Journal of Air Transport Management 73 (2018) 15–31

Contents lists available at ScienceDirect

Journal of Air Transport Management

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

Robust integrated maintenance aircraft routing and crew pairing T

a,∗ b c
Mohamed Ben Ahmed , Farah Zeghal Mansour , Mohamed Haouari
a
Faculty of Logistics, Molde University College, PO Box 2110, NO-6402 Molde, Norway
b
UR-OASIS, Ecole Nationale d’Ingénieurs de Tunis, Université de Tunis El Manar, 1002, Tunis, Tunisia
c
Department of Engineering Management & Systems Engineering, Old Dominion University, Norfolk, VA 23529, United States

A R T I C LE I N FO A B S T R A C T

Keywords: Given a daily flight schedule and a set of aircraft fleets, the integrated aircraft routing and crew scheduling
Aircraft routing problem requires finding a maintenance feasible set of aircraft routes and crew pairings such that each individual
Crew pairing flight is covered by exactly one aircraft route and one crew pairing. Although these problems are interdependent,
Robust airline planning they have been traditionally solved sequentially, where the aircraft routing problem, which is solved first, de-
fines a set of periodic aircraft rotations that impose some restrictions on short connections that are subsequently
accommodated by the crew pairing problem. A major drawback of this sequential approach is that it ignores
most of the interdependencies between the two problems. In particular, it fails to build robust solutions that are
resilient to unpredictable disruptions (like adverse weather, aircraft breakdowns, etc.) that translate into delayed
and canceled flights. In this paper, we propose an integrated robust model that incorporates the aircraft routing
and crew pairing problems within a single framework that aims at generating aircraft routes that are both robust
and cost-effective while accommodating technical constraints. A peculiar feature of the proposed model is that it
includes a polynomial number of variables and constraints. We solve the resulting integrated model by using a
general-purpose solver. Computational results obtained by using data from major airlines demonstrate the
benefits of the proposed robust model.

1. Introduction
On January 1, 1914, the world witnessed the first scheduled pas-
senger airline service linking St. Petersburg and Tampa, Florida.
Although, this line lasted three months only, it paved the way for to-
day's air transport prosperity. Since then, the use of commercial avia-
tion has grown more than seventy-fold, an achievement that is un-
matched by any other major form of transport. A disposal income and
living standards, reduced air travel costs, deregulation and globaliza-
tion enlarged the demand for air services, all of these and other factors
foreshadow a perennial appeal of the sector.
A proper appraisal of the economic importance of air industry re-
quires highlighting its contributions to the global business. Actually, in
2016, airlines worldwide carried upward 3.6 billions of passengers and
53 million tones of freight. It supported a total of 62.7 million jobs
globally, where 9.9 million of them are direct jobs. The total economic
impact of the worldwide aviation industry accounted for $ 2.7 trillion,
roughly 3.5% of world's gross domestic product (GDP).1 Actually, the
aviation industry contributes to the world GDP more than any other
sectors such as the pharmaceuticals, automotive and the textiles in-
dustries, and it is more than half the size of the global financial services
industry.
In the current context of volatile markets and shrinking prices,
airlines compete by using various Operations Research (OR) models
and algorithms to neatly manage their complex operations and pro-
cesses (Barnhart et al. (2003)). In this regard, typical airline processes
that intensively use OR techniques include:
• Schedule design (Warburg et al. (2008), Jiang and Barnhart (2009)
and Eltoukhy et al. (2017)): This planning process requires de-
termining which cities to fly to and at what times so as to generate a
schedule that offers the highest potential revenue. The generated
schedule constitutes the basis of the airline operations.
• Fleet assignment (Sherali et al. (2006) and Dožić and Kalić (2015)):
This planning process deals with assigning aircraft types, each
having a different capacity (number of seats), to the scheduled flight
legs, based on aircraft fleet sizes, operational costs, and expected
revenues.

∗
Corresponding author.
E-mail addresses: Mohamed.BenAhmed@himolde.no (M. Ben Ahmed), farah.zeghal@enit.rnu.tn, farah_zeghal@yahoo.fr (F. Zeghal Mansour),
mohamed.haouari@qu.edu.qa (M. Haouari).
1
http://www.atag.org/facts-figures.html.

