On The Vehicle Routing Problem: School of and Statistics

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Nonlinear

Analysis,

Theory,

Methods

Pergamon PII: SO362-546X(97)00127-2

& Applicafionr, Vol. 30. No. 7. pp. 42774288, 1997 Proc. 2nd World Congress of Nonlinear Analysts

Printed

0 1997 Elsevier Science Ltd in Great Britain. All rights reserved


0362-546X/97 $17.00 +O.OO

ON THE VEHICLE
N.R Achuthan,
Curtin

ROUTING
L. Caccetta

PROBLEM

and S.P. Hill

School of Mathematics and Statistics


University of Technology GPO Box WI987 Perth. Western Australia.

Key Words and Phrases : Branch and Bound, branch and cut, combinatorial optimization, mixed-integer linear programming, vehicle routing, vehicle scheduling.
1.

INTRODUCTION

Vehicle Routing Problemsare concerned with the delivery of somecommodities from one or more depotsto a numberof customerlocationswith known demand. Such problemsarise in many physical systems dealingwith distribution. For example,delivery of commodities such as mail, food, newspapers, etc. The specificproblemwhich arises is dependent uponthe type of constraints andmanagement objective. The constraintsof the problem arise from : the vehicle capacity; distance/timerestriction; number of customers serviced by a vehicle; and other practical requirements. The management objectivesusually or fleet size. In this paper we considerthe Vehicle Routing relatesto the minimization of cost/distance Problem(VRP) having the following features:

(0
(ii) (iii) (9 (4

a singlecommodityto bedistributedfrom a singledepotto customers with knowndemand. all customerdemands are serviced by onevehicle. eachvehiclehasthe same capacityandmakes onetrip. total distance travelled by eachvehiclecannotexceeda specified limit. travelledby all vehicles. the objective isto minimizethe total distance

The resultingproblemis calledthe Capacity and DistanceRestrictedVehicle Routing Problem(CDVRP). Relaxing the distancerestriction (iv) gives rise to the so called CapacitatedVehicle Routing Problem VW. TheseNP-hard [lo] problems have attractedconsiderable attentionover the pasttwo decades resultingin many exact and heuristicalgorithms[3,5,6,8]. A good deal of the recent work has beenmotivated by the successof Branch and Cut methodsin solving large Travelling SalesmanProblems (TSP) [7,11]. Unfortunately the success of this approachin solvingthe CVRP and CDVRP hasbeenrestrictedto smaller size problems. In fact, very few CVRPs with over 100 customers have beensolvedand the situationis much worsefor the hard CDVRP. Recently [1,2,4], we have developedand implemented a number of Branch and Cut algorithmsthat consistentlysolveup to 100customer problems.Our computational superiorityhasbeenachieved through: improvedsubtourelimination constraints 0 new cutting planes eliminating non-optimalsolutions througha description of the structureof an optimalsolution. l concurrentuseof equivalentsubtourelimination constraints.
l

4277

4278

Second World

Congress

of Nonlinear

Analysts

The objective of this paper is to briefly describe this work. We present some mixed integer Linear programming models in Section 2. We restrict ourselves to formulations that we have extensively tested computationally. Algorithms for the CVRP and the CDVRP arc presented in Section 3 and computational results are discussed in Section 4.
2. MODELS

We denote the depot by 1 and the set of customer locations by &= {2,3,...,n}. Let G = (N,E) be the graph representing the vehicle network, with N = { 1,2,...,n} and E = {(ij) : ij E N, i < j}. Further, we adopt the following notation : q m L : Demandofcustomerj,j E G. : Number of delivery vehicles. : Maximum allowable distance of a vehicle route.
Gj

Q Li

: Distance between locations i and j. : Common vehicle capacity. : Length of shortest (l,i)-path in G.

