Computers & Operations Research 57 (2015) 123–137

Contents lists available at ScienceDirect

Computers & Operations Research

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

A novel passenger recovery approach for the integrated airline

recovery problem
Stephen J Maher n
School of Mathematics and Statistics, University of New South Wales, Sydney, NSW 2052, Australia

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

Available online 3 December 2014 Schedule disruptions require airlines to intervene through the process of recovery; this involves
Keywords: modifications to the planned schedule, aircraft routings, crew pairings and passenger itineraries.
Airline recovery Passenger recovery is generally considered as the final stage in this process, and hence passengers
Passenger recovery experience unnecessarily large impacts resulting from flight delays and cancellations. Most recovery
Column generation approaches considering passengers involve a separately defined module within the problem formula-
Row generation tion. However, this approach may be overly complex for recovery in many aviation and general
transportation applications. This paper presents a unique description of the cancellation variables that
models passenger recovery by prescribing the alternative travel arrangements for passengers in the
event of flight cancellations. The results will demonstrate that this simple, but effective, passenger
recovery approach significantly reduces the operational costs of the airline and increases passenger flow
through the network. The integrated airline recovery problem with passenger reallocation is solved
using column-and-row generation to achieve high quality solutions in short runtimes. An analysis of the
column-and-row generation solution approach is performed, identifying a number of enhancement
techniques to further improve the solution runtimes.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction performed simultaneously. Solving the airline recovery problem as a

series of discrete but related stages greatly reduces the problem
The efficient flow of passengers is critical to all transportation complexity. These four stages are fundamental in the airline recovery
fields, especially when regular operations are affected by disrup- process and provide the major focus areas of research. A very good
tions. The response to such a disruption is termed recovery, which recent review of the current airline recovery literature can be found in
involves strategies such as delaying and cancelling services in an Clausen et al. [13].
effort to return to the originally planned state. Since transporta- The sequential solution approach for airline recovery involves
tion applications involve a number of interrelated resources, for solving each stage in isolation using the results from previous
example crew and aircraft in the airline context, each of these stages as input. Fixing the solution at each stage has the unfortu-
must be considered in recovery. Passenger are also of high nate effect of providing less flexibility to the problems solved later
importance, but their direct costs are not easily quantified. Con- in this process. Consequently, it is common for the solutions to
sequently, passengers are not usually considered until the final later stages to be suboptimal or even infeasible. The integration of
stages of recovery processes. Hence, it is common for passengers multiple recovery stages is proposed as a method to address the
to be significantly affected by any disruption in transportation possible suboptimality and infeasibility. The focus of this paper is
applications. an airline recovery problem that integrates schedule, aircraft, crew
The complete airline recovery process is a very large and complex and passenger recovery. A major contribution of this paper is the
problem that is commonly broken into a number of sequential stages. modelling approach employed to consider passengers in the recov-
In practice, the sequential stages that form the airline recovery process ery process.
are, the schedule, aircraft, crew and passenger recovery problems; Airline planning and recovery processes and the flow of passen-
however, it is common for the schedule and aircraft recovery to be gers is greatly affected by the network structure that is employed.
There are two common types of airline networks, which are labelled
as hub-and-spoke and point-to-point networks. The hub-and-spoke
n
Tel.: þ 61 2 9385 6910. network identifies a number of key major airports (hubs) with a
E-mail address: [email protected] large proportion of all flights scheduled between these and outlying

http://dx.doi.org/10.1016/j.cor.2014.11.005
0305-0548/& 2014 Elsevier Ltd. All rights reserved.
124 S.J Maher / Computers & Operations Research 57 (2015) 123–137

