Transportation Research Part C 68 (2016) 38–57

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

An integrated approach for airline scheduling, aircraft fleeting

and routing with cruise speed control
Hüseyin Gürkan a, Sinan Gürel b, M. Selim Aktürk c,⇑
a
Fuqua School of Business, Decision Sciences, Duke University, USA
b
Department of Industrial Engineering, Middle East Technical University, Turkey
c
Department of Industrial Engineering, Bilkent University, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: To place an emphasis on profound relations among airline schedule planning problems and
Received 12 October 2015 to mitigate the effect of unexpected delays, we integrate schedule design, fleet assignment
Received in revised form 5 March 2016 and aircraft routing problems within a daily planning horizon while passengers’ connection
Accepted 5 March 2016
service levels are ensured via chance constraints. We propose a nonlinear mixed integer
Available online 1 April 2016
programming model due to the nonlinear fuel consumption and CO2 emission cost terms
in the objective function, which is handled by second order conic reformulation. The key
Keywords:
contribution of this study is to take into account the cruise time control for the first time
Airline scheduling
Aircraft fleeting and routing
in an integrated model of these three stages of airline operations. Changing cruise times of
Cruise time controllability flights in an integrated model enables to construct a schedule to increase utilization of fuel
Second order cone programming efficient aircraft and even to decrease total number of aircraft needed while satisfying the
same service level and maintenance requirements for aircraft fleeting and routing. There is
a critical tradeoff between the number of aircraft needed to fulfill the required flights and
overall operational expenses. We also propose two heuristic methods to solve larger size
problems. Finally, computational results using real data obtained from a major U.S. carrier
are presented to demonstrate potential profitability in applying the proposed solution
methods.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Managing an airline is unavoidably expensive. One of the most basic costs is the price of purchasing the aircraft which
range from 70 to 400 mio USD per unit Boeing (2015). Moreover, among the operational expenses, fuel has been the largest
single cost term for the global airlines. According to IATA (2010) analysis on airline financial data, fuel expenses accounted
for $210 billion in 2013 for global airline industry. On the other hand, unexpected delays are endemic in airline operations
and to demonstrate their impact, the estimated total cost to the U.S. economy because of flight delays was as much as
$41 billion in 2007 Rebollo and Balakrishnan (2014). Naturally, the efficient utilization of such expensive resources, decreas-
ing operational expenses and higher robustness are objectives of any profitable airline. However, unfortunately these terms
are inversely correlated, in other words higher utilization might cause higher operational cost, less robustness and vice versa.
In this direction, generating a robust schedule with high aircraft utilization and less operational expenses at the same time is
quite crucial for any airline. Therefore, in this paper, we propose an integrated model for robust schedule design, aircraft

⇑ Corresponding author.

http://dx.doi.org/10.1016/j.trc.2016.03.002
0968-090X/Ó 2016 Elsevier Ltd. All rights reserved.
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 39

routing and fleeting with cruise speed control that aims efficient aircraft utilization and robustness within the consideration
of operational expenses such as fuel consumption, CO2 emission and spill cost.
Airline schedule design problem decides where to fly and in which frequency in consideration of market demand,
profitability, available resources and the competitors. Due to its broad scope, Barnhart et al. (2003) state that building flight
schedules from scratch is performed manually with limited optimization in the typical airline practice. Following the con-
struction or design of a flight schedule, fleet assignment problem tries to find the optimal assignment of aircraft types to
flights by considering the number of aircraft in each fleet and coverage of all flights. After fleet assignment decomposes flight
networks into subnetworks in terms of a particular fleet type, maintenance routing problem assigns individual aircraft to
these flights in consideration of the maintenance requirements. For a general review on airline schedule planning problems,
we refer the reader to Barnhart and Cohn (2004).
Since the different airline schedule planning problems are strongly related to each other, several integrated models are
proposed that take into account combinations of these problems to improve suboptimal solutions for the entire system.
Lohatepanont and Barnhart (2004) consider schedule design and fleet assignment in an integrated way in which a base
schedule and two flight lists including mandatory and optional flights are given. Starting from the base schedule they con-
sider deleting/adding flights from/to the base schedule with respect to given flight lists. In a similar fashion, Sherali et al.
(2013a) propose a model that integrates the schedule design and fleet assignment processes while considering flexible flight
times, schedule balance, and recapture issues, along with optional legs, path/itinerary-based demands, and multiple fare-
classes. Differently, they consider the flow of passengers along itineraries over the network together with flight scheduling
and fleeting decisions in order to maximize profits. Integrating three problems enables to improve local optimal solutions,
however tractability worsens as much as the scope of integration expands. Therefore, these integrated problems are modeled
and solved for a daily planning horizon.
In the literature, there are studies related to our work in some aspects such as daily planning horizon, passenger connec-
tion, cruise speed control or maintenance considerations. Duran et al. (2015) propose a robust airline scheduling model with
controllable cruise times. In their study, the tradeoff between the costs of cruise time change and idle time insertion is
considered while passengers’ connection service levels are ensured by chance constraints. Speed control is quite a recent
concept in solving airline scheduling problems. Aktürk et al. (2014) is the first study that makes use of speed control in
the context of airline schedule recovery from disruptions. Sherali et al. (2013b) propose an approach in which they integrate
the schedule design, fleet assignment, and aircraft-routing problems within the consideration of flight selection, departure
timing and maintenance requirements. For maintenance requirements, they use a limit on total flight time of each aircraft
that might be different for each fleet type. As a solution method, they use Benders’ decomposition and enhance the model via
valid inequalities. Haouari et al. (2013) propose a model for daily maintenance routing problem in which they ensure
maintenance feasibility by counter constraints on flight hours, take offs and number of days since the last maintenance
checks for each aircraft. They present a compact polynomial-sized representation for the general aircraft routing model
and they linearize and lift that representation. Moreover, in the study of Aloulou et al. (2013), a MIP model is proposed
for the robust aircraft routing problem without directly accommodating maintenance constraints however by considering
that the flights start and end in the single hub where maintenance checks are achieved overnight. Aloulou et al. (2013)
capture robustness by an objective function pertaining to aircraft and passenger connections.
What distinguishes our work from the studies above and makes challenging simultaneously is cruise speed control and
integration. To the best of our knowledge, this is the first study in which cruise speed/time is controlled within the integrated
robust schedule design, aircraft fleeting and routing problem. In our study, the fuel consumption and CO2 emission cost
functions are nonlinear functions in cruise time and involve binary variables. We have shown that these nonlinear functions
with binary variables can be transformed into a set of second order conic inequalities. Moreover, even if it is a special case of
our problem, Parmentier (2013) showed that aircraft routing problem by itself is an NP-complete problem. In addition to
aircraft routing problem, we consider robust airline scheduling and fleet type assignment problems in an integrated fashion
that involves a large number of decision variables. For that reason, when the number of flights and aircraft increases, the
problem size increases drastically. We also consider passengers’ connection service levels with chance constraints as well
as departure timing, idle time insertion and cruise speed control different from the aircraft routing problem. Changing cruise
time of flights in an integrated model enlarges the solution space and enables to construct a schedule with new flight
sequences, which could not be considered previously due to fixed cruise speed/time restriction. For two flights to be
connected consecutively by the same aircraft, there must be enough time gap between departure times of these flights. This
time gap is the sum of cruise time, non-cruise time, turnaround time and idle time. In other studies, the lower bound for this
gap is taken as fixed, however cruise speed/time change enables to control this lower bound on the gap between departure
times. By this means, in our study more flight connection alternatives could be generated.
The first contribution brought by our study is that aircraft utilization could be increased and even total number of aircraft
needed to cover a set of flights could be decreased while ensuring equivalent service level and maintenance requirements.
Due to having more alternatives on flight connections and compression of cruise time of flights, it is possible to increase the
number of flights to be performed by an efficient aircraft. While this increase in the utilization of fuel efficient aircraft could
reduce the minimum number of required aircraft to perform a set of flights. There is a critical tradeoff between the number
of aircraft needed to fulfill the required flights and the overall operational expenses, such as fuel consumption costs.
The second is the robustness issue. Since we have more alternatives on flight connections, it is possible to generate better
flight sequences in terms of robustness. For example, on a route having a flight with a great delay probability would require
40 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

an intervention for the following flights to be performed on time while removing the problematic flight from that sequence
could render that intervention unnecessary. Our study has more options to make this type of changes on routing decisions so
changing routes could improve the robustness.
Finally, we have used the recent advances in second order cone programming to handle the nonlinear mixed integer
programming formulation. Furthermore, in order to solve the large scale problems in a reasonable time, we propose two
heuristic algorithms. The first one is discretized approximation and cruise speed control algorithm, and the second one is
multi-stage triplet search algorithm.
In the remaining of the paper, in Section 2, the problem definition is presented. Section 3 is devoted to the problem
formulation and conic reformulation. In Section 4, heuristic methods are proposed and in Section 5 results of a computa-
tional study are presented.