We assume throughout that Q 5. max{q},

L$ i L for every j E& and the distance matrix C = (cij) is

symmetric and its elements satisfy the triangle inequality. Note that since C is Euclidean, we have L, = q,. For S G & we let 4s) = lower bound on the number of vehicles required to visit all locations of S in an optimal solution. Note that @) 2 1. We write s for the complement in & of S. For ij E N our decision variables are defined as

Xij

= 2, 0,

1,
f=

if a vehicle travels on a single trip between i and j, if i = 1 and(l,j,l)isaroute, otherwise.


c,,

Note that xi) only needs to be defined for i < j, and qi + q I Q and L, + LI + formulated the problem as follows : Minimize
CCijXij

5 L. Laporte et al. [9]

(2.1)

i<j Subject to c
Xlj

= 2m

(2.2)

Cxik i<k
CXijSlSi-C(S),

+ Cxkj j>k forall

~2,

foreachk

E0

(2.3)

S&with

3<1SIIn-2

(2.4)

i,j&

xij= m is a positive integer.

0,lor 2, for i = 1 o1 otherwise. ,,

(2.5) (2.6)

Second World Congress

of Nonlinear

Analysts

4279

In the above formulation, m can be fixed or variable. Constraints (2.2) and (2.3) are degree constraints on the depot and the customer locations, respectively. Constraints (2.4) eliminate subtours, that is they prohibit subtours that are free from the depot, or connected to the depot but violate the capacity or distance restrictions. In an actual implementation of the above formulation I (S) is only calculated when required. Laporte et al. [6] observed that (2.4) is equivalent to :
CXij

CXij

i&j& i<j

i&,jcS i<j

+ CXli i&i

2 2r(S)

(2.7)

forallS~@with3IISI <n-2. The constraints (2.7) are usually referred to as connectivity constraints. Using (2.7) for both fixed and variable m, we established [ 1] the following constraints each of which is equivalent to (2.4).

i ES,j ES i<j

Xij

CXij

i ES,jES i<j

+ C Xii iES

2 2L(S),

(2.8)

CXij+CXli-mllSi-r(~). i,jsS i-cj id

(2.9)

Note that for an optimal solution to a subproblem, in a Branch and Cut Method, (2.4) and (2.9) need not be satisfied nor violated simultaneously for a given S. In the absence of efficient methods to detect violating subtours, heuristics detecting either (2.4) or (2.9) are employed. This results in sharper lower bounds and hence less branches. Consider the formulation (2.1)-(2.6). Observe that a violation of (2.4) for S c 0 implies a violation of (2.9) for S Whenever S u { 1) forms a tour and S does not provide a violating constraint, a violation of (2.4) for P c S is possible. In such a case the following remark [2] provides a valid cutting plane similar to (2.9) for P = S\ P. This constraint is computationally useful whenever S does not provide a violating constraint. It is convenient henceforth to use the following notation in our description. For S,T c & let E[S,1[1= {(ij) : i E S, j E T, i <j} u {(ij) : i E T, j E S, i <j}. Remark 2.1 : Let PcScz&, we have
CXlj

F=S\P

and s=&\S.

ThenforafeasiblesolutionX=(x,j)oftheVRP

CXij

CXij

CXij

5Iq+IPl-C(P>.

(2.10)

jEP

i,jEP

i,jES

EF,Sl

Note that for any feasible solution (x,,) of the CARP satisfying (2.9) for a tour, the slack in (2.9) with respect to P is at least equal to the slack in (2.10) with respect to P and S. Our computational procedure takes advantage of this. In [2] we established a new cutting plane that is based on the fact that if the combined demand of any three disjoint subsets exceeds the vehicle capacity, then there is a restriction on the number of edges involving these subsets in any feasible solution. The result is :

4280

Second

World

Congress

of Nonlinear

Analysts

Theorem 2.1 : Let S, Tr,Tz ,..., Tk c &?, be such that :

(4
(b) (4 (d)

k22

and

qi > Q, c i&uTP uTq f j;

forevery I -< p z q 5 k;

Ti nTj =+,fori Sr\Ti=+,l<i<k; k and T= UTi. i=l

(2.11)

Then for any feasible solution (x3 of the CVRP we have :

3 Cxij i,j&
Remark

+ C

Xij + i Cxij p=l i,jeTP WC-U

4 31SI-2+Il-k

(2.12)

2 2 : For k = 3, we have the sharper constraint:

2 CXij+ i,j&
Remark

Xij+i

CXij

I2lSI+I~-4.