airports (spokes) and very few flights occurring between the network. Additionally, this approach may be overly complex for
spokes. This network is characterised by passenger itineraries, a airlines operating on a point-to-point network, where itineraries
set of flights booked to travel between an origin and destination, most commonly contain a single flight. The work presented in this
generally involving a transit through at least one hub. Alternatively, paper introduces an alternative passenger recovery approach with
the point-to-point network is characterised by direct flights a specific focus on airlines operating point-to-point networks. The
between most airports. As such, it is not common for passenger modelling approach developed in this paper is a contribution to
itineraries to involve more than one flight. The difference between the airline passenger recovery literature and can be employed for
these network types greatly affects the treatment of passengers alternative transportation applications, such as railway operations
during recovery. The attention towards point-to-point carriers in recovery.
this paper is in response to the growing number of low-cost carriers One of the first models dedicated to the recovery of passenger
providing this network structure and its relationship with other itineraries is presented by Bratu and Barnhart [10]. In [10], two
transportation applications, such as railway operations. different optimisation models are described, analysing the trade-
off between operating and passenger recovery costs, while also
1.1. Airline recovery considering aircraft rerouting and the use of reserve crew. McCarty
[21] presents an alternative passenger recovery approach that fits
The practicality of the sequential airline recovery process has within the sequential recovery framework. The approach by [21]
driven the attention of researchers to review and develop indivi- attempts to identify recovered itineraries for passengers in the
dual recovery stages. Examples of the aircraft recovery problem, event of a delay on one or many flights. The integration of the
including schedule recovery decisions, are presented by Jarrah passenger and aircraft recovery problems is presented by Jafari
et al. [17], Cao and Kanafani [11,12], Bard et al. [6], Eggenberg et al. and Zegordi [16]. The recovery problem in [16] uses the modelling
[15] and Rosenberger et al. [23]. The crew recovery problem has approach of Abdelghany et al. [2] and Abdelghany et al. [3], where
not received as much research attention compared to the aircraft the recovery horizon is partitioned into sets of resource indepen-
problem, but many advancements have been made. Key examples dent flights. The integrated aircraft and passenger recovery pro-
of the crew recovery problem are presented by Wei et al. [28], blem was also the focus of the 2009 ROADEF Challenge [1] that
Stojković et al. [26], Lettovsky et al. [19] and Abdelghaney et al. [2]. resulted in the development of many exact and heuristic solution
While the developments in each of the recovery stages have approaches. The winning solution for this challenge is given by
improved the overall recovery process, the issues of suboptimality Bisaillon et al. [9], implementing a large neighbourhood search
and infeasibility still remain. These issues are only adequately heuristic that identifies high quality solutions in short runtimes.
addressed through the development of an airline recovery pro- Finally, the integrated airline recovery approaches of Lettovsky
blem that integrates two or more of the sequential stages. [18] and Petersen et al. [22] both consider the recovery of passenger
The development of the integrated airline recovery problem itineraries as a dedicated subproblem in the Benders' decomposi-
has been limited due to the problem complexity and difficulty in tion framework.
achieving fast solution runtimes. However, improvement in solu- The alternative passenger recovery approach presented in this