2. Problem definition

Our problem is to solve robust airline schedule design, aircraft fleeting and routing problems within a daily planning hori-
zon for a given set of flights and a set of aircraft in an integrated manner while considering maintenance requirements and
passengers’ connection service levels. In the proposed model, we determine the departure time, cruise time, an inserted idle
time, if necessary, and aircraft fleet type for each flight along with the routing decisions. For each aircraft, the model
determines whether that aircraft is going to be used and, if it is used, the flight sequence to be flown for that particular air-
craft. These decisions are made in consideration of the idle time insertion cost, fuel consumption and CO2 emission costs,
spill cost and aircraft usage cost simultaneously.
The feasible set of solutions satisfy maintenance requirements, passengers’ connection service levels along with the flight
connection constraints in cases where two flights are performed consecutively by the same aircraft. For maintenance
requirements, we adopt two basic rules for each aircraft. The first one is that we limit the total cruise time for each aircraft
on a day. A similar approach is used by Sherali et al. (2013b), they use a limit (kt ) on total flight hours of each aircraft on a
day. The second one is that the first and the last airport of an aircraft is predetermined on a day similar to the study of
Aloulou et al. (2013). In case when an aircraft is used, in order that aircraft to follow its ordinary maintenance checks, we
secure that aircraft takes off from/lands to the particular airports at the beginning/end of a day. For passengers’ connection,
we use service levels with chance constraints and ensure necessary time for passengers to make connection with a particular
probability, denoted as service level, between each pair of flights which is feasible for passengers’ connection. Furthermore,
we provide minimum turnaround time between flights which are performed by the same aircraft.
Similar to the study of Duran et al. (2015), in our model we take the flight time as the sum of cruise time and non-cruise
time. Non-cruise time consists of the taxi-in and taxi-out stages as well as climb and descend stages of a flight, which include
uncertainty depending on the airport congestion or weather conditions while cruise time refers to the time which is control-
lable by with speeding up the aircraft. Consequently, we take cruise time as a decision variable and non-cruise time as a
given parameter. On the other hand, there are different methods in the literature to investigate the role of fuel cost on
aircraft operating costs as discussed by Ryerson and Hansen (2013). In this study, we have used the following function pro-
posed by Aktürk et al. (2014) and EUROCONTROL (2012) to calculate fuel consumption in kg for a given cruise time of flight i
operated by aircraft type t:

t 1 1 t 3 t 2
F ti ðf i Þ ¼ ci;t
1  t þ ci;t
2  þ ci;t i;t
3  ðf i Þ þ c 4  ðf i Þ ð2:1Þ
fi t 2
ðf i Þ

where ci;t i;t i;t i;t

1 ; c 2 ; c3 and c4 correspond to aircraft properties and the coefficients related to physical conditions of a flight for
each flight i and aircraft type t pair as discussed in Aktürk et al. (2014). These coefficients for 399 different aircraft types are
available in EUROCONTROL (2012). Fuel cost has been the largest operational expense for the global airlines. The EU emis-
sions trading system (EU ETS) covers airlines as well as power plants and industrial plants in 31 countries. By introducing
CO2 emission costs, these trading systems aim at cutting CO2 emissions by airline companies. Consequently, the fuel con-
sumption and CO2 emission costs are calculated by multiplying fuel consumption with their corresponding cost coefficients
ðcfuel þ cCO2 Þ.
In this study, similar to the study of Duran et al. (2015), passengers’ connections are taken into account to develop a
robust schedule such that mis-connections of passengers are minimized when a disruption occurs. Between two flights i
and j, if the origin airport of the flight j is same as the destination airport of the flight i and the departure time of flight j
is later than the arrival time of the flight i, the time needed for the passengers’ connection is TPij . It is ensured that the
percentage of the passengers’ connection satisfied between flights ði; jÞ is greater than the minimum desired service level
for each connecting flight pair, denoted as cij , through a set of probabilistic chance constraints as discussed later.

2.1. Numerical example

In order to elaborate our problem definition and the model mechanics, in this section, we provide a numerical example.
First, we present a published schedule in Table 1, which shows the planned departure time and cruise time for each flight,
and idle time after each flight, if necessary, to satisfy the passenger connections. In our proposed approach, a new schedule
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 41

Table 1
Published schedule.

Tail# Flight # From To Departure time Cruise time Idle time

N3ELAA 2057 ORD SJU 08:30 270 35.6
2078 SJU ORD 14:25 315 –
N3DUAA 2099 ORD LAX 07:00 250 34.9
1972 LAX ORD 12:40 225 24.4
1972 ORD RDU 17:45 95 –
N412AA 2345 ORD DFW 17:15 135 2.5
2374 DFW ORD 20:40 110 –
N4XGAA 2079 ORD SAN 08:45 250 13.5
1438 SAN ORD 14:00 230 58.9
346 ORD LGA 19:50 115 –

with a new aircraft fleeting and routing is generated by considering idle time cost, fuel consumption and CO2 emission costs,
spill cost and daily usage cost simultaneously. In accordance with the marketing requirements, the departure times could be
changed within ±15 min. The proposed schedule, given in Table 2, is feasible in terms of maintenance requirements because
all aircraft land to the same airports as in the published schedule, and we also limit the flight hours of each aircraft regarding
maintenance requirements. Further, passengers’ service levels are ensured by the chance constraints in the proposed sched-
ule. For this numerical example same service level is adopted with the published schedule for each passengers’ connection.
There are major changes between two schedules such as the number of aircrafts used, routings of aircraft, idle time inser-
tion and cruise times. The total number of aircraft used in the published schedule was four, while the proposed schedule
clearly illustrates that the same passenger connection service levels can be attained with three aircraft. This improvement
can be achieved by compressing the cruise time of flight 1438 around 14%, from 230 min to 198.4 min in the proposed sched-
ule. In addition, the departure time of flight 2057 is changed from 8:30 to 8:45 in the proposed schedule. Fig. 1 represents the
time space networks of the schedules. The horizontal axis represents the time and the vertical axis represents the airports.
For each flight, there is an angular arrow nearby and a flight number. The horizontal arrows represent the turn around and
idle time of an aircraft in the corresponding airport. The route of each aircraft is drawn by a different line style. For example,
the solid arrows represent the operations of tail number N4XGAA.
We compare the performance of the published schedule with the proposed schedule in terms of the improvement in total
cost by using the following formula:

Cost of the Published Schedule  Cost of the Proposed Schedule

Cost Improvement ¼ 100 
Cost of the Published Schedule
The cost calculation of the published schedule is shown in Table 3. The fuel consumption and CO2 emission costs calcu-
lations are explained in former sections. Idle time costs are calculated by the multiplication of unit idle time cost of each
aircraft and the idle time amount of that aircraft after each flight. For a given demand realization, spill costs are calculated
by multiplication of the number of spilled passengers and the cost incurred when a spill occurs. In the published schedule,
we assume that there is no spill cost. However, when we assign an aircraft with a smaller number of seats than the number
of already assigned passengers, a spill cost can incur as shown in Table 4. Daily usage costs are the costs incurred when the
corresponding aircraft are used. Total cost of the published schedule is $419,821.3.
As mentioned before, the total number of aircraft used is decreased from 4 to 3. In the proposed model, the most expen-
sive aircraft in terms of daily usage cost is N3ELAA which is not a fuel efficient aircraft in comparison to N412AA and
N3DUAA. Fig. 2 illustrates the differences in efficiency by providing the fuel consumption and CO2 emission costs that would
occur if flight 1438 was performed by these aircraft. Our model generates a schedule in which N3ELAA is not used due to its
high daily usage cost and low fuel efficiency while the other aircraft are utilized more. Although the fuel efficiency of aircraft

Table 2
Proposed schedule.

Tail# Flight # From To Departure time Cruise time Idle time

N3DUAA 2079 ORD SAN 08:38 249.5 8.4
1438 SAN ORD 13:45 198.4 –
1972 ORD RDU 18:00 94.8 –
N412AA 2099 ORD LAX 06:55 249.9 –
1972 LAX ORD 12:25 221.7 –
2345 ORD DFW 17:26 133 –
2374 DFW ORD 20:55 110 –
N4XGAA 2057 ORD SJU 08:45 275.9 1.2
2078 SJU ORD 14:10 293.6 –
346 ORD LGA 20:05 115 –
42 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

SJU

SAN

LGA

ORD

DFW

RDU

LAX

06:00 07:00 08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00 22:00 23:00 24:00

(a) Published schedule

SJU

SAN

LGA

ORD

DFW

RDU

LAX

06:00 07:00 08:00 09:00 10:00 11:00 12:00 13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00 22:00 23:00 24:00

(b) Proposed schedule

Fig. 1. Time space networks.

Table 3
Cost calculation for the published schedule.

Tail# Flight # Fuel and CO2 Idle time Spill Daily usage
N3ELAA 2057 7611 5334 – 90,000
2078 8879.5 – –
N3DUAA 2099 5923.4 4955.8 – 85,200
1972 5331 3461.9 –
1972 2250.9 – –
N412AA 2345 2319 347.2 – 84,000
2374 1889.6 – –
N4XGAA 2079 6509.4 1944 – 86,400
1438 5988.7 8481.6 –
346 2994.3 – –
Total 49696.8 24524.5 0 345,600

N4XGAA is worse than N3ELAA, the model prefers to use N4XGAA since its daily usage cost is smaller than N3ELAA. As a
direct consequence of better aircraft utilization, the cost of daily usage is improved by 26%.
Another effect of higher efficient aircraft utilization is on the fuel consumption and CO2 emission costs. Since utilizations
of efficient aircraft are increased and better flight aircraft assignment is achieved, total cost of fuel consumption and CO2
emission is improved around 10%. This improvement is achieved even if the cruise times of flights 2078, 1438 and 2345
are decreased by 7%, 14% and 2%, respectively. As an example of better flight aircraft assignment, in the published schedule
flight 1438 is performed by aircraft N4XGAA with $5988.7 fuel and emission cost, while in the proposed schedule same flight
is performed by aircraft N3DUAA with $5650 fuel and emission costs though the cruise time is compressed by 14% in albeit
of additional fuel and emission costs. Since N3DUAA is more efficient than N4XGAA, the extra fuel and emission costs caused
by that compression which is necessary to assign 1438 to N3DUAA can be compensated. In the proposed schedule, we
increase the utilization of N412AA, which is the most fuel efficient aircraft. Furthermore, the necessity of idle time insertion
after flights to satisfy the desired minimum service levels for passenger connections can be eliminated by changing routing
decisions, cruise time and departure time.
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 43

While there is an improvement in sum of daily usage cost, idle time cost, fuel consumption and CO2 emission costs, the
spill cost is increased in the proposed schedule. The reason of this increase is that the fleet assignment in the published
schedule is generated in consideration of aircraft capacity and passenger demand, however, in the proposed model the other
cost terms are also considered. The improvement in the other cost terms overcomes the increase in the spill cost. While in
the published schedule the spill cost is 0, the proposed model has a spill cost $2022.1. The spill of passengers are occurred in
flights 2079, 1438, 2099 and 1972 in the proposed schedule.
When we compare the overall results, the total cost is improved 27% in the proposed schedule. The costs of both sched-
ules are calculated at the planning stage. After the realizations of non-cruise times of each flight, the actual cost of the idle
time and delay costs can be seen. For this purpose a simulation study can be done to compare the realized costs of the sched-
ules. However, with this numerical example, we aim to show how the mechanics of the proposed model work.

3. Problem formulation

Our main aim in this paper is to integrate schedule design, fleeting and routing decisions along with aircraft cruise speed
decisions so that the total cost is minimized while the passenger connection service level constraints are satisfied. In this
section, we first present the mathematical formulation of the problem. Consequently, by the help of this model, we can easily
demonstrate the interdependencies among these interrelated problems. Afterwards, we show the conic reformulation of the
objective function and give the reformulation of the model.

Parameters
Idlet : Unit cost of the idle time of aircraft t
Capt : Seat capacity of aircraft t
Demi : Passenger demand of flight i
Dailyt : The cost incurred when aircraft t is used
TAtij : Turntime needed to prepare aircraft t between flights i and j
TP ij : Turnaround time needed to connect passengers between flights i and j
Cspli : Opportunity cost of spilled passengers of flight i
kt : The total available cruise time of aircraft t on a day
cij : minimum desired passengers’ connection service level between flights i and j
cfuel : Cost of fuel per kg of aircraft fuel consumption
cCO2 : Cost of emission per kg of aircraft CO2 emission
u l
f i ; f i : Upper and lower limit of the cruise time of flight i
u l
di ; di : Upper and lower limit of the departure time of flight i
eb : Airport congestion coefficient for airport b
Or i : Origin airport of flight i
Dni : Destination airport of flight i
NC i : The random parameter denoting the non cruise time of flight i

Decision variables
xtij : 1 if flights i and j are performed by the same aircraft t consecutively and is 0 ow.
yti : 1 if flight i is the first flight performed by aircraft t and is 0 ow.
zti : 1 if flight i is the last flight performed by aircraft t and is 0 ow.
di : Departure time of flight i
sti : Idle time of aircraft t after flight i
t
f i : Cruise time of flight i performed by aircraft t

In the proposed nonlinear mixed integer programming model, we utilize the following sets. T represents the set of aircraft
that can be operated. F is the set of flight legs which have to be covered by using the aircraft in T. For each flight i in F, there
are two sets which are U i , upstream flights and Di , downstream flights. U i denotes the set of flights which can follow flight i in
terms of departure time and origin, destination airports pair. It is easy to check origin destination pairs, since the destination
airport of a flight in U i has to be same with the origin airport of i. For departure time, there has to be a sufficient time for
connection so that a flight in U i and flight i might be performed by the same aircraft consecutively. If flight j is in U i then,
flight i is in D j . A denotes the set of flights ði; jÞ such that flight j 2 Di , i.e. flight i 2 U j . If ði; jÞ is in A, it means that an aircraft
can fly these flights consecutively. On the other hand, the connection ði; jÞ can also be avoided removing this pair from the set
A, whenever there are some conditions on ði; jÞ restricting the chance of being performed by the same aircraft. In our numer-
ical examples, we only consider the most necessary requirements which are the consistency between the arrival time/airport
of the former flight i and the departure time/airport of the latter flight j. B denotes the set of airports to be considered in the
model. F te denotes the set of flights for an aircraft t in T such that the flights in F te can be the last flight of aircraft t since they
44 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

terminate at the convenient airports in terms of maintenance feasibility. F ts denotes the set of flights for an aircraft t in T such
that the origin airport of flights in F ts are the airport from which that aircraft starts the planning day. Pi is the set of flights
such that the passengers in flight i might have connection.
For each i 2 F and t 2 T, we redefine the fuel consumption as follows:
8 0 1
>
>
3
2  X
> c 1t þ c 1 þ c f
> i;t i;t i;t t i;t
þ c4 f i
t
if @yti þ xtji A ¼ 1
>
> 1 f 2
ðf ti Þ
2 3 i
< i
j2U i
t
F ti ðf i Þ ¼ 0 1 ð3:1Þ
>
> X
>
>
>
>
:0 if @yti þ xtji A ¼ 0
j2U i

t
so that if aircraft t is not assigned to flight i, then F ti ðf i Þ ¼ 0.
The proposed nonlinear mixed integer programming model is provided below:
0 1
XX X XX XX XX
min @ x þy A Cspli maxð0;Demi Capt Þþ
t t t
ðcfuel þcCO2 ÞF ti ðf i Þþ yti Dailyt þ sti Idlet ð3:2Þ
ji i
t2T i2F j2U i i2F t2T i2F t2T i2F t2T
X X
s:t: xtji þyti  xtij zti ¼ 0 8i 2 F; t 2 T ð3:3Þ
j2U i j2Di
X
yti 6 1 8t 2 T ð3:4Þ
i2F
0 1
X X
@y þ
t
xji A ¼ 1 8i 2 F
t
ð3:5Þ
i
t2T j2U i
X
IF xtij ¼ 1
t2T
X t X 8ði;jÞ 2 A ð3:6Þ
THEN; dj di TAtij  f i E½NC i  sti ¼ 0
t2T t2T
" #
X t
Pr NC i 6 dj di  f i TP ij P cij 8i 2 F; j 2 Pi ð3:7Þ
t2T
X t
f i 6 kt 8t 2 T ð3:8Þ
i2F
0 1
X
IF @yt þ i
tA
xji ¼ 1
j2U i
l t u
THEN; f i 6 f i 6 f i 8i 2 F; t 2 T ð3:9Þ
ELSE
t
fi ¼0
sti ¼ 0
t
8
yi ¼ 0 t 2 T; i 2 F nF ts ð3:10Þ
8
zti ¼ 0 t 2 T; i 2 F nF te ð3:11Þ
l u
di 6 di 6 di 8i 2 F ð3:12Þ
sti P 0 8i 2 F; t 2 T ð3:13Þ
xtij 2 f0;1g 8ði;jÞ 2 A; t 2 T ð3:14Þ
yti 2 f0;1g 8i 2 F; t 2 T ð3:15Þ
The objective function (3.2) is the sum of the spill cost, fuel consumption and CO2 emission cost, the daily aircraft usage
cost and idle time cost, respectively. Constraint (3.3) is network balance equations. Constraints (3.4) and (3.5) are typical
assignment constraints, such that each aircraft can be used for at most one flight sequence (string or path), and each flight
can be performed by exactly one aircraft. Constraint (3.6) ensures that if two flights are performed by the same aircraft con-
secutively, time between the departures of these flights have to be greater than the sum of cruise time, non-cruise time and
turnaround time. This time difference corresponds to idle time at the end of the first flight. Hence, when we add the idle time
variable to (3.6), we obtain an equality constraint since the idle time is always nonnegative due to constraint (3.13). Con-
straint (3.7) is the chance constraint which ensures the service level of passengers’ connection. This probabilistic constraint
guarantees that the percentage of the passengers’ connection satisfied between flights ði; jÞ is greater than the minimum
desired service level for each connecting flight pair, denoted as cij , such that the probability of the time between arrival
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 45

Table 4
Cost calculation for the proposed schedule.

Tail# Flight # Fuel and CO2 Idle time Spill Daily usage
N3DUAA 2079 5923.3 1192.8 453.1 85,200
1438 5650 – 431.5
1972 2250.8 – –
N412AA 2099 4294.5 – 443.9 84,000
1972 3866.4 – 693.6
2345 2320.4 – –
2374 1889.6 – –
N4XGAA 2057 7035.2 172.8 – 86,400
2078 8260.5 – –
346 2994.3 – –
Total 44,485 1365.6 2022.1 255,600

7000

6000

5000

4000

3000

2000

1000

0
N3DUAA N412AA N4XGAA N3ELAA

Fig. 2. The fuel and CO2 cost of flight 1438 with each aircraft type.

of flight i and departure of flight j being greater than the required connection time TP ij is at least cij . Due to constraint sets
(3.6) and (3.7), there is a critical tradeoff between the passenger connection service levels defined in the constraint set and
the decision variables sti (the idle time of aircraft t after flight i, i.e. slack time), which are penalized in the objective function
of the proposed model. When sti values are increased, it is more likely to catch the connecting flight, i.e., higher passenger
connection service levels. Constraint (3.8) ensures that total cruise time of an aircraft does not exceed a predetermined time
limit in order to ensure maintenance feasibility. If a flight i is performed by aircraft t then constraint (3.9) limits cruise time
t
change; cruise time of a flight cannot exceed the upper and lower bounds, else the corresponding variables f i and sti are set to
zero. The aim of constraints (3.10) and (3.11) is to sustain maintenance policy intended in the published schedule, in this
fashion first and last airport for each aircraft is determined. Constraint (3.12) sets the upper and lower bounds for departure
time of each flight due to marketing requirements. As discussed above, it is guaranteed that the idle time is always nonneg-
ative due to constraint (3.13). Constraints (3.14) and (3.15) guarantee that xtij and yti are binary variables and due to (3.3) all zti
are also binary variables.