(2.13)

E[S>Jl

p=l i,jETp

2.3 : For k = 2 or L (S) t 2, the subtoureliminationconstraints(2.4) provide strongercutting planesthan (2.12) and (2.13). However, for L (S) = 1 andk 2 3 violations of (2.12) and (2.13) often occur in the absence of violationsof (2.4).

In the current literature, cutting planesusedare basedon feasibility requirements, Recently [2], we developednew cutting planesbasedon a specifiedstructureof an optimal solution. Any feasiblesolution which doesnot satisfy thesecutting planes neednot be considered whilst searching for an optimal solution. Thus the new cutting planeswill, in effect, reducetire number of feasible solutionsthat need to be considered.Our procedures ensure that an optimalsolutionis obtained. The resultsare :
2.2 : There exists an optimal solutionX = (Xij) of the CVRP (2.1) to (2.6) with variable m satisfyingthe following setof constraints : Theorem
CXij

i,j6S forallSz& with21 ISI

+C jd

Xlj

2 ISI+ 1.

(2.14)

~101 and Cqi


i&S

<Q.

2.3 : There exists an optimal solutionX = (Xij) of the CVFCP (2.1) to (2.6) with variable m satisfying(2.14) andthe following setof constraints : Theorem

Second World

Congress

of Nonlinear

Analysts

4281

CXij+C i,j& id forahSc&with2<ISI

Xii <ISI+ I

2(Cqi+6)I(Q+l+S) iEs

9 I

(2.15)

< l&l

and Cqi i&

>Q. 6=0,1accordingasQisoddoreven.

2.4 : Observe that (2.14) will eliminate somefeasible solutions and perhapssomeoptimal solutions from further consideration. However, the aboveresultensures the existence of an optimalsolution with the specifiedstructure. We notefurther that (2.14) may not be valid for a fixed m satisfyingm > c (& Remark Remark 2.5 : It follows from Theorem2.2 and 2.3 that thereexists an optimal solutionof the CVRP (2.1) to (2.6) with variablem suchthat :

if

Cqi i&

IQ

thenm=landotherwise(i.e.

Cqi is&

>Q)

m<min{n, 2(Cqi+b)/(Q+1+5) ! i&

} J

where6 is 0 or 1 accordingasQ is oddor even. This upperboundon m corresponds to an upperboundon the numberof bins requiredin the Bin Packing problem,and is achievable. Theorems2.1 - 2.3 were proved in the context of CVRP but are, in fact, valid for the CDVRP. The effectiveness of a branchand cut procedurefor solvingthe aboveproblemsis highly dependent upon being that goodvaluesfor I (S) can be generated in a ableto generate tight constraints of type (2.4). This requires computationallyeffective manner. For the CARP,

For the distance restrictions,asobserved in [9] for the Non-Euclidean caseand in [4] for the Euclideancase, 1 (S) is not availablefor S c & from the solutionof the LP subproblem.We remarkthat (2.4) is valid for S = &, in which casethe optimalobjectivefunction value of the LP subproblem canbe usedin the calculation of 1 (&I Node k of the searchtree, developed by our branchand cut algorithm,can be represented by a subsetof feasiblesolutions definedby : F(k) = {x = (xlj) : x satisfies the degree constraints, subtoureliminationconstraints andother cutting planes introducedup to nodek} At nodek, 3 problems are identified:

Problem

1 : Find

f = min[ 2cijxij:x

eF(k)j

Problem

2 : For S c & find f, = min{g(x) : x E F(k)}, where

4282

Second World

Congress

of Nonlinear

Analysts

g(X) = CCiXij
i<j i,j&

+CCljXlj
jd

+M

Ccijxi
i,jeE[S,S]

and M is a large positive integer. Problem 3 : For S c & find HkO,S)=m~{ CCijYij+CCUy?i:yEF(k,V,S)f, j& i,j&

i<j

where F(k,V,S) denotes the set of feasible solutions of the restricted CDVRP obtained from F(k) by restricting the constraints to the subset S u { 1}, using V vehicles. Let S c & be such that F(k,V,S) and F(k,V, s ) are non-empty. We proved in [4 J that if x* is an optima1 solution of the LP relaxation of Problem 2, then g(x*) I I&(V,S). At node k, c (S) may be calculated as :

(2.16)

For the Euclidean case the value of g(x*) and hence L L(S) generated as above usually yields a gcod bound. Unfortunately, the same cannot be said for the non-Euclidean case. In this case and when the generated g(x*) is not satisfactory, we utilize the cutting plane provided by the following result established in [4] : Theorem 2.4 : Let S c & and X = (xi) be a feasible solution of the distance restricted VRP. Then
C Xij I ISI -+( CCijXij + CCljXlj + C(Cij + Lj)Xij).

i,j& i<j

i,j& i<j

je

ES%

Remark 2.6 : In the non-Euclidean case, where the relaxed subproblem has an integer solution satisfying the capacity restrictions, if a tour S violates the distance restriction, then the following constraint may be applied to remove the infeasibility : 3 C Xij + Cxlj i,jcS j& i<j < 3(SI-2, (2.18)

Second World Congress 3. BRANCH

of Nonlinear

Analysts

4283

AND CUT ALGORITHMS

h this s&ion we briefly outline our work on the implementation of a Branch and Cut algorithm for solving the CARP and CDVRP utilizing the cutting planes discussed in Section 2. The basic steps are s~~~eflowdiagramofFigure3.1. In brief, the important features of our method are : . The initial relaxed problem does not include the subtour elimination constraints and m is
bounded as in (Remark 2.5).

Paessens [ 121Modified SavingsAlgorithm is usedto generate an initial upper bound. Laporteet al [9] procedures are usedto setforced variablesandpurgeineffective constraints. o The relaxedLP subproblems are solvedusingCPLEX [ 131. l We use 6 searchprocedures [2,4] for finding violations of cutting planes(2.4), (2.9), (2.10), (2.12), (2.13), (2.17) and(2.18). l Up to 5 Gomory cuts are usedat the root nodeand the Land Powell rule is usedfor selecting frirdional variablesfor branching. All castmints are retained whenbacktrackingin the search tree asthey remainvalid. For de& of these procedures we refer to [ 1,2] for the CVRP andto [4] for the CDVRP.
l

SZl

Figure 3.1 : Flow Diagram

of Algorithm

4284

Second World

Congress

of Nonlinear

Analysts

4. COMPUTATIONAL

RESULTS

We implemented our algorithm in C on a SLJN SPARC II workstation operating at 28.5 MIPS and our analysis represents the most extensive testing ever carried out. For the CVRP, our computational results are based on 1650 simulated test problems with a variable number of vehicles m and 24 standard literature problems with fixed m. The 1650 test problems are generated as follows : . Number of customers range from 15 to 100 in increments of 5. l The co-ordinates of the customer locations are integers and are randomly generated from a square of length 1000 units. The qs are then calculated to 2 decimal places. l The customer demands are generated randomly from the integers 1 to 100. l The common vehicle capacity is determined as in Laporte et al [9], that is

