the process. The details
of such reduction tests and their impact on computing time
can be found in Salhi and Sari.2

Computational

referredto as H, Q and T for the 50%, 25% and 10% of


backhauls. This is done for the single and multi-depot
problems.
For the base of simultaneous pickup and delivery, the
same coordinatesets and demandmatriceswere used again.
For each customer, a we calculated a ratio ra as
min((xa/ya), (Ya/Xa)). Then, the new demand level of
customer a is qa =ra - ta and its supply level is
Pa = (1 - ra) * tam where ta is the original demand. This

way the original demand is split up between delivery and


pickup loads. Anotherset of supply and demandlevels was
created by exchanging the demand and supply figures of
every other customer. The average results are computed
and referred to as X and Y for the two sets of data
respectively.

results and conclusions

The heuristics described in this study are written in Fortran


77 and executed on a VAX 4000-500 computer at the
University of Birmingham. They are evaluated using
empirical testing.

Data generation
We used the problems given by Christofides et at9 to
generate our single-depot data (50 to 199 customers) and
the ones given by Gillett and Johnsonl' for multi-depot data
(2 to 5 depots, 50 to 249 customers). There are 14 problems
in the first case and 11 in the second.
For the case of mixed pickups and deliveries we have
generated three VRPB problems for each VRP, declaring
every second, fourth or tenth customer on the list a backhaul and assigning it supply figure equal to the original
demand figure, in other words we let p(a) = q(a). For each
of the three classes the average results are computed and

Analysis of results

The computational results are summarised in Table 1.


Overall 70 single depot problems and 55 multi-depot
problems were tested. For simplicity we report only the
average values, in terms of total cost and computingtimes,
grouped by the percentage of backhauls for the categories
of single and multiple depot data. The detailed results are
presented in the Appendix. These results are obtained for
the case of o = 1.5 and /3= 0.8 which were selected after
some preliminarytesting. The number of vehicles is not
tabulatedas the different methods gave, on average, about
the same number of vehicles (11 and 17 for single and
multi-depotproblems respectively).
The cluster insertionmethod, though it does not produce
substantiallybetter results, does offer a positive improvement with only a slight increase in computing time. The
proposed heuristics produce better results than the 1-INS

Table 1 Averagerouting costs and computingtimes


Computertime (s)

Solution quality

10% (T)
25% (Q)
50% (H)
Single-depot
(X)
simultaneous (Y)
average
10% (T)
25% (Q)
mixed
50% (H)
Multi-depot
(X)
ae
simultaneous
(Y)
mixed

average

I-INS

2-INS

SC-INS

LC-INS

1-INS

2-INS

SC-INS

LC-INS

1015
1040
1061
1101
1105
1064
2008
2052
2136
2237
2173

1014
1038
1052
1099
1104
1061
2008
2050
2112
2237
2173

1013
1034
1045
1096
1098
1057
2008
2054
2088
2230
2172

1011
1035
1047
1097
1093
1056
2008
2050
2099
2230
2160

2.5
2.8
3.1
3.9
3.9
3.2
9.0
7.6
10.7
18.5
11.9

2.6
2.9
3.3
4.3
4.3
3.4
9.8
7.9
11.8
28.3
13.3

2.8
3.1
3.6
4.9
4.8
3.8
10.9
8.7
13.3
39.4
14.8

2.8
3.1
3.6
4.9
4.8
3.8
10.9
8.7
13.4
39.4
14.9

2121

2116

2110

2109

11.5

14.2

17.4

17.5

Rows referto the averagesoverthe differentproblemsets. (T,Q,H) referto averagesover single depotproblems(14 in each case)
and multiple depot problems(11 in each case) when 10%,25% and 50% backhaulsare considered.'X' and 'FYreferto the two
sets of simultaneouspickup-and-deliveryproblems. Similarly the average values are recorded for single and multiple depot
problems.

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

hearistic1039
insertion
cluster
andGNagy-A
S Saihi

procedure except for a few small problems where no


improvementwas found. The improved results are highlighted in bold in the Appendix. The proposed cluster
insertion heuristic appears to behave even better for the
clustereddata, that is the last 4 problems in Christofideset
al.9
It is worth noting that it is not easy to establish whether
there is a clear relationship between the goodness of the
initial (linehaul)routingsolution and the qualityof the final
solution. This may be due to the fact that the linehaul
solution does not take the location of backhaul customers
into account.
Table 1 shows that, in most cases, the computing time
increases with the percentage of backhauls as was to be
expected. However, this trend is distortedin Table 1 by a
small number of instances requiring large cpu times (see
GJ8 and GJl0 in the Appendix).