https://doi.org/10.1016/j.jairtraman.2018.07.007
Received 15 March 2018; Received in revised form 21 July 2018; Accepted 31 July 2018
Available online 20 August 2018
0969-6997/ © 2018 Elsevier Ltd. All rights reserved.
M. Ben Ahmed et al.
• Aircraft routing (Lacasse-Guay et al. (2010) and Al-Thani et al.
(2016)): This planning process includes the determination of the
sequence of flight legs to be flown by each individual aircraft so as
to cover each flight exactly once while satisfying maintenance re-
quirements.
• Crew scheduling (Gopalakrishnan and Johnson (2005) and
Kasirzadeh et al. (2015)): This planning process involves the as-
signment of qualified crews to each flight leg while satisfying nu-
merous complex work rules.
These problems are typically solved sequentially, where the solution
of one problem serve as an input for the following. Clearly, this divide-
and-conquer strategy offers the significant advantage of considerably
reducing the computational burden, but at the cost of producing sub-
optimal solutions. Nevertheless, over the last decade, several authors
successfully addressed the foregoing problems in some integrated
fashion (Shao et al. (2017)). The basic premise behind such approaches
is to catch the interdependencies that exist between the various plan-
ning stages, and produce more cost-effective solutions.
A glaring fact in this context, is that optimized schedules are scar-
cely implemented as planned. The reason behind this paradoxical si-
tuation is that several random (uncontrollable) adverse incidents fre-
quently disrupt scheduled flights. These disruptions are usually
attributed to several causes such as aircraft breakdowns, air-traffic
congestion, crew shortages, aircraft arriving late, and inclement
weather conditions, etc., and translate into delayed and canceled
flights, and thereby additional financial burden and profit loss.
Actually, the number of disrupted flights has spiraled during the last
years along with the air industry expansion. In this context, it has been
reported that 18.58% of US airlines flights were delayed by more than
15 min in2016.2
The sizable expenses generated by flight delays are arguably
straightforward to quantify, consisting of (1) cost of flights' operations
(e.g. additional fuel and maintenance costs), (2) passengers' delay costs
(including passengers' accommodation and meals), (3) crews' overtime
payment, and (4) revenue losses that are incurred by, refleeting deci-
sions, cancelled routes as well as passengers that are reaccommodated
on different airlines (AhmadBeygi et al. (2008)). Cook and Tanner
(2015) estimated that a 1 min of delay may cost up to €1.75 per pas-
senger. It is noteworthy that these estimates downplay the cost of the
damages inflicted to airlines' reputation, which is translated into loss of
future business.
In response to these challenges, airlines are seeking to implement
novel tools and techniques for building robust schedules that are less
vulnerable to unpredictable disruptions when they occur (see Chtourou
and Haouari (2008), Ben Ahmed et al. (2017b) and Ladier and Alpan
(2016) among others). In this context, robustness can be achieved
through embedding a schedule with specific patterns, that allow it to
exhibit (i) resilience (or, flexibility), which is defined as the ability to
easily recover after a disruption, or (ii) stability which refers to the
ability to absorb or mitigate flight delays with limited impact on the
downstream flights. During the last, the issue of designing robust airline
schedules has been intensely investigated by many authors. We refer to
Muter et al. (2013), Cadarso and Marín (2013), Jamili (2016), Ben
Ahmed et al. (2017a) and Yan and Kung (2018), for the most recent
ones.
In this paper, we address the robust integrated aircraft routing and
crew pairing problem (RIARCP). This model achieves the integration of
two stages: it requires simultaneously determining periodic main-
tenance feasible aircraft routes and crew pairings. In contrast to pre-
vious integrated models, our model includes a polynomial number of
variables and constraints. Therefore, it can be directly solved using a
general-purpose solver without requiring the implementation of
2
https://transtats.bts.gov/OT_Delay/OT_DelayCause1.asp?pn=1.
16
Journal of Air Transport Management 73 (2018) 15–31
sophisticated branch-and-price or branch-and-cut algorithms.
Furthermore, a major goal of our model is to produce a cost-efficient
robust integrated solution. In this case, robustness is achieved through
embedding two effective modeling features: (i) connections having very
short buffer times are aggressively avoided both in aircraft routes and
crew schedules as well, and (ii) connections that are simultaneously
covered by aircraft routes and crew schedules are promoted. In so
doing, a delay that occurs on a given flight has a limited chance to
propagate to downstream flights and thereupon cause severe disrup-
tions. Therefore, our model aims at deriving schedules that are less
sensitive to reactionary delays (i.e. delays caused by late arrival of
aircraft or crew from previous flight). It is well-documented that reac-
tionary delays represent the major cause of delays3 that have been
plaguing airlines worldwide, and whose impact is more dramatic. As
shown in Table 1 below, because of the so-called snowball effect, a
reactionary delay can be, under specific circumstances, up to seven
times greater in magnitude than the source delay.
More precisely, we make the following contributions:
• We propose a polynomial-size mixed-integer nonlinear program-
ming model for the robust integrated aircraft routing and crew
pairing problem. To the best of our knowledge, this is the first
contribution in this regard. Our model builds upon the aircraft
routing model of Haouari et al. (2013) and the crew pairing model
of Zeghal Mansour et al. (2018). We apply the Reformulation-Lin-
earization Technique (RLT) of Sherali and Adams (1990, 1994) to
provide an equivalent lifted linear mixed-integer programming
formulation.
• We assess the empirical performance of the proposed model through
an extensive computational study that was carried on realistic in-
stances based on data provided by major airlines. In particular, we
provide evidence that instances with up to 336 flights, and 95 air-
craft can be optimally solved within reasonable CPU times.
• We evaluate the performance of the derived robust solution, using a
Monte-Carlo simulation study, and we show that our solutions
substantially outperform solutions produced by a non-robust ap-
proach.
The remainder of this paper is organized as follows. In Section 2, we
review the literature pertaining to integrated airline operations models.
In Section 3, we present a formal problem description along with the
associated underlying aircraft and crew graphs. In Section 4, we pro-
vide a detailed description of the proposed integrated model. This
model is first introduced as a nonlinear mixed-integer program. Next,
the Reformulation-Linearization Technique (RLT) is employed to derive
an equivalent linear model. In Section 5, we report the results of our
computational experiments that were carried out on realistic data from
major airlines. Finally, in Section 6, we provide some concluding re-
marks and outline directions for future research.
2. Literature review
Since the pioneering applications of Operations Research in the
airline industry in the early 1960s, it has been realized that solving the
integrated aircraft routing and crew pairing problem is a cumbersome
task that goes well beyond the capabilities of the optimization tech-
nology that was available at these times. However, prompted by the
joint development of sophisticated optimization techniques, speeded-up
computers, together with enhanced commercial optimization solvers,
several approaches have been proposed during the last 15 years in the
literature to address numerous variants of integrated aircraft routing
and crew pairing models. In this section, we shed light on the most
3
Understanding the Reporting of Causes of Flight Delays and Cancellations,
http://www.bts.gov/help/aviation/html/understanding.html.
M. Ben Ahmed et al.
Table 1
Basic reactionary multipliers by delay magnitude (Cook and Tanner (2015)).
Delay (min) 5 15 30 60 90 120 180 240
Basic multiplier 1.52 1.70 1.97 2.51 3.05 3.60 4.68 5.77
relevant contributions in this regard. We shall start by reviewing non-
robust integrated models, and next focus on robust ones.
2.1. Non-robust integrated models
A first attempt to partially integrate the schedule planning, aircraft
routing and crew pairing problem was proposed in Klabjan et al.
(2002), where, the crew pairing problem is solved while ensuring the
feasibility of aircraft routes through plane-count constraints. In this
case, maintenance checks are assumed to be performed overnight when
all aircraft are grounded. Later, Cohn and Barnhart (2003) proposed a
path-based formulation to integrate crew pairing and maintenance
routing problems. The proposed solution approach requires im-
plementing a branch-and-price algorithm to solve the integrated pro-
blem, which can be easily extended to take into account some restric-
tions on short connections. Yao et al. (2005) addressed an integrated
aircraft routing and crew pairing model while also considering flexible
departure times during peak demand days. The authors show that a
flexible departure time strategy has a positive impact on schedules
profitability and allows a more efficient aircraft utilization. Sandhu and
Klabjan (2007) investigated an integrated fleet assignment and crew
paring problem while (partially) accommodating aircraft routing con-
siderations. However, for the sake of computational tractability, aircraft
maintenance restrictions were ignored. Furthermore, Papadakos (2009)
successfully integrated three main planning stages: fleet assignment,
aircraft routing and crew pairing. This tour de force was achieved by
applying an enhanced Benders decomposition technique combined with
accelerated column generation. Recently, Shao et al. (2017) presented a
novel modeling approach for solving the integrated fleet assignment,
aircraft routing, and crew pairing problem. Additionally, their model
conservatively includes some pertinent operational considerations, such
as itinerary-based demands and other mandated aircraft and crew re-
quirements. They solved the integrated model using a Benders de-
composition approach along with several acceleration algorithmic ex-
pedients.
Furthermore, Salazar-González (2014) investigated an integrated
fleet assignment, aircraft routing, crew pairing problem, and crew
rostering problem. The author proposed a heuristic approach that re-
quires decomposing the genuine problem into subproblems that are
solved using a commercial solver. Recently, Cacchiani and Salazar-
González (2017) proposed two distinct MILP formulations to model the
integrated fleet assignment, aircraft routing and crew pairing problem.
The first model rely on path variables to describe aircraft routes and
crew pairings. Whereas, the second one uses arcflow variables to model
aircraft routes and path variables to describe crew pairings. Both
models aim at minimizing an aggregated cost function, and were solved
using exact algorithms. Computational experiments demonstrated that
arc-path formulation significantly outperforms the arc-arc formulation,
and provided substantial improvement on tractability and optimality.
2.2. Robust integrated models
In their pioneering paper, Cordeau et al. (2001) investigated an
integrated aircraft routing and crew scheduling problem where the
schedule robustness was achieved through forcing crews to follow air-
craft on short connections. In fact, crews that change form one aircraft
to another are likely to be delayed, particularly when there is not en-
ough connection time. For instance, if an inbound flight endured a
disruption, then both the operating aircraft and the aircraft to which
Journal of Air Transport Management 73 (2018) 15–31
the crew are assigned will arrive late. This model was extended by
Mercier et al. (2005), who penalizes aircraft swap along restricted
300 connections. Later, this model was further enhanced by Mercier and
Soumis (2007), where flight departure time can be suitably shifted
6.85 backward or forward within a 5-min gap. The problem was modeled
using a path formulation and solved using a hybrid solution strategy
that combines a Benders decomposition approach and a dynamic con-
straint generation procedure.
At this stage, it is noteworthy that numerous contributions have
drawn inspiration from the seminal work of Cordeau et al. (2001) and
successfully built airline schedules that promote crews following air-
craft. In particular, Weide et al. (2010) investigated the robust in-
tegrated aircraft routing and crew pairing problem. Starting from an
initial solution, they iteratively solve both aircraft routing and crew
pairing problems, to deliver an integrated aircraft and crew schedule.
The incumbent solution yielded a small number of crews changing
aircraft and incurred less costs than others implemented in practice.
Gao et al. (2009) addressed an integrated fleet assignment and crew
pairing model, where station purity is accommodated by restricting the
number of fleet types and crew bases allowed to serve each airport. A
clear advantage of imposing station purity is that restricting the number
of crews allowed to serve at each airport may promote finding solutions
that exhibit an increased number of move-up crews opportunities.
On the other hand, Dunbar et al. (2012) investigated an iterative
approach to accurately estimate propagated delays and to yield robust
plans for aircraft and crews. Dunbar et al. (2014) extended the former
approach by considering stochastic delay information. In this regard,
the authors developed two new algorithms that embed delay scenarios
within the aircraft routing and crew pairing problems, which are solved
in an integrated fashion. They additionally proposed an heuristic-based
approach that modifies flight departure times in order to minimize total
propagated delay costs. Also, Dück et al. (2012) investigated an in-
tegrated stochastic crew scheduling and aircraft assignment model that
involves delay propagation through aircraft routes as well as crew
schedules. This model by a heuristic column-based iterative approach.
Recently, Ruther et al. (2017) studied the operational variant of the
integrated maintenance routing and crew pairing problem. In practice,
routes and crew pairing are optimized over a short planning horizon
few days before operation. Besides, exact aircraft locations and the
maintenance and flying history of each individual aircraft are accom-
modated beforehand. The robustness of the resulting integrated model
is enhanced by promoting the assigning of crews and aircraft on the
same connections when the connection time is not long. The proposed
model is solved using a sophisticated branch-and-price model with
many pricing problems.
3. Problem description
3.1. Problem definition, terminology, and notation
In the sequel, an aircraft type refers to a certain model of aircraft. All
aircraft of the same type share the same cockpit configuration and
number of seats in each compartment. An aircraft family is comprised of
the union of all aircraft types that share the same cockpit configuration
and cockpit rating. For instance, the popular Airbus A320 family in-
cludes several aircraft types that consist of variants of A318, A319,
A320 and A321. A particular cockpit crew should be allocated to only
one aircraft family.
Each aircraft should undergo a preventive maintenance check be-
fore accumulating a specified number of flying hours since the last
maintenance check. This mandatory preventive maintenance requires
the aircraft to remain grounded at a maintenance station for a specified
time duration. A maintenance feasible aircraft route (or, rotation), is a
sequence of legs to be served by a single aircraft that satisfies the fol-
lowing restrictions: (i) the departure station of the first leg and the
arrival station of the last leg must correspond to the same maintenance
17
M. Ben Ahmed et al.
station; (ii) the time elapsed between the arrival time of the last leg and
the departure time of the first leg must exceed the time required for
performing a maintenance check; and (iii) the total flying time should
not exceed a specified time limit.
A duty period is a single workday of a crew that is comprised of a
sequence of flight legs with short rest periods, or sits, separating them.
It also includes briefing and debriefing periods at the beginning and end
of the period. A pairing is a sequence of duty periods with overnight
rests between consecutive periods. Each pairing begins and ends at the
same station (the crew base). Because of the restrictive work-rules
building optimal legal duties and pairings is a highly complex problem.
A feasible pairing should satisfy the following constraints: (i) Minimum
and maximum sit-time between consecutive flights. However, if both
flights are served by the same aircraft then the sit-time can be shorter
than the legal minimum and equal to the aircraft turn-time. We refer to
such a connection as a short connection; (ii) Maximum flying time be-
tween two consecutive rests; (iii) Maximum number of landings be-
tween two consecutive rests; (iv) Maximum time in a pairing away from
base; (v) Maximum number of duties in the pairing. Furthermore, each
crew should be granted a specified minimum rest time after a com-
pleting a pairing.
In this paper, we investigate the following integrated daily aircraft
routing and crew pairing problem. We are given a set L of daily flights to
be operated by a single aircraft family that consists of a set of K aircraft
types. We denote by NAk the number of aircraft of type k. We assume
that the fleeting decisions were made at a preceding stage, and we
denote by Lk the set of legs that will be served by an aircraft of type k.
For each flight j ∈ L , we denote by t j the corresponding flying time, DTj
and ATj the departure and arrival times, respectively, and DSj and ADj
the departure and arrival stations, respectively. It is noteworthy that all
time parameters are expressed in minutes and thereby are defined
within the time interval [0,1440) .
The Robust Integrated Aircraft Routing and Crew Pairing Problem
(RIARCP) requires finding a set of aircraft routes and crew pairings such
that the following constraints are accommodated:
• Each flight in L is covered by exactly one aircraft route and exactly
one crew pairing.
• Each aircraft route and each crew pairing should be periodic and
thereby repeats itself every day.
• Each aircraft route is maintenance feasible.
• The total number of required aircraft of each type k should not ex-
ceed the available size of the corresponding subfleet.
• Each pairing along with the corresponding duties should satisfy all
the aforementioned legal and structural constraints.
• The total number of required crews should not exceed the available
number of crews.
• Each short connection that is included in a pairing should be cov-
ered by an aircraft route, and vice-versa.
• The operational costs of crew pairings should not exceed a preset
value. By appropriately setting this latter value, this constraint en-
sures that the generated solution is cost-effective.
Prior to describing the objective function, we introduce the concept
of a critical connection. This concept refers to a feasible aircraft or crew
connection where the idle time is larger than the minimum legal time
but smaller than a specified threshold (say, 80 min. However, in prac-
tice, this parameter should be adjusted empirically). To clarify the re-
levance of critical connections, consider the following simple example
where a crew pairing covers a connection having a corresponding sit-
time of 45 min, while the minimum sit-time is 40 min. Assume that no
aircraft route was assigned to this connection. Therefore, at this con-
nection a crew must be transferred from an arriving aircraft to a dif-
ferent departing aircraft. In case where the arriving aircraft was subject
to a 40-min delay, then two subsequent flights might be immediately
delayed. In contrast, if the crew was following the aircraft then only one
18
Journal of Air Transport Management 73 (2018) 15–31
subsequent flight might be immediately delayed. Due to the fact that
each single flight delay may cause the delay of multiple downstream
flights, and thereby seriously damage the schedule adherence, it is
worth covering each critical connection that is included in a crew
pairing by an aircraft route. Furthermore, assume that the aircraft turn
time is 30 min and that an aircraft route includes a 35 min connection.
Although being feasible, it is easy to realize that this latter connection is
highly sensitive to minor delay. That is, if the aircraft is arriving late by
just 15 min, then this will cause a delay of the immediate subsequent
flight.
To promote the derivation of a robust schedule, we propose an
objective function that amounts to maximizing the sum of the rewards
minus the sum of penalties, where:
• A reward is granted to each connection (whether it is critical or not)
that is both covered by a crew pairing and an aircraft route. In so
doing, solutions where the crew is following the aircraft are pro-
moted.
• A penalty is incurred for each covered critical connection. In so
doing, both aircraft and pairing connections that exhibit short buffer
times are prevented. These buffer times aim at absorbing un-
predictable delays. However, if the crew follows the aircraft on a
critical connection, then the corresponding crew penalty is can-
celled.
Remark 1. Traditionally, airlines do not evaluate the cost of a crew
schedule in monetary ways. Rather, they apply a particular metric denoted
by the flight-time credit (FTC) (Schaefer et al. (2005)). The FTC of a
pairing is the difference between its total duration and the its total block time
(flying time). In what follows, we shall estimate the FTC of a pairing as the
sum of idle times (that is, sum of sit-times and layovers).
Remark 2. The model that is presented in this paper differs to a great extent
from the one addressed by Cacchiani and Salazar-González (2017), who
likewise provided a compact formulation to solve the integrated aircraft
routing and crew pairing problem. First and foremost, Cacchiani and
Salazar-González (2017) proposed an acyclic/non-periodic formulation
for the problem, assuming that aircraft locations at the beginning and the end
of the planning horizon are known in advance. Conversely, our model aims
at producing cyclic (steady state) schedules, that can be effortlessly
reproduced/duplicated over a long period. Besides, they assumed that
maintenance tasks are performed overnight, and imposed restrictions on
the number of aircraft that can undergo maintenance at once. In contrast,
our model accommodates planned maintenance checks that have to be
scheduled, either overnight or by daytime, before accumulating a maximum
number of flying hours. Finally, the authors considered building single-day
crew pairings. Therefore, several crews' work rules were ignored (namely,
restrictions on layover duration and total pairing duration). Whereas, our
model incorporates all of the aforementioned work rules, without requiring
the implementation of column generation/branch-and-price algorithm to
solve it.
3.2. Underlying graphs
The proposed integrated mixed-integer programming model is
based on two types of associated directed graphs: one for aircraft routes
and one for crew pairings. The former, was first described in Haouari
et al. (2013), and the latter was recently introduced by Zeghal Mansour
et al. (2018). For the sake of making the paper self-contained, both
graphs are described below.
3.2.1. Aircraft routing graphs
For each sub-fleet k, we define a digraph GkAR = (V k , Ak ) where each
node j ∈ V k corresponds to a daily flight leg that belongs to Lk . An arc
(i, j ) ∈ Ak if and only if the following two compatibility conditions hold:
(i) The arrival station ASi of flight i coincides with the departure station
DSj of flight j, and (ii) the total connection time is greater than or equal
M. Ben Ahmed et al.
to the minimum aircraft turn-around time. For the sake of computa-
tional convenience, arcs that exhibit a connection time larger than 24 h
are not included. The weight of the arc (i, j ) ∈ Ak is t j. Hereafter, we
denote by A ≡ ∪k = 1, … , K Ak the set of all aircraft connections.
For each sub-fleet k, we denote by Sk the set of maintenance stations
where an aircraft of type k can be serviced, Mk the maintenance check
k
duration, tmax the maximum flying time between two consecutive
maintenance checks (note that t j ≤ tmax , ∀ j ∈ Lk ) and τk the aircraft
turn-time.
The set of arcs Ak is given by the union of six arc subsets A1k , A2k , A3k ,
and A4k that are defined as follows:
• An arc (i, j) ∈ A k
1 if and only if a maintenance check could be
planned between the arrival of flight i and the departure of flight j,
and both flights are required to be served consecutively on the same
day. Hence, (i, j ) ∈ A1k ⇔ (1) ASi ∈ S; and (2) ATi + Mk ≤ DTj..
• An arc (i, j ) ∈ A2k if and only if a maintenance check could be
planned between the arrival of flight i and the departure of flight j,
and the same aircraft is required to serve flight leg j the day after
serving flight leg i. Hence, (i, j ) ∈ A2k ⇔ (1) ASi ∈ S; and
(2) DTj < ATi + Mk ≤ DTj + 1440.
• An arc (i, j ) ∈ A3k if and only if a maintenance check could not be
planned between the arrival of flight i and the departure of flight j,
and both flights are required to be served consecutively on the same
day. Hence, (i, j ) ∈ A3k ⇔ (1) ASi ∉ S or (2) DTj < ATi + Mk .
• An arc (i, j ) ∈ A4k if and only if a maintenance check could not be
planned between the arrival of flight i and the departure of flight j,
and the same aircraft is required to serve flight leg j the day after
serving flight leg i. Hence, (i, j ) ∈ A4k ⇔ (1) ASi ∉ S or
ATi + M > DTj + 1440; and (2) DTj < ATi + τk ≤ DTj + 1440.
We denote by AM k
≡ A1k ∪ A2k the set of maintenance arcs for sub-
fleet k, and ANM ≡ Ak \AM
k k
, the set of non-maintenance arcs.
Furthermore, we observe that arcs that belong to A2k ∪ A4k correspond
to wraparound ground connections that link pairs of flights that are
flown on two consecutive days. On the other hand, connections that
belong to A1k ∪ A3k correspond to pairs of consecutive flights that depart
on the same day. Let nak denote the number of wraparound flights in Lk
(that is, those flights whose arrival time occurs on the day after their
departure day). We make the following observations:
Claim 1. A cycle (j1 , j2 , …, jp , j1 ) in GkAR corresponds to a rotation of an
aircraft of type k that consecutively covers flight legs j1 , j2 , …, jp and returns
to j1 in a cyclic fashion.
Claim 2. A cycle (j1 , j2 , …, jp , j1 ) in GkAR that covers wg wraparound ground
connections and wf wraparound flights spans (wg + wf ) consecutive days.
3.2.2. Crew pairing graph
Let LD ⊂ L denote the set of flights that depart from the base station,
and LA ⊂ L the set of flights that arrive to the base station. Define sij as
the sit-time between consecutive flights i and j that are included in a
same duty period, and lij as the layover time between two consecutive
flights i and j that belong to two consecutive duty periods within a same
pairing. Furthermore, we provide the following additional notation.
• s / s : Minimum/Maximum sit-time between two consecutive
min max
flights within a same duty,
• : Maximum flying time within a duty,
φ max
• λ : Maximum duty duration. We assume that the longest flight
max
duration is shorter than the maximum duty duration (i.e., t j ≤ λmax ,
j ∈ V ),
• l min / l max : Minimum/Maximum layover duration between two con-
secutive duties within a same pairing, (note that l min is also the
minimum rest time after a completing a pairing),
• ηmin / ηmax : Minimum/Maximum duration of a pairing (i.e. time away
19
Journal of Air Transport Management 73 (2018) 15–31
from base),
•μ max
: Maximum number of landings within a duty,
•υ max
: Maximum number of duties within a pairing,
The crew pairing graph GCP = (V , B ) is defined as follows. The node
set V is derived by adding a dummy start node to V . That is, we have
V ≡ V ∪ {0}, where the dummy node 0 represents both the start and the
end of a pairing. Similar to the aircraft routing graph, an arc (i, j ) is
associated with each feasible connection between flights i and j that can
be served consecutively by the same crew in a pairing. Hence, an arc
(i, j ) ∈ B if and only if the following two compatibility conditions hold:
(i) The arrival station ASi of flight i coincides with the departure station
DSj of flight j, and (ii) the total connection time is greater than or equal
to the minimum sit-time and smaller than the maximum layover
duration. In addition, we define for each flight node j ∈ LD that departs
from the crew base an arc (0, j ), and or each flight node j ∈ LA that
arrives to the crew base an arc (j, 0). The correspond arc subsets will be
referred to as BD and BA, respectively.
The arcset B includes subsets B1 and B2 that are defined as follows:
1 An arc (i, j ) ∈ B1 if and only if legs i and j ∈ L can be consecutively
served by a same crew during the same duty period. That is, each arc
(i, j ) ∈ B1 corresponds to a short rest period within a duty. Hence,
(i, j ) ∈ B1 ⇔ (1) the sit-time sij is bounded by aircraft turn-time and
the maximum allowed sit-time (i.e., sij ∈ [mink = 1, K {τk }, smax ]); (2) the
maximum flying time with a duty is satisfied (i.e., ti + t j ≤ φmax ), and
(3) the maximum duty duration is satisfied (i.e., ti + t j + sij ≤ λ max ),
where the sit-time sij is computed as follows:
• sij = DTj − ATi , if DTj > AT,i in this case the arrival of flight i and the
departure of flight j occur on the same day,
• sij = DTj + 1440 − ATi , if DTj < ATi , in this case the arrival of flight i
and the departure of flight j occur on two consecutive days.
2 An arc (i, j ) ∈ B2 : An arc (i, j ) ∈ BL connects two consecutive flights i
and j ∈ L that can be performed consecutively by the same crew in
two consecutive duty periods within the same pairing. Therefore,
each arc (i, j ) ∈ B2 corresponds to a layover within a multi-day
pairing, where the layover time lij is computed as follows:
• lij = DTj − ATi , if ATi + l min ≤ DTj ≤ ATi + l max , in this case the ar-
rival of flight i and the departure of flight j occur on the same day.
• lij = DTj + 1440 − ATi , if ATi + l min ≤ DTj + 1440 ≤ ATi + l max , in
this case the arrival of flight i occurs one day before the departure of
flight j.
• lij = DTj + 2880 − ATi , if ATi + l min ≤ DTj + 2880 ≤ ATi + l max , in
this case the arrival of flight i occurs two days before the departure
of flight j.
Hence, (i, j ) ∈ B2 ⇔ (1) The layover duration is bounded by the
minimum and the maximum layover time (i.e., lij ∈ [l min , l max ]), and (2)
the maximum duration of a pairing is satisfied (i.e., ti + t j + lij ≤ ηmax ).
Furthermore, we define the set of short connections B S ⊂ B1 that
include all those connections that exhibit sit-times that are shorter than
minimum sit-time.
We observe that:
Claim 3. A cycle (0, j1 , j2 , …, jp , 0) in GCP corresponds to a pairing that
consecutively covers flight legs j1 , j2 , …, jp .
4. A compact mixed-integer programming model
4.1. Critical connections
For each arc (i, j ) ∈ Ak (k = 1, …, K ) we define the aircraft planned
idle time Iija as the difference between the departure time of flight j and
the aircraft ready time. That is,
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

DTj − ATi − τk , if DTj ≥ ATi + τk subject to

Iija = ⎧ ∀ (i, j ) ∈ Ak ,
⎨ DT + 1400 − ATi − τk , otherwise
⎩ j
∑ x ij = 1, ∀ k = 1, …, K , j ∈ Lk ,
k = 1, …, K . (1) i :(i, j ) ∈ Ak (6)

To promote the generation of robust aircraft routes, aircraft con- ∑ x ji = 1, ∀ k = 1, …, K , j ∈ Lk ,

nections having short buffer times are penalized. Toward this end, we i :(j, i) ∈ Ak (7)
define a quadratic penalty γija for each aircraft connection (i, j ) whose k
uj x ij = t j x ij , ∀ k = 1, …, K , j ∈ Lk , (i, j ) ∈ AM , (8)
idle time is shorter than a preset aircraft connection cushion time I a .
More precisely, γija is given by k
uj x ij = (ui + t j ) x ij , ∀ k = 1, …, K , j ∈ Lk , (i, j ) ∈ ANM , (9)
(I a − Iija)2 , if Iija < Ia
γija =⎧ ∀ (i, j ) ∈ Ak , k = 1, …, K . ∑ x ij ≤ NAk − nak , ∀ k = 1, …, K ,
⎨
⎩ 0, otherwise (2) (i, j ) ∈ A2k ∪ A 4k (10)
Similarly, we define for each arc (i, j ) ∈ B1 that corresponds to a
short connection whose sit-time sij ∈ [smin, smax ], the planned crew idle ∑ yij = 1, ∀ j ∈ L,
i :(i, j ) ∈ B (11)
time as follows

DTj − ATi − smin, if DTj ≥ ATi + smin ∑ yji = 1, ∀ j ∈ L,

Iijc = ⎧ ∀ (i, j ) ∈ B1. i :(j, i) ∈ B (12)
⎨
⎩ DTj + 1400 − ATi − smin, otherwise (3)

The set of critical connections is defined as follows

∑ y0j − ∑ yj0 = 0,
j ∈ LD j ∈ LA (13)
B C = {(i, j ) ∈ B1 \B S : Iijc < I c },
∑ (sij + lij ) yij ≤ (1 + ε ) FTC ,
where I c is a preset crew connection cushion time. i :(j, i) ∈ B (14)
In case where a critical connection is covered by a crew pairing μj yij = yij , ∀ (i, j ) ∈ B2 ∪ BD , (15)
without being simultaneously covered by an aircraft route (i.e, the crew
does not follow the plane), then a penalty γijc is incurred. This penalty is μj yij = (μi + 1) yij , ∀ (i, j ) ∈ B1, (16)
given by
φj yij = t j yij , ∀ (i, j ) ∈ B2 ∪ BD , (17)
γijc = (I c − Iijc )2 , ∀ (i, j ) ∈ B C , (4)
φj yij = (φj + t j ) yij , ∀ (i, j ) ∈ B1, (18)
4.2. A nonlinear MIP model λj yij = t j yij , ∀ (i, j ) ∈ B2 ∪ BD , (19)