where a is a parameter chosen in the interval [O,l]. We use a-values of 0,0.33, 0.67 and 1. The smaller the u-value the more difficult the problem. For each choice of a andn we generate 30 problems. For CDVRP our computational results are based on 8 literature problems plus 4590 test problems of which 1965 are Euclidean. The Euclidean test problems are generated as follows : l Number of customers range from 15 to 50 in increments of 5. l The parameter a is chosen as : 0, 0.33, 0.67, 1.0. We generate 15 problems for each value of n and a (a = 0 is only considered for n 5 30). . The maximal allowable distance L for a vehicle route is initially specified as 1500; 1750; 2000; 2250; and an effective CO (104. As n increases, we increase the values of L. Each test problem is considered for each choice of L. l The co-ordinates of the customer locations are integers and are randomly generated from a square of length 1000 units. The CiIS are then calculated to 2 decimal places. To ensure feasibility, customer locations are chosen so that c,, 5 750 for every j. l The customer demands q,s are generated randomly from the integers 1 to 100. The Non-Euclidean test problems are generated with the number of customers ranging from 15 to 60, as for the Euclidean case, except for the following modifications : . The values of L used for all problems are fixed as : 1500; 1750; 2000; 2250; to (103.
l

The distances c, are generated as follows : (4 (b) For i < j, cI1is a random real number (rounded to 2 decimal places) chosen from the interval (O,lOOO]. To increase the likelihood of feasibility, we determine L,, the length of the shortest path from the depot to location i and accept the c,,s only if: (9 cl, I750; or

SecondWorld Congress of NonlinearAnalysts (ii) for cl, > 750, there exists at least one i # j such that L, + '21) + Lj 5 1500 and qi + q < Q.

4285

Detailed computational results can be found in [2] for the CVRP and in [4] for the CDVRP. Table 4.1 provides a brief comparative analysis for the 1650 simulated CVRPs of Branch and Cut procedures using (2.4) only [l], (2.4) and (2.9) [2] and (2.4), (2.9), (2.10), (2.12), (2.13) and (2.14) [4]. In our experimentation we observed that violations of (2.15) were infrequent and consequently it was excluded, For a comparative analysis of the 24 literature problems we refer to [2]. Table 4.2 presents our results for the Euclidean CDVRP simulated test problems for 1& 1 = 25 and 45 using (2.4), (2.9), (2.10), (2.12), (2.13), (2.14), (2.17) and (2.18). Table 4.3 presents similar results for the Non-Euclidean CDVRP simulated test problems for (0 1 = 30 and 60. Full results are in [4].

ALGORlTHM
us~(2.4),(2.9~~.10~2.12~ (2.14) *Root cllat a (2.13)md

using

ALGORITHM (2.4) and (2.9) Avg. Node Solved out or *Root %

ALGORITHM
using(2.4)

No.

AVE. Time

Avg. Node

solved out or

Root %

Avg. Time

Avg. Th?

99.1 99.4

251.4 189.5

365.1 462.1

0 25 29

* Percentage of the original lower bound at the root node in terms of the optimal objective function value. Table 4.1 : CVRP Comparative Analysis on Simulated Problems

4286

Second World

Congress

of Nonlinear

Analysts

CPU TIME lel a


0

(SECS)

Average Number of
EUm+
MU

No. of Problems Solved


out of 15 Avg.
NO.

Calculation of fD*r
%
succcsa Exact

ROM
l %

Avg.

Mh

Max.

Child
Node

25

l/3

1500 1750 2000 2250 co 1500 1750 2000 2250 Is00 1750 2000 2250 (D 1500 1750 2000 2250 2Go 2750 3000 02 2500 2750 3000 2zlo 2750 3000 co

2r3

l/3

45

213

96.8 96.4 96.6 96.6 96.6 90.8 93.3 98.0 98.9 99.5 87.2 91.8 95.9 97.3 99.6 89.2 89.1 93.5 93.4 100.0 99.0 98.9 98.9 98.9 99.8 98.3 98.8 99.5 97.6 98.7 98.9 99.8

450.3 1181.9 984.1 912.2 923.5 499.0 690.9 107.2 48.9 9.0 815.9 851.9 466.0 128.0 1.5 872.4 1077.8 530.5 303.3 0.7 612.0 832.3 850.6 844.2 9.8 568.8 455.9 158.2 597.5 848.3 408.3 38.7