tion approaches and computing capabilities has prompted a recent paper attempts to directly provide passengers with alternative travel
surge of interest in this problem. An early proposal for solving the arrangements following flight cancellations. This involves redistribut-
integrated recovery problem is provided by Lettovsky [18], which ing passengers from cancelled flights to alternative operating flights,
is based upon a Benders' decomposition framework. Lettovsky [18] which may have been delayed, to ensure passengers arrive at their
presents a model that integrates the complete airline recovery desired destination. This is achieved by introducing variables that
problem; however, only the crew recovery problem has been describe both flight cancellations and the optimal redistribution of
implemented [19]. The modelling approach for the complete passengers. Such passenger reallocation methods are applicable to
integrated recovery problem using Benders' decomposition is many transportation applications where passengers book single seg-
further explored by Petersen et al. [22]. In [22] a set of experi- ment journeys, which is observed with low-cost airlines and railway
ments for this approach are presented that achieve optimality operations. To the best of the author's knowledge, the modelling of the
within 30 min. An alternative, novel approach for the integrated cancellation variables to describe passenger reallocation options has
recovery problem is presented by Abdelghany et al. [3] as an not been previously considered. A contribution of this paper is the
extension to the crew recovery problem of Abdelghany et al. [2]. development of a simple approach that efficiently recovers passengers
This approach integrates aircraft, pilots and flight attendants with while significantly reducing the expected recovery costs.
fast solution runtimes achieved by partitioning the schedule into
sets of resource independent flights. While this partitioning 1.3. Solution approaches
process improves solution runtimes, a trade-off with the solution
quality is observed through the overestimation of the optimal Solution runtimes are a critical consideration of airline recovery
recovery costs. Finally, Maher [20] presents an airline recovery problems, with reductions achieved through various approxima-
problem, integrating schedule, aircraft and crew. The focus of [20] tion and decomposition approaches. Such approximation approaches
is to improve the solution runtimes while maintaining a high include the selection of specific recovery policies [17,28,26], approx-
solution quality using an exact solution approach. This is achieved imating flight arrival and departure times [6,15] and the selection of
with the application of a column-and-row generation framework affected aircraft [23] or crew [19,2]. In addition, decomposition
that is developed in [20]. techniques, such as column generation [24–26,15] and Benders'
decomposition [18,22], have also been employed to improve the
1.2. Passenger recovery solution runtimes. Finally, a heuristic approach is employed to solve
the integrated airline recovery problem presented by Bisaillon et al.
The vast majority of literature related to passenger recovery [9]. While these approaches successfully reduce the runtimes of the
focuses on airlines operating a hub-and-spoke network. As such, recovery problems, the heuristic approaches and techniques such as
the passenger recovery process must reconstruct all disrupted Benders' decomposition do not guarantee integer optimality.
itineraries, which may consist of multiple flight legs. This is a very An investigation into an exact solution approach to solve the
complex and difficult task, since the number of possible itineraries integrated airline recovery problem is presented by Maher [20]. In
is potentially much larger than the number of flights in the [20], a general framework for column-and-row generation is
 S.J Maher / Computers & Operations Research 57 (2015) 123–137 125

developed and evaluated using the integrated airline recovery network. The cancellation variables describe the cancellation of a
problem as an example. This framework is a direct alternative to flight and the alternative flights that the passengers are rebooked
Benders' decomposition, providing a guarantee of near integer onto. The specific details regarding the modelling of the cancellation
optimal solutions. In addition, column-and-row generation is variables are given in Section 2.3. The three variable types are linked
presented as a solution approach that reduces runtimes without in the IRP-PR by the flight delay and cancellation decisions and
the need to restrict the set of included crew and aircraft. While the specific flights allocated to each aircraft and crew. An important
general column-and-row generation framework presented by aspect of the IRP-PR is the use of all aircraft and crew resources in the
Maher [20] describes an approach to solve problems with multiple recovery problem, allowing for the optimal allocation of the available
secondary variables and linking constraints, the application in resources.
[20] does not adequately evaluate such a framework. Hence,
the application of column-and-row generation to a more general 2.1. Network structure
problem is not discussed.
In this paper, the consideration of passengers in the integrated The IRP-PR is solved for a single day flight schedule, attempting
airline recovery problem describes a problem with two sets of to completely recover from a disruption before the end of that day.
secondary variables and linking constraints. As such, the imple- The airline recovery process is executed immediately following a
mentation of column-and-row generation discussed in this paper schedule disruption, which is simply defined as anything prevent-
provides a more in-depth analysis of the framework presented in ing at least one flight from departing on time. As such, the flight
[20]. The discussion regarding the implementation and application schedule that is used in the IRP-PR, called the recovery schedule
of the framework from [20] to a more general problem is a ND, contains only a subset of all flights from the original flight
contribution of this paper. schedule N. The construction of the recovery schedule ND uses a
recovery window, defining the time period during which recovery
1.4. Outline of this paper actions can be performed. The recovery window commences imme-
diately after the disruption and concludes at an arbitrary time that
The mathematical model for the integrated recovery problem specifies when operations must return to plan. So the set ND is
with passenger reallocation (IRP-PR) is given in Section 2 with a defined to contain all the flights in N that depart after the disruption
description of the modelling approaches used for the included occurs but prior to the end of the recovery window.
recovery policies. The column-and-row generation framework In the IRP-PR, recovery actions are only permitted on flights
presented in Maher [20] is employed to solve the IRP-PR, with a contained in the set ND, implying that the flights in N n N D must
description of its implementation provided in Section 3. This depart as scheduled. To ensure that the flights contained in N n N D
description will include a review of column generation and the are assigned crew and aircraft in the recovered solution, the
integration with row generation in Sections 3.1 and 3.2 respec- concepts of carry-in and carry-out activities are introduced. The
tively. Section 4 will report the results from experiments, demon- carry-in and carry-out activities define the expected position of
strating the cost savings and improvement in passenger flow that each crew and aircraft at the start and end of the recovery period.
can be achieved by this passenger recovery approach. Finally, a These positions, generally stated as flights in the schedule, are
summary of the findings will be provided in Section 5. identified from the sequences of flights assigned to each crew and
aircraft in the planning solution. The carry-in activities for crew
and aircraft, contained in NKin and NRin, are the flights in each
2. Integrated recovery problem with passenger reallocation sequence that are performed immediately prior to the disruption.
These activities may also include origination airports if the crew or
The integrated airline recovery problem with passenger reallo- aircraft have not commenced the days work prior to the disrup-
cation (IRP-PR) integrates the schedule, aircraft and crew recovery tion. A carry-out activity for crew and aircraft, contained in NKout
problems with consideration to passenger flows through the and NRout, are the flights in each sequence that are performed
recovered network. The integration of aircraft and crew recovery immediately after the end of the recovery window. Similar to the
ensures that the optimal solution to both problems is found with carry-in activities, a carry-out activity may be defined as a termination
consistent flight delay and cancellation decisions. The passenger airport if the crew or aircraft is expected to finish working within the
considerations in the IRP-PR involves explicitly modelling the recovery window. The set of all origination and termination airports t
possible flight options that passengers can be rebooked onto in are contained in TK and TR for crew and aircraft respectively.
the event of a flight cancellation. Flight delays are implemented in the IRP-PR using the techni-
Passenger reallocation is a key aspect of the IRP-PR and the que of flight copies. The implementation of flight copies employed
novel modelling approach is a major contribution of this paper. in this paper is identical to that presented in Maher [20]. This
Following a flight cancellation passengers are either (i) rebooked technique models flight delays by creating multiple copies of each
onto another flight operated by the airline, (ii) rebooked onto flight within the schedule, each with a later departure time. The
another flight operated by a different airline, or (iii) provided with set of flight copies v defined for each flight jA N are contained in
a refund, including some compensation, and must rebook their the set Uj. Since the flights jA N n N D must depart as originally
own flight. Traditional airline recovery approaches ignore passen- scheduled, the set of copies only includes the copy representing
gers until the end of the recovery process, which is equivalent to the original departure time, i.e. U j ¼ f0g. For the disruptable set of
only modelling case (iii). Consequently, schedule recovery deci- flights, jA ND , the set of allowable copies contains additional
sions, such as delaying and cancelling flights, may be suboptimal copies vi ; i ¼ 1; …; n to represent flight delays, i.e. U j ¼ f0; v1 ; …; vn g.
for passenger flow. This paper presents an effective and efficient For convenience, the notation jv is used to define the flight-copy pair
modelling approach for case (i) in the integrated problem. The representing the delayed departure of flight j given by copy v. Using
modelling approach employed involves reallocating passengers to this definition, all flight-copy pairs for the original and recovery
D
alternative operating flights and calculating the flight cancellation schedule are contained in N^ and N^ respectively, which includes all
costs to include this reallocation process. flights and origination and termination locations.
The IRP-PR introduces three types of decision variables, described Every flight in the schedule is defined by an origin and destina-
as the aircraft, crew and cancellation variables. The aircraft and crew tion with respective departure and arrival times. Implementing
variables detail the movement of these resources through the flight delays using flight copies provides a number of alternative departure
126 S.J Maher / Computers & Operations Research 57 (2015) 123–137