Conclusions

In this paper we presenteda numberof insertionheuristics


for the vehicle routingproblemwith pickupsand deliveries.
The idea of the 1-insertion proposed by Casco et a18 is
extendedto the case of clusterinsertion.These ideas, though
simple, provedto find good solutions for both the single and
the multiple depot problems. To our knowledge, this is the
first time the simultaneous and the mixed backhauling
problems were investigated for the case of multiple
depots. The methods presentedhere have been tested and
we found thatthey are all able to solve VRPB problemsin a
relativelyshort time.
The cluster insertion heuristic can be used to provide
good starting solutions to meta-heuristics such as tabu
search or simulated annealing based methods. The authors
are currentlyexploring such a researchavenue.

Appendix

Detailed resultsfor pickup-and-deliveryproblems


Rows correspond to the different solution methods. Columns refer to the 14 single depot test problems and the 11 multidepot test problems. CMT refers to the problems initially given by Christofides et al.9 and GJ refers to those problems given
by Gillett and Johnson.'0 For each initial problem we have five new problems, namely three mixed problems, each with
10%(T), 25%(Q), 50%(H) of backhauls, and two simultaneous problems (X and Y). Therefore, we have 125 problems in
total, 70 of them single-depot and 55 multi-depot. Entries in the table are routing cost, number of vehicles and computing
time (in seconds). The best results are highlighted in bold, except when no improvement was made to those found by the 1INS.
Results for single-depot problems
1-INS
2-INS
SC-INS
LC-INS