3.1. Reformulation of the model

In the proposed model, we have a nonlinear cost terms in the objective function and a probabilistic chance constraint.
Therefore, we conduct a conic reformulation of the model so that we can utilize the recent advances in second order cone
programming to solve the proposed formulation using a commercial solver.
In the objective function, the cost functions involve nonlinearity due to controllable cruise time associated with the
changing the cruise speed of the aircraft. To handle nonlinearity, nonlinear mixed integer optimization often requires too
much computation time. On the other hand, it may not result in exact solutions. In order to shorten the solution time
and obtain an optimal solution, in this section we show the conic quadratic reformulation of the fuel consumption function
t
F ti ðf i Þ cost as discussed in Aktürk et al. (2009) and Günlük and Linderoth (2010). To simplify the presentation, we drop the
indices of the variables and parameters.
(
3 2

c1 1f þ c2 ðf1Þ2 þ c3 ðf Þ þ c4 ðf Þ if w ¼ 1
Fðf Þ ¼
0 if w ¼ 0

P 
where w ¼ wti ¼ yti þ t
j2U i xji , for simplicity.
46 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

n o
Fðf Þ is discontinuous and therefore its epigraph EF ¼ ðf ; tÞ 2 R2 : Fðf Þ 6 t is non-convex. In the next proposition, we
describe how the convexity of EF is obtained. A more detailed information can be found in Aktürk et al. (2009) and
Günlük and Linderoth (2010).

Proposition 1. The convex hull of EF can be expressed as

t P ðc 1  q þ c 2  d þ c 3  / þ c 4  v Þ ð3:16Þ
w 6qf2
ð3:17Þ
2
w4 6 f  d  1 ð3:18Þ
4
f 6w /f 2
ð3:19Þ
f 6v w
2
ð3:20Þ
in the constraint set. Moreover, each inequalities (3.16)–(3.20) can be represented by conic inequalities.

Proof. Perspective of a convex function Fðf Þ is wFðf =wÞ (Hiriart-Urruty and Lemaréchal, 2001). Since each of the nonlinear
3 2
terms 1
f
; f12 ; f and f is a convex function for f P 0, then epigraph of the perspective of each term can be stated as
w2
f
6 q; w4
f2
6 d; f3
w2
6 /, and f2
w
6 v , respectively, where q; d; / and v are the variables required for defining epigraphs. Since
w; f P 0, they can be written as stated in the proposition.
Finally, observe that (3.17) and (3.20) are hyperbolic inequalities, (3.18) can be restated as two hyperbolic inequalities
w2 6 u  f and u2 6 d  1
and (3.19) can be restated as
2
f 6 u  w and u2 6 /  f
which can be written as a conic quadratic inequality. h
We now have a linear fuel cost in the objective function with a set of conic inequalities given in (3.22)–(3.25) as in Aktürk
et al. (2014). Furthermore, we take the flight time as the sum of cruise and non-cruise times. In our integrated model, the
cruise time of each flight is a decision variable, whereas the non-cruise time of each flight i 2 F; NC i , is a given parameter.
Arikan et al. (2013) performed an extensive study on airline flight schedules across several U.S. airlines and showed that the
log-Laplace distribution provides a good fit to the block time of a flight. Therefore, in our computational study, the non-cruise
time of each flight, NC i , is taken from a log-Laplace distribution. Consequently, probabilistic chance constraints could be rep-
resented in a closed form and written as second order conic inequalities in (3.26) as discussed in Duran et al. (2015). F 1
X ðcij Þ
is the value of quantile function for log-Laplace distribution at cij . After the reformulation, we could solve the following
mathematical model using a commercial solver such as IBM ILOG CPLEX.
0 1
XX X
min @ xt þ yt A  Cspli :maxð0; Demi  Capt Þ
ji i
t2T i2F j2U i
XX
1 qi þ c 2 di þ c 3 /i þ c4 v i Þ
þ ðcfuel þ cCO2 Þ  ðci;t t i;t t i;t t i;t t ð3:21Þ
i2F t2T
XX XX
þ yti  Dailyt þ sti  Idlet
i2F t2T i2F t2T
0 12
X
s:to @yti þ xtji A 6 qti  f i
t
8i 2 F; t 2 T ð3:22Þ
j2U i
0 14
X t 2
@yt þ xtji A 6 ðf i Þ  dti  1 8i 2 F; t 2 T ð3:23Þ
i
i
j2U
0 12
t 4
X
6 @yti þ tA t
ðf i Þ xji  /ti  f i 8i 2 F; t 2 T ð3:24Þ
j2U i
0 12
X
6 v ti
t 2
ðf i Þ  @yti þ xtji A 8i 2 F; t 2 T ð3:25Þ
j2U i
X t
dj  di  f i  TP ij P F 1
X ðcij Þ 8i 2 F; j 2 Pi ð3:26Þ
t2T

ð3:3Þ—ð3:6Þ; ð3:8Þ—ð3:15Þ
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 47

4. Heuristic methods

Although we could solve the small instances using the proposed integrated robust airline scheduling, aircraft fleeting and
routing model, we could face some numerical stability problems due to the large variability of the problem parameters.
Therefore, two heuristic methods are proposed to handle large size instances.

4.1. Discretized approximation and cruise speed control algorithm

To solve our problem faster than the integrated model, we propose discretized approximation model (hereafter DAM)
which excludes nonlinearity caused by the cost function of fuel consumption and CO2 emission. In this algorithm, the con-
trollable cruise times, which are defined by a continuous function in the integrated model, are approximated through a set of
discrete cruise time settings.
DAM is a mixed integer programming model which solves robust airline scheduling, aircraft fleeting and routing prob-
lems simultaneously. As distinct from the integrated model, in DAM, cruise times can only take a predetermined value from
an interval rather than any value from that interval. For example, while cruise time of a flight in the integrated model can
take any value between 85 and 115 min, in DAM the cruise time can be in increments of five minutes such as 85, 90, 95, 100,
105, 110 and 115. In this way, instead of a nonlinear cost function, we calculate the price of each cruise time option and use
linear terms to denote the fuel consumption and CO2 emission costs in the objective function. Consequently, DAM can be
solved faster than the integrated model.
In order to adapt these changes, we introduce the following additional parameters and a set of binary variables for each
flight and aircraft:

crstik : The kth cruise time option of flight i when it is performed by aircraft t
costtik : The cost of crstik which is equal to ðcfuel þ cCO2 Þ  Fðcrstik Þ
rtik : 1 if cruise time of flight i takes the kth setting for aircraft t and is 0 ow.

The formulation of DAM is as follows:

0 1
XX X
min @ xt þ yt A  Cspli  maxð0; Demi  Capt Þþ
ji i
t2T i2F j2U i
ð4:1Þ
XX X
p XX XX
r t
ik costtik þ yti  Dailyt þ sti  Idlet
i2F t2T k¼1 i2F t2T i2F t2T
0 1
X
p X
s: to rtik ¼ @yti þ xtji A 8i 2 F; t 2 T ð4:2Þ
k¼1 j2U i

X
p
rtik  crstik ¼ f ti 8i 2 F; t 2 T ð4:3Þ
k¼1

ð3:3Þ—ð3:6Þ; ð3:8Þ—ð3:15Þ; ð3:26Þ

In DAM, the objective function is the sum of spill cost, the costs of fuel consumption and CO2 emission, daily usage cost
and idle time cost. As different from the integrated model, the cost of fuel consumption and CO2 emission is represented by
P P Pp
k¼1 rik  cost ik . Constraints (4.2) and (4.3) ensure that when flight i is performed by aircraft t, the
t t
the linear term i2F t2T
cruise time of that flight takes one of the discrete cruise time settings. The remaining constraints are same as with the inte-
grated model.
We could solve the robust airline scheduling, aircraft fleeting and routing problem with DAM in a reasonable computation
time. However, since we have discretized cruise time in DAM, there is still a chance to improve that solution by considering a
continuous value of cruise time. Therefore, after fixing fleeting and routing with DAM, in order to decide continuous values of
cruise time, we solve the cruise speed control model (hereafter CSCM), which is suggested by Duran et al. (2015). CSCM is a
nonlinear second order cone programming model which solves robust airline scheduling problem by considering departure
timing, cruise time control and idle time insertion. Since the fleeting and routing decisions are made by DAM, CSCM only
deals with continuous decision variables. Therefore, even if it is a nonlinear model it can still be solved faster than the inte-
grated model.
In order to use the fleeting and routing generated by DAM as parameters of CSCM, we introduce the following additional
notation:

 ti : A binary parameter that is 1 if flight i is performed by aircraft t in the solution of DAM

w
A: Set of flights ði; jÞ such that ði; jÞ are two consecutive flights performed by the same aircraft in DAM
48 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

The formulation of CSCM is as follows:

XX

min w ti  ðcfuel þ cCO2 ÞFðf ti Þ þ sti :Idlet ð4:4Þ
i2F t2T
X X
s:to xj  xi  TAij  w ti  f ti  E½Ai   w ti  sti ¼ 0 8ði; jÞ 2 A ð4:5Þ
t2T t2T

l
 ti 6 f i 6 f i  w
fi  w
t u
 ti 8i 2 F; t 2 T ð4:6Þ
 ti
sti 6 M:w 8i 2 F; t 2 T ð4:7Þ
ð3:8Þ; ð3:12Þ; ð3:13Þ; ð3:26Þ

In the objective function of CSCM, the costs of idle time insertion, fuel consumption and CO2 emission are minimized.
Constraint (4.5) ensures the aircraft connection for the routing generated by DAM. If aircraft t performs flight i as a result
t
of DAM, constraint (4.6) keeps the cruise time of each flight between the upper and lower limits, otherwise f i value is set
to zero. Similarly, if aircraft t performs flight i as a result of DAM, then sti can be nonnegative, otherwise it is set to zero
by constraint (4.7).
In discretized approximation and cruise speed control algorithm (hereafter Heuristic 1), we solve two models sequen-
tially. In Fig. 3, we illustrate the flow chart of the algorithm. Initially we solve the DAM and fix the fleeting and routing deci-
sion, then we solve the CSCM in order to find the minimum sum of idle time, fuel consumption and CO2 emission costs, e.g.
corresponding cruise time, for a given fleeting and routing decision. In this way, instead of solving the integrated model,
which is a nonlinear MIP model; as a heuristic method we propose to solve first a MIP model, e.g., DAM, and then a nonlinear
model, CSCM, so that total solution time is quite smaller than the integrated model as discussed earlier.

4.2. Multi-stage triplet search algorithm

As another observation, when we remove the cost terms related to the continuous decision variables from the objective
function of the integrated model, we observe that solution time decreases drastically. This is due to two reasons. First, by
removing fuel consumption and CO2 emission costs, we exclude nonlinearity and get a MIP model as in the case of DAM.
Second, when we remove the cost term of the idle time insertion as well as the cost term related to fuel consumption, aircraft
fleeting and routing decisions are given more easily. To elaborate, without idle time, fuel consumption and CO2 emission
costs, the model only considers the tradeoff between daily usage and spill cost so it neglects the costs of idle time insertion
and cruise time change. Therefore, we introduce daily usage and spill costs model (hereafter DSCM) in which the idle time
insertion, fuel consumption and CO2 emission costs are removed. All constraints of the integrated model in Section 4.1 are
P P t t P P
t2T ðc fuel þ cCO2 Þ  F i ðf i Þ and t2T si  Idlet are extracted from the objective
t
valid in DSCM, however the cost terms i2F i2F
function.
Although we get a non-dominated solution regarding daily usage and spill costs by DSCM, the generated solutions might
have high cost of idle time and fuel consumption since DSCM takes idle time and fuel as free. In order to enhance a solution
generated by DSCM, the cost of fuel and idle time should be taken into account. Thus, we take the solution of DSCM and then
by fixing fleeting and routing of DSCM, we solve CSCM which is presented in the previous section.
Even if solving CSCM after DSCM can improve the cost of idle time and fuel, this improvement is local since CSCM is
solved within the fixed fleeting and routing of DSCM. By driving DSCM to generate fleeting and routing which are also favor-
able regarding the costs of idle time and fuel, the sequential solution of DSCM and CSCM can find better solutions. For this

Set offlights,F
Start DAM
Set of aircraft,T

Proposed w̄it ∀i ∈ F,t ∈ T

CSCM
Schedule Ā ⊆ (FxF )

End

Fig. 3. Discretized approximation and cruise speed control algorithm.

 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 49

purpose, initially we introduce triplet which is a constituent of a fleeting and routing in an airline schedule, then present a
search algorithm over triplets to improve the solutions generated by the sequential solution of DSCM and CSCM.

Definition 1. A triplet, (i, j, t) is a collection of two consecutive flights i and j and the aircraft t which performs them. A triplet
has the cruise times and idle time information regarding the flights in the triplet.
The cost of a triplet is calculated as follows: the fuel consumption and CO2 emission costs are calculated for two flights
and the idle time cost of the aircraft between these flights are summed and then the minimum required fuel consumption
and CO2 emission costs for these flights are extracted. The minimum required fuel consumption and CO2 emission costs are
calculated by assuming these flights are performed by the most efficient aircraft in the set of aircraft at optimum speed.
Therefore, the cost of each triplet can be thought as the improvement capacity on the cost of that triplet.
In DSCM, each decision variable xtij refers to a triplet so by fixing xtij ¼ 0 beforehand, we avoid DSCM to have triplet ði; j; tÞ
in the generated solution. In this manner, we present multi-stage triplet search algorithm in which we employ a variable-
fixing idea, and solve DSCM and CSCM models sequentially.

Algorithm 1: Two Stage Algorithm

Heuristic 2: Multi-Stage Triplet Search Algorithm

The multi-stage triplet search algorithm, as illustrated in Heuristic 2, takes a set of aircraft and a set of flights with given
beam size b and depth size d as an input. As an initialization step at the root node, we apply two stage algorithm, given in
Algorithm 1, without a triplet. Afterwards, the set of triplets, S, is generated from the root node solution. After initialization
step, a procedure which repeats itself for each depth starts. At the beginning of each depth, we have a set of triplets coming
from the previous step, S, note that for the first depth, set of triplets comes from the root node solution. From S, we choose
first b triplets with highest cost and then remove them from the solution set. For each chosen triplet, b times at total for a
depth, we apply two stage algorithm to get a solution and add the generated triplets to S for the next depth. Since we apply
this procedure d times, we get one solution from the root node and b solutions at each depth. In sum, we eventually report
the best solution among b  d þ 1 feasible solutions generated by the multi-stage triplet search algorithm.
50 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

5. Computational study

In this section, we compare the performances of three solution methods which are the integrated approach and two
heuristics, in terms of different airline cost components against the published schedule along with the required CPU times.
In this computational study, there are four experimental factors, and their corresponding levels are given in Table 5.
The first factor is the fuel cost, which is the price of jet fuel per ton. The fuel prices are taken as $1.8/gallon for the lower
setting and $3.6 gallon for the higher setting, or accordingly $600/ton and $1200/ton, respectively. Another important factor
is the opportunity cost of spilled passengers of flight i; Cspli . We again consider two different levels of base spill cost, denoted
as Factor B. The base spill cost is used as a multiplier to calculate the spill cost of a flight per passenger as discussed in a
recent study conducted by Duran et al. (2015). One of the main contributions of the integrated approach is to improve
the aircraft utilization by solving the aircraft scheduling, routing and fleeting problems simultaneously along with the cruise
speed control decisions. On the other hand, there is a significant tradeoff between the aircraft utilization and overall
operational expenses, such as fuel and carbon emission costs. Due to a possible increase in the utilization of a fuel efficient
aircraft, total fuel consumption cost could be decreased for the entire flight network in albeit of additional spill costs. Con-
sequently, we need to quantify the aircraft utilization along with the operational expenses in our objective function. We have
used the parameter Idlet for each aircraft type t as a surrogate parameter to evaluate a possible tradeoff between the
aforementioned aircraft utilization and operational expenses in our proposed mathematical model. Since it is difficult to
accurately estimate these parameters, we have taken the daily usage cost of each aircraft type, Dailyt , as one of the
experimental factors. The third experimental factor, m, is used for the calculation of the daily usage cost of an aircraft.
In our study, we consider the daily usage cost of an aircraft which refers to the sum of the fixed operating costs and the
lost opportunity cost when that aircraft is used as discussed in Section 2. We calculate daily usage cost related to the unit
idle time cost since it is rational that a valuable aircraft has both high unit idle time cost and daily usage cost. For each air-
craft type, we set its daily usage cost to low and high levels, respectively, to observe its impact on the overall results as
follows.

Dailyt ¼ m  Idlet  60 ð5:1Þ

The last factor is the minimum desired service level of the passengers‘ connections. As distinct from the first three factors,
this factor has three levels. While the first three factors have linear effects on the objective function, the effect of service level
on the objective function is not as straightforward as the other factors since it is not a direct cost term of the objective func-
tion. Therefore, in order to reveal its effect in the experimental design, we choose to adopt three levels on this factor.
In this study, we use the flight network, which is in the work of Aktürk et al. (2014). The published schedule of a major U.
S. airline is provided in Table 6. The flight information were taken from the Bureau of Transportation Statistics (BTS) database
(BTS, 2010). In the published schedule, each column represents the tail number, flight number, departure and arrival airport,
departure time, flight block time and arrival time of each flight respectively. From the published schedule, we take the set of
flights and set of aircraft as input in our model.
Among these 114 flights, we generate three different flight networks. The first one is 23 flight network; the second is 35
flight network and the last one is 114 flight network as given in Table 6. We conduct all computational study over these flight
networks. Furthermore, we have taken 3 replications for each factor combination, e.g. out of 2  2  2  3 combinations in a
full-factorial experimental design, resulting in 72 randomly generated runs for each flight network.

5.1. Analysis on the integrated model

We start the analysis on the results of our computational study by pointing out the significance of our integrated
approach. For this purpose, we compare the cost of the published schedule, which is generated by the sequential approach,
and the proposed schedule, which is the output of our proposed integrated robust airline scheduling, aircraft fleeting and
routing model with cruise speed control.
We first analyze the 23 flight network in terms of operational efficiency. In particular, this small network consists of 23
flights that are performed by seven aircraft. The tail numbers of these aircraft are N3DUAA, N5DXAA, N412AA, N531AA,
N4XGAA, N544AA and N3ELAA. The corresponding flights are available in Table 6. In this schedule, the average number

Table 5
Factor values.

Factors Description Levels

Low(0) Medium(1) High(2)
A Fuel cost $600 – $1200
B Base spill cost $15 – $60
C m 1 – 10
D Service level 75% 85% 95%
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 51

Table 6
Published schedule for 114 flight network.

Tail # Flight # From To Dep. time Dur. Tail # Flight # From To Dep. time Dur.
N530AA 398 ORD LGA 06:15 134 N3ETAA 1704 ORD EWR 06:35 125
319 LGA ORD 09:25 170 1883 EWR ORD 09:30 160
2329 ORD DFW 13:35 155 810 ORD DCA 13:10 105
2364 DFW ORD 17:00 150 2013 DCA ORD 15:45 135
N459AA 394 ORD LGA 06:50 135 2013 ORD LAS 19:00 250
321 LGA ORD 10:00 170 N3DYAA 1063 ORD LAX 08:50 275
366 ORD LGA 13:55 140 874 LAX ORD 14:30 255
347 LGA ORD 17:15 170 874 ORD BOS 19:45 135
N531AA 2303 ORD DFW 06:45 155 N3DRAA 1021 ORD LAS 08:30 245
2336 DFW ORD 10:10 140 1544 LAS ORD 13:25 215
1053 ORD AUS 13:25 170 1544 ORD DCA 18:00 105
336 AUS ORD 17:00 165 N5DXAA 1048 ORD MIA 07:35 190
336 ORD LGA 20:40 125 1763 MIA ORD 11:55 200
N4XGAA 2079 ORD SAN 08:45 270 1899 ORD MIA 16:20 185
1438 SAN ORD 14:00 250 N454AA 2441 ORD ATL 06:30 120
346 ORD LGA 19:50 135 1986 ATL ORD 09:15 135
N598AA 1341 ORD SFO 07:50 295 1872 ORD MCO 12:25 160
348 SFO ORD 13:30 265 1131 MCO ORD 15:50 185
1521 ORD TUS 19:15 235 N4YMAA 1137 ORD MSY 08:20 145
N439AA 2455 ORD PHX 07:10 240 1768 MSY ORD 11:30 150
358 PHX ORD 11:55 210 1768 ORD PHL 15:05 125
358 ORD LGA 16:25 145 1697 PHL ORD 18:00 155
371 LGA ORD 20:00 155 N467AA 1823 ORD PBI 09:20 175
N475AA 407 ORD STL 06:20 70 2067 PBI ORD 13:00 200
755 STL ORD 08:35 75 2067 ORD STL 17:15 70
755 ORD SAT 10:45 180 1186 STL ORD 19:10 80
408 SAT ORD 14:30 160 N536AA 2305 ORD DFW 07:45 160
408 ORD PHL 18:05 125 2344 DFW ORD 11:35 140
N3EEAA 876 ORD BOS 06:35 130 1201 ORD STL 14:50 65
413 BOS ORD 09:35 185 1815 STL ORD 17:00 80
413 ORD SNA 13:45 275 1815 ORD SLC 19:15 270
1262 SNA ORD 19:10 230 N420AA 1686 ORD RDU 06:50 110
N4YDAA 451 ORD SFO 09:45 295 2435 RDU ORD 09:25 135
554 SFO ORD 15:45 265 2435 ORD PHX 12:35 235
N3ERAA 496 ORD DCA 06:45 100 1206 PHX ORD 17:15 205
1715 DCA ORD 09:15 130 N546AA 1462 ORD EWR 08:00 140
1715 ORD LAS 12:25 255 1387 EWR ORD 11:25 160
1708 LAS ORD 17:20 220 1397 ORD MCO 15:00 160
N5CLAA 1425 ORD SNA 08:25 280 1221 MCO ORD 18:25 175
556 SNA ORD 14:00 240 N4WPAA 2311 ORD DFW 09:05 155
1940 ORD MIA 19:25 180 2348 DFW ORD 12:35 140
N535AA 2460 ORD RSW 06:45 165 1797 ORD STL 15:50 70
564 RSW ORD 10:20 185 1982 STL ORD 18:00 80
1446 ORD EWR 14:55 165 1339 ORD SAN 20:15 270
1411 EWR ORD 18:45 165 N5EBAA 2375 ORD EGE 08:10 175
N3DMAA 568 ORD FLL 07:25 175 2378 EGE ORD 12:25 165
711 FLL ORD 11:10 195 1677 ORD SNA 18:40 270
2021 ORD SJU 15:25 275 N3DUAA 2099 ORD LAX 07:00 270
N544AA 2463 ORD MCI 06:25 90 1972 LAX ORD 12:40 245
754 MCI ORD 08:40 90 1972 ORD RDU 17:45 115
2321 ORD DFW 11:15 155 N3ELAA 2057 ORD SJU 08:30 290
2356 DFW ORD 14:40 140 2078 SJU ORD 14:25 335
2487 ORD DEN 17:50 165 N3DTAA 2363 ORD HDN 09:50 170
N3EBAA 1565 ORD MSP 06:40 90 2318 HDN ORD 13:40 170
779 MSP ORD 09:00 85 N412AA 2345 ORD DFW 17:15 155
779 ORD SAN 11:35 260 2374 DFW ORD 20:40 130
1358 SAN ORD 16:45 235
1358 ORD BOS 21:50 125

of flights operated by a single aircraft is approximately 3. The average utilization of these aircraft is around 73%. Note that we
consider the utilization of an aircraft as the ratio of the total time spent in the air, i.e., total cruise time to the total available
cruise time for that aircraft denoted by kt in our problem formulation.
52 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

We next track the changes of the proposed model at these performance measures in order to interpret the impact of the
integrated approach as well as the cruise time flexibility. Table 7 illustrates the values for these measures of the proposed
schedule with respect to the different factor values. As the results over all instances row indicates that the integrated model
yields higher aircraft utilization while decreasing the total number of aircraft required to perform these flights. Moreover, we
provide the average cruise time change per flight because it is worth to note that when the daily usage cost factor is high (i.e,
C ¼ 2), the model achieves to decrease the total number of aircraft in all instances by controlling cruise time more
aggressively.
It is clear that the integration and cruise time change increase the efficiency of the flight schedule in comparison to
sequential approach however it is intricate to evaluate how valuable this increase because of the network effect and the non-
trivial cost function of the fuel consumption. As a next step, we provide how these changes correspond to cost improvement.
We calculate the cost improvement for total cost and other cost terms for each factor setting as follows:

Cost of Alternative  Cost of Our Model

Improvement ¼ ð5:2Þ
Cost of Alternative
As it is shown in Table 8, our integrated model provides around 17% average improvement on the total cost of the pub-
lished schedule over all factor levels. Moreover, from Table 8, the tradeoff between daily usage cost and fuel consumption
and CO2 emission costs can be deduced. When daily usage cost is high, the improvement over the fuel and CO2 emission costs
is limited around 4% at maximum and the cost improvement over daily usage cost is around 14% at minimum. That means in
order to decrease total number of aircraft used, the model compresses the cruise time of flights in albeit of additional fuel
consumption. Similarly, when the cost of fuel consumption and CO2 emission is high, the average improvement on the daily
usage cost decreases from 9% to 7% approximately. This is an important advantage of the proposed model, since we could
now evaluate a possible tradeoff between the number of aircraft needed to cover all flights and the corresponding overall
operational expenses.

5.1.1. Scenario analysis

The extensive computational study results intrigue several questions on model behavior and its performance. In this part,
some insights into model dynamics are provided by realizing different scenarios.

Table 7
Efficiency of the proposed model.

Avg. aircraft util. (%) Avg. # of flights per route Avg. cruise time change per # of aircraft used
flight (%)
Min. Avg. Max. Min. Avg. Max. Min. Avg. Max. Min. Avg. Max.
A 0 72 79 84 3.29 3.61 3.83 0.82 1 1.30 6 6.42 7
2 72 78 84 3.29 3.56 3.83 0.83 1 1.09 6 6.5 7
B 0 72 78 84 3.29 3.56 3.83 0.95 1.02 1.30 6 6.5 7
2 72 79 84 3.29 3.61 3.83 0.82 0.98 1.09 6 6.42 7
C 0 72 73 84 3.29 3.33 3.83 0.82 0.94 1.06 6 6.92 7
2 84 84 84 3.83 3.83 3.83 1.04 1.06 1.3 6 6 6
D 0 72 79 84 3.29 3.58 3.83 0.82 1.01 1.09 6 6.46 7
1 72 79 84 3.29 3.58 3.83 0.82 1.01 1.30 6 6.46 7
2 72 79 84 3.29 3.58 3.83 0.82 0.99 1.05 6 6.46 7
All ins. 72 79 84 3.29 3.58 3.83 0.82 1 1.30 6 6.46 7

Table 8
Cost improvement over the published schedule.