4.2.1. Additional parameters λj yij = (λi + sij + t j ) yij , ∀ (i, j ) ∈ B1, (20)

ρ: reward (or, prize) that is collected if a connection is simulta- νj y0j = y0j , ∀ j ∈ LD , (21)
neously covered by an aircraft route and a crew pairing.
ε: tolerance (or, margin) on the solution's crew operational costs in νj yij = νi yij , ∀ (i, j ) ∈ B1, (22)
comparison with the baseline cost.
νj yij = (νi + 1) yij , ∀ (i, j ) ∈ B2, (23)
FTC: The schedule's baseline cost.
ηj y0j = t j y0j , j ∈ LD , (24)
4.2.2. Decision variables
ηj yij = (ηi + sij + t j ) yij , ∀ (i, j ) ∈ B1, (25)
x ij : binary variable that takes 1 if flights i ∈ L and j ∈ L are served
consecutively by the same aircraft, and 0 otherwise ηj yij = (ηi + lij + t j ) yij , ∀ (i, j ) ∈ B2, (26)
yij : binary variable that takes 1 if flights i ∈ L and j ∈ L are served
ηj ≥ ηmin yj0 , ∀ j ∈ LD , (27)
consecutively by the same crew, and 0 otherwise
z ij : binary variable that takes 1 if a crew follows an aircraft on a ηj yj0 + l min ≤ 1440dj , ∀ j ∈ LA , (28)
connection (i, j ) ∈ A ∩ B
uj : total accumulated flying hours for an aircraft since its last
∑ dj ≤ nbcrew ,
maintenance check after serving flight leg j ∈ L (j,0) ∈ BA (29)
μj : total number of landings after serving flight leg j ∈ L
φj : total accumulated duty flight duration after serving flight leg yij ≤ x ij , ∀ (i, j ) ∈ B S , (30)
j∈L
λj : total accumulated duty duration after serving flight leg j ∈ L 0 ≤ z ij ≤ x ij , ∀ (i, j ) ∈ A ∩ B, (31)
νj : total accumulated number of duties after serving flight leg j ∈ L
0≤ z ij ≤ yij , ∀ (i, j ) ∈ A ∩ B, (32)
ηj : total accumulated time away from base after serving flight leg
j∈L t j ≤ uj ≤ tmax , ∀ j ∈ L, (33)
dj : integer variable that corresponds to the duration (in days) of the
pairing that ends with flight j (if any), j ∈ LA . 1 ≤ μj ≤ μmax , ∀ j ∈ L, (34)

The problem is formulated as follows. t j ≤ φj ≤ φmax , ∀ j ∈ L, (35)

RIARCP: Maximize ∑ ρz ij − ∑ γija x ij − ∑ γijc (yij − z ij ) t j ≤ λj ≤ λ max , ∀ j ∈ L, (36)

(i, j ) ∈ B ∖ B S (i, j ) ∈ A (i, j ) ∈ BC

(5) 1 ≤ νj ≤ υmax , ∀ j ∈ L, (37)

20
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

t j ≤ ηj ≤ ηmax , ∀ j ∈ L, (38) 4.8. Short and critical connections

x ij ∈ {0,1}, ∀ (i, j ) ∈ A, (39) Constraints (30) enforces crew to follow aircraft on short connec-
tions. Finally, Constraints (31)–(32) enforce that if z ij = 1 then a crew
yij ∈ {0,1}, ∀ (i, j ) ∈ B, (40) follows the aircraft on the critical connection (i, j ) ∈ B C .

z ij ∈ {0,1}, ∀ (i, j ) ∈ A ∩ B, (41) 4.9. Model linearization

dj ≥ 0, integer, ∀ j ∈ LA . (42)
4.3. Objective function
The objective function (5) maximizes the total reward that is col-
lected from connections where crews are following aircraft, minus the
sum of the penalties that are incurred by including critical connections
in aircraft routes and penalties that are incurred by critical connections
that are covered by crew pairings without being covered by any aircraft
route.
Following Haouari et al. (2013) and Zeghal Mansour et al. (2018)
who proposed compact models for the aircraft routing and the crew
pairing problems, respectively, we enhance the solvability of the pro-
posed robust integrated model by applying the Reformulation Linear-
ization Technique of Sherali and Adams (1990, 1994) to derive a tight,
equivalent linear model representation of the problem.
Toward this end, we define new variables that shall be substituted
to the nonlinear terms. These variables are defined as follows:
αij = ui x ij , ∀ (i, j ) ∈ A, (43)
βij = μi yij ∀ (i, j ) ∈ B, (44)

δij = φi yij , ∀ (i, j ) ∈ B, (45)

4.4. Aircraft routes feasibility
ζij = λi yij , ∀ (i, j ) ∈ B, (46)
Constraints (6)–(10), together with (33) and (39), define maintenance
feasible aircraft routes. Indeed, for each k (k = 1, …, K ) Constraints (6)–(7) σij = νi yij , ∀ (i, j ) ∈ B, (47)
require that in each aircraft route, each flight leg has exactly one pre-
decessor and one successor, respectively. The nonlinear constraints (8)–(9) ϑij = ηi yij , ∀ (i, j ) ∈ B. (48)
together with (33) enforce that the total flying time restriction is satisfied.
To alleviate the paper presentation, the linearization process is de-
Note that the nature of these constraints precludes a cyclic rotation with
scribed in Appendix A.
no maintenance visit. Constraints (10) requires that the total number of
Upon linearization, the resulting equivalent model reads as follows.
aircraft in service should not exceed the fleet size.
RIARCP − RLT: Maximize ∑ ρz ij − ∑ γija x ij
(i, j ) ∈ B \ B S (i, j ) ∈ A
4.5. Duties feasibility
− ∑ γijc (yij − z ij )
(i, j ) ∈ BC (49)
Constraints (11)–(12) require that in each pairing, each flight has
exactly one predecessor and one successor (possibly, the dummy flight subject to:
0). Constraints (13) are flow conservation constraints that impose that
the number of start arcs is equal to the number of end arcs. Hence, ∑ x ij = 1, ∀ k ∈ K , j ∈ Lk ,
together with (40), these three restrictions enforce building cyclic crew i :(i, j ) ∈ Ak (50)
pairings. Constraints (15), (16) and (34) guarantee the restriction on
the maximum number of landings within a duty. Constraints (17), (18) ∑ x ji = 1, ∀ k ∈ K , j ∈ Lk ,
i :(i, j ) ∈ Ak (51)
and (35) enforce the restriction on the maximum flying time within a
duty. Constraints (19), (20) and (36) impose the restriction on the
∑ αji = t j + ∑ αij, ∀ k ∈ K , j ∈ Lk ,
maximum duty duration. i :(j, i) ∈ Ak k
i :(i, j ) ∈ ANM (52)
k
4.6. Pairings feasibility ti x ij ≤ αij ≤ (tmax − t j ) x ij , ∀ k ∈ K , (i, j ) ∈ ANM , (53)
k
ti x ij ≤ αij ≤ tmax x ij , ∀ k ∈ K , (i, j ) ∈ AM , (54)
Constraints (21)–(23) and (37) ensure the restriction on the max-
imum number of duties within a pairing. Similarly, Constraints
(24)–(26) and (38) enforce the restriction the maximum time away
∑ ⎛⎜1 − ∑ xij⎞⎟ ≤ NAk − nak , ∀ k ∈ K,
from base. Constraints (27) express the minimum duration of a pairing.
Finally, constraints 10 ensures that the expected FTC does not exceed
⎝ i :(i,j)∈A1k ∪A3k ⎠
j ∈ Lk (55)

the preset maximum value. ∑ yij = 1, ∀ j ∈ L,

i :(i, j ) ∈ B (56)
4.7. Number of available crews
∑ yji = 1, ∀ j ∈ L,
i :(j, i) ∈ B (57)
Constraints (28), (29) and (42) require that the total number of
crews in service should not exceed the available crews qualified for the ∑ y0j − ∑ yj0 = 0,
considered aircraft family. The validity of these latter constraints follow (0, j ) ∈ BD (j,0) ∈ BA (58)
from the fact that the if a pairing ends with flight j then the corre-
sponding duration is ηj yj0 . After adding the post-pairing (mandatory) ∑ sij yij + ∑ lij yij ≤ (1 + ε ) FTC ,
rest we get a total duration of (ηj yj0 + l min ) minutes which yields a total i :(i, j ) ∈ B1 i :(i, j ) ∈ B2 (59)
ηj yj0 + lmin
duration of dj = ⎡ 1440 ⎤ days. Since, the all flights are scheduled ∑ βji = 1 + ∑ βij , ∀ j ∈ L,
⎢ ⎥
daily, then this pairing requires dj crews. (j, i) ∈ B (i, j ) ∈ B1 (60)

21
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Table 2
Description of the data sets.
Data Set A318 A319 A320 A321 E190 E170 F70 B737 B738 B739 Total Turn Time (min) Sit Time (min)

I1 26 26 35 30
I2 42 42 35 30
I3 18 25 5 48 45 40
I4 2 35 9 46 45 40
I5 12 18 10 40 45 40
I6 12 12 35 30
I7 18 27 30 20 95 45 40
I8 17 26 12 60 45 40

∑ δji = t j + ∑ δij ∀ j ∈ L, dropped from Model 50–84, and the new objective function reads as
(j, i) ∈ B (i, j ) ∈ B1 (61) follows:

∑ ζ ji = ∑ sij yij + t j + ∑ ζij, ∀ j ∈ L, NR: Minimize ∑ sij yij + ∑ lij yij

(i, j ) ∈ B (i, j ) ∈ B1 (i, j ) ∈ B1 (62) i :(i, j ) ∈ B1 i :(i, j ) ∈ B2 (85)

∑ σji = ∑ yij + ∑ σij, ∀ j ∈ L,

(j, i) ∈ B (i, j ) ∈ BD ∪ B2 (i, j ) ∈ B1∪ B2 (63) 5. Computational experiments

∑ ϑji = ∑ sij yij + ∑ lij yij + t j + ∑ ϑij , ∀ j ∈ L,

In this section, we present computational experiments that were
(j, i) ∈ B (i, j ) ∈ B1 (i, j ) ∈ B2 (i, j ) ∈ B1∪ B2
carried out on instances based on publicly available data, which per-
(64) tains to the following major international airlines: KLM, Air France,
ϑj0 + l min ≤ 1440dj , ∀ j ∈ LA , (65) Alitalia, Finnair and Qatar Airways. All runs where made on an Intel i5
dual core 3.1 GHz processor computer having 8 GB of RAM, where the
∑ dj ≤ nbcrew , models where implemented in OPL and solved using CPLEX 12.2 with
j ∈ LA (66) default settings. To reduce the computation time, the relative optim-
ality gap was set to 2%.
yij ≤ x ij , ∀ (i, j ) ∈ B S ∩ A, (67)

0≤ z ij ≤ x ij , ∀ (i, j ) ∈ B C ∩ A, (68) 5.1. Problem instances

0≤ z ij ≤ yij , ∀ (i, j ) ∈ B C ∩ A, (69) The inputs for the RIARCP come from daily fleet assignment solu-
tions obtained from international airline companies. The data includes,
yij ≤ βij ≤ (μmax − 1) yij , ∀ (i, j ) ∈ B1, (70) for each aircraft family, the number of available aircraft, a daily flight
yij ≤ βij ≤ μmax yij , ∀ (i, j ) ∈ B2 ∪ BA, schedule with the flight numbers, departure and arrival times, origin
(71)
and destination airports and the operating aircraft type. Moreover, they
ti yij ≤ δij ≤ (φmax − t j ) yij , ∀ (i, j ) ∈ B1, (72) also specify the respective turn times, to be imposed between two
consecutive flights. The data sets comprise eight sets of instances,
ti yij ≤ δij ≤ φmax yij , ∀ (i, j ) ∈ B2 ∪ BA. (73) hereafter referred to by Ii , i = 1..8. They concern midsized- and narrow-
body aircraft such as Airbus 320 and Boeing 730 families. Table 2
ti yij ≤ ζij ≤ (λ max − sij − t j ) yij , ∀ (i, j ) ∈ B1, (74) displays for each data set the number of aircraft types, as well as the
number of available aircraft.
ti yij ≤ ζij ≤ λ max yij , ∀ (i, j ) ∈ B2 ∪ BA, (75) Each of the first five sets of instances comprises seven instances.
yij ≤ σij ≤ (νmax − 1) yij , ∀ (i, j ) ∈ B2, They differ by few flights that are scheduled only several days a week.
(76)
The last three instances sets comprise one instance that repeats daily. In
yij ≤ σij ≤ νmax yij , ∀ (i, j ) ∈ B1 ∪ BA, (77) the sequel, we denote each instance by Iij , i = 1..8 and j = 1..7 .
Furthermore, since the actual number of crews was not available, we
ti yij ≤ ϑij ≤ (ηmax − sij − t j ) yij , ∀ (i, j ) ∈ B1, (78) estimated this parameter by inflating the lower bound that is derived by
solving the linear programming relaxation of the non-robust model
ti yij ≤ ϑij ≤ (ηmax − lij − t j ) yij , ∀ (i, j ) ∈ B2, (79) with the objective of minimizing the total number of crews.
The main characteristics of the different instances are described in
max(ηmin , t j ) yj0 ≤ ϑj0 ≤ ηmax yj0 , ∀ (j, 0) ∈ BA, (80)
Table 3, where we indicate for each instance set the number of flights, the
x ij ∈ {0,1}, ∀ (i, j ) ∈ A, (81) average of flight durations and their standard deviation, respectively.
In addition, we assumed that each aircraft must undergo at least one
yij ∈ {0,1}, ∀ (i, j ) ∈ B, (82) maintenance check before accumulating 100 h of flying time. The
duration of a maintenance check was set to 8 h.
z ij ∈ {0,1}, ∀ (i, j ) ∈ B ∩ A, (83) Regarding duty and pairing parameters, we relied on the standard
dj ≥ 0, integer , ∀ j ∈ LA . rules mentioned in the literature and set the corresponding parameters
(84)
to the following values: the maximum number of landings within a
duty, μmax = 4, the maximum flying time within a duty, φmax = 8 hours,
4.10. The non-robust variant of the IARCP problem the maximum duty duration, λmax = 12 hours, the maximum sit time
between two consecutive flights within a same duty, smax = 4 hours, the
In a preprocessing step, we start first by solving a non-robust version maximum number of duties within a pairing, υmax = 3, the maximum
of our integrated model with the objective of computing the optimal time away from base, ηmax = 96 hours, the minimum duration of a
crew operational costs. Toward this end, Constraints 59 and 68–69 are pairing, ηmin= 5 h, the minimum layover duration l min = 9 hours and

22
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Table 3
Characteristics of the data instances.
Class Instance 1 2 3 4 5 6 7

I1 No. of flights 62 72 76 80 78 78 86
Average duration (min) 72.50 74.44 74.14 75.88 75.06 75.58 76.86
Standard deviation (min) 13.95 14.60 13.66 14.03 14.18 14.17 15.91
I2 No. of flights 130 154 154 156 156 142 146
Average duration (min) 92.08 91.98 92.63 92.56 93.24 93.49 93.53
Standard deviation (min) 17.18 16.96 17.13 17.60 18.21 18.40 18.15
I3 No. of flights 210 212 212 212 214 192 216
Average duration (min) 114.71 114.53 114.53 114.53 114.16 116.93 113.98
Standard deviation (min) 31.56 31.41 31.41 31.41 31.41 32.45 31.26
I4 No. of flights 140 142 118 148 150 146 154
Average duration (min) 162.97 162.07 158.08 158.04 157.16 159.40 155.37
Standard deviation (min) 99.22 98.69 98.23 98.25 97.74 98.39 96.78
I5 No. of flights 82 88 80 92 89 80 80
Average duration (min) 136.84 133.66 131.95 146.64 149.14 156.20 154.44
Standard deviation (min) 43.69 46.32 50.53 38.44 38.99 33.25 33.81
I6 No. of flights 60
Average duration (min) 86.58
Standard deviation (min) 16.79
I7 No. of flights 336
Average duration (min) 115.42
Standard deviation (min) 36.67
I8 No. of flights 136
Average duration (min) 126.25
Standard deviation (min) 32.29

the maximum layover duration, l max = 33 hours. Finally, the tolerance • PM : The number of all crews that follows aircraft.
3
on the solutions crew operational costs was set to ε = 5%. • PM : The number of critical aircraft connections.
4

5.2. Computational results We run both NR and RIARCP-RLT models on a testbed of 38 instances.
A summary of the computational results is provided in Table 4, where we
To analyze the performance of our approach, we provide the fol- provide, for each instance, the required CPU time is sec., the solutions' cost
lowing measures: (FTC ) along with the different aforementioned performance metrics.
Table 4 shows that the RIARCP problem can be solved within rea-
• PM : The number of pairing connections that are shorter than
30
c sonable CPU times, even for large instances. It also shows that this
30 min. problem is computationally more tractable than the NR variant.
• PM : The number of pairing connections that are shorter than
c
40
In Table 5, we provide for each class of instances, and for all per-
40 min. formance metrics, but PM2 and PM3, the cumulative percentage devia-
• PM : The number of pairing connections that are shorter than
50
c tion 100 ×
PM (NR) − PM (RIARCP )
PM (NR)
, where PM (NR) and PM (RIARCP ) are
50 min.
• PM : The number of aircraft connections that are shorter than
30 computed over all of the instances of each class, respectively.
a
Looking at Table 5, a striking observation is that, most of measures
30 min.
• PM : The number of aircraft connections that are shorter than
40
a
of performances were substantially improved. A greater achievement
can be noticed in metrics related to crew pairings. Indeed, the number
40 min.
of pairing connections that are shorter than 30 min (PMc30 ) was im-
• PM : The number of aircraft connections that are shorter than
50
a
proved by 86.67–100%, the number of pairing connections that are
50 min.
shorter than 40 min (PMc40 ) was reduced by 42.86–100%, and the
• PM : The number of critical pairing connections.
1
number of pairing connections that are shorter than 50 min (PMc50 ) was
• PM : The number of crews that follows aircraft on critical pairing
2
reduced by 26.67–100%.
connections.
With regard to aircraft connections, the number of critical aircraft

Table 4a
Summary of results for NR and RIARCP-RLT on class I1 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 4.43 38680 1 9 12 22 0 1 20 14 14 14
RIARCP 3.67 40210 0 3 4 9 9 43 18 5 5 5
2 NR 9.31 43710 6 13 17 24 0 6 29 20 20 23
RIARCP 6.52 45860 0 4 5 12 12 46 22 7 7 8
3 NR 9.09 47840 4 12 20 28 0 4 38 21 22 23
RIARCP 7.85 50230 0 6 7 11 11 45 21 9 9 10
4 NR 14.11 49380 5 13 18 26 0 5 38 24 25 27
RIARCP 7.19 51650 0 9 10 18 18 49 32 11 11 12
5 NR 9.66 49055 1 10 16 30 0 1 40 27 31 31
RIARCP 12.69 51390 0 7 9 16 16 50 29 12 13 13
6 NR 11.43 48485 2 10 16 26 0 2 34 25 26 28
RIARCP 9.07 50870 0 4 5 12 12 47 25 8 8 8
7 NR 21.06 55700 3 10 14 26 0 3 46 28 31 33
RIARCP 13.23 58400 0 11 13 26 26 56 40 17 18 18

23
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Table 4b
Summary of results for NR and RIARCP-RLT on class I2 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 94.78 80255 19 31 39 26 0 19 62 40 42 46
RIARCP 47.91 84120 0 10 17 35 35 87 47 20 21 23
2 NR 194.12 94295 21 33 45 34 0 21 95 65 70 74
RIARCP 427.54 99005 0 20 29 49 49 86 80 41 44 44
3 NR 817.84 94325 18 34 41 38 0 18 89 60 63 67
RIARCP 100.48 98720 0 20 29 49 49 92 79 40 42 44
4 NR 3042.37 97015 19 34 46 36 0 19 98 62 70 73
RIARCP 243.04 101380 0 20 30 53 53 102 87 43 46 46
5 NR 313.64 94750 23 37 45 37 0 23 96 60 68 73
RIARCP 163.55 98800 0 19 29 55 54 92 89 42 46 46
6 NR 694.00 86515 11 16 29 38 0 11 83 47 58 63
RIARCP 131.67 90475 0 8 18 39 39 93 65 31 32 34
7 NR 905.76 89750 21 34 47 33 0 21 83 54 62 66
RIARCP 122.08 94085 0 19 30 39 39 83 66 41 41 41

Table 4c
Summary of results for NR and RIARCP-RLT on class I3 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 1410.57 119230 29 30 39 37 0 59 140 106 114 123

RIARCP 969.16 125160 0 0 17 66 66 90 130 74 80 92
2 NR 1092.01 119960 31 35 48 32 0 66 143 105 114 117
RIARCP 555.96 125140 1 1 21 60 60 98 136 76 80 86
3 NR 854.77 120090 30 36 45 34 0 66 145 105 111 122
RIARCP 928.06 126090 0 0 19 70 70 101 130 72 77 86
4 NR 1180.59 120340 27 30 44 36 0 57 144 104 110 116
RIARCP 225.81 125495 0 0 22 68 68 101 134 80 83 91
5 NR 2837.12 122020 25 31 46 38 0 56 143 105 111 117
RIARCP 396.52 127860 0 0 23 71 70 98 134 77 86 89
6 NR 1093.96 112160 26 33 43 30 0 59 112 85 89 92
RIARCP 167.76 117725 0 0 20 50 50 88 95 65 69 72
7 NR 1663.40 121780 30 32 45 32 0 62 147 96 104 112
RIARCP 755.41 127865 0 0 21 69 69 96 136 80 83 92

Table 4d
Summary of results for NR and RIARCP-RLT on class I4 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 93.35 81309 9 13 18 28 0 22 75 46 56 62
RIARCP 44.85 84875 0 0 0 15 15 73 40 12 17 22
2 NR 121.72 82515 14 20 23 21 0 24 81 45 58 64
RIARCP 75.85 86120 0 0 0 16 16 70 43 17 22 28
3 NR 134.81 84304 16 20 24 24 0 26 82 42 51 57
RIARCP 281.09 88490 0 0 0 18 18 72 48 16 19 23
4 NR 125.03 82519 18 20 22 22 0 28 88 48 60 65
RIARCP 125.47 86525 0 0 0 19 19 73 52 16 22 27
5 NR 131.41 84584 12 18 21 26 0 30 79 42 54 60
RIARCP 231.77 88794 0 0 0 20 20 77 48 16 18 23
6 NR 108.86 83214 11 12 15 31 0 23 77 47 59 64
RIARCP 277.40 87145 0 0 0 18 18 75 46 15 18 22
7 NR 150.14 88654 17 21 24 22 0 38 90 50 63 69
RIARCP 184.04 93085 0 0 0 16 16 75 51 14 18 24

connections (PM4 ) was tremendously decreased, by the same token, the
remaining metrics have significantly decreased in magnitude. In fact,
the number of critical aircraft connections (PM4 ) was reduced by
8.11–42.66%, PMc30 was reduced by 25.87–66.88%, PMc40 was reduced
by 25.90–66.58%, while PMc50 was decreased by 23.90–61.68%.
In Table 6, we provide, for each class of instance Ii , i = 1..8, the total
number of crews to follow aircraft (PM3 ), the ratio
∑I ⊂ I PM3
R1 = 100 ×
ij i
, the total number of crews to follow aircraft on
total connections
∑I ⊂ I PM2
ij i
critical pairing connections (PM2 ), and the ratio R2 = 100 × total criticals .
Table 6 demonstrates that the RIARCP model efficiently promotes
solutions where crews following aircraft. Indeed, we observe that while
for the non-robust solution the percentage of connections where crews
follow the aircraft typically ranges between 4 and 28%, this percentage
varies between 45.78 and 71.67% for the robust solution. In other
words, the generated robust solutions exhibit a significant number of
connections are covered by crews following the aircraft. Even more
remarkably, we see from the last column of Table 6 that for the vast
majority of critical connections, crews were forced to stay in the robust
solutions on the same aircraft. This is in sharp contrast with the non-
robust model where no critical connection was simultaneously covered
by a crew pairing and an aircraft route.
Furthermore, we see from Table 4, that the FTC gap between the NR
and the RIARCP problems is marginal (strictly less than 5%). Hence, the

24
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Table 4e
Summary of results for NR and RIARCP-RLT on class I5 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 6.77 60200 3 7 14 27 0 10 32 23 26 27
RIARCP 5.11 62919 0 2 7 20 17 52 26 14 16 17
2 NR 13.45 62414 1 4 13 34 0 5 35 21 24 27
RIARCP 6.23 64218 0 0 3 21 21 58 28 12 14 16
3 NR 6.61 55549 2 5 11 24 0 7 32 21 23 25
RIARCP 4.18 58019 0 1 4 21 18 48 21 13 15 17
4 NR 11.98 65489 2 5 11 33 0 7 41 26 28 29
RIARCP 6.88 67053 1 1 7 26 26 60 34 18 21 24
5 NR 8.94 58958 3 6 10 22 0 9 33 20 25 27
RIARCP 6.73 61774 0 1 5 23 23 54 30 13 15 18
6 NR 7.74 56459 2 4 10 27 0 6 30 19 23 25
RIARCP 5.36 57349 0 2 7 22 19 54 22 11 13 15
7 NR 5.09 59250 2 6 13 26 0 8 32 22 26 27
RIARCP 5.85 59615 1 1 6 20 16 48 22 9 10 12

Table 4f
Summary of results for NR and RIARCP-RLT on class I6 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 5.52 34865 8 12 15 7 0 8 32 27 28 29
RIARCP 3.93 36585 0 3 11 20 20 43 25 14 15 17

Table 4g
Summary of results for NR and RIARCP-RLT on class I7 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 6377.01 194345 65 76 79 24 0 65 197 124 137 144

RIARCP 2112.82 204060 0 0 4 64 64 163 150 74 83 88

Table 4h
Summary of results for NR and RIARCP-RLT on class I8 –based instances.
Instance CPU (in s ) FTC c c c PM1 PM2 PM3 PM4 a a a
PM30 PM40 PM50 PM30 PM40 PM50

1 NR 57.01 79580 14 20 23 22 0 34 61 36 42 49
RIARCP 41.12 83390 0 0 2 25 25 70 40 19 21 25

proposed solutions are both robust and cost-effective. of delay ζ j assigned to each delayed flight j is computed as follows:
ζ j = (ζmax − ζmin ) r j + ζmin
5.3. Monte-Carlo simulation study
where ζmin and ζmax are the minimum and maximum primary delays,
To further demonstrate the robustness of the solutions derived by
RIARCP-RLT models, we carried out a Monte-Carlo simulation proce-
Table 6
dure that was proposed by Aloulou et al. (2013). The proposed proce- Comparison between models NR and RIARCP-RLT with respect to the number
dure is based on a 3-step methodology that can be described as follows: of crews following aircraft.
Total connections PM3 R1 (%) PM1 PM2 R2 (%)
Step 1: Random generation of primary delays. Primary delays were
arbitrary generated following a beta distribution Be (α, β ) . The amount I1 NR 532 22 4.14 182 0 0.00
RIARCP 532 336 63.16 104 104 100.00
Table 5 I2 NR 1038 111 10.69 242 0 0.00
Summary of performance deviations between NR and RIARCP-RLT models in RIARCP 1038 635 61.18 319 318 99.69
I3 NR 1468 425 28.95 239 0 0.00
percentage.
RIARCP 1468 672 45.78 454 453 99.78
Class c c c PM1 PM4 a a a I4 NR 998 221 22.14 174 0 0.00
PM30 PM40 PM50 PM30 PM40 PM50
RIARCP 998 515 51.60 122 122 100.00
I1 100.00 42.86 53.10 42.86 23.67 56.60 57.99 58.66 I5 NR 591 52 8.80 193 0 0.00
I2 100.00 47.03 37.67 −31.82 15.35 33.51 37.18 39.83 RIARCP 591 374 63.28 153 140 91.50
I3 99.49 99.56 53.87 −89.96 8.11 25.78 25.90 23.90 I6 NR 60 8 13.33 7 0 0.00
I4 100.00 100.00 100.00 29.89 42.66 66.88 66.58 61.68 RIARCP 60 43 71.67 20 20 100.00
I5 86.67 78.38 52.44 20.73 22.13 40.79 40.57 36.36 I7 NR 336 65 19.35 24 0 0.00
I6 100.00 75.00 26.67 −185.71 21.88 48.15 46.43 41.38 RIARCP 336 163 48.51 64 64 100.00
I7 100.00 100.00 94.94 −166.67 23.86 40.32 39.42 38.89 I8 NR 136 34 25.00 22 0 0.00
I8 100.00 100.00 91.30 −13.64 34.43 47.22 50.00 48.98 RIARCP 136 70 51.47 25 25 100.00

25
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Table 7a
Comparison between NR and RIARCP-RLT model on class I1 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation (%) TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 541.56 * 300.02 * 500.61 * 284.55 * 71.89 79.45

RIARCP 240.92 55.51 20.54 93.15 206.23 58.80 14.17 95.02 83.11 88.98

2 NR 423.17 * 170.34 * 373.96 * 151.84 * 78.33 86.10

RIARCP 425.22 −0.48 170.88 −0.32 388.95 −4.01 164.95 −8.63 80.31 85.68

3 NR 615.11 * 291.82 * 563.91 * 312.26 * 74.77632 82.75

RIARCP 305.36 50.36 35.42 87.86 261.48 53.63 24.97 92.00 82.16 88.61

4 NR 532.35 * 218.97 * 472.19 * 193.92 * 75.23 84.04

RIARCP 343.93 35.39 47.41 78.35 290.68 38.44 31.13 83.95 80.24 87.96

5 NR 467.44 * 185.3 * 404.14 * 153.8 * 74.68 84.59

RIARCP 314.99 32.61 45.46 75.47 265.95 34.19 30.3 80.30 81.15 88.62

6 NR 398.37 * 116.43 * 356.47 * 108.05 * 76.86 87.12

RIARCP 366.67 7.96 94.4 18.92 301.12 15.53 62.21 42.42 82.38 87.95

7 NR 404.9 * 106 * 338.86 * 76.22 * 77.87 87.00

RIARCP 356.86 11.86 60.09 43.31 301.92 10.90 42.09 44.78 81.12 88.43

Average deviation 27.60 56.68 29.64 61.41

Table 7b
Comparison between NR and RIARCP-RLT model on class I2 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation (%) TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 635.46 * 166.08 * 528.48 * 116.9 * 76.47 86.46

RIARCP 529.16 16.73 63.07 62.02 451.74 14.52 44.67 61.79 82.24 88.47
2 NR 861.93 * 283.21 * 720.47 * 208.09 * 73.32 84.49
RIARCP 698.49 18.96 126.61 55.29 603.58 16.22 100.33 51.79 80.51 86.99
3 NR 827 * 250.3 * 706.64 * 196.46 * 76.19 84.72
RIARCP 705.06 14.74 133.95 46.48 611.29 13.49 108.56 44.74 80.47 86.81
4 NR 878.85 * 300.05 * 747.97 * 235.54 * 74.37 84.06
RIARCP 700.82 20.26 129.85 56.72 604.72 19.15 102.6 56.44 80.80 87.20
5 NR 857.5 * 280.32 * 721.66 * 211.08 * 74.54 84.72
RIARCP 702.59 18.07 130.95 53.29 602.07 16.57 98.87 53.16 80.31 87.22
6 NR 706.52 * 184.43 * 591.33 * 131.31 * 76.00 85.94
RIARCP 603.72 14.55 83.94 54.49 523.15 11.53 66.13 49.64 81.94 87.68
7 NR 1034.56 * 469.07 * 912.39 * 408.13 * 74.69 84.90
RIARCP 650.77 37.10 130.95 72.08 556.23 39.04 99.33 75.66 79.96 87.31
Average Deviation 20.06 57.20 18.65 56.17

Table 7c
Comparison between NR and RIARCP-RLT model on class I3 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation (%) TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 1135.31 * 359.57 * 976.7 * 292.12 * 76.52 84.33

RIARCP 991.61 12.66 222.88 38.01 857.52 12.20 181.19 37.97 80.00 86.31

2 NR 1118.61 * 343.1 * 963.65 * 277.86 * 77.00 84.67

RIARCP 972.16 13.09 201.82 41.18 843.91 12.43 165.08 40.59 80.70 86.69

3 NR 1111.51 * 336.67 * 957.63 * 273.09 * 77.28 84.84

RIARCP 956.71 13.93 186.46 44.62 833.76 12.94 154.81 43.31 81.12 86.88

4 NR 1268.24 * 479.52 * 1106.79 * 406.98 * 75.84 83.58

RIARCP 984.56 22.37 212.45 55.70 856.79 22.59 175.57 56.86 80.50 86.51

5 NR 1113.88 * 315.68 * 959.44 * 255.11 * 77.52 85.07

RIARCP 1016.32 8.76 220.85 30.04 882.4 8.03 181.9 28.70 80.13 86.32

6 NR 955.61 * 254.77 * 829.44 * 210.55 * 78.81 85.53

RIARCP 865.42 9.44 169.38 33.52 751.96 9.34 139.71 33.65 81.09 86.93

7 NR 1111.84 * 312.07 * 961.94 * 254.39 * 77.90 84.99

RIARCP 1013.36 8.86 217.49 30.31 880.49 8.47 178.61 29.79 80.35 86.37

Average Deviation 12.73 39.05 12.28 38.70

26
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Table 7d
Comparison between NR and RIARCP-RLT model on class I4 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation (%) TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 946.6 * 410.32 * 880.57 * 405.87 * 79.83 86.11

RIARCP 819.45 13.43 294.59 28.20 736.53 16.36 271.23 33.17 83.49 88.09

2 NR 1420.7 * 795.3 * 1333.07 * 767.61 * 77.56 84.07

RIARCP 536.32 62.25 20.04 97.48 465.7 65.07 12.85 98.33 84.02 88.87

3 NR 973.92 * 412.1 * 904.03 * 405.91 * 81.16 87.46

RIARCP 852.81 12.44 308.37 25.17 774.01 14.38 291.75 28.12 83.57 88.14

4 NR 1649.37 * 1007.14 * 1568.97 * 986.84 * 78.52 84.12

RIARCP 976.96 40.77 414.93 58.80 906.06 42.25 407.41 58.72 83.41 88.08

5 NR 1214.77 * 580.62 * 1144.91 * 575.87 * 81.94 87.19

RIARCP 743.93 38.76 184.82 68.17 659.8 42.37 165.78 71.21 82.65 87.67

6 NR 1325.83 * 728.33 * 1247 * 711.05 * 79.54 85.14

RIARCP 969.97 26.84 397.6 45.41 902.76 27.61 393.62 44.64 83.55 88.03

7 NR 1350.93 * 665.43 * 1282.39 * 660.42 * 79.70 85.41

RIARCP 1084.63 19.71 479.99 27.87 1000.53 21.98 458.32 30.60 82.71 87.03

Average Deviation 30.60 50.16 32.86 52.11

Table 7e
Comparison between NR and RIARCP-RLT model on class I5 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation (%) TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 798.46 * 455.56 * 742.85 * 434.18 * 74.13 81.04

RIARCP 769.11 3.68 416.3 8.62 704.92 5.11 387.36 10.78 76.23 83.57

2 NR 626.82 * 287.77 * 564.37 * 265.45 * 76.47 84.59

RIARCP 376.48 39.94 56.76 80.28 309.4 45.18 30.73 88.42 80.48 88.28

3 NR 1258.43 * 859.25 * 1188.98 * 825.56 * 72.06 82.03

RIARCP 995.15 20.92 647.53 24.64 925.08 22.20 613.28 25.71 74.15 82.83

4 NR 2540.76 * 1937.78 * 2456.83 * 1888.92 * 66.76 77.00

RIARCP 712.8 71.95 362.14 81.31 647.43 73.65 335.72 82.23 77.14 84.46

5 NR 1156.13 * 782.36 * 1094.11 * 754.68 * 72.26 79.77

RIARCP 1207.43 −4.44 826.89 −5.69 1145.34 −4.68 799.2 −5.90 73.49 81.34

6 NR 2177.83 * 1678.76 * 2132.65 * 1667.58 * 71.75 77.78

RIARCP 1543.87 29.11 1102.59 34.32 1490.41 30.11 1083.34 35.04 73.70 81.04

7 NR 1642.13 * 1162.63 * 1588.75 * 1142.74 * 70.65 78.24

RIARCP 952.16 42.02 604.59 48.00 874.09 44.98 562.53 50.77 73.49 83.83

Average Deviation 29.02 38.78 30.93 41.01

respectively, and r j is a random parameter drawn from beta distribution ζ j = max (DTi + ti + τk + ζi − DTj, 0) . Consequently, using a push-
Be (α, β ) . Note that α and β are set as follows. Recall that the mean μ and back recovery strategy, that postpones flight's departure until its
the mode λ of the beta distribution are given by scheduled plane is ready, we can sequentially derive the updated
α α−1 departure times of all the scheduled flights.
μ= and λ= Step 3: Computation of performance measures. To evaluate the
α+β α+β+2
performance of the derived solutions, we pull out the following
Hence, given two estimates of ζμ and ζλ of the mean and mode of the performance measures:
primary delays, the corresponding parameters α and β are derived.
• TD : Total delay.
In our experiments, we delay the departure time of each flight with
• TD15: Total delay, when the delay is more than 15 min.
a probability of Pd = 0.15. In addition, we used the following parameter
• TSD : Total secondary (i.e. reactionary) delay.
setting: ζmin = 5 min , ζmax = 80 min , ζμ = 25 min , and ζλ = 15 min . Thus,
• TSD15: Total secondary delay, when the delay is more than 15 min.
the obtained values of α and β are 1.467 and 4.033, respectively.
• OT : On-time performance.

Step 2: Generation of reactionary delays. If a flight i that is planned

• OT15: On-time performance, when the delay is more than 15 min.

at time DTi , and is served by an aircraft of type k, exhibits a delay of Table 7, we report the obtained Monte-Carlo simulation study re-
ζi , then its subsequent downstream flight cannot be served earlier sults for both NR and RIARCP-RLT solutions. We also display for each
than DTi + ti + τk + ζi . Consequently, flight j that is planned at time class of instances, and for all performance metrics, but OT and OT15 , the
DTj and served immediately after i, may be delayed up to percentage deviation 100 ×
PM (NR) − PM (RIARCP )
.
PM (NR)

27
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Table 7f
Comparison between NR and RIARCP-RLT model on class I6 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation (%) TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 310.37 * 87.93 * 264.32 * 69.66 * 76.97 85.10

RIARCP 257.96 16.89 36.47 58.52 224.22 15.17 31 55.50 82.02 87.65

Average Deviation 16.89 58.52 15.17 55.50

Table 7g
Comparison between NR and RIARCP-RLT model on class I7 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 1579.81 * 327.73 * 1366.53 * 259.36 * 79.86 86.37

RIARCP 1414.46 10.47 167.56 48.87 1234.26 9.68 133.31 48.60 82.47 87.72

Average Deviation 10.47 48.87 9.68 48.60

Table 7h
Comparison between NR and RIARCP-RLT model on class I8 –based instances with respect to simulation metrics.
Instance TD Deviation (%) TSD Deviation (%) TD15 Deviation (%) TSD15 Deviation (%) OT (%) OT15 (%)

1 NR 637.03 * 161.27 * 561.03 * 142.96 * 81.19 86.87

RIARCP 506.3 20.52 41.41 74.32 437.89 21.95 14.7 89.72 83.90 88.86

Average Deviation 20.52 74.32 21.95 89.72

Looking at Table 7, a striking observation is that, all measures of
performance have markedly improved on most instances when em-
bedding robustness's features. The solutions delivered by RIARCP
models exhibit enhanced on-time performances with comparison to NR
ones. For instance, the on-time performance (OT ) and the 15 min on-
time performance (OT1 5) were improved by up to 11.23% and 9.53%,
respectively. Table 7 also shows that total delay (TD ) and total sec-
ondary delay (TSD ) were reduced in average by up to 30.6% and 58.52%,
respectively. This provides empirical evidence that RIARCP model
builds optimized schedules that are less vulnerable to delays, and
substantially outperforms NR schedules.
6. Conclusion
During the last decade, the global commercial airline industry has been
witnessing an unprecedented period of prosperity as the number of pas-
sengers keeps growing and airlines keep expanding their networks.
However, a negative side effect of this success is that the number of dis-
rupted flights has been soaring and thereby damaging airlines reputation
and profitability. To mitigate the impact of aircraft arriving late, we pro-
posed a robust integrated model that simultaneously builds aircraft routes
and crew pairing while accommodating maintenance constraints along
with crew-related work constraints. To generate robust solutions that are
less vulnerable to minor disruptions, the proposed model aggressively
penalizes both aircraft and pairings short connections. Furthermore, so-
lutions where the crews follow the aircraft are promoted as much as
possible. The proposed model is a mixed-integer nonlinear program that is
linearized using the Reformulation-Linearization technique (RLT). An ap-
pealing feature of the derived MIP model is that is includes a polynomial
number of variables and constraints and is thereby amenable to be directly
solved by invoking a general-purpose solve without resorting to sophisti-
cated optimization methodologies. To the best of our knowledge, this is
the first contribution in this regard. The singular simplicity and flexibility
in implementation of the proposed approach makes it very attractive. We
tested the empirical performance of the proposed model on instances
based on publicly available data which pertains to major international
airlines. We found that the model successfully delivers cost-effective so-
lutions that typically exhibit few short connections along with a large
number of connections where the crews follow the aircraft. We also show
that the solutions delivered by our model include up to 71.67% of con-
nections where crew remains on the same aircraft. Furthermore, we ob-
served that the required CPU times are moderate.
A first possible direction for future research is to incorporate addi-
tional realistic aircraft routing constraints (such as capacitated main-
tenance stations), and/or different crew pairing rules. The flexibility of
the proposed MIP model enables to readily accommodate these exten-
sions. Furthermore, a second possible direction for future research is the
additional integration of the fleet assignment problem within a single
compact model framework. The resulting integrated fleet assignment,
aircraft routing, and crew pairing model will make it feasible to gen-
erate effective robust solutions by simply invoking an MIP solver.
Acknowledgment
This research was made possible by NPRP Grant No. 06-818-5-094
from the Qatar National Research Fund (a member ofThe Qatar
Foundation). The statements made herein are solely the responsibility
of the authors.

Appendix A. Linearization process

To enhance the solvability of the proposed non-linear RIARCP problem, we apply the RLT of Sherali and Adams (1990, 1994) to derive a tight,
equivalent linear model representation of the problem. To begin with, consider constraints (8)–(9). We can linearize these constraints by using the
substitution:
αij = ui x ij , ∀ (i, j ) ∈ A, (A.1)

αij = uj x ij , ∀ (i, j ) ∈ A. (A.2)

28
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

Thus, (8) and (9) get transformed to the following:

k
αij = t j x ij , ∀ k ∈ K , j ∈ Lk , (i, j ) ∈ AM , (A.3)
k
αij = αij + t j x ij , ∀ k ∈ K , j ∈ Lk , (i, j ) ∈ ANM . (A.4)
Also, multiplying (33) by each x ij and x ji , ∀ (i, j ) ∈ A , ∀ (j, i) ∈ A , respectively and linearizing using (A.1)-(A.2), we obtain after rearranging
indices:
t j x ij ≤ αij ≤ tmax x ij , ∀ k ∈ K , (i, j ) ∈ Ak , (A.5)

ti x ij ≤ αij ≤ tmax x ij , ∀ k ∈ K , (i, j ) ∈ Ak . (A.6)

Furthermore, multiplying each of (6) and (7) by its corresponding uj and linearizing, we get after rearranging indices:

∑ αij = uj , ∀ k ∈ K , j ∈ Lk ,
i :(i, j ) ∈ Ak (A.7)

∑ αji = uj , ∀ k ∈ K , j ∈ Lk .
i :(j, i) ∈ Ak (A.8)
Finally, by (A.6)-(A.8), we obtain:

tj ∑ x ij ≤ uj ≤ tmax ∑ x ij , ∀ k ∈ K , j ∈ Lk .
i :(i, j ) ∈ Ak i :(i, j ) ∈ Ak (A.9)

Proposition 1. Constraints (A.3)-(A.9) along with can be replaced by the following restrictions, which, together, with (6) and (7), are equivalent to
these constraints even in the continuous (LP) relaxation sense:

∑ αji = t j + ∑ αij, ∀ k ∈ K,
i :(j, i) ∈ Ak k
i :(i, j ) ∈ ANM (A.10)
textj ∈ Lk , (A.11)
k
t j x ji ≤ αji ≤ (tmax − ti ) x ji , ∀ k ∈ K , (j, i) ∈ ANM , (A.12)
k
t j x ji ≤ αji ≤ tmax x ji , ∀ k ∈ K , (j, i) ∈ AM . (A.13)

Proof. Using the identities (A.8)-(A.8), we get:

∑ αij = ∑ αji, ∀ k ∈ K , j ∈ Lk .
i :(i, j ) ∈ Ak i :(j, i) ∈ Ak

This is equivalent to:

∑ αij + ∑ αij = ∑ αji, ∀ k ∈ K , j ∈ Lk .

k
i :(i, j ) ∈ AM k
i :(i, j ) ∈ ANM i :(j, i) ∈ Ak

We next eliminate the α− variables from theses restrictions using the identities (A.3)-(A.4), which yields:

∑ αji = t j + ∑ αij, ∀ k ∈ K , j ∈ Lk .
i :(j, i) ∈ Ak k
i :(i, j ) ∈ ANM

k
Furthermore, by substituting (A.4) into (A.5) and combining this with (A.6) for k∈ K, (i, j ) ∈ ANM , we obtain (A.12), and we retain the in-
k
equalities in (A.6) for k ∈ K , (i, j ) ∈ AM , as indicated in (A.13). ■
For the sake of convenience, we define the following parameter:
k
⎧tmax − ti, ∀ k ∈ K , (j, i) ∈ ANM ,⎫
btj = k
⎨ tmax , ∀ k ∈ K , (j, i) ∈ AM . ⎬
⎩ ⎭
Hence, (A.12)-(A.13) can be expressed as follows (after rearranging indices):
ti x ij ≤ αij ≤ bijt x ij , ∀ k ∈ K , (i, j ) ∈ Ak . (A.14)
Finally, we used (6) to equivalently rewrite (10). We obtain the following constraint:

∑ ⎛1 − ∑ xij⎞ ≤ NAk
⎜ ⎟ − nak , ∀ k ∈ K.
⎝ i :(i,j)∈A1 ∪A3 ⎠
j ∈ Lk (A.15)
Next, we linearize Constraints (15)–(16) using the following substitution:
βij = μi yij , ∀ (i, j ) ∈ B, (A.16)

βij = μj yij , ∀ (i, j ) ∈ B. (A.17)

Thus, (15)–(16) get transformed to the following:

29
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

βij = yij , ∀ (i, j ) ∈ B2 ∪ BD , (A.18)

βij = βij + yij , ∀ (i, j ) ∈ B1. (A.19)

Also, multiplying (33) by yji ((j, i) ∈ B \BD ) , we obtain:
yji ≤ μj yji ≤ μmax yji , ∀ (j, i) ∈ B \BD .

Then, linearizing these inequalities using (A.16), we obtain after rearranging indices i and j:
yij ≤ βij ≤ μmax yij , ∀ (i, j ) ∈ B \BD . (A.20)
Similarly, multiplying (33) by yij ((i, j ) ∈ B \BA) , we obtain:
yij ≤ βij ≤ μmax yij , ∀ (i, j ) ∈ B \BA. (A.21)
Moreover, multiplying each of (11) and (12) by its corresponding μj and linearizing, we get:

∑ βij = μj , ∀ j ∈ L,
(i, j ) ∈ B (A.22)

∑ βji = μj , ∀ j ∈ L.
(j, i) ∈ B (A.23)

Proposition 2. Constraints (A.18)-(A.23) can be replaced by the following restrictions (A.24)-(A.26), which, together with (11) and (12), are
equivalent to these constraints even in the continuous (LP relaxation) sense:

∑ βji = 1 + ∑ βij , ∀ j ∈ L,
(j, i) ∈ B (i, j ) ∈ B1 (A.24)

yij ≤ βij ≤ (μmax − 1) yij , ∀ (i, j ) ∈ B1, (A.25)

yij ≤ βij ≤ μmax yij , ∀ (i, j ) ∈ B2 ∪ BD . (A.26)

Proof. Using the identities (25)–(26), we get:.

∑ βij = ∑ βij , ∀ j ∈ L.
(i, j ) ∈ B (j, i) ∈ B

This is equivalent to:

∑ βij + ∑ βij + ∑ βij = ∑ βij , ∀ j ∈ L.

(i, j ) ∈ B1 (i, j ) ∈ B2 (i, j ) ∈ BD (j, i) ∈ B

We next eliminate the β − variables from these restrictions using the identities (A.18) and (A.19), which yields:

∑ (βij + yij ) + ∑ yij + ∑ yij = ∑ βij , ∀ j ∈ L.

(i, j ) ∈ B1 (i, j ) ∈ B2 (i, j ) ∈ BD (j, i) ∈ B

Using (11) we obtain:

1+ ∑ βij = ∑ βji , ∀ j ∈ L,
(i, j ) ∈ B1 (j, i) ∈ B (A.27)
which yields (A.24). ■
Similarly, we can linearize (17)–(26) and (28) in a likewise fashion by using the substitutions:
δij = φi yij , ∀ (i, j ) ∈ B, (A.28)

δij = φj yij , ∀ (i, j ) ∈ B. (A.29)

ζij = λi yij , ∀ (i, j ) ∈ B, (A.30)

ζij = λj yij , ∀ (i, j ) ∈ B. (A.31)

σij = νi yij , ∀ (i, j ) ∈ B, (A.32)

σij = νj yij , ∀ (i, j ) ∈ B. (A.33)

ϑij = ηi yij , ∀ (i, j ) ∈ B, (A.34)

ϑij = ηj yij , ∀ (i, j ) ∈ B. (A.35)

In so doing, constraints (17)–(26) and (28) can be written as follows:

∑ δji = t j + ∑ δij, ∀ j ∈ L,
(j, i) ∈ B (i, j ) ∈ B1 (A.36)

30
M. Ben Ahmed et al. Journal of Air Transport Management 73 (2018) 15–31

ti yij ≤ δij ≤ (φmax − t j ) yij , ∀ (i, j ) ∈ B1, (A.37)

ti yij ≤ δij ≤ φmax yij , ∀ (i, j ) ∈ B2 ∪ BA, (A.38)

∑ ζ ji = ∑ sij yij + t j + ∑ ζij, ∀ j ∈ L,

(j, i) ∈ B (i, j ) ∈ B1 (i, j ) ∈ B1 (A.39)

ti yij ≤ ζij ≤ (λ max − sij − t j ) yij , ∀ (i, j ) ∈ B1, (A.40)

ti yij ≤ ζij ≤ λ max yij , ∀ (i, j ) ∈ B2 ∪ BA, (A.41)

∑ σji = ∑ yij + ∑ σij, ∀ j ∈ L,

(j, i) ∈ B (i, j ) ∈ BD ∪ B2 (i, j ) ∈ B1∪ B2 (A.42)

yij ≤ σij ≤ (νmax − 1) yij , ∀ (i, j ) ∈ B2, (A.43)

yij ≤ σij ≤ νmax yij , ∀ (i, j ) ∈ B1 ∪ BA, (A.44)

∑ ϑji = ∑ sij yij + ∑ lij yij + t j + ∑ ϑij , ∀ j ∈ L,

(j, i) ∈ B (i, j ) ∈ B1 (i, j ) ∈ B2 (i, j ) ∈ B1∪ B2 (A.45)

ti yij ≤ ϑij ≤ (ηmax − sij − t j ) yij , ∀ (i, j ) ∈ B1, (A.46)

ti yij ≤ ϑij ≤ (ηmax − lij − t j ) yij , ∀ (i, j ) ∈ B2 (A.47)

ϑj0 + l min ≤ 1440dj , ∀ j ∈ LA . (A.48)

References
AhmadBeygi, S., Cohn, A., Guan, Y., Belobaba, P., 2008. Analysis of the potential for delay
propagation in passenger airline networks. J. Air Transport. Manag. 14 (5), 221–236.
Al-Thani, N.A., Ben Ahmed, M., Haouari, M., 2016. A model and optimization-based
heuristic for the operational aircraft maintenance routing problem. Transport. Res. C
Emerg. Technol. 72, 29–44.
Aloulou, M.A., Haouari, M., Mansour, F.Z., 2013. A model for enhancing robustness of
aircraft and passenger connections. Transport. Res. C Emerg. Technol. 32, 48–60.
Barnhart, C., Belobaba, P., Odoni, A.R., 2003. Applications of operations research in the
air transport industry. Transport. Sci. 37 (4), 368–391.
Ben Ahmed, M., Ghroubi, W., Haouari, M., Sherali, H.D., 2017a. A hybrid optimization-
simulation approach for robust weekly aircraft routing and retiming. Transport. Res.
C Emerg. Technol. 84, 1–20.
Ben Ahmed, M., Zeghal Mansour, F., Haouari, M., 2017b. A two-level optimization ap-
proach for robust aircraft routing and retiming. Comput. Ind. Eng. 112, 586–594.
Cacchiani, V., Salazar-González, J.-J., 2017. Optimal solutions to a real-world integrated
airline scheduling problem. Transport. Sci. 51 (1), 250–268.
Cadarso, L., Marín, Á., 2013. Robust passenger oriented timetable and fleet assignment
integration in airline planning. J. Air Transport. Manag. 26, 44–49.
Chtourou, H., Haouari, M., 2008. A two-stage-priority-rule-based algorithm for robust
resource-constrained project scheduling. Comput. Ind. Eng. 55 (1), 183–194.
Cohn, A.M., Barnhart, C., 2003. Improving crew scheduling by incorporating key main-
tenance routing decisions. Oper. Res. 51 (3), 387–396.
Cook, A.J., Tanner, G., 2015. European Airline Delay Cost Reference Values.
EUROCONTROL Performance Review Unit.
Cordeau, J.-F., Stojković, G., Soumis, F., Desrosiers, J., 2001. Benders decomposition for
simultaneous aircraft routing and crew scheduling. Transport. Sci. 35 (4), 375–388.
Dožić, S., Kalić, M., 2015. Three-stage airline fleet planning model. J. Air Transport.
Manag. 46, 30–39.
Dück, V., Ionescu, L., Kliewer, N., Suhl, L., 2012. Increasing stability of crew and aircraft
schedules. Transport. Res. C Emerg. Technol. 20 (1), 47–61.
Dunbar, M., Froyland, G., Wu, C.-L., 2012. Robust airline schedule planning: minimizing
propagated delay in an integrated routing and crewing framework. Transport. Sci. 46
(2), 204–216.
Dunbar, M., Froyland, G., Wu, C.-L., 2014. An integrated scenario-based approach for
robust aircraft routing, crew pairing and re-timing. Comput. Oper. Res. 45, 68–86.
Eltoukhy, A.E., Chan, F.T., Chung, S., 2017. Airline schedule planning: a review and fu-
ture directions. Ind. Manag. Data Syst. 117 (6), 1201–1243.
Gao, C., Johnson, E., Smith, B., 2009. Integrated airline fleet and crew robust planning.
Transport. Sci. 43 (1), 2–16.
Gopalakrishnan, B., Johnson, E.L., 2005. Airline crew scheduling: state-of-the-art. Ann.
Oper. Res. 140 (1), 305–337.
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.
Jamili, A., 2016. A robust mathematical model and heuristic algorithms for integrated
aircraft routing and scheduling, with consideration of fleet assignment problem. J.
Air Transport. Manag. 58, 21–30.
Jiang, H., Barnhart, C., 2009. Dynamic airline scheduling. Transport. Sci. 43 (3), 336–354.
Kasirzadeh, A., Saddoune, M., Soumis, F., 2015. Airline crew scheduling: models, algo-
rithms, and data sets. EURO Journal on Transportation and Logistics 6 (2), 111–137.
Klabjan, D., Johnson, E.L., Nemhauser, G.L., Gelman, E., Ramaswamy, S., 2002. Airline
crew scheduling with time windows and plane-count constraints. Transport. Sci. 36
(3), 337–348.
Lacasse-Guay, E., Desaulniers, G., Soumis, F., 2010. Aircraft routing under different
business processes. J. Air Transport. Manag. 16 (5), 258–263.
Ladier, A.-L., Alpan, G., 2016. Robust cross-dock scheduling with time windows. Comput.
Ind. Eng. 99, 16–28.
Mercier, A., Cordeau, J.-F., Soumis, F., 2005. A computational study of benders decom-
position for the integrated aircraft routing and crew scheduling problem. Comput.
Oper. Res. 32 (6), 1451–1476.
Mercier, A., Soumis, F., 2007. An integrated aircraft routing, crew scheduling and flight
retiming model. Comput. Oper. Res. 34 (8), 2251–2265.
Muter, İ., Birbil, Ş.İ., Blbl, K., Şahin, G., Yenign, H., Taş, D., Tzn, D., 2013. Solving a
robust airline crew pairing problem with column generation. Comput. Oper. Res. 40
(3), 815–830.
Papadakos, N., 2009. Integrated airline scheduling. Comput. Oper. Res. 36 (1), 176–195.
Ruther, S., Boland, N., Engineer, F.G., Evans, I., 2017. Integrated aircraft routing, crew
pairing, and tail assignment: branch-and-price with many pricing problems.
Transport. Sci. 51 (1), 177–195.
Salazar-González, J.-J., 2014. Approaches to solve the fleet-assignment, aircraft-routing,
crew-pairing and crew-rostering problems of a regional Carrier. Omega 43, 71–82.
Sandhu, R., Klabjan, D., 2007. Integrated airline fleeting and crew-pairing decisions.
Oper. Res. 55 (3), 439–456.
Schaefer, A.J., Johnson, E.L., Kleywegt, A.J., Nemhauser, G.L., 2005. Airline crew sche-
duling under uncertainty. Transport. Sci. 39 (3), 340–348.
Shao, S., Sherali, H.D., Haouari, M., 2017. A novel model and decomposition approach for
the integrated airline fleet assignment, aircraft routing, and crew pairing problem.
Transport. Sci. 51 (1), 233–249.
Sherali, H.D., Adams, W.P., 1990. A hierarchy of relaxations between the continuous and
convex hull representations for zero-one programming problems. SIAM J. Discrete
Math. 3 (3), 411–430.
Sherali, H.D., Adams, W.P., 1994. A hierarchy of relaxations and convex hull character-
izations for mixed-integer zero-one programming problems. Discrete Appl. Math. 52
(1), 83–106.
Sherali, H.D., Bish, E.K., Zhu, X., 2006. Airline fleet assignment concepts, models, and
algorithms. Eur. J. Oper. Res. 172 (1), 1–30.
Warburg, V., Hansen, T.G., Larsen, A., Norman, H., Andersson, E., 2008. Dynamic airline
scheduling: an analysis of the potentials of refleeting and retiming. J. Air Transport.
Manag. 14 (4), 163–167.
Weide, O., Ryan, D., Ehrgott, M., 2010. An iterative approach to robust and integrated
aircraft routing and crew scheduling. Comput. Oper. Res. 37 (5), 833–844.
Yan, C., Kung, J., 2018. Robust aircraft routing. Transport. Sci. 52 (1), 118–133.
Yao, Y., Zhao, W., Ergun, O., Johnson, E., 2005. Crew pairing and aircraft routing for on-
demand aviation with time window. SSRN Electronic Journal Available at. https://
ssrn.com/abstract=822265.
Zeghal Mansour, F., Haouari, M., Sherali, H.D., 2018. A new compact formulation for the
daily crew pairing problem. Transport. Sci in press.

31