CMT1T
541/5/0.6
541/5/0.7
541/5/0.7
541/5/0.7

CMT1Q
557/5/0.9
557/5/1.2
557/5/1.4
557/5/1.4

CMT1H
594/6/0.9
594/6/1.2
594/6/1.4
594/6/1.4

CMT1X
601/6/1.5
601/6/2.0
601/6/3.0
601/6/3.0

CMT1Y
603/5/1.5
603/5/2.1
603/5/3.0
603/5/3.0

1-INS
2-INS
SC-INS
LC-INS

CMT2T
839/10/0.9
839/10/0.9
839/10/1.1
839/10/1.1

CMT2Q
871/11/1.0
860/11/1.1
860/11/1.1
860/11/1.1

CMT2H
900/12/1.0
900/12/1.1
873/12/1.3
900/12/1.3

CMT2X
873/10/1.2
873/10/1.3
903/11/1.7
903/11/1.7

CMT2Y
924/12/1.2
924/12/1.3
924/12/1.3
924/12/1.3

1-INS
2-INS
SC-INS
LC-INS

CMT3T
911/10/1.3
911/10/1.5
911/10/1.8
903/10/1.8

CMT3Q
918/9/1.3
918/9/1.6
918/9/1.8
918/9/1.9

CMT3H
930/10/1.5
930/10/1.8
915/9/2.0
918/9/2.7

CMT3X
923/10/1.6
923/10/1.9
923/10/2.3
923/10/2.3

CMT3Y
923/10/1.7
923/10/1.9
923/10/2.3
923/10/2.3

1-INS
2-INS
SC-INS
LC-INS

CMT4T
1111/13/3.4
1111/13/3.4
1111/13/3.4
1111/13/3.4

CMT4Q
1178/15/3.7
1164/14/3.9
1164/14/4.0
1164/14/4.0

CMT4H
1178/15/4.1
1164/14/4.3
1164/14/4.5
1164/14/4.5

CMT4X
1178/15/4.3
1178/15/4.3
1178/15/4.3
1178/15/4.3

CMT4Y
1178/15/4.3
1178/15/4.3
1178/15/4.4
1178/15/4.3

1-INS
2-INS
SC-INS
LC-INS

CMT5T
1423/18/7.8
1423/18/7.8
1423/18/7.9
1423/18/7.8

CMT5Q
1477/19/8.3
1477/19/8.4
1477/19/8.7
1477/19/8.7

CMT5H
1509/19/9.1
1509/19/9.1
1509/19/9.3
1509/19/9.3

CMT5X
1509/19/12.3
1509/19/12.5
1509/19/12.7
1509/19/12.8

CMT5Y
1477/19/12.0
1477/19/12.7
1477/19/13.0
1477/19/12.9

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

oftheOperational
Research
Society
Vol.
50,No.10
1040 Journal

1-INS
2-INS
SC-INS
LC-INS

CMT6T
571/6/0.6
571/6/0.6
571/6/0.7
571/6/0.7

CMT6Q
594/6/0.9
594/6/0.9
600/6/1.2
600/6/1.2

CMT6H
594/6/0.9
594/6/1.2
594/6/1.4
594/6/1.3

CMT6X
601/6/1.5
601/6/2.0
601/6/3.0
601/6/3.0

CMT6Y
609/6/1.5
609/6/1.6
609/6/1.7
609/6/1.7

1-INS
2-INS
SC-INS
LC-INS

CMT8T
911/10/1.3
911/10/1.5
911/10/1.8
911/10/1.8

CMT8Q
918/9/1.3
918/9/1.6
918/9/1.8
918/9/1.9

CMT8H
930/10/1.5
930/10/1.8
915/9/2.0
915/9/2.0

CMT8X
923/10/1.6
923/10/1.9
923/10/2.3
923/10/2.3

CMT8Y
923/10/1.7
923/10/1.9
927/10/2.2
927/10/2.2

1-INS
2-INS
SC-INS
LC-INS

CMT9T
1164/14/3.4
1164/14/3.4
1164/14/3.4
1164/14/3.4

CMT9Q
1178/15/3.7
1178/15/3.8
1178/15/4.0
1178/15/3.9

CMT9H
1178/15/4.1
1164/14/4.3
1164/14/4.5
1164/14/4.5

CMT9X
1215/15/4.6
1215/15/4.6
1215/15/4.6
1215/15/4.6

CMT9Y
1215/15/4.6
1215/15/4.6
1215/15/4.7
1215/15/4.6

1-INS
2-INS
SC-INS
LC-INS

CMT1OT
1423/18/7.7
1423/18/7.8
1418/18/7.8
1418/18/7.8

CMT1OQ
1477/19/8.4
1477/19/8.5
1477/19/8.7
1477/19/8.7

CMT1OH
1509/19/9.1
1509/19/9.1
1509/19/9.3
1509/19/9.3

CMT1OX
1573/21/12.1
1573/21/14.3
1573/21/17.0
1573/21/17.0

CMT1OY
1573/21/12.1
1573/21/14.4
1573/21/16.9
1527/20/17.2

1-INS
2-INS
SC-INS
LC-INS

CMT11T
1087/7/2.4
1087/7/2.4
1075/7/2.4
1075/7/2.4

CMT11Q
1101/8/2.5
1101/8/2.5
1075/7/2.5
1087/7/2.6

CMT11H
1133/8/2.6
1133/8/2.7
1120/8/2.9
1120/8/2.9

CMT11X
1546/11/3.1
1530/11/3.2
1500/11/3.4
1500/11/3.4

CMT11Y
1557/11/3.0
1557/11/3.0
1500/11/3.5
1500/11/3.6

1-INS
2-INS
SC-INS
LC-INS

CMT12T
853/10/1.4
845/10/1.7
845/10/2.1
827/10/2.3

CMT12Q
873/11/1.7
873/11/1.9
843/10/2.8
853/10/2.7

CMT12H
900/12/2.3
873/11/2.7
850/11/3.5
850/11/3.6

CMT12X
853/10/3.0
853/10/3.7
820/10/4.9
831/10/4.9

CMT12Y
888/11/3.1
888/11/3.6
888/11/4.9
873/11/4.8

1-INS
2-INS
SC-INS
LC-INS

CMT13T
1600/12/2.4
1600/12/2.4
1600/12/2.4
1600/12/2.4

CMT13Q
1613/13/2.5
1613/13/2.5
1613/13/2.5
1613/13/2.6

CMT13H
1589/12/2.6
1546/11/2.7
1546/11/2.7
1546/11/2.7

CMT13X
1613/13/2.8
1613/13/2.8
1613/13/3.0
1613/13/3.0

CMT13Y
1593/12/2.7
1593/12/2.7
1589/12/2.8
1589/12/2.8

1-INS
2-INS
SC-INS
LC-INS

CMT14T
866/11/1.4
866/11/1.7
866/11/2.2
866/11/2.2

CMT14Q
899/11/1.7
885/11/1.8
873/11/1.9
873/11/1.9

CMT14H
900/12/2.3
878/11/2.7
866/11/3.5
866/11/3.5

CMT14X
954/13/3.1
954/13/3.9
954/13/5.1
954/13/5.0

CMT14Y
961/13/3.0
954/13/4.0
920/12/5.3
920/12/5.3

Results for multi-depot problems


1-INS
2-INS
SC-INS
LC-INS

GJ1T
614/11/0.3
614/11/0.3
614/11/0.3
614/11.0.3

GJ1Q
674/14/0.2
666/13/0.2
666/13/0.2
666/13/0.2

GJ1H
655/13/0.3
636/12/0.3
619/12/0.4
619/12/0.4

GJ1X
674/14/0.2
674/14/0.2
674/14/0.2
674/14/0.2

GJ1Y
615/12/0.2
615/12/0.2
614/12/0.2
614/12/0.2

1-INS
2-INS
SC-INS
LC-INS

GJ2T
497/5/0.3
497/5/0.3
497/5/0.3
497/5/0.3

GJ2Q
550/6/0.3
550/6/0.3
550/6/0.4
550/6/0.4

GJ2H
562/6/0.7
562/6/0.9
562/6/1.1
562/6/1.1

GJ2X
569/6/1.1
569/6/1.5
569/6/2.3
569/6/2.3

GJ2Y
519/5/0.3
519/5/0.3
5 19/5/0.3
519/5/0.3

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

SSalhi
insertion
hearistic1041
andGNagy-A
cluster

1-INS
2-INS
SC-INS
LC-INS

GJ3T
662/11/0.7
662/11/0.7
662/11/0.7
662/11/0.7

GJ3Q
670/12/0.7
670/12/0.8
670/12/0.8
670/12/0.8

GJ3H
688/12/1.2
671/11/1.4
662/11/1.8
662/11/1.8

GJ3X
734/13/1.1
734/13/1.3
734/13/1.5
734/13/1.5

GJ3Y
737/14/1.1
737/14/1.2
737/14/1.4
737/14/1.4

1-INS
2-INS
SC-INS
LC-INS

GJ4T
1055/15/1.0
1055/15/1.0
1055/15/1.0
1055/15/1.0

GJ4Q
1168/18/1.1
1168/18/1.1
1168/18/1.1
1168/18/1.1

GJ4H
1084/16/1.0
1055/15/1.1
1055/15/1.1
1055/15/1.1

GJ4X
1193/18/1.6
1193/18/1.6
1193/18/1.6
1193/18/1.6

GJ4Y
1162/18/1.6
1162/18/1.7
1162/18/1.7
1162/18/1.7

1-INS
2-INS
SC-INS
LC-INS

GJ5T
794/8/3.4
794/8/3.4
794/8/4.2
794/8/4.2

GJ5Q
828/9/2.9
828/9/2.9
828/9/2.9
828/9/3.0

GJ5H
863/9/2.9
863/9/3.0
853/9/3.2
853/9/3.2

GJ5X
909/10/13.3
909/10/17.7
909/10/26.5
909/10/26.5

GJ5Y
912/8/13.3
912/8/18.6
912/8/26.5
912/8/26.5

1-INS
2-INS
SC-INS
LC-INS

GJ6T
914/16/0.6
914/16/0.6
914/16/0.6
914/16/0.6

GJ6Q
978/16/1.7
978/16/1.7
988/16/2.2
988/16.2.2

GJ6H
1034/19/1.7
1034/19/2.2
1034/19/2.6
1034/16/2.6

GJ6X
954/19/0.7
954/19/0.7
954/19/0.7
954/19/0.7

GJ2Y
1003/16/2.8
1003/16/2.9
1003/16/3.1
1003/16/3.1

1-INS
2-INS
SC-INS
LC-INS

GJ7T
992/16/1.1
992/16/1.1
992/16/1.1
992/16/1.1

GJ7Q
968/15/1.5
952/15/1.5
940/15/1.6
940/15/1.6

GJ7H
958/16/1.3
932/15/1.3
932/15/1.3
93215/1.3

GJ7X
973/16/1.5
973/16/1.5
973/16/1.5
973/16/1.5

GJ7Y
973/16/1.5
973/16/1.5
973/16/1.5
973/16/1.5

1-INS
2-INS
SC-INS
LC-INS

GJ8T
4674/25/26.1
4674/25/26.1
4674/25/31.9
4674/25/31.9

GJ8Q
4877/30/22.1
4877/30/22.1
4877/30/22.1
4877/30/22.9

GJ8H
5188/30/40.6
5188/30/43.5
5188/30/43.5
