THE FLEET SIZE AND MIX VEHICLE

ROUTING PROBLEM

BRUCEGOLDEN?,ARJANGASSAD&LARRYLEVY§and FILIP GHEYSENS~

University of Maryland at College Park, College Park. MD 20742. U.S.A.

Scope and Purpnse-The effective routing of vehicleshassaved and will continue to save government and
industry many millions of dollars a year. Sophisticated multi-vehicle routing procedures have increased
productivity, improved operations, aided in long-range planning, assisted in contract negotiations, and
helped to control the financial impact of adverse weather conditions on vehicle utilization. In this paper, we
consider the problem of determining an optimal fleet size and mix with respect to both acquisition and
routing costs. Several efficient computerized procedures are proposed and evaluated.

Abstract-In this paper, we address the problem of routing a fleet of vehicles from a central depot to
customers with known demand. Routes originate and terminate at the central depot and obey vehicle
capacity restrictions. Typically, researchers assume that all vehicies are identical. In this work, we relax the
homogeneous fleet assumption. The objective is to determine optimal fleet size and mix by minimi~ng a
total cost function which includes fixed cost and variable cost components. We describe several efficient
heuristic solution procedures as well as techniques for generating a lower bound and an underestimate of
the optimal solution. Finally, we present some encouraging computational results and suggestions for
further study.

INTRODUCTION
In many organizations, the management of dist~bution activities constitutes a major
decision-making problem. This problem has assumed increased urgency due to the significant
contribution of distribution cost to total costs, especially in view of the recent dramatic
escalation of fuel costs.
Most firms require vehicles for pick-up and/or delivery purposes to service a network of
supply or demand locations. The efficient utilization and routing of this fleet of vehicles lies at
the heart of almost all distribution-routing problems. In particular, a natural question facing a
distribution manager is: How many and what size vehicles are needed in order to accommodate
demand at minimal cost? This quesfion, which is di~cult to answer due to the enormous
number of potential combinations of fleet mix and routing patterns, gives rise to the problem we
formulate and address in this paper.
Variants of the vehicle routing problem generally share the following characteristics. A set
of routes must be designed for the vehicles, originating from and terminating at a central depot.
Each route visits a set of customer locations with known demands, the sum of which must not
exceed the capacity of the vehide assigned to that route. The routing costs associated with the

tBruce L. Golden is Chairman of the Department of Management Science and Statistics at the University of Maryland.
His research interests include network optimization, mathematical programming, and applied statistics, and he has
published numerous articles in these fields. He is currently an Associate Editor of Nefworks, a member of ORSA’s Long
Range Planning Committee, and Chairman of ORSA’s Transportation Science Dissertation Prize Committee. Dr. Golden
received his B.A. from the University of Pennsylvania in Mathematics. Later he earned his M.S. and Ph.D. in Operations
Research from M.I.T.
tArjang A. Assad is an Assistant Professor in the Department of Management Science .and Statistics at the
University of Maryland. His interests include operations management. mathematical programming, and network modeling
of t~nsportation/distribution systems and he has conducted applied research projects with a number of firms, and the U.S.
Departments of T~nsportation and Energy. He received his Ph.D. in Management Science from M.I.T.
5Larry J. Levy is a telecommunications analyst with Contel i~ormation Systems (formerly Network Analysis
Corporation) in Vienna, Virginia. His work includes the design and analysis of computer and communications contigura-
tions. Mr. Levy received his Masters degree in Applied Mathematics at the University of Maryland and his Bachelors
degree in Mathematics at Stevens Institute of Technology. His research at the University of Maryland included the study
of the traveling salesman and vehicle routing problems.
!Filip Gheysens is pursuing a doctoral degree in management science at the University of Gent in Belgium. In 1982,
Filip won a C.I.M. Fellowship from Belgium to spend the year as a visiting graduate student at the University of Maryland.
Filip’s major research interests include network location and vehicle routing. He received his Masters degree in
Engineering from the University of Gent.

CAOR Vol. I I. No. I-D 49

LO BRUCEGOLDEN et a/.

vehicles form one component of total distribution cost. The other essential component consists
of fleet acquisition and maintenance costs. It is possible to distinguish among three vehicle
routing cases that frequently arise in practice:
(1) The vehicle routing problem (VRP). This case involves a predetermined number of
vehicles, all of the same capacily. This number may be viewed as an upper bound on the size of
the vehicle fleet that will actually be utilized. A key assumption is that the fixed (or acquisition)
costs have already been incurred and that only variable (or routing) costs.need to be considered
explicitly. The objective is to minimize total routing cost which is a function of the total
distance traveled by the fleet of vehicles. This problem has received extensive research
attention and arises in newspaper delivery, school bus routing, municipal waste collection, fuel
oil delivery, and truck dispatching in a variety of industries (see Bodin and Golden[3]).
(2) The fleet size problem. The routing decision is frequently preceded and dominated by
the question of fleet size, that is, the number of vehicles to purchase or lease in order to satisfy
demand. In this case, fixed vehicle costs and variable routing costs both need to be considered.
The vehicles in the fleet are assumed to have identical operating characteristics (e.g. cost and
capacity). The goal is to determine an economical fleet size. In some cases, complicated
tradeoffs between purchasing and leasing need to be addressed. Ball ef a/.[l] have recently
conducted preliminary research in this direction. Their application involves a large phar-
maceutical company subject to union regulations that impose maximum route-time restrictions
on the vehicles. The solution methodology developed is of wide applicability and can be used in
the context of barges, railcars, and aircraft as well as trucks.
(3) The fleet size,and mix problem. This case is a generalization of the previous one in that
the assumption that all vehicles have the same cost and capacity is relaxed. Thus, in addition to
fleet size, the mix of vehicles in the fleet must be optimized.
The goal of this paper is to tackle the third problem listed above. Without loss of generality,
we consider a situation in which the distributor decides, for the purposes of convenience and
reliability, to lease, rather than acquire, vehicles for a certain period of time. The number of
vehicles of each type needed to handle the system demand effectively must be determined
together with a coherent routing policy for such vehicles. We assume that each vehicle type has
a fixed leasing cost and a variable routing cost that is proportional to the distance traveled.
The variable cost component encompasses the costs of fuel, maintenance, and manpower.
While the mathematical formulation of this problem is readily available (see the next section),
no efficient technique for obtaining near-optimal solutions to this problem is currently known.
Computer simulation is sometimes used to evaluate various specific options, but more analytic
techniques still need to be explored. We have begun moving along these lines and in this paper
we propose several analytic solution procedures that integrate the fleet composition and routing
sub-problems in a reasonable manner.

FORMULATION

In order to clarify and make more precise the problem that we address in this paper, we
present a mathematical programming formulation below. The problem as stated is a delivery
problem. In the case where pick-ups are made instead, an equivalent problem results.
Minimize

subject to

(j=l,...,n)

&‘:p-&ii=0 (k=l,...,T;
p,=l,...,n)
 The sleet size and mix vehicle routing problem .(I

ro = 0

ri - ri 2E(dj + UT) 2 Xi-

&=I
UT (i=O,...,n;

x$E{O, 1) forall i, j, k (j=l,...,n)

where n = number of customers, T = number of vehicle types, al, = capacity of vehicle type k
(a,-La*C ..I < aT), fk = fixed acquisition cost of vehicle type k vi <fz <. . . <fr>, di =
demand of customer j, Cir= cost of travel from customer i to customer j (we assume symmetry
and that cost is independent of the vehicle type used), ri = commodity flow variable associated
with customer i, xi = 1 if vehicle type k travels from customer i to customer j and = 0 otherwise.
It is assumed that an infinite supply of each vehicle type is available. The central depot is
denoted by 0 and the term i x& represents the number of vehicles of type k used. Therefore,
j-l
the first double-sum in the objective function gives the total fixed or acquisition cost; the next
triple-sum gives the total variable or routing cost.
The first two sets of constraints ensure that each customer is visited exactly once and that a
vehicle arriving at a customer location also leaves that location.
The next three sets of constraints guarantee that vehicle capacities are not exceeded. The
variable ri gives the total demand that a vehicle has serviced on its route after it reaches
customer i (the demand of customer i is included). Thus, (**) states that the cumulative
demand at any customer location is bounded by the capacity of the vehicle serving that
customer. The constraints r. = 0 and (*) properly define the variables ri (i = 1,. . . , n). This is
r
easily seen after observing that Z x$ equaIs either 0 or 1. Moreover, these constraints serve as
subtour-breaking constraints. ‘FL
This vehicle routing formulation is rather compact and it permits the incorporation of
unlimited choice among vehicle types while not greatly increasing the number of decision
variables.
SAVINGS HEURISTICS
The most obvious approach for solving the fleet size and mix problem is to adapt the Clarke
and Wright savings technique[S] which was originally designed for the VRP. The Clarke and
Wright procedure is extremely simple and flexible and has been. shown to solve VRP’s quite
effectively. Let the index 0 denote the central depot and let n be the number of customers. For
the VRP, where all vehicles have the same capacity, the Clarke and Wright aigorithm proceeds
as follows. Recall that a subtour is a closed loop (trip) based at the depot visiting a subset of the
n customers. The endpoints of a subtour are the two customers adjacent to the depot. Initially,
each of the n customers is serviced by a single vehicle. This results in n subtours with a total
routing cost of 2 2 coi. (At this point, each subtour has only one endpoint.) The Clarke and
i=t
Wright procedure then proceeds to combine two subtours into one at each step, thereby
eliminating one vehicle. Two distinct subtours containing customers i and j can be combined if
(a) both i and j are endpoints of their respective subtours, and
(b) the combined subtour containing both i and j has a total demand not greater than the
vehicle capacity.
The savings due to such a combination is Sij= Cot + Coj- Cij and is computed for all
1 I i, j s n (i# j). The Clarke and Wright procedure makes combinations in the order indicated
by the savings (from largest to smallest), until no further feasible combinations exist.
This procedure is deficient for the fleet size and mix problem due to the fact that it ignores
fixed vehicle costs. Since an infinite number of vehicles is assumed to be available, feasibility is
certainly no problem. As a result, the savings method will tend to combine subtours until the
capacity of the largest vehicle is reached. Unfortunately, a fleet composed almost entirely of
52 BRUCE GOLDEN et ai.

largest vehicles may not be the least expensive. This naive application of the Clarke and Wright
approach to the fleet size and mix problem will be denoted by CW.
We now seek to extend the concept of savings to include fixed vehicle costs. Every subtout
has a total cost equal to the sum of its routing costs and the cost of the vehicle servicing its
customers. When two subtours combine to form a larger subtour, the total cost of this larger
subtour can be calculated in a similar fashion. To determine which two subtours should
combine at any point in the procedure, we need to find the subtours which generate the greatest
total savings when combined. These savings must be calculated anew at every step, but all the
other attractive features of the CW approach remain. This method, which will be referred to as
the Combined Savings (CS) approach, is the basis for the remaining savings algorithms that we
consider in this paper.
In order to give a precise representation of this algorithm, some additional notation is
necessary. Let F(z) be the fixed cost of the smallest vehicle that can service a demand of size z
for a subtour. This defines a step function F on IO,m] with F(0) = 0, Now consider a subtour I
which has node i as one of its endpoints and a subtour J which has customer j as one of its
endpoints. If the total demands on these two subtours are Ziand Zj,then the combined savings of
joining customers i and j is

Sij = Sij + F( Zj) + F(Zj) - Ftzi + Zj)*

The algorithm terminates when Siiis negative for all eligible i, j pairs.
The CS approach is a logical extension of the CW approach to the fleet size and mix
problem. There is, however, an element of imprecision to the algorithm since Siirepresents only
the immediate savings gained by joining customers i and j. Consider the following example.
Suppose that subtours I, J and K each have total demand of 100and are each serviced by a
vehicle with a capacity of 100.Suppose further that the next largest vehicle has a capacity of
300and is therefore considerably more expensive than the M-capacity vehicle. It might be that
having one 300-capacityvehicle service the combination of the I, f and K subtours is cheaper
than having these routes individuallyserviced by three lOO-capacityvehicles. But this combined
route .willnot necessarily form under the CS approach. Since subtours can only combine two at

Table 1. Summary of the savings al~o~thms

Algorithm Savings Savings Formula

cw 'Oi + 'Oj _ 'ij

'ij

cs "ij sij + F(z~) + F(z~) - Ffsi ' sj'

*
FfP$ + zj) - Zi - zji
00s 9..
13 %j +

ROS “ij + 6(w) f'(P$ + Zj) - zi - zjl

%j

ROS-y "Ij "ij + (1-y) cij

where F(z) = the fixed cost of the smallest vehicle that can

service a demand of t

P(z) = the capacity of the smallest vehicle that can

service a demand of s

F'(z) = the fixed cost of the largest vehicle that has a

capacity less than or equal to s

w - P(Zi + 2.1 - P(max{zi,zjIf

3
0 ifw=O
6(w) =
11 ifwz0
 The !leet size and mix vehicle routing problem 53

a time, it is quite possible that the combined savings for any pair of customers, say i and j
(endpoints of subtours I and J), will be negative due to the large fixed cost of the 3OO-capacity
vehicle. In this situation, the CS approach ignores the fact that the 300-capacity vehicle would
have an unused capacity of 100 units which could be used to absorb additional subtours later
on.
Situations such as the one ‘above motivated us to develop variants of the CS approach,
referred to as opportunity savings algorithms. Basically, these algorithms consider a total
savings based on savings in routing costs, savings in fixed vehicle costs, and opportunity
savings. The first two savings are determined in the same manner as in the CS algorithm
Opportunity savings is a function of the unused capacity of the vehicle servicing the combined
subtours.
The first opportunity savings algorithm that we devised is labeled Optimistic Opportunity
Savings (00s). In this algorithm, opportunity savings is somewhat arbitrarily defined to be the
cost of the smallest vehicle that can service the entire unused capacity of the new vehicle. For
example, if the available vehicles have capacities of 50, 100 and 200 units, the total savings
incurred by combining two subtours with total demands of 40 and 80 units into one with a total
demand of 120 units would be Sii plus the opportunity savings (the cost of a IO&capacity
vehicle-the smallest vehicle that can service the unused capacity of 80 units). The 00s algorithm
optimistically assumes that a subtour S with a total demand equal to the unused capacity on
subtour R always exists and that at some future time subtour S will combine with subtour R
and be absorbed by the vehicle servicing subtour R. The opportunity savings represents the
fixed vehicle cost savings that would occur if this future combination materialized. The savings
in routing cost from this future combination cannot be estimated as easily and is therefore not
included in the 00s algorithm. Clearly, one expects the 00s algorithm to combine subtours
more readily than the CS algorithm.
We now provide a formal description of the 00s algorithm. Let P(z) be the capacity of the
smallest vehicle that can service a subtour with a total demand of z. Thus, in the previous
example, P(40) = 50, P(80) = 100 and P(l20) = 200. The 00s algorithm is identical to the CS
approach, except that the savings upon linking customers i and j becomes

Sri = Sii + F(Zi) + F(,?j) - F(Zi + Zj) + F(P(Zi + Zj)- -5 - zi).

Optimism regarding the 00s algorithm does not seem warranted. Indeed, the computational
results of Table 2 show that, overall, the CS approach outperforms the 00s algorithm.
Actually, both methods are severely flawed. Essentially, the CS approach under-combines and
the 00s algorithm over-combines subtours.
A remedial algorithm that comes to mind is the method of Realistic Opportunity Savings
(ROS). This method is motivated by two basic assumptions. First, the sole purpose of
opportunity savings should be to encourage the use of larger vehicles when it seems profitable
to do so. Thus, opportunity savings should not be included in the savings formula unless the
combining of two subtours requires the use of a larger vehicle. If the vehicle which would
service the combination of the subtours is larger than each of the vehicles presently used, we
say that a vehicle threshold has been crossed. Opportunity savings should be included only if a
vehicle threshold is crossed. Second, based on our computational experiments, the total demand
of a subtour tends to be close to a vehicle capacity. Therefore, instead of assuming that the
vehicle which might be absorbed in some future combination is the smallest vehicle that can
service the unused capacity, it is perhaps more appropriate to use the largest vehicle that can be
squeezed into the unused capacity. Note that this is the vehicle just smaller than the 00s
vehicle, except when the unused capacity equals a vehicle capacity, in which case both
algorithms use the same vehicle.
A full description of the ROS algorithm follows. Let F’(z) be a function analogous to F(z),
except that F’(z) represents the fixed cost of the largest vehicle whose capacity is less than or
equal to z. If z < al, F’(z) = 0. Let I and / denote two subtours with total demands of z; and Zj
and let i and j be endpoints of subtours I and J. Assume, without loss of generality, that Zi2 z,
If F(zi t Zj) = F(z,), then Sij is the total savings from linking customers i and j as in the CS
approach. If, however, F(Zi t Zj)> F(zi), then a vehicle threshold has been crossed and
54 BRUCEGOLDEN efal.

Table 2. Fleet mix costs (fixed + routing)

NO. of Nodes 12 12 20 20 20 20 30 30 30
VC/TC .49 .66 .61 .93 .62 .92 .04 .60 .65 .
Best Known 602 722 965 6446 1013 6522 7290 2349 2220 23
cw 640 796 1119 7022 1061 9343+ 7900 2430 2220 24
cs 636 034 1044 7911+ 1060 7016 7039 2709 2270 25
00s 610. 792 1024 7306 1101 9401+ 0227 2590 2412 25
ROS 636 760 1042 7369+ 1052 7016 7475 2672 2396 25
ROS-y 610. 760 1009 6937 1040 6522* 7452 2468 2266 24
SGT 620 722* 965* 6910* 1027 7391+ 7527 2360 2240 23
MGT 642 736 909 7345+ 1056 7356 7550 2305 2262 24
MGT+2-opt 634 722' 967 7304+ 1031 7330 7506 2303 2253 24
MGT+OROPT 634 722* 966 7300+ 1025 7334 7391 2370 2226 23
MGTS 632 734 901 6933 1037 7019+ 7414 2374 2245 23
(MGT+~-~P~)~ 622 722' 967 6930+ 1020 6904 74G6 2370 2226 23
(HGT+~ROPT)~ 622 722* 966 6930+ 1013* 6974 7309" 2367" 2220' 23
Lower Bound 554 672 790 6260 066 6375 7060 2204 2110 22
Underestimate 556 670 936 6300 950 6390 7165 2204 2124 22
Best-LB ($1 0.66 7.44 22.15 2.97 16.97 2.31 3.37 6.50 5.21 5.
Best-Underestimate ($1 0.27 6.49 3.10 0.91 6.63 2.07 1.06 2.05 4.52 7.
Mix--Best Known B'C' A'C AB'CE' A" ABDE= A6 B'C'D' C'D' DE' C'
Mix--Lb B'C' B'C C'E A‘ CJE2 A6 A'B'C' B'CD' DE' BD
Mix--Unuerestimate BZCz B'C C'E AS E' A' B'C*D C'D' DE' C'

Problem 11 12 13 14 15 16 17 10 19
No. of Nodes 30 30 50 50 50 50 75 75 100
VC/TC .70 .73 .74 .93 .70 .74 .57 .60 .06
Best Known 4763 4112 2430 9132 2640 2022 1703 2432 0721
cw 5420 4276 2550 12000 2005 3026 1960 3447 11319
cs 4959 4243 2650 9609 2763 2970 2043 2677 0741
00s 5203 4315 2629 10154 2949 2902 2102 2507 10233
ROS 4979 4243 2616 9609 2763 2949 2000 2612 0741
ROS-y 4953 4221 2566 9170 2763 2094 1950+ 2520 0741
SGT 4199 4173 2449 9637 2722 2055 1015 2479 9203
MGT 4965 4260 2494 9174 2742 2912 1037 2520 9411
MGT+2-opt 4016 4240 2455 9137 2670 2079 1700 2467 9239
MGTtOROPT 4013 4240 2451 9132* 2640* 2061 17%3* 2432* 0721'
MGT' 4020 4260 2471 9174 2706 2045 1037 2510 9290
(MGT+ 2-0ptJs 4763* 4136* 2442 9137 2640 2025 1700 2456 9220
(MGT+OROPT)~ 4763* 4136* 2430* 9132. 2640* 2022* 1703* 2432" 0721*
Lower Bound 4636 3070 2119 0074 2264 2504 1300 2002 0290
Underestimate 4632 3002 2304 9054 2501 2631 1703 2237 0537
Best-LB (%) 2.74 6.03 15.05 2.91 16.61 12.70 29.20 21.40 5.20
Best-Underestimate (%) 2.03 5.92 2.27 0.06 5.56 7.26 4.70 0.72 2.16
Mix--best Known BC' E' A'B'CF' A'B A‘B5 AB6C2 AB'C' AB'C'D' AIS
Mix--LB A2Cs CE'F AC'F' A'B A'B'C 8'0 ABC' AB'C'D' A"
Mix--Underestimate BCS CE’F CrD2F3 A'B A" B'C C' ABCl' A"
 The fleet size and mix vehicle routing problem

opportunity savings need to be considered. In this case, the total savings would be

S)j = jij + F(P(Zi i Zj)- Zi - Zj).

After total savings has been computed, the ROS algorithm proceeds in the same way as the CS
approach.
The computational results of Table 2 support the claim that the ROS algorithm is superior to
the 00s and CS algorithms. In thirteen of the twenty problems tested, the ROS method
performed at least as well as the other two algorithms. Its average and worst case behavior on
the twenty problems is also superior to all other savings methods tested.
Still, none of the above methods are powerful enough that single application will con-
sistently generate good solutions. It is therefore necessary to vary these algorithms so that they
produce a number of different solutions for each problem, with the least costly of these being
selected as a final solution. A method used which introduces variety into the ROS algorithm is
referred to as ROS-y.
The ROS-y algorithm uses a route shape parameter that changes the Clarke-Wright savings
expression to Sii= Coi-I-cot - y * CijmIn our tests, y was allowed to vary from 0.0 to 3.0, by
tenths, for a total of 31 trials.
The logic behind the ROS-y approach is fairly straightfo~ard. A major flaw of the savings
algorithms is that once a node is assigned to a tour, it cannot be reassigned to another at a later
point. Two linkings may seem almost equivalent at an early stage of the algorithm and yet the
choice between these can have enormous impact upon the final solution. Varying y allows the
approach to incorporate the variety which is essential to finding a good solution. The Realistic
Opportunity Savings algorithm with route shape parameter yields the best results among all
savings algorithms that we have developed and tested. Although it is outperformed by another
approach, the ROS-y provides a valid and successful approach to the problem and an
interesting extension to the concept of savings.
The savings algorithms described in this section are conveniently summarized in Table 1.

GIANT TOUR ALGORITHMS

These algorithms represent an approach to solving the fleet size and mix problem
significantly different from the savings methods. Giant tour algorithms are examples of “route
first-cluster second” heuristics. Similar approaches can be found in papers by Newton and
Thomas[i2], Bodin and Berman[2], and Stern and Dror[I6].
Giant tour (GT) algorithms are two-step procedures. First, some method is used to generate
a tour that visits all customers. This tour is then partitioned into subtours, each satisfying the
problem constraints. The subtours are composed of contiguous segments from the original tour
with both ends connected to the depot.
A key to the success of the GT approach is the partitioning procedure. As we shall see, the
problem of partitioning the tour can be transformed into a number of shortest path problems.
Therefore, the variou~giant tour algori~ms differ mainly in the way in which the initial tour is
formed. In the algorithms we tested, the initial tour is formed by solving a traveling salesman
problem (TSP). A 2-opt procedure (see [8]) was chosen to solve the TSP, although other
intelligent heuristics would have been equally acceptable.
This two-step procedure is easily motivated. Solving the TSP guarantees that adjacent
customers will be reasonably close as far as routing cost is concerned, but the formation of the
tour in no way considers the effect of clustering customer demands on the fixed vehicle costs
(i.e. the “packing” of routes). The petitioning component enables us to include the packing
factor in the analysis and compare a huge number of feasible solutions in an effective manner.

THE SINGLE PARTlTION GT ALGORITHM

This algorithm (SGT) starts with a tour that begins at the central depot, visits each customer
exactly once, and then returns to the depot. Suppose the TSP tour can be written as
o-s(l)-s(2)-***- s(n)-0. The “distance” between customers s(k) and s(m), denoted by
COST (k, m), is the total cost of having a vehicle service customers s(k), s(k + I), . . . , s(m - 1)
for k < m, in that order (let COST (k, n + 1) denote the distance between a customer and the
56 BRUCE
GOLDEN
ef al.

depot). The vehicle used is, of course, the smallest that can accommodate such a subtour. That
is,

COST(k m) = CO.s(k)
+ z; Csm(,tI) + c,on-I).o+ F($;d,,r,).
If the total demand of customers s(k), s(k + I), . . . , s(m - 1) is greater than the largest vehicle’s
capacity, the distance between these two customers in the shortest path problem is set to infinity.
Now that the “distances” have been defined, we seek the shortest path from customer s( 1) (that
is, the first customer in the tour) to the central depot. Since all arcs point forward, customers
are labeled in the order in which they appear in the tour. If the solution to the shortest path
problem requires p arcs, then the SGT solution will have p subtours, derived from the arcs in
the manner explained above. Figure 1 should help clarify these points.
Like the savings algorithms, single application of the GT approach will sometimes produce
unsatisfactory results depending on the sequence in which customers appear in the giant tour.
Variety is introduced by starting from several good traveling salesman tours, applying the
shortest path procedure to each tour, and then selecting the best solution. In particular, we
started from 25 random tours and generated 25 traveling salesman tours that were two-optimal.
For the problems tested, a sample size of 25 proved to be large enough to provide the variety
necessary to give good results and small enough to ensure the algorithm’s efficiency.

THE MULTIPLE PARTITIONGT ALGORITHM

The starting tour in this algorithm (MGT) visits each customer but not the depot. Since, in
the previous algorithm, the starting tour visits the depot only once and the final solution visits
the depot as many.times as there are vehicles, we don’t lose much by this alteration. In fact,
by excluding the depot we are able to consider a much greater number of partitions, all on the
same TSP tour.
Suppose the TSP tour can be written as s(l) - s(2) -. . . - s(n). Consider then the string
s( 1) - s(2) - ***- s(n) - s(n + 1) - ***- s(n + M) where s(n + i) is a duplication of node s(i)
for i=l,..., M with M defined by

M-l

x dXc,,< largest vehicle capacity I 2 d,(,,.

r=l r=l

The “distance” between customers s(k) and s(m) is defined as before. We can define S(i) to be
the length of the shortest path between s(i) and s(n f i) for i = 1,. . . , M. Then, the path
corresponding to min S(i) gives us the optimal partition of the initial tour. We only need to
i=I....,M

TSP Tour Fleet Size and Mix Solution when

shortest path is s(l)-s(3)-s(6)-0
CoST(3r5) = =o,s (3) + =s(3),s (4) +

cs(4),0 + F(d3 + d4)

Fig. I. SGT algorithm.

 The fleet size and mix vehicle routing problem 51

minimize over i = 1,. , . , A4 since each path musf pass through at least one node in the set
{s(I), s(2), * *. >Mf%
As with the SGT algorithm, a single application of MGT may produce unsatisfactory results.
Again variety is introduced by starting from a number of TSP tours. Because of the more
powerful partitioning, a sample size of 5 has proved to be effective. We refer to this variant of
the multiple partition GT algorithm as MGT’.

IMPROVEMENT ALGORITHMS

Although MGTS gives quite reasonable results, improvement algorithms are always worth
considerjng. Indeed, the TSP tour guarantees that adjacent customers are reasonably close, but
does not take into account the fact that each vehicle has to start and end its trip at the depot.
Moreover, although the partitioning procedure compares a large number of feasible solutions,
this number is still only a small fraction of all feasible solutions. As a result, the solution
obtained by a partitioning algorithm could easily benefit from an improvement post-processor.
The two improvement algorithms presented here are adaptations of TSP improvement
heuristics (see Golden et al.181for background). We first take the starting solution (as produced,
say, by the MGT algorithm) and construct from it an equivalent TSP tour by including as many
copies of the depot as there are vehicles in the starting solution. To illustrate, if we take the
three-subtour solution from Fig. 1 as a starting solution and refer to the copies of the depot as
cl, c2 and ~3, an equivalent TSP tour becomes

cl - s( 1) - s(2) - c2 - s(3) - s(4) - s(5) - c3 - s(6).

To allow the improvement algorithm to use more vehicles than used at present, we insert two
additional depot nodes in the tour. Using Fig. 1 once again, the TSP tour might become

Implicit in such a configuration is the fact that the “distance” between two depot copies is zero
units. The improvement algorithms that we appeal to are adaptations of TSP improvement
heuristics in that they consider fixed costs as well as routing costs and they check each time for
feasibility. As we shall see, in our computational experience, the MGT algorithm gives the best
“one run” results, Consequently, we use this solution as our starting solution.
The first improvement algorithm (MGT + 2-opt) is an adaptation of the 2-opt algorithm, All
exchanges of two arcs are tested until there is no feasible exchange that improves upon the
current solution.
The second (MGT -t OROPT) is an adaptation of the modified 3-opt procedure proposed by
Or [ 131.This procedure, which we will refer to as OROPT, considers all 2-exchanges and those
3-exchanges that would result in a string of one, two, or three currently adjacent nodes being
inserted between two other nodes. By thus limiting the number of exchanges that need to be
considered, OROPT requires signi~cantly fewer calculations than does the 3-opt procedure,
The 2-opt and OROPT post-processors reduce the gap between the MGT solution and the
best known solution for the test problems considered, as we shall see. Since MGT’ gives results
comparable to (MGT + a-opt), variety is again achieved by starting from several (in this case,
five) TSP tours. This leads to (MGT + 2-opt)5 and (MGT t OROPT)S.

PATHOLOGICAL EXAMPLES
Any time a heuristic algorithm is proposed for a problem, a natural question is how poorly
the algorithm can perform in the worst case. For certain heuristics, such as the first-fit
decreasing algorithm for bin packing (see Johnson et al. [lo]), researchers have been able to
prove that worst case behavior is not bad at all. For many other heuristics, however,
pathological examples do exist and they are important in that they indicate particular weaknes-
ses inherent in the various heuristics (for example, see Papadimitriou and Steiglitz[I4]). In this
section, we consider the performance of the ROS and GT algorithms on some perverse
problems.
sx BHUCE GOLDEN etoi.

In Example I, we pick up the ROS algorithm at a point where three subtours remain, each of
which has total demand of 100 (actually, any three numbers between 90 and 100 will work just
as well). Each of these subtours is serviced by a vehicle with a capacity of 120 and a fixed cost
of $100. The only other vehicle type has a capacity of 300 and a cost of $300. Only the six
endpoints of the three subtours are available for combining. The cost of traveling between the
depot and each of these customers is $50. The travel cost between each of the endpoints is
negligible compared to the other costs and can, thus, be considered to be $e. In addition, the
costs between ail adjacent customers in the present subtours can also be assumed to be
negligible-not too unrealistic a supposition (see Fig. 2 for clarification).
The dollar savings for any pair of endpoints is {2(50)- e}+{2(100)- 300) + 0 = - e. The
opportunity savings is zero because 300-2(100) = 100) = 100< 120. Since ail savings are negative,
the ROS procedure stops with a solution that has an approximate cost of 3(%50+ $50 + $100) =
$600. Suppose, however, that the three subtours were combined and serviced by a single vehicle of
capacity 300. The cost of this solution would be approx. $50 + $50 + $300 = $400. Thus, the ROS
solution is off by a factor of approx. 50% (on another pathological example that we have
constructed, the ROS solution is off by 100%).Note that the GT approach, irrespective of the tour
formed in the first phase of the algorithm, will easily find the optimal solution. It is probably
possible to construct examples for which the ROS procedure performs far worse than it does here.
The preceding example, however, is one that could easily occur in practice. The example iflustrates
a basic flaw of the ROS and other savings algorithms-the inability to consider the combining of
more than two subtours at a time. This fundamental weakness in the ROS algorithm is one that the
GT approaches seek to avoid.
Example 2 is somewhat more contrived. Suppose there are 4n customers in addition to the
depot, where n is any positive integer. Let 2n of these have demands of 49 and each be denoted
by S and let 2n have demands of 51 and be denoted by L. The one vehicle type available has a
capacity of 100 and a cost of $C Ail depot-customer trips cost $X Ail trips between two
customers with equal demands cost $e, where e is much fess than C and *X. Finally, ail trips
between customers with unequal demands cost $2~.
The least expensive traveling salesman tour first visits all customers with demands of 49 and
then visits ail customers with demands of 5 I. The best partitioning of this tour has 2 subtours of
the form depot - S - L -depot, (n - 1) subtours of the form depot - S - S - depot, and 2(n - 1)
subtours of the form depot - L - depot. The approximate cost of this solution is $(6n - 2)X +
$(3n - 1)C The optimal solution, of course, has 2n subtours of the form depot - S - L - depot
and costs approximately $4nX + $2nC’.So MGT is about 50% above optima&y in performance
as n gets farge. See Fig. 3 for details.
In this case, the ROS approach fares no better. Since a S- S combination has an associated
savings E dollars better than a S - L combination, the ROS approach will pair type S customers
together until only type L customers remain. Thus, the solution illustrated in Fig. 3(b) which is
about 50% above the optimal solution with an approximate cost of $6nX t $3nC will result
from applying the ROS algorithm.

Fig. 2. ROS pathological example.

 The fleet size and mix vehicle routing problem 59

Depot

Fig. 3(a). MGT solution to example 2.

Fig. 3(b). ROS solution to example 2.

2n

Fig. 3(c). Optimal solution to example 2.

This example exploits the fundamental flaw of the GT approach, the lack of consideration of
customer demands in the formation of the giant tour, and carries it to an extreme. The example
is certainly not a problem one would expect to encounter in the real world.
There are also problems for which the ROS approach performs well whereas the GT
approach performs poorly. Consider Example 3 which is Example 2 but with an additional
vehicle type. The vehicle has a capacity of 50 and a cost of W/2. The best partitioning is the
same as in the previous example. So is the optimal solution. For this example, though, the ROS
algorithm computes a savings of --E for S - S combinations and a savings of C/2 - 2r for S - L
combinations. It therefore obtains the optimal solution without difficulty.
The significance of all this is somewhat diminished by the fact that the variability introduced
into each of the algorithms and the use of post-processors are likely to reduce or even eliminate
the errors in the examples cited. Creating pathological examples for the ROS-7 algorithm or for
the GT approach followed by one of the improvement algorithms is, clearly, a far more
complex problem and one that we do not address in this paper. The value of the examples
60 BRUCE GOLDEN et 01.

presented is that they illustrate the sort of configurations that might present difficulties for the
various algorithms, they help to motivate new algorithms, and they underscore the necessity of
introducing variability and/or post-processors into the heuristic procedures.

LOWER BOUND PROCEDURE

In a later section, we will present the results of testing the various algorithms on twenty
sample problems. Since, to our knowledge, very little work has been reported on the fleet size
and mix vehicle routing problem, we have devised our own set of test problems. In the absence
of optimal solutions for these problems, there is no yardstick with which we can readily
measure the accuracy of our heuristics. With this in mind, the following important question
arises: How can the quality of the solutions generated be assessed?
One method frequently used is to obtain lower bounds on the objective value of the original
problem and compare these with the heuristic solutions. There are numerous ways of finding
lower bounds for fleet size and mix vehicle routing problems. Unfortunately, most of the
easily-derived bounds are virtually worthless in that they fail far below the optimal objective
value. The basic difficulty is that the two components that contribute to total cost, namely, the
routing cost and the fixed vehicle cost, are highly interdependent. Any method that analyzes
these two components separately should be expected to yield inferior results.
Indeed, a minimum routing cost solution would most likely be based on a totally different
vehicle mix than would a minimum fixed cost solution. The sum of these two minimum values
would provide a lower bound for the problem, but it would probably not be very informative.
The lower bound procedure we present here, trades off fixed costs against routing costs in
the following way. First, we order the customers in decreasing distance from the depot. Then
we search for a “lower bound” fleet mix using the following ideas. The total cost for each
vehicle is greater than or equal to the sum of its fixed cost and the cost of moving from the
depot to the farthest customer not yet fully served and back to the depot. Each vehicle serves
as much as possible of the demand not yet satisfied (note that splitting of a customer’s demand
between two vehicles is allowed here). An additional assumption is that the total capacity of
available vehicles larger than a certain type must be at least as large as the total demand of
customers who must be served by these vehicles. As we shall see, this “lower bound” fleet mix
can be determined by solving a shortest path problem.
The procedure we present provides a lower bound by simplifying the original problem in two
ways. As far as routing costs are concerned, it takes into account only the radial component. As
for the fleet mix, since splitting of demand is allowed, there is no guarantee that the vehicle mix
is feasible for the original problem.
Assume without loss of generality, that the customers are ordered in decreasing distance
from the depot. Thus, customer I is farthest from the depot, customer 2 is second farthest, and
so on. Let P = i d, be the total demand and define do = 0 and C(i) = 2Cojwhere
/=I

!i
I=0
d,si< 2
/=o
d, for i=O,...,P-1.

Also, let Sk be the sum of demands of customers for whom vehicle type k is the smallest one
that can service the demand (sr+, = 0). For the sake of convenience we assume that
SI, 82,* * - 7ST are all positive. Now, we are ready to construct the following network. The first
node is denoted by 0 and is the origin. The other nodes are labeled from 1 to P, which is the
destination. For each vehicle type k, there is an arc from node i to node min {P, i + ok} with a
T+I
length of fk + C(i) if and only if i 2 c sl. ln other words, the first ST nodes in the network
/=k+I
represent demands that require the largest vehicle type, the next ST-1nodes represent demands
that can be served by either the largest or the second largest vehicle type, and so on. With
regard to routing costs, the first d, nodes represent the demand of the farthest customer, the
next d2 nodes represent the demand of the second farthest, and so on. The shortest path from
the origin to the destination gives us the required lower bound. Each arc used in the shortest
 The fleet size and mix vehicle routing problem 61

path corresponds to a vehicle in the lower bound mix. The example below should help to clariiy
the procedure.
In Fig. 4, we construct the appropriate shortest path network from a fleet size and mix
vehicle routing problem where there are four customers and three vehicle types and the data
are as follows:

This gives

s,- 1, sz=5, s3=5 and P =fl.

Also,

Only arcs starting from reachable nodes are drawn in Fig. 4. In generai, the node set can be
greatly consolidated; however, we do not pursue this matter here.
As might be expected, the lower bound procedure gives excehent estimates of total fixed
costs. Due to the fact that it only takes into account the radial component of routing costs, it
may greatly underestimate the routing costs. Therefore, we have developed another procedure
which we refer to as an underestimate procedure.

UNDERESTIMATEPROCEDURE
Another way to make routing costs and fixed costs ‘~dependent” on each other is first to fix
the number of vehicles, m, to be used and then to find the minimum total cost for that many
vehicles. Since it is not known a priori how many vehicles are needed, we iterate on m and
then choose the result which gives the smallest total cost.
Once the number of vehicles is fixed, well-known methods come to mind for estimating the
minimum routing and fixed vehicle costs. A lower bound on minimum routing costs could be found
by solving the m-traveling salesman problem over all the customers, where m is the number of
vehicles. Unfortunately, since the m-TSP is NP-hard, heuristics must be used to solve the problem
for a near-optimal answer. Since capacity constraints are ignored in this operation, the answer will
remain a lower bound on routing costs in the vast majority of cases. However, to be precise, we use
the term “underestimate” to describe this answer.
To estimate the minimum fixed vehicle costs we construct the same network as for our
lower bound procedure, however, we totally ignore routing costs. Next we compute the shortest
path from node 0 to node P with exactly m arcs. A variant of the Ford-Bellman-Moore shortest
path algorithm (see Dreyfus[6]) due to Saigal[lS] easily handles this problem.
Once routing and vehicle costs are determined for all possible values of m, the smallest sum
of these two components is used as an underestimate for the optimal objective value. If this

Fig. 4. Lower bound nettiork.

62 BRUCEGOLDEN et ~11.

underestimate is a good approximation to the optimal objective value, then its vehicle mix is
expected to be quite close to the vehicle mix of the optimal solution. This is not always the case
due to the specifics of the routing cost estimation procedure. In many cases, the m-TSP
solution includes M - I single customer subtours (usually, consisting of the M - I customers
closest to the depot) and one subtour serving the rest of the customers. The depot is, thus,
adjacent to only m + I customers, rather than the 2m that we mentioned earlier. This
phenomenon leads to recommended vehicle mixes with more vehicles than are absolutely
necessary. A simple way to overcome this difficulty is to require that all m-TSP subtours have
at least two customers (this is easily handled within the context of the 2-opt procedure). This is
another reason why the numbers produced are not lower bounds, since a solution with single
customer subtours could conceivably be least expensive. However, we again emphasize that the
numbers derived are underestimates, that is, conservative estimates of the optimal objective
value along with highly accurate vehicle mixes but infeasible routings. As such, they are valuable
indicators of the quality of the solutions generated by the heuristic algorithms developed in this
paper. These underestimates figure to be inaccurate only for those problems where the optimal
solution has many single customer subtours.
In addition, the underestimate procedure can be used in other ways. Once a vehicle mix is
suggested, the fleet size and mix vehicle routing problem reduces to a vehicle routing problem
of the type described by Fisher and Jaikumar[6]. The underestimate procedure can therefore be
utilized to supply this mix. Of course, it is entirely possible that this mix is infeasible or is too
tight to produce a good solution. Therefore, it would probably be necessary to obtain several
vehicle mixes and solve a VRP for each one. The idea obviously needs more work, but it seems
worth pursuing.
It should be mentioned that one can calculate a true lower bound using a variation of the
underestimate procedure. Iterate on the number of vehicles as before and determine a lower
bound on fixed vehicle cost in the same manner as previously. To determine a lower bound on
routing cost, distinguish between depot-customer travel and customer~ustomer travel. Add
together the costs of the m cheapest depot-customer trips and multiply the sum by two. This
represents the cost of visiting the depot exactly m times. Next, find the cost of the first n - m
insertions into a minimal spanning tree over the set of customers {i/i = I, 2,. . . , n}. Adding
these two routing costs gives a valid lower bound on the total routing cost. Now, if we sum the
lower bounds on routing and vehicle costs, we have a reasonable lower bound on the optimal
objective value, but it will not be nearly as i~ormative as the underestimate in that the resulting
lower bound will tend to be quite loose.

COMPUTATIONALRESULTS
The data associated with twenty sample problems are given in the Appendix. Computational
results on these problems are summarized in Tables 2-1. The third row of Table 2 gives the
ratio of vehicle cost to total cost in the best known solution-this is a compact parameter for
classifying problems, which indicates the diversity of the sample problems studied. The fourth
line records the cost of the best known solution for each problem. The rows that folfow
correspond to the various algorithms, the lower bound, and the underestimate that we have
discussed in this paper. The next two lines display the difference (in %) between the best known
solution value and, respectively, the lower bound and the underestimate. Finally, the vehicle
mixes corresponding to the best known solution, the lower bound, and the underestimate are
presented. An asterisk indicates a best generated solution for a particular problem. A cross
indicates that the algorithm performed poorest on that problem. The percentage above the best
known solution value in that case is listed under the column headed “Worst” in Table 3. The
column labeled “Average” in Table 3 contains the average percentage above the best known
solution value for the twenty problems. In Table 4, the percentages above the best known
solution values are given for all problems and all algorithms.
Clearly, the (MGT + OROPT)’ algorithm is superior to the other algorithms in terms of
accuracy. In most cases, however, it performs only marginally better than (MGT + Z-opt)‘. In
sixteen of the twenty problems tested, the (MGT+ OROPT):' algorithm produced the best
generated solution. Only on problems I, 4 and 6 did the algorithm perform poorly. For problems
 The fleet size and mix vehicle routjng problem

Table 3. Worst and average performance of algorithms

Worst (9) Average ($1

cw 43.25 14.10

cs 22.73 8.11

00s 44.14 12.03

ROS 14.32 6.39

ROS-y 9.81 3.59

SGT 13.32 2.80

HGT 13.95 4.50

MGT+ 2-Opt 13.31 3.04

MGT+ OROPT 13.25 2.13

MGTS 7.62 2.71

(MGT+ 2-opws 7.51 1.53

(MGT + OROPT) s 7.51 1.03

Table 4. Percentage above best known solution values

-
Problem 1 2 3 4 5 6 7 8 9 10

cw 6.31 10.25 15.96 21.35 4.14 43.25 9.45 3.45 0.36 2.19

cs 5.65 15.51 8.19 22.73 4.64 7.57 7.41 18.73 2.61 8.99

00s 2.66 9.70 6.11 13.34 8.69 44.14 12.73 10.26 8.65 8.66

Flos 5.65 6.37 7.98 14.32 3.85 7.57 2.43 13.75 7.93 9.57

ROS-y 2.66 6.37 4.56 7.62 3.46 0.00 2.11 5.07 2.07 2.36

SGT 2.99 0.00 0.00 7.32 1.38 13.32 3.l4 0.8% 1.26 1.31

MGT 6.64 1.94 2.49 13.9s 4.24 12.79 3.45 1.53 1.89 5.41

MGT+Z-opt 5.32 0.00 0.21 13.31 1.78 12.51 2.85 1.45 1.49 3.76

MGT+OROPT 5.32 0.00 0.10 13.25 1.18 12.45 1.27 0.89 0.27 0.13

MGTS 4.98 1.66 1.66 7.56 2.37 7.62 1.59 1.06 1.13 0.68

NGT+2-cpt)' 3.32 0.00 0.21 7.51 0.69 7.08 k.48 0.89 0.27 0.51

(MGT+OROPT~$ 3.32 0.00 0.10 7.51 0.00 6.93 1.25 0.71 0.00 0.08

Problem 11 12 13 14 15 16 17 18 19 20
-
cw 13.96 3.99 4.59 31.41 9.28 7.23 10.38 41.74 29.19 11.78

cs 4.12 3.19 8.70 6.10 4.66 5.53 14.58 10.07 0.23 2.93

005 10.92 4.94 7.83 11.19 11.70 5.67 22.38 6.37 17.34 17.31

ROS 4.53 3.19 7.30 6.10 4.66 4.50 12.17 7.40 0.23 2.34

ROS-y 3.99 2.65 5.25 0.50 4.66 2.55 9.81 3.62 0.23 2.34

SGT 0.76 1.48 0.45 5.53 3.11 1.17 1.79 1.93 6.44 1.86

MGT 4.24 3.79 2.30 0.46 3.86 3.19 3.03 3.62 7.91 3.27

MGT+2-opt 1.11 3.31 0.70 0.95 1.44 2.02 0.28 1.44 5.94 1.84

MGT+OROPT 1.05 3.31 0.53 0.00 0.00 1.38 0.00 0.00 0.00 1.43

MG'i' 1.36 3.79 1.35 0.46 2.50 0.82 3.03 3.21 6.52 2.00

mG‘r+ I-opt) * 0.00 0.58 0.16 0.05 0.30 0.11 0.28 0.99 5.72 0.50

@iGT+ORQPT)s 0.00 0.58 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
64 BRUCEGOLDEN et al.

4 and 6 the reason is easily understood. Since vehicle costs are so important and capacity
constraints for the optimal mix are so tight, the problem reduces to a packing problem. The GT
procedures, however, begin with a TSP tour which is the result of looking at only the routing
component. The best of the savings algorithms, ROS-y, is outperformed by the GT with
post-processor approach. Compared with the SGT and MGT approaches without post-proces-
sors, ROS-y performs reasonably well.
The computational results for the lower bound are as expected. The higher the VC/TC ratio,
the better the bound. In other words, the procedure gives excellent estimates for the vehicle
costs, but due to the fact that only radial distances are taken into account, it does not estimate
routing costs as effectively. This observation is confirmed if we compare the vehicle mixes
associated with the best known solution and the lower bound. In six cases, they are exactly the
same and in most other cases they are very similar; the main reason for differences is probably
that splitting of demand is allowed in the lower bound procedure.
The accuracy of the underestimate procedure is quite satisfying. In no problem is the best
known solution as much as 10% above the underestimate. Eleven of the solutions are within 5%
of this value. Again the vehicle mixes are remarkably close to those for the best known
solutions, especially when the ratio VClTC is high. In nine cases, the mixes are identical and in
most other cases they are very similar. Differences here are due to the fact that splitting is
allowed in the underestimate procedure and that single customer subtours are forbidden. This is
particularly evident in Problem 13, where we have six single customer subtours in the best
known solution. In general, our computational work leads us to believe that the underestimates
are, in most cases, lower bounds that are capable of indicating optimal or nearly optimal vehicle
mixes.

CONCLUSIONS
The fleet size and mix vehicle routing problem is an important and practical problem
which has not yet received much attention in the literature. In this paper, several general
solution techniques have been described. One of these, the giant tour with post-processor
approach, has yielded excellent results on a wide variety of test problems. Moreover, this
procedure would have no difficulty in dealing with more realistic situations where, for example,
routing costs are dependent upon vehicle type. Indeed, both the partitioning procedures and the
exchange procedures can be easily modified to apply to this variation of the vehicle size and mix
problem. In addition, the lower bounding and underestimate procedures developed in the paper
are of value not only in assessing the quality of heuristic solutions, but also in providing a basis
for the determination of a nearly optimal fleet mix.
As for future research, it would be. interesting to see if the “generalized assignment” VRP
heuristic of Fisher and Jaikumar[7] can be modified to incorporate the more flexible fleet
constraints of the fleet size and mix vehicle routing problem. Finally, a realistic variation of the
problem that might be investigated is the problem in which there are limits on the number of
vehicles of each type.
It is hoped that this paper has succeeded in providing a first set of answers to the many
questions it has raised and that this effort will inspire further investigation of the fleet size and
mix vehicle routing problem.

REFERENCES
I. M. Ball, B. Golden, A. Assad and L. Bodin, Planning for truck fleet size in the presence of a common carrier option.
(To be published).
2. L. Bodin and L. Berman, Routing and scheduling of school buses by computer. Transpn Sci. 13, 113-129 (1979).
3. L. Bodin and B. Golden, Classification in vehicle routing and scheduling. Networks 11.97-108 (1981).
4. N. Christolides and S. Eilon, An algorithm for the vehicle dispatching problem. Oper, RPS. Quart 20, 309-318 (1969).
5. G. Clarke and J. Wright, Scheduling of vehicles from a central depot to a number of delivery points. Ops Res. 12,
568-581 (1964).
6. S. Dreyfus. An appraisal of some shortesl-path algorithms. Ops Res. 17. 395-412 (1969).
7. M. Fisher and R. Jaikumar, A generalized assignment heuristic for vehicle routing. Networks 11, 109-124 (1981).
8. B. Golden, L. Bodin, T. Doyle and W. Stewart, Approximate traveling salesman algorithms. Opers Rex 211,694-711
(1980).
9. B. Golden, T. Magnanti and H. Nguyen, Implementing vehicle routing algorithms. Nehwrks 7, 113-148 (1977).
 The fleet size and mix vehicle routing problem 65

IO. D. Johnson, A. Demers. J. Ullman, M. Garey and R. Graham, Worst-case performance bounds for simple one-
dimensional oacking algorithms. SIAM I. Computing 3, 297-325 (1974).
II. L. Levy. The fleet mix vehicle routing problem. Masters Thesis, Applied Mathematics, University of Maryland at
College Park (1980).
12. R. Newton and W. Thomas, Bus routing in a multi-school system. Compvf. Ops Res. 1,213-222 (1974).
13. 1. Or, Traveling salesman-type combinatorial problems and, their relation to the logistics of regional blood banking.
Doctoral dissertation, Northwestern University (1976).
14. C. Papadimitriou and K. Steiglitz, Some examples of difficult traveling salesman problems. Ops Res. 26,436443 (1978).
IS. R. Saigal, A constrained shortest route problem. Ops Res. 16. 205-209 (1968).
16. H. Stern and M. Dror. Routing electric meter readers. Cornput Ops Res. 6, 209-223 (1979).

APPENDIX: PROBLEM DATA

Problems #I and 2-12 node problem from [5]

Xl t2

Vehicle Capacity cost Vehicle Capacity cost

A I5 20 A 30 60
B 50 40 90
C Zt 100 ! 110 300

Problems t3 and 4-depot and first 20 nodes of 50 node problem from [4]

#3 #4

Vehicle Capacity cost Vehicle Capacity cost

A 20 20 A 60 1000
; 40
30 :; 150
80 3000
1500

D 70 120
E I20 225

Problems #5 and 6-same as X’S 3 and 4 except depot is now at 20.20

Problems t7-12-30 node problem from [5]

$7 #8 #9

Vehicle Capacity cost Vehicle Capacity cost Vehicle Capacity cost

A 40 150 A 10 IS A 40 30
s 140
100 800
500 : 150
50 200
50 100 100
: 140
D 200 1200 D 400 600 D 200 ::
E 300 2000 E 300 400

#IO #II t12

Vehicle Capacity cost Vehicle Capacity cost Vehicle Capacity cost

A 40 30 A 30 60 A 30 40
B 100 I00 B 80 200 50 80
C 140 I60 C 700 : 75 I50
D 200 240 D I500 D 120 300
I80 500
: 250 800

Problems #I3 and M-depot and first 50 nodes of 75 node problem from [4]

#I3 #I4

Vehicle Capacity cost Vehicle Capacity cost

A 20 20 A 120 1000
B 30 35 B I60 I500
C 40 50 C 300 3500
D 70 120
E I20 225
F 200 400
CAOR V,tI. I I. No I--E
Problems #IS and 16-50 node problem from [4]

#IS X16
--
Vehicle Capacity cost Vehicle Capacity cost

A 50 100 A 40 ;:
: 100
160 450
250 s 140
80 400

Problems #I7 and M-75 node problem from (41

t17 #I8

Vehicle Capacity cost Vehicle Capacity cost

A 50 25 A 20
SO :;
: 200
120 150
80 CB 100 100
D 350 320 D 150 180
E 250 400
F 400 800

Problems #I9 and 20-100 node problem from [4]

t19 #20

Vehicle Capacity Cost Vehicle Capacity cost

A 500 60 100

: ;:
300 2lOO
1200 C
; 200
140 500
300