Fuel and CO2 (%) Idle time (%) Daily usage (%) Total (%)
Min. Avg. Max. Min. Avg. Max. Min. Avg. Max. Min. Avg. Max.
A 0 4.2 8.3 14.5 89.3 93.8 96 0 8.6 14.7 14.9 18 22.8
2 4.2 9.8 14.5 89.3 93.5 95.7 0 7.4 14.7 13.4 16 19.8
B 0 4.2 9.3 14.5 89.3 93.4 95.7 0 7.4 14.7 14.5 17.9 22.8
2 4.2 8.2 14.5 89.3 93.8 96 0 8.6 14.7 13.4 16 21.2
C 0 4.2 13.4 14.5 89.2 93.1 96 0 1.23 14.7 13.7 18.9 22.8
2 4.2 4.2 4.2 91.8 94 95.7 14.7 14.7 14.7 13.4 15.1 16.6
D 0 4.2 8.8 14.5 92.2 94.6 96 0 8 14.7 13.5 17.1 22.8
1 4.2 8.8 14.5 91.3 93.9 95.4 0 8 14.7 13.4 17 22.6
2 4.2 8.8 14.5 89.3 92.3 93.8 0 8 14.7 13.4 16.9 22.3
All ins. 4.2 8.8 14.5 89.3 93.6 96 0 8 14.7 13.4 17 22.8
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 53

What if cruise time change is not allowed?

We continue the analysis on the results of our computational study by indicating what would happen without cruise time
control. By this means, we present the analysis on the integrated model and compare its performance with other integrated
models, which do not consider cruise speed control. For a fixed cruise time setting, we also consider two models. Both models
are integrated ones such that airline scheduling, aircraft fleeting and routing decisions are made within the consideration of
passengers’ service level and maintenance requirements. The first model tries to minimize the sum of spill cost, fuel consump-
tion and CO2 emission cost, idle time cost and daily usage cost without cruise time control. The second one tries to minimize
the sum of spill cost, idle time cost and daily usage cost but considers neither fuel consumption nor CO2 emissions cost nor
cruise time control. Hereafter, Alternative1 refers to the model, which considers fuel and CO2 emission cost but not cruise time
control, and Alternative2 refers to the model, which considers neither fuel nor CO2 emissions cost nor cruise time control.
In Table 9, we present the total cost improvements provided by the cruise time control strategy over 23 flight network.
When all the results are considered, it is seen that cruise time control can achieve 8% and 10% average cost improvement
over Alternative1 and Alternative2, respectively. When we examine the effect of different factor levels, the cost improvement
of cruise time control is remarkably increasing depending on factor C. That means, when higher utilizations of aircraft are
more valuable, cruise time control is getting more crucial. The essence of the matter is that cruise time control can decrease
the total cost around 9% and this average improvement can increase up to 11% when the daily usage cost is higher.
What if the same service levels with the published schedule are used?
Another critical question that might arise is whether the cost improvements over published schedule are achieved by the
effect of service level changes in the proposed schedule or not. While comparing the proposed schedule and the published
schedule previously, we take the service level as a factor of the experimental design. However, in this part, we use the same
minimum service levels with the published schedule on each passenger connection in the proposed schedule.
In Table 10, it is seen that there is around 14% improvement on the average over the published schedule even though we
use the same service levels on passenger connections in the proposed schedule. This improvement is achieved by considering
cruise time change along with the fleeting and assignment decisions in the integration of these subproblems. Therefore, it
can be deduced that although the same service levels of the published schedule are used, it is possible to improve the total
cost by using this novel approach.

5.2. Analysis on heuristic methods

In this section, we analyze the proposed heuristic methods in terms of gap and improvement. Initially, the gap between
the integrated model and the heuristic methods are examined with 23 flight network. We calculate the gap between the
optimal solution of the integrated model and the solution of heuristic methods as follows:

Table 9
Total cost improvement of cruise speed control.

Imp. over Alt1 (%) Imp. over Alt2 (%)

Min. Avg. Max. Min. Avg. Max.
A 0 3.9 8.8 12.4 4.9 9.9 13.0
2 4.7 7.9 10.4 7.7 11.2 13.1
B 0 4.7 8.2 12.4 7.7 10.6 13.0
2 3.9 8.5 12.1 4.9 9.2 12.3
C 0 3.9 5.7 7.7 4.9 8.1 10.8
2 9.7 11.0 12.4 10.4 11.7 13.0
D 0 4.2 8.5 12.4 5.3 10.0 13.0
1 4.1 8.4 12.4 5.1 10.0 13.0
2 3.9 8.3 12.4 4.9 9.8 13.0
All ins. 3.9 8.4 12.4 4.9 9.9 13.0

Table 10
Cost improvement over the published schedule with same service levels.

Fuel and CO2 (%) Idle time (%) Daily usage (%) Total (%)
Min. Avg. Max. Min. Avg. Max. Min. Avg. Max. Min. Avg. Max.
A 0 4.1 5.8 14.5 52.7 60.6 62.2 0 12.3 14.7 10.2 14.2 16
2 4.1 9.3 14.5 52.8 57.5 62.2 0 7.4 14.7 9.7 13.3 15.4
B 0 4.1 7.6 14.5 52.8 59.1 62.2 0 9.8 14.7 13.3 14.7 16
2 4.1 7.5 14.5 52.7 59 62.2 0 9.8 14.7 9.7 12.8 14.5
C 0 4.2 11 14.5 52.7 55.9 62.2 0 4.9 14.7 9.7 13.8 16
2 4.1 4.2 4.2 62.2 62.2 62.2 14.7 14.7 14.7 12.1 13.7 15
All ins. 4.1 7.6 14.5 52.7 59 62.2 0 9.8 14.7 9.7 13.7 16
54 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

Table 11
Gap of heuristic methods over 23 flight network.

Heuristic 1 (%) Heuristic 2 (%)

Min. Avg. Max. Min. Avg. Max.
A 0 0.000 0.000 0.001 0.000 0.3 2.1
2 0.000 0.000 0.002 0.000 0.4 3.7
B 0 0.000 0.000 0.002 0.000 0.5 3.7
2 0.000 0.000 0.002 0.000 0.2 2.1
C 0 0.000 0.000 0.002 0.000 0.7 3.7
2 0.000 0.000 0.000 0.000 0.000 0.001
D 0 0.000 0.000 0.001 0.000 0.4 3.7
1 0.000 0.000 0.000 0.000 0.3 3.7
2 0.000 0.000 0.002 0.000 0.2 2.1
All ins. 0.000 0.000 0.002 0.000 0.3 3.7

Table 12
Improvement of heuristic methods over 35 flight network.

Heuristic 1 (%) Heuristic 2 (%)

Min. Avg. Max. Min. Avg. Max.
A 0 0.000 3.4 14.9 0.041 3.3 14.9
2 0.000 2.3 14.0 0.291 2.1 13.6
B 0 0.000 2.6 14.9 0.149 2.4 14.9
2 0.000 3.0 13.9 0.291 3 13.6
C 0 0.000 2.8 13.9 0.291 2.5 13.6
2 0.000 2.9 14.9 0.149 2.9 14.9
D 0 0.000 3.1 14.9 0.291 2.9 14.9
1 0.000 3.1 13.9 0.128 2.9 13.6
2 0.000 2.4 13.2 0.149 2.2 13.2
All ins. 0.000 2.9 14.9 0.291 2.8 14.9

Table 13
Gap between the heuristic methods and best lower bound over 35 flight network.

Heuristic 1 (%) Heuristic 2 (%)

Min. Avg. Max. Min. Avg. Max.
A 0 5.783 2.148 0.000 6.002 2.218 0.000
2 6.203 2.655 0.000 6.690 2.859 0.000
B 0 6.203 2.601 0.000 6.690 2.828 0.000
2 5.707 2.202 0.002 5.722 2.248 0.003
C 0 6.203 4.200 0.000 6.690 4.429 0.000
2 1.231 0.603 0.002 1.406 0.647 0.003
D 0 5.094 2.091 0.000 5.650 2.244 0.000
1 5.059 2.279 0.000 5.346 2.407 0.000
2 6.203 2.835 0.188 6.690 2.964 0.188
All ins. 6.203 2.470 0.002 6.690 2.611 0.003

Heuristic  Optimal
Gap ¼ ð5:3Þ
Optimal
As it can be seen from Table 11, the gaps of the heuristic methods are less than 1% on the average at all instances. It is
worthwhile to point out that Heuristic 1 has 0.002% maximum gap over all factor levels and all instances. Although, Heuristic
2 has maximum gap 3.7% over all instances, when factors C and B are at their high levels, it has the maximum gap of 0.001%
and 2.1%, respectively. This is due to the fact that Heuristic 2 fixes the routings and fleeting decisions by just considering
daily usage and spill cost. When these cost terms dominate the other cost terms on routings and fleeting decisions, ignoring
fuel and CO2 emission and idle time costs at routing and fleeting gets less significant. Therefore, Heuristic 2 causes smaller
gaps when factors B and C are at their high levels and factor A is at the low level.
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 55