5188/30/43.5

GJ8X
5367/30/17.4
5367/30/46.4
5326/30/52.2
5326/30/52.2

GJ8Y
4816/27/23.8
4816/27/23.8
4804/27/24.7
4804/27/24.7

1-INS
2-INS
SC-INS
LC-ISN

GJ9T
4087/26/17.1
4087/26/17.1
4087/26/17.1
4087/26/17.1

GJ9Q
4087/26/19.2
4087/26/21.4
4136/26/27.8
4087/26/27.8

GJ9H
4136/28/18.0
4087/26/18.9
4087/26/19.8
4087/26/19.8

GJ9X
4426/26/74.8
4426/26/99.7
4426/26/150
4426/26/150

GJ9Y
4501/31/25.6
4501/31/27.8
4501/31/27.8
4501/31/27.8

1-INS
2-INS
SC-INS
LC-INS

GJ1OT
4002/26/31.3
4002/26/36.5
4002/26/36.5
4002/26/36.5

GJ1OQ
3931/28/17.0
3931/28/17.5
3931/28/18.4
3931/28/18.0

GJ1OH
4166/31/22.4
4166/31/24.6
4041/31/29.1
4166/31/29.1

GJ1OX
4446/31/78.3
4446/31/104
4446/31/157
4446/31/157

GJ1OY
4309/30/25.3
4309/30/30.1
4309/30/35.3
4183/29/35.9

1-INS
2-INS
SC-INS
LC-INS

GJ11T
3794/26/16.9
3794/26/20.6
3794/26/26.5
3794/26/26.6

GJ 1Q
3840/28/17.1
3840/28/17.6
3840/28/18.5
3840/28/18.0

GJ11H
4164/31/27.8
4039/29/32.7
3933/29/42.4
3933/29/43.6

GJ11X
4357/31/13.5
4357/31/36.0
4323/31/40.5
4323/31/40.5

GJ11Y
4357/31/36.0
4357/31/38.2
4357/31/40.5
4357/31/40.5

Acknowledgements-The authors are grateful to both referees for their


suggestions which improved the content as well as the presentationof
this paper.

References
1 Laporte G (1992). The vehicle routing problem: an overview of
exact and approximate algorithms. Eur J Opl Res 59: 345-358..
2 Salhi S and Sari M (1997). A multi-level composite heuristic for
the multi-depot vehicle fleet mix problem. Eur J Opl Res 103:
95-112.

3 Min H (1989). The multiple vehicle routing problem with


simultaneous delivery and pick-up points. Trans Res 23A:
377-386.
4 Golden BL, Baker EK, Alfaro JL and Schaffer JR (1985). The
vehicle routing problem with backhauling: two approaches.
Working paper MS/S 85/017, University of Maryland, College
Park.
5 Casco DO, Golden BL and Wasil EA (1988). Vehicle routine
with backhauls: models, algorithms, and case studies. In:
Golden BL and Assad AA (eds). Vehicle Routing: Methods
and Studies. Elsevier: Amsterdam, pp 127-147.

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions

Vol.50,No.10
Society
Research
oftheOperational
1042 Journal

6 MosheiovG (1994). The travellingsalesmanproblemwith pickup and delivery.Eur J Opl Res 79: 299-310.
7 Min H, CurrentJ and Schilling D (1992). The multiple depot
vehicle routingproblemwith backhauling.JBus Log 13: 259288.
8 TothP and Vigo D (1996). A heuristicalgorithmfor the vehicle
routingproblemwith backhauls.In: Bianco L and TothP (eds).
Analysis. SpringerVerlag:
AdvancedMethodsin Transportation
Berlin, pp 585-608.

9 Christofides N, Mingozzi A and Toth P (1979). The vehicle


routing problem. In: Christofides N, Mingozzi P, Toth P and
Sandi C (eds). Combinatorial Optimization. Wiley: Chichester,
pp 315-338.
10 Gillett BE and Johnson JG (1976). Multi-terminal vehicledispatch algorithm. Omega 4: 711-718.

ReceivedJune 1998;
accepted June 1999 after one revision

This content downloaded from 69.245.131.37 on Mon, 7 Apr 2014 11:59:40 AM


All use subject to JSTOR Terms and Conditions