13.9 10.7 15.8 15.5 15.9 10.8 16.3 1.0 0.8 0.8 54.1 38.3 0.5 0.5 0.5 11.0 110.3 4.8 1.3 0.3 15.1 13.1 13.0 12.7 6.4 6.6 3.8 3.3 169.2 3.1 3.2 1.6

2205.8 3390.2 3473.0 3176.9 2786.8 2 130.4 3065.4 423.6 271.4 41.8 2584.9 2764.8 1999.5 560.6 5.6 2499.7 2779.1 1615.7 1350.4 1.8 2247.1 2811.2 2804.2 2804.0 13.3 2862.9 2891.1 1135.1 1025.8 3061.1 1423.7 493.9

50.3 56.3 55.7 56.4 56.8 22.4 24.4 24.6 24.0 23.0 21.9 18.4 16.3 16.5 12.7 20.8 19.0 15.4 14.4 7.7 34.9 33.0 32.8 32.9 18.5 23.3 22.7 21.7 21.0 16.8 16.9 13.8

616.3 1627.2 1212.2 1120.3 1041.8 1330.0 1881.3 275.5 146.8 14.4 2166.9 4383.4 2089.6 617.7 2.1 2244.4 3668.9 2334.7 1568.7 0.4 677.0 781.6 830.4 821.5 6.0 1036.4 758.0 218.5 1590.0 1792.0 918.6 170.8

14 13 12 12 12 11 14 14 15 15 9 7 10 13 15 10 11 9 14 15 8 11 11 11 2 9 10 15 2 5 7 15

286.5 63.7 9.5 0.3 0 2873.2 1021.4 114.9 36.7 0 3378.1 3290.1 1177.1 343.9 0 5945.1 3868.5 1673.1 854.6 0 34.6 4.1 0.1 0 0.5 373.3 217.8 0 1051.0 830.8 452.9 0

45.3 44.6 57.9 75.0 63.4 67.0 69.7 71.1 72.0 83.4 87.9 89.2 78.8 83.7 89.6 96.9 69.7 37.8 100.0 0.0 90.5 88.8 95.4 93.2 95.9

32.8 39.9 57.0 50.0 27.0 31.4 27.8 45.1 31.8 41.6 46.3 49.4 29.3 36.3 52.9 59.6 52.7 33.3 100.0 0.0 48.0 60.7 55.1 51.9 60.0

* Percentage

lower bound at the root node in terms of the optimal objective function value. + The average of the maximum number of cutting plane constraints present at any one stage. +t The average of the number of times g(r*) is calculated (Avg. No.); Percentage of times this calculation is useful (% success); Percentage of times g(x*) = g(x) (% Exact). Table 4.2 : Euclidean CVDRPs I& 1 = 25,45.

of the original

Second

World

Congress

of Nonlinear

Analysts

4287

CPU TIME

(SECS)

l&l
a 0

Average Number of Elim MPL


58.0 52.3 51.6 51.2

T
1500 1750 2000 2250 00 1500
1750

Root* % 92.0
91.5 93.0 93.3 93.2 98.6 98.7 98.9 98.9 98.9 99.7 99.5 99.6 99.6 99.6 97.3 98.2 99.2 99.3 100.0 98.8 98.3 98.5 98.5 98.5 99.8 99.8 99.9 99.9 99.9 99.0 99.4 99.8 99.7 99.8

Avg.
3054.7 1125.8 1133.9 1038.4
1187.1

Min.
3054.7 726.1 780.2 550.7 531.3 0.6 0.6 0.6 0.6 0.6 0.4 0.4 0.4 0.4 0.4 0.6 0.4 0.3 0.3 0.3 14.8 35.6 6.3 6.3 6.2 1.9 1.9 1.9 1.9 1.9 5.3 2.6 2.2 1.8 1.8

MaX.

1 Child Nodes
3444.0 1027.3 978.0 896.4

No. of Problems Solved out of 15

l/3

2000 2250 30 213 m 1500 1750 2000 2250 al 1500 1750 2000 2250 Is0 1750 2000 2250 m 1500 1750 2000 2250 cm 1500 1750 2000 2250 m

l/3

60

213

19.6 16.8 30.0 29.3 29.2 1.4 1.4 1.1 1.1 1.1 133.3 65.2 41.7 209.4 0.6 1614.2 1759.4 1668.5 1753.1 1777.8 24.5 9.2 10.8 10.8 10.8 497.0 446.7 63.8 6.6 6.8

3054.7 1843.1 1957.5 1549.8 2246.4 71.8 61.7 114.7 102.7 102.2 4.4 4.8 3.0 3.0 3.0 710.9 488.3 276.6 2393.0 1.5 3084.3 2868.4
3181.1 3159.9 3161.0 210.1

(M)
(3,3) 5 5 5 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 10 11 12 13 13 15 15 15 15 15 10 6 9 12 15

13.8 14.5 14.5

49.5 102.8 101.1

4.0 2.1 2.1 2.1 932.8 556.5 351.7 1873.1

17.5 16.9 17.0

1862.0 1693.8 1826.9

37.4 33.2 33.2 33.2 2013.6 1857.4 538.3 23.0 34.5

7.9 8.1 8.1 8.1 6.9 6.3 5.1 4.7 4.7

6.9 10.7 10.7 10.7 1109.6 990.0 122.7 7.3 7.3

* Percentage

lower bound at the root node in terms of the optimal objective function value. + The average of the maximum number of cutting plane constraints present at any one stage. i+ First value in brackets indicates the number of problems solved optimally; the second figure denotes the number of infeasibilities detected. Other values are the number of problems solved optimally. Table 4.3 : Non-Euclidean CVDRPs I# 1 = 30,60.

of the original

4288

Second World Congress

of Nonlinear

Analysts

REFERENCES

1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.


13.

ACHUTHAN, N.R., CACCETTA, L. and HILL, S.P., A new subtour elimination constraint for the vehicle routing problem, EJ0.R. 91,573-586 (1996). ACHUTHAN, N.R., CACCETTA, L. and HILL, S.P., An improved branch and cut algorithm for the capacitated vehicle routing problem. (submitted). ACHUTHAN, N.R., CACCETTA, L., CARLTON, M. and SARWADI, Comparison of vehicle routing heuristics, (submitted). ACHUTHAN, N.R., CACCETTA, L. and HILL, S.P., The vehicle routing problem with capacity and distance restrictions (submitted). BODIN, L.D., GOLDEN, B.L., ASSAD, A. and BALL, M., Routing and scheduling of vehicles and crews : the state of the art. Comp and Oper. Rex. 10,69-2 11 (1983). CHRISTOFIDES, N., Vehicle routing, in The Travelling Salesman Problem (E. Lawler et al, Editors), 43 l-448 (1985). CROWDER H. and PADBERG, M., Solving large-scale symmetric travelhng salesman problems to optimality, Management Science 26,495-509 (1980). LAPORTE, G., The vehicle routing problem : An overview of exact and approximate algorithms, EJ0.R. 59,213-247 (1992). LAPORTE, G., NORBERT, Y. and DESROCHERS, M., Optimal routing under capacity and distance restrictions, Operations Research 33, 1050-1073 (1985). LENSTRA, J.K. and RINNOOY KAN, A.H.G., Complexity of vehicle routing and scheduling prob1enqNetwork.r 11,221-227 (1981). PADBERG, M. and RINALDI, G., A branch and cut algorithm for the resolution of large scale symmetric travelling salesman problems, Si%4 Review 33, 60-l 00 (199 1). PAESSENS, H., The savings algorithm for the vehicle routing problem, E.J. 0. R. 34, 336-344 (1988). Using the CPLEX Library and CPLEX Mixed Integer Library, CPLEX Optimization Inc., 930 Tahoe Blvd # 802-297, Incline Village, NV 89451, U.S.A., (1993).

You might also like