and arrival times for each flight j given by the set Uj. The nodes of the permissible overnight airport. Second, an upper bound on the
connection network are represented by each flight-copy pair con- number of working and flying hours are set to manage the effects
D
tained in N^ and the arcs are supplied by the connections between of fatigue on crew. In the IRP-PR the maximum number of flying
each of the nodes. A feasible connection between two flight-copy hours and working hours are set to 8 and 13 respectively, which
pairs ðiu ; jv Þ exists if (i) the origin of flight j is the same as the have been adopted from Barnhart et al. [8]. Finally, disruptions can
destination of flight i, and (ii) the difference between the departure occur at any point during the day, as such it is important to
of flight-copy jv and the arrival of flight-copy iu is greater than a account for the work history performed by each crew prior to the
minimum ground time. For aircraft the ground time is called the turn disruption. This is achieved by constructing personalised duties for
time and for crew it is the sit time. Since there are different conditions each crew that satisfy the working rules with respect to any work
that dictate the feasible connections for crew and aircraft, the performed before and after the disruption.
connection networks for the disruptable schedule are defined as CK
and CR respectively. 2.3. Cancellation variables—passenger recovery by reallocation
The minimum turn time for aircraft is shorter than the mini-
mum sit time for crew, as such all permissible connections for Passenger reallocation is a process whereby passengers booked
crew are also permissible for aircraft. The set of connections with a on cancelled flights are rebooked onto alternative flights. These
ground time between the minimum turn and sit times are defined alternative flights are termed reallocation options and may still be
as short connections, given by E ¼ C R n C K . While the ground time operating as scheduled or delayed. This process accounts for the
for these connections is less than the minimum sit time, crew can number of seats remaining on the reallocation options and the
be assigned to connection ðiu ; jv Þ A E if an aircraft is also assigned to total length of delay to be experienced by the reallocated passen-
this connection. This is a critical aspect of the integration between gers. The consideration of passenger reallocation allows the
the aircraft and crew recovery problems which greatly improves optimisation process to make cancellation and delay decisions
upon the solution quality achieved by the comparative sequential while attempting to improve passenger flow. In the traditional,
recovery process. sequential airline recovery process, passenger reallocation is per-
formed using a greedy approach that will be shown to be sub-
2.2. Flight string variables optimal in Section 4.2.1.
The conventional approach used to model flight cancellations is
A flight string formulation introduced by Barnhart et al. [7] is with a single variable per flight that equal 1 to indicate a cancella-
used to model the crew and aircraft variables in the IRP-PR. Flight tion and 0 otherwise. In the IRP-PR, the variable zjp is defined to
strings constructed for all crew K and aircraft R form a sequence of indicate the cancellation of flight j but also the reallocation of
connected flights from a carry-in to a carry-out activity. All flight passengers to alternative flights as given by scheme p. Since there
strings representing the flow of each crew k and aircraft r through are multiple reallocation options following the cancellation of each
the connection network are contained in the sets Pk and Pr flight j, the set Pj is defined to contain each passenger reallocation
respectively. The variables xkp equal 1 if flight string p is operated scheme p. The cost of a reallocation scheme p for flight j is djp,
by crew k and the variables yrp are similarly defined for aircraft. The which attempts to quantitatively describe the reduced passenger
cost of a flight string for crew k, ckp, represents the additional work satisfaction resulting from the delay experienced by the reallocated
performed, measured in terms of work hours, as a result of a passengers and not providing alternative travel arrangements to the
disruption and the subsequent recovery actions. Similarly, the cost remaining.
of a flight string for aircraft r, crp, describes the cost of delaying any To conveniently describe the passenger reallocation schemes in
flights that are contained in string p. the IRP-PR, the sets Npost
i and Nj are defined to contain the alternative
The parameters akjp and arjp equal 1 if flight j is included on flight flights for disrupted passengers. The alternative flights for passengers
string p for crew k and aircraft r respectively. To indicate the use of following the cancellation of flight i are identified by having the same
k
short connection ðiu ; jv Þ on flight string p, the parameters biu jv p and origin-destination (O-D) pair as i and depart within a time window
r
biu jv p equal 1 for crew k and aircraft r respectively. Also, ajp and arv
kv
jp commencing from the departure time of i. Using this definition, the
equal 1 to indicate that flight j on string p is delayed by a length of set Npost
i contains all flights j that are travel alternatives for passen-
time given by copy v for crew k and aircraft r respectively. In gers booked on flight i. Conversely, the set Nj contains all flights i for
addition to defining a sequence of flights performed throughout which flight j is a travel alternative.
the day, an aircraft flight string also specifies the termination Each of the reallocation schemes p for the cancellation vari-
airport with the parameters ortp that equal 1 if the flight string p for ables can be likened to the solution of a knapsack problem. For
aircraft r terminates at overnight airport t. example, following the cancellation of flight i the size of the
An important aspect of crew flight strings is the legality with knapsack is given by the number passenger booked on that flight,
respect to various work rules. There are numerous rules that Pax(i), and the items to enter the knapsack are the available seats
dictate the construction of legal crew flight strings, related to r j ¼ Maxcap PaxðjÞ on alternative flights j A N posti . In the model
duties, pairings and schedules. A review of the crew scheduling constraints, the number of passengers reallocated to the flight-
process, including a detailed overview of the relevant crew work copy pair jv following the cancellation of flight i on reallocation
rules is given in Barnhart et al. [8]. Since the IRP-PR is solved for a scheme p is given by hvijp.
single day schedule, only the crew duty rules are considered in
this paper. 2.4. Recovery policies
The crew duty rules modelled in the IRP-PR are station
compatibility requirements and an upper bound on the number The recovery policies implemented in the IRP-PR are flight
of flying and working hours. First, the station compatibility delays and cancellations, crew deadheading, the use of reserve
requirements state the airports where crew may start and finish crew and the generation of new crew duties and aircraft routes.
their work day. In particular, crew are employed from one of many This set of recovery policies is representative of the actions taken
crew bases located throughout the network. To ensure station by airlines in practice. In Section 2.1, the concept of flight copies is
compatibility, it is ideal for crew to originate and terminate their presented as the approach used in the IRP-PR to model flight
duties at the same base. However, this is not always achieved as a delays. The cost of flight delays attempts to model the effect of the
result of the flight schedule design and crew may terminate at a disruption on passenger satisfaction. This is included in the
 S.J Maher / Computers & Operations Research 57 (2015) 123–137 127

w
objective function through the cost of aircraft flight strings, crp. An ∑ ∑ vrw r
lp yp r Al 8 l A L; 8 w A W; ð8Þ
unfortunate effect of performing flight delays is that airport arrival r A Rp A P r

capacities, the number of permitted arrivals in a given time period,

k r
may be violated. This aspect related to recovery is considered by ∑ ∑ biu jv p xkp  ∑ ∑ biu jv p yrp r 0 8 ðiu ; jv Þ A ED ; ð9Þ
k A K p A Pk r A Rp A P r
defining L as the set of all airports and W as the set of time
windows that partition the complete recovery period. