Omega 64 (2016) 102–114

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

Applications

Obtaining the optimal fleet mix: A case study about towing tractors
at airports$
Jia Yan Du a, Jens O. Brunner b,n, Rainer Kolisch a
a
TUM School of Management, Technische Universität München, Arcisstr. 21, 80333 Munich, Germany
b
Faculty of Business and Economics, Universität Augsburg, Universitätsstr. 16, 86159 Augsburg, Germany

art ic l e i nf o a b s t r a c t

Article history: Planes do not have a reverse gear. Hence, they need to be towed by tractors when leaving the gate.
Received 7 August 2014 Towing tractors differ with respect to investment as well as variable costs and plane type compatibility.
Accepted 15 November 2015 We propose a model which addresses the problem of a cost minimal fleet composition to support towing
Available online 17 December 2015
service providers in their strategic investment decisions. The model takes into account a maximum
Keywords: lifetime, a minimum duration of use, an overhaul option and a sell option. In a case study with a major
Airport operations management European airport (our cooperating airport) we generate a multi-period fleet investment schedule. Fur-
Turnaround processes thermore, we introduce a 4-step approach for demand aggregation based on flight schedule information.
Fleet composition problem We analyze the impact of demand variation, flight schedule disruptions and cost structure on the optimal
buy, overhaul and sell policy. The scenario analyses demonstrate the robustness of the investment
schedule with respect to these factors. Ignoring the existing fleet, a green field scenario reveals saving
potentials of more than 5% when applying this model.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction assignment significantly impacts service quality as well as

Planes do not have a reverse gear. Hence, they need assistance
to leave the gate. Furthermore, over long distances it is often more
economical and ecological to use towing tractors (see [2]). Towing
can be distinguished between (i) push-back: the fully boarded
plane is pushed backwards from the gate to the taxiway; (ii)
repositioning: the empty plane is towed from one parking position
to another; and (iii) maintenance towing: the empty plane is towed
to the hangar area for maintenance or repairs.
Towing tractors differ with respect to technical compatibility
with plane types, investment costs and variable costs. From the
perspective of a towing service provider, there are two key
questions:
1. What is the cost optimal assignment of towing jobs to towing
tractors in daily operations? The towing service provider is
responsible for carrying out all towing jobs on time. The
assignment of tractors to towing jobs is part of their daily
operations. Today most towing service providers apply manual
planning tools, often resulting in inefficient schedules. The
☆
This manuscript was processed by Associate Editor Salazar-Gonzalez.
n
Corresponding author.
E-mail addresses: jia-yan.du@tum.de (J.Y. Du),
jens.brunner@unikat.uni-augsburg.de (J.O. Brunner),
rainer.kolisch@tum.de (R. Kolisch).
operating costs. In the short-term the available fleet for the
assignment is given by the existing tractors. This operative
planning problem is covered in Du et al. [7].
2. What is the cost optimal fleet composition and respective (dis-)
investment strategy? On a strategic level the towing service
provider is responsible for deciding on the fleet size and mix
and thereby determining in each period (typically of 6 month
length) how many tractors are to be bought, overhauled or sold.
This decision impacts investment costs, operating costs, as well
as the service level. This paper addresses the strategic planning
problem. We introduce a model that generates a cost optimal
schedule for a heterogeneous set of towing tractors considering
a long-term horizon of e.g. 10 years with a period length of
6 months. We formulate the model using an extended formula-
tion, which is solved by standard column generation technique
(see [6]). We allow the fleet size and mix to change from period
to period. The model includes aspects like a selling option, a
general overhaul option, minimum duration of use, maximum
lifetime and the technical compatibility of tractor types with
plane types. For the strategic problem addressed in this paper, a
“schedule” refers to an investment schedule which determines
in which period to buy, overhaul and sell tractors of
different types.
In the literature determining the number of vehicles for a
homogeneous fleet is referred to as Fleet Sizing Problem (FSP),

http://dx.doi.org/10.1016/j.omega.2015.11.005
0305-0483/& 2015 Elsevier Ltd. All rights reserved.
 J.Y. Du et al. / Omega 64 (2016) 102–114
while a Fleet Composition Problem (FCP) addresses the problem of
deciding on the fleet size and mix simultaneously for a hetero-
geneous set of vehicles (e.g. see [8]). FSP and FCP literature can be
categorized in those considering routing (e.g., see [15]) and those
ignoring routing. Our proposed model does not include routing
aspects since we focus on a long-term strategic perspective. At a
strategic level, demand, costs and revenue uncertainties related to
fleet operations are high, thus taking into account routing aspects
on a detailed level is ineffective (see [10]).
Kirby [11] and Wyatt [18] are among the first to address the
FSP. Kirby [11] investigates the wagon fleet size of a railway sys-
tem. He concludes that the fraction of days to hire external vehi-
cles should be equal to the ratio of costs for external vehicles and
fixed costs of internal vehicles. Wyatt [18] considers a fleet of
barges. He extends the idea of Kirby [11] by adding variable costs.
Papers investigating a heterogeneous fleet are Gould [9], Lox-
ton et al. [12] and Redmer et al. [14]. In contrast to this work, their
fleet composition is determined for a single period or is constant
in all periods.
New [13], Schick and Stroup [16], Etezadi and Beasley [8],
Couillard and Martel [5], Wu et al. [17] and Burt et al. [4] examine a
planning horizon of multiple periods and allow fleet composition
to change over time. New [13] presents a linear programming
model which minimizes the operating costs of an airline fleet by
deciding on the timing of investment and disposal of planes.
Schick and Stroup [16] propose a model to address the multi-year
aircraft fleet planning problem which takes passenger demand
requirements, a minimum and maximum flight frequency and
aircraft balance equations into account. The authors discuss the
application of their model in a real life environment. Etezadi and
Beasley [8] propose a mixed integer program to determine the
optimal fleet composition of vehicles which serve several custo-
mers from a central depot. Their model minimizes the fixed and
variable costs of own and hired vehicles, while ensuring a suffi-
cient number of vehicles in each period to cover the distance to
and capacity for all customers. Couillard and Martel [5] introduce a
stochastic programming model to tackle the FCP for road carriers.
The model determines the cost minimal purchase, sell and rental
policy for a set of heterogeneous trucks, while demand is subject
to seasonal fluctuations. It considers among others the age of
vehicles in the fleet as well as tax allowances for owning a vehicle.
Wu et al. [17] apply the FCP to the specifics of the truck-rental
industry. The authors introduce a linear programming model
which decides on truck investment and divestment, demand
allocation and repositioning of empty trucks. The solution proce-
dure applies Benders decomposition and Lagrangian relaxation in
a two stage approach. It can solve instances with 3 tractor types,
60 periods and 25 locations within 12 h. The work of Burt et al. [4]
investigates the FCP for the mining industry. The proposed integer
program determines the optimal buy and sell policy for trucks and
loaders used in a mining location. A unique aspect of this model is
the consideration of compatibility between trucks and loaders. In a
case study the authors determine the optimal solution for a pro-
blem with eight trucks, 20 loaders and 13 periods within 2.5 h.
Although the discussed papers are relevant, none of these papers
capture a general overhaul option and a minimum duration of use.
To the best of our knowledge, the towing fleet investment
decision has not been investigated yet in the FCP literature. New
[13], Etezadi and Beasley [8], Couillard and Martel [5], Wu et al.
[17] and Burt et al. [4] come closest to this work. A general over-
haul option and the minimum duration of use are not included in
any of the models. Furthermore, technical compatibility is in most
cases not taken into account. Yet, these aspects are essential when
determining the optimal investment strategy in a real-world
towing setting at airports. Our work closes the gap and con-
tributes to the FCP literature by introducing a model which takes
103
into account a maximum lifetime, a minimum duration of use, an
overhaul option and a sell option. Furthermore, our main con-
tributions are a 4-step approach for demand aggregation and
demonstrating the application of the model in an extensive
case study.
The remainder of this paper is organized as follows: in the next
section we introduce the problem and explain the mathematical
formulation and the solution approach. Section 3 presents an
approach to aggregate demand using flight schedule information
and describes how the existing fleet can be incorporated in the
model. In Section 4 we demonstrate how the model can be applied
in a real-world setting. For this, we determine the schedule at our
cooperating airport. Additional scenario analyses are conducted to
investigate the impact of demand and costs deviations. We con-
clude with a summary of the main findings and outline directions
for future research in Section 5.
2. Model and solution approach
The presented model generates a cost optimal multi-period
schedule for a set of heterogeneous towing tractors. It considers a
planning horizon of j T j periods. The model determines for each
tractor type b the number of required tractors in each period t in
order to satisfy a demand DM d;t of each demand pattern d in
period t. A demand pattern is an aggregation of simultaneous
towing jobs taking into account plane type and overlapping tractor
type compatibility. The demand for one period is expressed by a
set of demand patterns (see Section 3.1). To fulfill demand, a
tractor type b has to be technically compatible with demand pat-
tern d, i.e. CP b;d ¼ 1. The model takes into account the existing
fleet. NEb denotes the number of existing tractors of type b. The
fleet size and mix is adjusted from period to period by buying new
tractors, overhauling or selling existing ones. A general overhaul is
required, if a tractor is used beyond its maximum duration of use
DU. A general overhaul extends a tractor's lifetime by additional
AD periods. A tractor can be sold on the market before reaching its
maximum lifetime DU (without general overhaul) or DU þ AD
(with general overhaul). However, a tractor has a minimum
duration of use of MU periods, before it can be sold. MU does not
reflect a technical feature of a tractor, but rather is set by the
management. Buying, using, overhauling and selling a tractor in
period t is associated with investment costs ICt, variable costs VCt,
overhaul costs OCt, and sales revenue SRt, respectively. In the case
of a planning horizon of up to 10 years, costs and revenue are
time-dependent by, amongst others, taking into account discount
rates. Therefore, all cost and revenue parameters are time-indexed.
Both cost changes and the discount rate are incorporated in the
cost data, and do not explicitly appear in the model. The para-
meters DU, AD, MU, ICt, VCt, OCt and SRt are tractor type specific
and assume different values for each tractor type b.
We formulate the model using an extended formulation and
propose a Column Generation Heuristic (CGH) as solution proce-
dure. We do not present the compact mixed-integer linear pro-
gramming model for the problem. In simple terms, column gen-
eration decomposes the problem into a Master Problem (MP) and
a Subproblem (SP), which generates feasible columns (i.e. sche-
dules). A feasible column a A AðbÞ represents one schedule for a
specific tractor type b. The schedule defines in which periods the
tractor is in use and accordingly when to buy, overhaul and sell the
tractor. AðbÞ is the set of all schedules associated with tractor type
b. Each schedule a A AðbÞ is associated with total schedule costs of
TC b;a that are a function of IC t ; VC t ; OC t ; and SRt.
MP determines the fleet size and mix by selecting the sche-
dules to follow. It minimizes the costs while ensuring demand
satisfaction. Only a subset of all feasible schedules are considered
104 J.Y. Du et al. / Omega 64 (2016) 102–114

(we term the problem as Restricted Master Problem (RMP)) and Constraints (1c) take the existing fleet into account. We create
new columns are added iteratively. The procedure starts with a one additional SP for each existing tractor and adapt decision
small subset of columns and solves the Linear Programming (LP) variables and parameter settings (see Section 3.3), i.e. j Bj is the
relaxation of RMP. By inserting the dual value information of RMP number of general tractor types plus the number of existing
constraints into the objective function of SP, the most promising tractors. Constraints (1c) enforce NEb number of schedules of
column is generated. A column with negative reduced costs is tractor type b to be selected. NEb is greater than or equal to 1 for SP
added to RMP and RMP is reoptimized. The procedure terminates representing existing tractor types and NEb is 0 for all other SP.
with a valid lower bound for the problem (i.e. no absent column Variable definitions are given in (1d) and (1e).
with negative reduced costs exists). In a second step, all generated The dual solution of RMP is obtained by relaxing the integrality
columns which have been inserted in RMP thus far are taken and conditions and solving RMP with a subset of columns. Let δd;t Z 0
RMP is solved as Integer Program (IP) to generate a feasible solu- denote the dual values of constraints (1b), then δt Z 0 is defined as
tion, i.e. an upper bound. X
δt ¼ δd;t 8 t A T : ð2Þ
Master problem: the notation and mathematical formulation of dAD
RMP are as follows:
Sets: And let δ~ b Z 0 denote the dual values of constraints (1c). Then the
reduced costs of column a associated with tractor type b is
!
B set of tractor types (index b) X
c b;a ¼ TC b;a  ~
δt  X b;a;t þ δ b ð3Þ
AðbÞ set of schedules associated with tractor type b (index a)
tAT
D set of demand patterns (index d)
T set of periods (index t) with TC b;a representing the total costs of schedule a for tractor
type b defined as
X X
Parameters: TC b;a ¼ VC t  X b;a;t þ IC t  Y buy
b;a;t
þ Ocost
b;a;t
t AT t AT
XX
TC b;a costs of schedule a associated with tractor type b  SRt~  U b;a;t;t~ : ð4Þ
CW costs associated with auxiliary variable wd;t t A T t~ A T

CP b;d 1, if tractor type b is compatible with at least one plane X b;a;t ; Y buy ; Ocost
b;a;t b;a;t and U b;a;t;t~ represent the variable solution values
type associated with demand pattern d, 0 otherwise in SP. A detailed description of the cost components is given in the
X b;a;t 1, if schedule a for tractor type b covers period t, explanation of SP's objective function (5a).
0 otherwise Subproblem (b): One SP is created for each tractor type b and
DM d;t demand of demand pattern d in period t each existing tractor. Each SP generates schedules for tractor type
NEb number of existing tractors of tractor type b b. The following additional notation is used to formulate SP(b):

Variables Parameters

λb;a number of type b tractors assigned to schedule a VCt variable costs in period t
wd;t uncovered demand associated with demand pattern d ICt investment costs in period t
and period t SRt sales revenue per remaining use period, if tractor is sold
X X XX in period t
Minimize TC b;a  λb;a þ CW  wd;t ð1aÞ OCt general overhaul costs in period t
b A B a A AðbÞ d A Dt A T DU maximum duration of use without general overhaul
X X AD maximum additional duration of use after a general
s:t: X b;a;t  λb;a þ wd;t Z DMd;t 8 d A D; t A T ð1bÞ overhaul
b A B:
a A AðbÞ
CP b;d ¼ 1 MU minimum duration of use before tractor can be sold
X
λb;a Z NEb 8 bA B ð1cÞ Variables
a A AðbÞ

λb;a Z 0 and integer 8 bA B; a A AðbÞ ð1dÞ xt 1, if tractor is used in period t, 0 otherwise (we set
x0 ¼ 0; x j T j ¼ 0)
wd;t Z 0 and integer 8 d A D; t A T ð1eÞ ybuy
t 1, if tractor is bought in period t, 0 otherwise
yov 1, if tractor is overhauled, 0 otherwise
The objective function (1a) of the master problem minimizes ysell 1, if tractor is sold in period t, 0 otherwise
t
the total costs of the selected schedules. The first sum adds the ut;t~ remaining lifetime if tractor is bought in period t and
costs of all schedules TC b;a which are selected. Variable λb;a sold in period t~
denotes the number of tractors associated with type b which are ocost costs of general overhaul if tractor is bought in period t
t
bought, overhauled and sold according to schedule a. The (aux-
X X X
iliary) variables wd;t in the second term count the uncovered Minimize VC t  xt þ IC t  ybuy þ ocost
t t
demand for pattern d in period t. We use the variables to initialize tAT tAT t AT
!
RMP. They guarantee feasibility in the course of the column gen- XX X
eration procedure. The use of one not available tractor for one  SRt~  ut;t~  δt  xt þ δ~ b ð5aÞ
t A T t~ A T tAT
period is penalized with costs CW. CW is set to a value higher than
the most expensive column costs, i.e. CW 4 TC a;b 8 b A B; a A AðbÞ.
Demand constraints (1b) ensure that the demand in each per- s:t: xt  1 þ ybuy
t r1 8t AT ð5bÞ
iod is fulfilled for each demand pattern, i.e. there must be suffi-
cient numbers of compatible tractors to satisfy demand. xt  xt  1  ybuy
t r0 8t AT ð5cÞ
 J.Y. Du et al. / Omega 64 (2016) 102–114 105

Constraints (5j), (5k) and (5l) track the remaining lifetime of a

 xt  1 þ ysell
t r0 8t AT ð5dÞ
tractor. The remaining lifetime depends on the period t in which
the tractor is bought and the period t~ in which the tractor is sold.
xt þ ysell
t r1 8t AT ð5eÞ
ut;t~ is enforced to 0 for all periods a tractors is not bought (5j) or
sold (5k). In the periods in which the tractor is bought or sold ut;t~
 xt þ xt  1  ysell
t r0 8t AT ð5f Þ
is limited to an upper bound of DU þ AD. For the period t in which
X the tractor is bought and the period t~ in which the tractor is sold,
yov r ybuy
t ð5gÞ ut;t~ is set to the remaining lifetime: AD þ DU  t~ þt (with general
t AT
overhaul) or DU  t~ þ t (without general overhaul) (5l).
X X Constraint (5m) ensures that the tractor is bought not more
t~  ysell
t~
 t  ybuy
t Z ðDU þ 1Þ  yov ð5hÞ
t~ A T t AT than once. Constraints (5n) enforce the use of a tractor once it was
bought. The variable x is set to 1 for the period a tractor is bought
ocost Z OC t þ DU  ðybuy þ yov  1Þ 8 t A f1; …; j T j  DU g ð5iÞ and the following periods until the minimum usage duration MU
t t
or the end of the planning horizon is reached. For example, if a
ut;t~ r ðDU þ ADÞ  ybuy
t 8 t; t~ A T ð5jÞ tractor is bought in period 1 and the minimum duration of use MU
is 4 periods, x takes the value of 1 for periods 1, 2, 3 and 4. If the
ut;t~ r ðDU þ ADÞ  ysell 8 t; t~ A T ð5kÞ tractor is bought in period 8 and j T j ¼ 11, then x is set to 1 for
t~
periods 8, 9 and 10. Note, the last period is added only to sell
remaining tractors, i.e. DM d;j T j ¼ 0, 8 d A D and consequently
ut;t~ r AD  yov þ DU  ðt~  ysell
t~
 t  ybuy
t Þ
x j T j ¼ 0.
þ t~  ð1  ybuy
t Þ 8 t; t~ A T ð5lÞ Constraints (5o) and (5p) ensure that the maximum duration of
X use is not exceeded. In constraints (5o) the maximum duration of
ybuy
t r1 ð5mÞ use equals the initial lifetime DU of a tractor plus the additional
tAT
duration of use AD in case of a general overhaul. The constraints
  set the variable x in period t þDU þAD to 0 if a general overhaul
xt~ Z ybuy 8 t A f1; …; j T j  1g; t~ A t; …; minft þ MU  1; j T j  1g
t takes place in period tþDU. For example, if DU¼AD ¼ 4 and the
ð5nÞ tractor is bought in period 1 (ybuy ¼ 1), then constraints (5o)
1
enforce x9 to take the value 0. In the second case no general
1 xt þ DU þ AD Zyov þ ybuy
t 1 8 t A f1; …; j T j 1  AD  DU g overhaul takes place (5p), therefore the maximum duration of use
ð5oÞ corresponds to the initial lifetime DU of the tractors. Variable
definitions are given in (5q)–(5r).
1 xt þ DU Z ybuy
t  yov 8 t A f1; …; j T j  1  DU g ð5pÞ Table 1 gives an example of a schedule a for tractor type b. The
table shows the values of the decision variables and the total
xt ; ybuy
t ; yov ; ysell
t A f0; 1g 8t AT ð5qÞ schedule costs TC b;a in the periods 1–10. The maximum lifetime
DU is 4 periods and the additional lifetime after a general overhaul
ocost
t ; ut;t~ Z 0 8 t; t~ A T ð5rÞ is 4 periods. According to the schedule, the tractor is used in
The objective function (5a) minimizes the reduced costs of a periods 2–7. Variable ybuy is set to 1 when there is a shift from 0 to
new schedule. It consists of five terms that are discussed below. 1 in the x-variables, here ybuy 2 ¼ 1. Accordingly, ysell is set to 1 if
The first term takes into account the variable costs of all periods in there is a shift from 1 to 0 in the x-variables, i.e. ysell 8 ¼ 1. The
which a tractor is used. The second term considers the investment maximum lifetime is reached and requires a general overhaul, i.e.
costs ICt of a tractor bought in period t while the third term cal- yov ¼ 1. The remaining lifetime of the tractor, which is bought in
culates the overhaul costs ocost t due to a general overhaul in period period 2 and sold in period 8, is 2 periods, i.e. u2;8 ¼ 2, ut;t~ ¼ 0 in all
t. Selling a tractor results in additional earnings that are handled in other periods t; t~ .
the fourth term. Those earnings depend on the remaining lifetime The solution of SP(b) is a new schedule (column) a given as
ut;t~ and the sales revenue per remaining use period SRt~ when the 2 3
TC b;a
tractor is sold in period t~ . We assume that all tractors are sold at 6! 7
6X 7
the end of the planning horizon. Whereas the first four terms 4 b;a 5
consider the real value (i.e. cost including earnings for selling) of a 1b
new schedule, the fifth term handles the dual information (i.e.
where TC b;a is the total costs of schedule a associated with tractor
dual values δt and δ~ b obtained from solving the relaxed RMP). !
Constraints (5b), (5c) and (5n) detect a shift of the x variable type b. X b;a is a vector with j T j elements indicating in which
from 0 to 1 and thereby determine the buying period. These periods the tractor is used, or in other words containing the values
constraints are the linearization of ybuy ¼ xt  ð1  xt  1 Þ. Accord- of the decision variable xt of SP(b), i.e. X b;a;t ¼ xt 8 t A T . 1b is a unit
t
ingly, constraints (5d), (5e) and (5f) determine the selling period vector with length j Bj and value 1 at position b and 0 else. To find
and linearize ysell
t ¼ xt  1  ð1  xt Þ.
Constraints (5g), (5i) and (5h) set the rules for a general over- Table 1
Example schedule a for tractor type b.
haul: a general overhaul can take place if the tractor has been
bought previously (5g) and before the tractor is being sold (5h). t 1 2 3 4 5 6 7 8 9 10
We force yov ¼ 0, if the tractor is sold before the maximum dura-
tion of use DU is reached. Constraints (5i) determine the overhaul x 0 1 1 1 1 1 1 0 0 0
ybuy 0 1 0 0 0 0 0 0 0 0
costs. A general overhaul takes place at the end of the tractor's
yov 0 0 0 0 0 1 0 0 0 0
lifetime, i.e. in period t þDU if the tractor has been bought in 0 0 0 0 0 0 0 1 0 0
ysell
period t. Therefore, ocost
t is forced to OC t þ DU if ybuy
t ¼ 1 and yov ¼ 1.
yov 1
Here, OC t þ DU denote the costs of a general overhaul if this over-
u2;8 2
haul takes place in period tþDU. ocost t is set to 0 in all other periods TC b;a 1,305,289
since costs are minimized in the objective function.
106 J.Y. Du et al. / Omega 64 (2016) 102–114
a feasible solution, all columns that have been generated when
solving the LP relaxation of RMP are used and RMP is solved as IP.
3. Demand and fleet related input data
This section addresses the generation of appropriate and rea-
listic input data (for future periods) to apply the model presented
in Section 2. In the following, we introduce a procedure for
demand aggregation based on a given flight schedule and present
an example of demand forecasting. Moreover, we explain how the
existing fleet can be incorporated in the model.
3.1. Demand pattern generation
Constraints (1b) in MP ensure in each period a sufficient
number of compatible tractors to satisfy demand. The demand for
a period is expressed by a set of demand patterns. One period
equals a winter or summer flight schedule, i.e. 6 months. In the
following, we explain the four steps of demand pattern generation
for one period from a given flight schedule.
Step 1: Select a representative peak day in the period: Within a
summer or winter flight schedule the departure times, parking
positions and plane types are similar each day. Non-peak days
differ from peak-days in that not all flights are scheduled while
scheduled flights usually depart from the same gate at the
same time.
Step 2: Calculate the time window of tractor occupancy for each
towing job of the selected day: The occupancy time comprises travel
time, waiting time and processing time. The upper part of Fig. 1
visualizes the outcome of step 2. The horizontal axis refers to the
time of the day, each row refers to one towing job. The bars show
the occupation time of a tractor for each job.
Step 3: Determine the number of simultaneous jobs for each time
interval and plane type and derive the maximum of the day: The
lower part of Fig. 1 shows for each plane type (rows) and each time
interval (columns) the number of simultaneous jobs. The last
column at the right displays the maximum for each plane type of a
day. In this example, the maximum for plane type A is 3. The
model generates one demand constraint for each plane type and
period. For example, the demand constraint for the first row
ensures that the number of tractors compatible with plane type A
is greater or equal to 3.
Step 4: Ensure the aggregated demand of plane types is satisfied
for overlapping tractor compatibilities: In some cases it is not suf-
ficient to ensure demand satisfaction for each plane type sepa-
rately. One tractor might be used for several jobs at the same time
if tractor compatibility overlaps. There are three cases:
 Case 1: The tractor compatibilities are disjoint sets. In Table 2
plane type A is compatible with tractor types 1 and 2, while
plane type B is compatible with tractor type 3. Here one
constraint per plane type is sufficient, i.e. no additional con-
straints are required. This results in two relevant demand
constraints.
Fig. 1. Step 2 and step 3 – tractor occupancy time per job and maximum number of simultaneous jobs.
 Case 2: The compatibility structure of one plane type is a subset
of the compatibility structure of another plane type. In Table 3
plane type A is compatible with all tractor types, and plane type
B is only compatible with tractor type 2. In this case one
additional demand constraint is required to ensure the number
of tractors compatible with plane types A or B (here tractor
types 1, 2 and 3) to be greater or equal to the sum of the
maximum number of simultaneous jobs for plane types A and B
(here 4). The constraint for plane type A becomes redundant.
This results in two relevant demand patterns: B and A þB.
 Case 3: There is an overlap of tractor compatibility without
subset structure (see Table 4). Here plane type A is compatible
with tractor types 1 and 2 and plane type B is compatible with
2 and 3, i.e. both plane types are compatible with 2. This results
in three relevant demand patterns: A, B and A þ B.
3.2. Demand forecasting
A high quality demand forecasting as input is essential for a
reliable schedule as output. Demand, and thus demand patterns,
are primarily influenced by three factors: (i) the total number of
towing jobs per day, (ii) the plane type mix and (iii) the number of
towing jobs at each time of the day (in this paper called “temporal
distribution”). The forecasting of the number of towing jobs for the
case study in Section 4 is based on the following approach and
assumptions.
The forecast for the number of towing jobs (i.e. influencing
factor i) is based on the forecast for the number of flight move-
ments. We assume a constant ratio between the number of flight
movements and the number of push-backs, repositionings and
maintenance towings (0.8 , 0.11 and 0.09 for the winter schedule
and 0.83, 0.09 and 0.08 for the summer schedule). The ratios are
calculated from past data of our cooperating airport. The forecasts
for the number of towing jobs per day are displayed in Table 5. The
table gives the forecast of the number of towing jobs for one peak
day in summer 2013–2022 and winter 2014–2023. Note that
winter 2014 refers to winter 2013/2014. We assume an annual
Table 2
Step 4 – compatibility structure case 1.
Tractor type 1 2 3
Plane type A ✓ ✓
Plane type B ✓
Table 3
Step 4 – compatibility structure case 2.
Tractor type 1 2 3
Plane type A ✓ ✓ ✓
Plane type B ✓
 J.Y. Du et al. / Omega 64 (2016) 102–114
increase of 2% in flight movements and respectively the number of
towing jobs is expected to increase by the same size per year.
Assumptions on changes in the plane mix and temporal dis-
tribution of towing jobs (i.e. influencing factors ii and iii) are based
on the standard flight schedule 2020 of our cooperating airport.
The standard flight schedule 2020 does incorporate a change in
the plane mix. Other relevant sources for the future development
of the fleet mix are Airbus [1] and Boeing Commercial Airplanes
[3]. Real processing times are used as input data for flight schedule
summer 2013 and winter 2014. For flight schedule 2020 we take
the average processing time per job type based on historic data.
Furthermore, we assume the temporal distributions of main-
tenance towings and repositionings to remain the same as in
summer 2013 and winter 2014. Based on these assumptions the
temporal distribution of towing jobs during a day is given in Fig. 2
(summer flight schedule 2013), Fig. 3 (winter flight schedule 2014)
and Fig. 4 (flight schedule 2020).
This information is generated for each plane type. We assume a
linear growth of the temporal distribution and change in plane
type mix from summer 2013 and winter 2014 towards 2020 and a
constant distribution thereafter. Since the distribution 2020 is not
differentiated into winter and summer, we assume the same dis-
tribution for both seasons. This assumption seems reasonable
since the temporal distribution of the schedule summer 2013 and
winter 2014 resemble each other (see Figs. 2 and 3).
By combining (i) total number of jobs per day, (ii) plane type
mix and (iii) temporal distribution, all relevant input data are
available to generate the demand patterns for the forecasting
periods.
3.3. Consideration of the existing fleet
An implementable schedule takes into account the existing
fleet. We distinguish between three tractor categories. The pool of
tractors which can be bought on the market is category 1. For each
tractor type, we create one SP. Category 2 are the pool of existing
tractors without decision options. These are primarily tractors
with a remaining lifetime of one period. We insert these tractors
directly as columns in MP and ensure their selection. Category
3 are existing tractors with decision options on overhaul or selling.
We create one additional SP for each and adapt parameter and
decision variable settings. For instance, if a tractor has been bought
in summer 2012, we set ybuy1 ¼ 1, i.e. the tractor has to be bought in
the first period of the planning horizon (in the case study summer
2013). Furthermore, we decrease the lifetime DU and investment
costs ICt for those two periods, which are not in the model's
planning horizon. Constraints (1c) in MP ensure the selection of NE
number of columns from this SP in the final solution.
Table 4
Step 4 – compatibility structure case 3.
Tractor type 1 2
Plane type A ✓ ✓
Plane type B ✓ ✓
Table 5
Forecasting of numbers of towing jobs per day.
2013 2014 2015 2016
Summer 554 566 579 592
Winter 552 564 576
107
4. Case study
This section illustrates the application of the model for our
cooperating airport. To alleviate the end of horizon effect, we con-
sider a planning period of 30 years (i.e. j T j ¼ 60) but investigate
only the first 10 years in more detail. In total, we consider 13 dif-
ferent tractor types. Two of them, tractor types T12 and T13, cannot
be overhauled (i.e. ADT12 ¼ 0 and ADT13 ¼ 0), since they are so-called
towbarless tractors which in contrast to towbar tractors, for tech-
nical reasons, do not have an overhaul option. Furthermore, we
consider the existing fleet which results in 9 additional types and
thus subproblems (see Section 3.3). We start by setting up a basic
scenario in Section 4.1. In our sensitivity analysis we consider
demand and waiting time scenarios (see Section 4.2), cost scenarios
(see Section 4.3), and management scenarios (see Section 4.4).
Finally, we investigate a green field scenario in Section 4.5. An
overview of all 25 test instances with summary statistics is given in
Table 6. Columns 1–7 give input statistics for each instance. Column
1 gives the ID and column 2 the number of tractor types, i.e. number
of subproblems. The third column shows the selling price as the
percentage of investment cost while column 4 indicates the change
of overhaul cost in percent compared to the basic scenario. The
minimum duration of use is shown in column 5. Column 6 gives the
change in demand as a percentage whereas column 7 shows the
waiting time considered. The final 5 columns give general output
statistics. In particular, column 8 shows the final IP value, column
9 gives the relative gap, column 10 the total runtime, column 11 the
number of tractors bought in the entire planning period, and column
12 the average number of usage periods. The first row shows the
statistics for the basic scenario. The next four rows consider the
waiting time and demand instances whereas rows 6–19 show
results for different cost scenarios (here we vary the selling price and
overhaul cost). The output for the management scenarios is given in
rows 20–24. The last row gives the statistics for the green field case.
IBM ILOG CPLEX Optimization Studio 12.5.1 is used to code and solve
the RMP and the SPs. All computations are performed on a 2.7 GHz
PC (Intel(R) Core(TM) i7-4600U CPU) with 16 GB RAM running
under the Windows 7 operating system. For instance, for the basic
scenario with a problem size of 22 tractor types (existing and new),
60 periods and 76 demand patterns, the CGH obtains a near optimal
solution (with a optimality gap of 0.006%) within 109 min (see row
1 in Table 6). In general, all instances could be solved to near
optimality (gap is less than 0.01%). Total runtime varies between 48
and 156 min with an average of 110 min.
4.1. Basic scenario
The parameters in the basic scenario are derived from real-
world data and are given in Table 8. Tractors are linearly depre-
ciated over the maximum duration of use. However, when selling
the tractor the received revenue is only a given percentage of the
book value. For the base scenario this percentage is 70%. As an
example, if a tractor with acquisition cost of 100 (in k euros) and a
maximum duration of use of 10 years is sold after 6 years, the book
3 value is 40 and the selling price is 0.7  40 ¼28. Dividing the sales
revenue by the remaining use periods gives SRt as used in the
objective function (5a) of the subproblem. Table 7 provides the
sales revenue and the sales revenue per remaining usage period
2017 2018 2019 2020 2021 2022 2023
605 617 630 643 656 668
589 601 614 626 638 651 663
108 J.Y. Du et al. / Omega 64 (2016) 102–114

Fig. 2. Demand curve of a peak day in summer 2013.

Fig. 3. Demand curve of a peak day in winter 2014.

Fig. 4. Demand curve of a peak day in 2020.

 J.Y. Du et al. / Omega 64 (2016) 102–114 109

Table 6
Input and output statistics for all 25 test instances.

ID # Trct Selling price Change overhaul MU Demand change Waiting time IP soln Gap (%) Time # Trct bought Avg. usage
(%) cost (min) (min) (periods)

1 22 70 – 6 – 10 63,789,578 0.006 109 144 8.0

2 22 70 – 6  10% 10 57,753,428 0.018 102 132 7.8
3 22 70 – 6 þ 10% 10 70,994,738 0.051 103 157 7.9
4 22 70 – 6 – 0 42,301,748 0.000 106 86 7.7
5 22 70 – 6 – 15 93,460,298 0.000 64 225 7.7
6 22 100 – 6 – 10 62,045,318 0.000 115 165 7.0
7 22 90 – 6 – 10 62,836,338 0.002 129 142 8.1
8 22 80 – 6 – 10 63,355,718 0.006 156 144 8.0
9 22 60 – 6 – 10 64,105,848 0.021 116 142 8.1
10 22 50 – 6 – 10 64,344,028 0.039 120 142 8.1
11 22 40 – 6 – 10 64,521,018 0.038 98 136 8.5
12 22 30 – 6 – 10 64,661,748 0.023 102 136 8.5
13 22 20 – 6 – 10 64,790,348 0.000 110 136 8.5
14 22 10 – 6 – 10 64,919,448 0.000 106 135 8.5
15 22 0 – 6 – 10 65,041,338 0.001 91 135 8.5
16 22 70  20% 6 – 10 63,119,778 0.013 135 143 8.1
17 22 70  10% 6 – 10 63,490,558 0.015 119 143 8.0
18 22 70 þ 10% 6 – 10 63,955,718 0.018 130 144 8.0
19 22 70 þ 20% 6 – 10 64,074,548 0.006 133 145 8.0
20 22 70 – 1 – 10 63,059,828 0.014 121 154 7.4
21 22 70 – 2 – 10 63,208,088 0.021 124 145 7.9
22 22 70 – 4 – 10 63,505,478 0.020 138 144 8.0
23 22 70 – 8 – 10 63,916,778 0.025 108 142 8.1
24 22 70 – 10 – 10 64,958,888 0.000 48 132 8.8
25 13 70 – 6 – 10 61,125,370 0.027 60 127 6.9

Table 7 From the figures, two conclusions can be drawn: first, there is a
Absolute sales revenue and sales revenue per remaining usage period for a max- clear preference for certain tractor types, namely T10, T12 and T13.
imum life time of 10 years.
The fleet mix is dominated by these three tractor types starting
Usage periods 1 2 3 4 5 6 7 8 9 10 from period 7. From period 13 onwards, the fleet consists of only
these three tractor types. In particular, T10 and T12 are char-
Remaining lifetime periods 9 8 7 6 5 4 3 2 1 0 acterized by high flexibility in terms of technical compatibility,
Sales revenue [k euro] 63 56 49 42 35 28 21 14 7 0
while T13 has comparatively low investment and variable costs.
Sales revenue per remaining life- 7 7 7 7 7 7 7 7 7 0
time period [k euro] Second, the current number of existing tractors is too high: in the
first period, the number of tractors used is predetermined by the
existing fleet. With a reduction in the existing fleet size due to
Table 8 aging, the total fleet size decreases in periods 2 and 3. As demand
Parameter settings and input data for the basic scenario.
increases the fleet size grows again in period 4. Demand with
Selling price of book value SPt 70% of investment costs respect to number of jobs is lower in winter than summer (see
General overhaul costs OCt 60–80% of the investment Table 5). However, processing time per job is longer during winter,
Minimum duration of use MU 6 periods thus total tractor occupation time is higher during winter. The zig-
Maximum duration of use DU 10 periods
zag-pattern in periods 1–14 (winter 2020) results from this sea-
Additional lifetime after overhaul AD 10 periods
Waiting time per job 10 min sonal variation (winter vs. summer). The fleet size stabilizes after
period 14, since we assume one temporal distribution for both
seasons from 2020 onwards.
for all periods of the example. In our case study, we varied the
selling price as percentage of the book value as well as other 4.2. Demand and waiting time scenarios
parameters (see Table 8). The variable cost per hour is fixed and
Demand increases by 2% per year in the basic scenario. Fur-
does not change over time. Consequently, the operating costs
thermore, an average waiting time of 10 min is added to each job,
(average hours of use within a period year multiplied by variable
based on historical data. Both factors impact the fleet size. We
cost per hour) are fixed too. Furthermore, time dependent oper-
analyze in the following the impact of an increase and a decrease
ating costs are not relevant, since the pushback company has a
in demand by 10% (scenarios Dþ 10% and D  10%), an increase and
service contract with a service provider and pays for each minute a decrease of the average waiting time of 15 min (WT15) and of
the motor is running, independent of the age of the tractor. 0 min (WT0).
However, our model is capable to consider such cases. Waiting time scenarios: Fig. 8 compares the waiting time sce-
In the basic scenario, 60 tractors are bought, 13 tractors are sold narios with the basic scenario. The bars show the total costs TC of
and 6 tractors are overhauled. In the optimal solution for the first the schedules in percentage of the basic scenario costs (right
10 years (i.e. 20 periods), a tractor is used on average 6.1 periods. vertical axis). The lines show the number of tractors to be bought,
Fig. 5 visualizes the number of tractors to be bought and sold in overhauled and sold (left vertical axis). Waiting time itself is an
periods 1–20. Fig. 6 shows the number of tractors in use per period indicator of the robustness of a schedule regarding disruptions in
and tractor type and Fig. 7 displays the number of tractors to be the flight schedule and daily operations. The fleet size decreases
used per period differentiating with respect to the categories of considerably when ignoring waiting time (WT0). However, with-
existing versus new tractors (see Section 3.3). out buffer time there is a high risk of push-back delays due to
110 J.Y. Du et al. / Omega 64 (2016) 102–114

Fig. 5. Number of tractors bought and sold per period (basic scenario).

Fig. 6. Number of tractors used per period and tractor type (basic scenario).

Fig. 7. New vs. existing number of tractors used per period (basic scenario).

Fig. 8. Waiting time scenarios vs. basic scenario.

 J.Y. Du et al. / Omega 64 (2016) 102–114
Fig. 9. Demand scenarios vs. basic scenario.
Fig. 10. Delta of D  10% scenario and basic scenario.
flight schedule disruptions. Increasing the average waiting time
from 10 min to 15 min (WT15) creates a larger buffer for disrup-
tions. The comparison of WT0 with the basic scenario can also be
interpreted from the following perspective: The towing service
provider faces about 34% higher investment costs due to tractor
occupation from schedule disruptions in daily operations.
Demand scenarios: Fig. 9 shows the results of the demand
scenarios. An increase or decrease of demand by 10% is roughly
equivalent to a cost increase or decrease of 10%. A demand
increase does not expose the towing service provider to any risks,
since new tractors can be bought anytime. However, a demand
decrease might lead to a suboptimal fleet, if part of the investment
schedule has already been realized. The greater the differences
between the various scenarios, the higher the risk of a suboptimal
decision. The later in the planning horizon the differences occur,
the higher the chance of revising the schedule without losing
optimality.
Fig. 10 shows the difference of number of tractors to be bought
(dark gray bars) and sold (light gray bars) per period between the
scenario D  10% and the basic scenario. A positive number in the
chart means more tractors are bought or sold in the D  10% sce-
nario. In the basic scenario 4 tractors more are bought whereas the
number of sold tractors is the same in both scenarios. Looking at
the first 10 periods (i.e. 5 years), the net difference for buying and
selling is almost zero. In comparison, the next five years exhibit
different decisions, i.e. buying fewer tractors in the low demand
scenario D  10%. So, buying or selling more tractors if demand
does not develop as expected are decisions that easily can be
111
carried out. Thus, the schedule in the case study is rather robust
with respect to demand variations, while waiting time or schedule
disruptions have a greater influence on the optimal fleet size.
4.3. Cost scenarios
In this section we investigate how the ratio between invest-
ment costs and selling prices, and the ratio between investment
costs and general overhaul costs influence the schedule.
Selling price scenarios: in the selling price scenarios (S100%–
S0%), we vary the revenues for selling tractors. The percentage
number in the scenario label indicates the selling price that can be
realized on the market in percentage of the initial investment
costs. In scenario S100% we assume that a tractor can be sold to
the market without any loss of value, while scenario S40% assumes
a loss of 60% in value if a tractor is sold. Scenario S0% reflects the
scenario of not having a selling option at all. Fig. 11 summarizes
the results for the selling price scenarios. In the basic scenario we
assume that the selling price equals 70% of the initial investment
costs (i.e. a 30% loss of value). Decreasing the selling price has
limited impact on the schedule and costs (see scenarios S60–S0%
in Fig. 11). Compared to the basic scenario, the total costs increases
at most by 2%. In contrast to the negligible changes in scenarios
S60–S0%, an increase in selling prices (scenarios S80–S100%) does
change buying and selling behavior considerably. In scenario
S100% in which we assume that tractors can be sold to the market
without any loss of value, the number of tractors to be sold almost
triples, and accordingly the number of tractors to be bought
112 J.Y. Du et al. / Omega 64 (2016) 102–114
Fig. 11. Selling price scenarios vs. basic scenario.
Fig. 12. Overhaul cost scenarios vs. basic scenario.
increases by about 15%. Without loss of value when selling a
tractor, the fleet more frequently adapts to better fit changing
demand.
Overhaul costs scenarios: Analogously to the selling price, we
vary the general overhaul costs in scenarios OV  20% up to
OV þ20%. Here the percentage number in the scenario label indi-
cates an increase or reduction of the general overhaul costs com-
pared to the basic scenario. In the basic scenario the general
overhaul costs are between 60% and 80% of the investment costs
for the different tractor types. In scenario OV  10% the overhaul
costs are set to 50–70% of the investment costs. The scenario
results are displayed in Fig. 12. As expected, the number of over-
hauls increases slightly with decreasing overhaul costs, while total
costs remain almost constant.
4.4. Fleet management scenarios
In the fleet management scenarios, we investigate the impact
of the fleet management policy on the schedule. The minimum
duration of use MU does not reflect a technical feature of a tractor,
but rather a management decision. A small MU value allows more
flexibility to adjust the fleet composition more frequently, while a
high MU value results in less fleet management effort and greater
stability in daily operations since schedulers and drivers do not
have to adapt to a new fleet that often. In the basic scenario the
minimum duration of use is set to 6 periods (this value has been
set by management of our cooperating airport). The analysis
shows that setting MU to different values, i.e. 1, 2, 4, 8 or 10 per-
iods, affects total costs by less than 2% (see Fig. 13). High flexibility
decreases costs by more than 1% which amounts to almost a
quarter of a million euros in our planning horizon of 30 years.
However, allowing vehicles to be sold after one period does not
seem reasonable from a fleet management effort perspective. On
the other side, low flexibility with MU ¼10 increases cost by about
1.8% or 1.16 million euros. Based on our results, the current policy
with MU¼ 6 seems to be a good balance between changing the
fleet too often and total costs. It is noticeable, that with MU¼ 10
there are more overhauls than tractors sold in the first 10 years. An
explanation is that after an overhaul the tractor can be sold right
away giving more flexibility in making decisions.
4.5. Green field scenario
In the green field scenario we ignore the existing fleet and
assume that the fleet is built from scratch. Compared to the basic
scenario, the total costs in the green field scenario decreases by
4.2% (see Fig. 14). This equals the savings potential which can be
achieved in the long run. Fewer tractors are bought in the green
field scenario compared to the basic scenario. The reasons are
those tractors in the existing fleet with a remaining lifetime of 1 or
2 periods. Fig. 15 shows that without an existing fleet, tractor types
T10, T12 and T13 dominate the fleet composition from the first
period on. In the entire planning horizon, the fleet mix consists of
only these three tractor types. The spikes in periods 6–14 can be
 J.Y. Du et al. / Omega 64 (2016) 102–114 113

Fig. 13. Fleet management scenarios vs. basic scenario.

Fig. 14. Green field scenario vs. basic scenario.

Fig. 15. Number of tractors used per period and tractor type (green field scenario).

explained due to different demand in winter and summer where
summer demand is lower. This behavior vanishes over time since
the demand in winter and summer converges.
5. Conclusion
This paper addresses the fleet composition problem of towing
tractors at airports at a strategic level. The set-covering
formulation derives a multi-period schedule for a set of hetero-
geneous towing tractors. The model optimizes fleet size and mix
by determining the time of buying, overhauling and selling trac-
tors. The model takes into account restrictions such as technical
compatibility of tractor types with plane types, a minimum and a
maximum duration of use. The model incorporates an existing
fleet. The inclusion of aspects such as an overhaul option better
captures investment decisions in real-world situations. Further-
more, we introduce a 4-step approach to aggregate demand using
114 J.Y. Du et al. / Omega 64 (2016) 102–114
flight schedule information. The proposed column generation
heuristic solves all scenario instances to near optimality with an
average runtime of 110 min.
We illustrate the application of the model for a major European
airport (our cooperating airport) in a case study. Using the model
we derive optimal schedules for a basic scenario and a number of
additional scenarios. The results demonstrate the robustness of
the schedule of the basic scenario towards changes in demand,
flight disruptions and varying cost parameters. About 5% saving
potentials are identified if the model is applied in the long-run.
The fleet composition model might be adapted to other appli-
cation areas, such as other vehicles at airports (e.g. de-icing trucks,
passenger transport buses, towable passenger boarding stairs), as
well as road vehicles (e.g. street buses, trucks). Moreover, the
model might be extended by incorporating temporary unavail-
ability of tractors due to maintenance cycles or a general overhaul.
Also, variable general overhaul periods and lifetime dependent
variable costs might be considered in a model extension.
References
[1] Airbus SAS. Global market forecast 2012–2031. Technical report; 2012.
[2] Airport Authority Zürich Airport, Operational towing survey. Technical report.
Zürich; 1992.
[3] Boeing Commercial Airplanes, Current Market Outlook 2012–2031. Technical
report; 2012.
[4] Burt C, Caccetta L, Welgama P, Fouché L. Equipment selection with hetero-
geneous fleets for multiple-period schedules. Journal of the Operational
Research Society 2011;62(8):1498–1509.
[5] Couillard J, Martel A. Vehicle fleet planning the road transportation industry.
IEEE Transactions on Engineering Management 1990;37(1):31–36.
[6] Desaulniers G, Desrosiers J, Solomon MM. Column generation. New York:
Springer; 2005.
[7] Du JY, Brunner JO, Kolisch R. Planning towing processes at airports more
efficiently. Transportation Research Part E: Logistics and Transportation
Review 2014;70:293–304.
[8] Etezadi T, Beasley J. Vehicle fleet composition. Journal of the Operational
Research Society 1983;34(1):87–91.
[9] Gould J. The size and composition of a road transport fleet. Operational
Research Quarterly 1969;20(1):81–92.
[10] Hoff A, Andersson H, Christiansen M, Hasle G, Lokketangen A. Industrial
aspects and literature survey: fleet composition and routing. Computers &
Operations Research 2010;37(12):2041–2061.
[11] Kirby D. Is your fleet the right size. Journal of the Operational Research Society
1959;10:252.
[12] Loxton R, Lin Q, Teo KL. A stochastic fleet composition problem. Computers &
Operations Research 2012;39(12):3177–3184.
[13] New C. Transport fleet planning for multi-period operations. Operational
Research Quarterly 1975;26(1):155–166.
[14] Redmer A, Zak J, Sawicki P, Maciejewski M. Heuristic approach to fleet com-
position problem. Procedia—Social and Behavioral Sciences 2012;54:414–427.
[15] Salhi S, Sari M, Saidi D, Touati N. Adaptation of some vehicle fleet mix heur-
istics. Omega 1992;20(5):653–660.
[16] Schick G, Stroup J. Experience with a multi-year fleet planning model. Omega
1981;9(4):389–396.
[17] Wu P, Hartman J, Wilson G. An integrated model and solution approach for
fleet sizing with heterogeneous assets. Transportation Science 2005;39(1):87–
103.
[18] Wyatt J. Optimal fleet size. Journal of the Operational Research Society
1961;12(3):186–187.