Afterwards, we compare the performances of heuristic methods with the integrated model over 35 flight network to see a
possible impact of a larger flight network. Due to computational requirements of the integrated model, we put a time limit of
5400 s for the exact approach and report the best incumbent solution generated by the integrated model in a given time
limit. Consequently, we calculate the improvement of solutions of heuristic methods over the best solution as follows:
Cost of the Best Incumbent Solution  Cost of the Heuristic Method
Improvement ¼ ð5:4Þ
Cost of the Heuristic Method
As reported in Table 12, both heuristic methods now surpass the best incumbent solution of the integrated model on the
average due to the time limit of 5400 s. While Heuristic 1 achieves 2.9%, Heuristic 2 achieves 2.8% average improvement over
all instances. The negative improvement in Heuristic 2 means that the best solution generated in 5400 s by the integrated
model has smaller objective value than Heuristic 2 for certain instances.
In Table 13, we compare the cost of the heuristic methods and the cost of the best upper bound obtained in 5400 s. How-
ever, the optimal value of the integrated model is between the best upper bound and the best lower bound. Therefore, we
also provide a comparison of the performance of the proposed heuristic methods against the best lower bound obtained by
the commercial solver in 5400 s. The gap between the cost of the heuristic methods and the best lower bound value is found
by the following equation.
Best Lower Bound  Cost of the Heuristic Method
Gap ¼ ð5:5Þ
Cost of the Heuristic Method
It is worthwhile to note that the gap between the heuristic methods and the best lower bound is approximately 3% on the
average for both heuristic methods. Together with the analysis over the best incumbent solution, this observation implies
that the heuristic methods perform quite close to the integrated model over the 35 flight network as well as for the 23 flight
network.
Finally, we analyze the heuristic methods over the 114 flight network. As discussed earlier, the integrated model cannot
find a feasible solution in six hours so we compare the performance of the heuristic methods over the cost of published
schedule in Table 14. It can be seen that both heuristic methods provide around 15% average improvement over all instances
in comparison to the published schedule.
In the proposed model, for constraint (3.26), we take the minimum desired service levels for passenger connections, cij ,
equal to the levels of factor D, e.g. 75%, 85% and 95%, respectively. On the other hand, we could also calculate the probability
of successful passengers’ connection for each connecting passenger, denoted as bij , by replacing Constraint (3.7) with the fol-
lowing constraints:
" #
X t
Pr NC i 6 dj  di  fi  TPij 6 bij 8i 2 F; j 2 Pi
t2T

bij P cij 8i 2 F; j 2 Pi

Table 14
Improvement of heuristic methods over the published schedule in 114 flight network.

Heuristic 1 (%) Heuristic 2 (%)

Min. Avg. Max. Min. Avg. Max.
A 0 10.6 17.9 25.9 10.6 16.7 25.9
2 9.8 14.6 19.8 9.2 12.3 15.9
B 0 11.3 17.1 25.9 9.8 14.9 23.5
2 9.8 15.4 24.5 9.2 14.1 22.7
C 0 14.6 21.0 25.9 12.2 18.3 23.5
2 9.8 11.5 12.4 9.2 10.7 11.9
D 0 9.9 16.4 25.9 9.4 14.6 23.5
1 9.8 16.3 25.9 9.3 14.6 23.4
2 9.8 16.0 25.3 9.2 14.2 22.6
All ins. 9.8 16.2 25.9 9.2 14.5 23.5

Table 15
CPU time analysis.

Integrated model (in sec) Heuristic 1 (in sec) Heuristic 2 (in sec)
Min. Avg. Max. Min. Avg. Max. Min. Avg. Max.
23 FN 18 968 4048 0.8 1.1 1.8 10.5 12.4 15.3
35 FN 4243.4 5381.1 5400 3.2 5.7 7.8 31.6 40.2 54.7
114 FN – – – 724.8 5173.4 5400 1702.6 3254.3 5400
56 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57

Consequently, the achieved service levels of the proposed schedule are around 98%, 98% and 99%, respectively, on the
average. However in the published schedule, each connection has a different service level and some of them are actually
lower than even 75%, which is the lowest level of factor D in the computational study.
In summary, both heuristic methods have quite small optimality gaps over 23 flight network; while heuristic methods
surpass the best solution of the integrated model in 5400 s over 35 flight networks. Moreover, heuristic methods work quite
faster than the integrated model in terms of the CPU time. When the performance of the heuristic methods are compared, it
can be said that Heuristic 1 surpasses the Heuristic 2 in terms of optimality gap and improvement over the best solution and
published schedule over all flight networks. In Table 15, even for the CPU time, Heuristic 1 is more favorable than Heuristic 2
in 23 and 35 flight networks, however in 114 flight the CPU time of heuristic2 is 2000 s less than the CPU time of Heuristic 1
on the average, which could be attributed to the higher number of binary variables due to the discretization step in Heuristic
1. In other words, in 114 flight network even if the improvement of Heuristic 2 is 1.8% smaller than the improvement of
Heuristic 1, Heuristic 2 works 2000 s faster on the average.

6. Conclusion

For an integration of robust airline scheduling, aircraft fleeting and routing problems, this novel consideration of cruise
speed/time control enables us to make following contributions. First, aircraft utilization could be evaluated in terms of con-
sidering the entire flight network. We also show that total number of aircraft needed to cover a set of flights could be
decreased while ensuring equivalent service level and maintenance requirements. Therefore, there is a significant tradeoff
between aircraft utilization and overall operational expenses. Moreover, due to a possible increase in the utilization of a fuel
efficient aircraft, total cost of fuel consumption could be decreased for the entire flight network. Second, the robustness can
be achieved with smaller cost. To elaborate, on a route having a flight with a great delay probability would require idle time
insertion for the following flights to be performed on time; while removing the problematic flight from that sequence could
render that intervention unnecessary. Due to cruise speed control, our study has more options to make this type of changes
on routing decisions.
In addition to the contributions due to cruise speed control, we put forward the following contributions regarding the
solution of the problem. First, we propose a nonlinear mixed integer programming model and present its second order conic
reformulation. Furthermore, to solve large scale problems in a reasonable computation time, we propose two heuristic meth-
ods, which are discrete approximation and cruise speed control algorithm, and multi-stage triplet search algorithm. Finally,
we conduct a computational study in which we analyze these three proposed solution methods and compare their perfor-
mances among each other and against the published schedule.
Since we deal with the integration of three airline schedule planning problems, one immediate future research direction
can be integration of the last problem, crew assignment problem. Although there are different studies proposed to integrate
the crew assignment problem to one of these three stages; as far as we know, there is not a study which integrates these four
problems within the consideration of passengers’ service levels and cruise speed control.

Acknowledgments

The authors thank the editor and three anonymous referees for their constructive comments and suggestions that signif-
icantly improved this paper.

References

Aktürk, M.S., Atamtürk, A., Gürel, S., 2009. A strong conic quadratic reformulation for machine-job assignment with controllable processing times. Oper. Res.
Lett. 37 (3), 187–191.
Aktürk, M.S., Atamtürk, A., Gürel, S., 2014. Aircraft rescheduling with cruise speed control. Oper. Res. 62 (4), 829–845.
Aloulou, M.A., Haouari, M., Mansour, F.Z., 2013. A model for enhancing robustness of aircraft and passenger connections. Transport. Res. Part C: Emerg.
Technol. 32, 48–60.
Arikan, M., Deshpande, V., Sohoni, M., 2013. Building reliable air-travel infrastructure using empirical data and stochastic models of airline networks. Oper.
Res. 61 (1), 45–64.
Barnhart, C., Cohn, A., 2004. Airline schedule planning: accomplishments and opportunities. Manuf. Serv. Oper. Manage. 6 (1), 3–22.
Barnhart, C., Belobaba, P., Odoni, A.R., 2003. Applications of operations research in the air transport industry. Transport. Sci. 37 (4), 368–391.
Boeing, Jet Prices, 2015. <http://www.boeing.com/boeing/commercial/prices/> (Visited July 2015).
BTS, 2010. Airline On-Time Performance Data. <http://www.transtats.bts.gov/DL_SelectFields.asp?Table_ID=236&DB_Short_Name=On-Time> (Visited June
2015).
Duran, A.S., Gürel, S., Aktürk, M.S., 2015. Robust airline scheduling with controllable cruise times and chance constraints. IIE Trans. 47 (1), 64–83.
EUROCONTROL, 2012. User Manual for the Base of Aircraft Data (bada) Revision 3.10, Technical Report 12/04/10-45, EUROCONTROL Experimental Centre
Centre de Bois des Bordes B.P.15 F – 91222 Brtigny-sur-Orge CEDEX FRANCE.
Günlük, O., Linderoth, J., 2010. Perspective reformulations of mixed integer nonlinear programs with indicator variables. Math. Program. 124 (1–2), 183–
205.
Haouari, M., Shao, S., Sherali, H.D., 2013. A lifted compact formulation for the daily aircraft maintenance routing problem. Transport. Sci. 47 (4), 508–525.
Hiriart-Urruty, J.-B., Lemaréchal, C., 2001. Fundamentals of Convex Analysis. Springer.
IATA, 2010. Airline Fuel and Labor Cost Share. <http://www.iata.org/whatwedo/Documents/economics/Airline_Labour_Cost_Share_Feb2010.pdf> (Visited
June 2015).
Lohatepanont, M., Barnhart, C., 2004. Airline schedule planning: integrated models and algorithms for schedule design and fleet assignment. Transport. Sci.
38 (1), 19–32.
 H. Gürkan et al. / Transportation Research Part C 68 (2016) 38–57 57

Parmentier, A., 2013. Aircraft Routing: Complexity and Algorithms, Master’s thesis, Ecole des Ponts ParisTech.
Rebollo, J.J., Balakrishnan, H., 2014. Characterization and prediction of air traffic delays. Transport. Res. Part C: Emerg. Technol. 44, 231–241.
Ryerson, M.S., Hansen, M., 2013. Capturing the impact of fuel price on jet aircraft operating costs with leontieff technology and econometric models.
Transport. Res. Part C: Emerg. Technol. 33, 282–296.
Sherali, H.D., Bae, K.-H., Haouari, M., 2013a. A benders decomposition approach for an integrated airline schedule design and fleet assignment problem with
flight retiming, schedule balance, and demand recapture. Ann. Oper. Res. 210 (1), 213–244.
Sherali, H.D., Bae, K.-H., Haouari, M., 2013b. An integrated approach for airline flight selection and timing, fleet assignment, and aircraft routing. Transport.
Sci. 47 (4), 455–476.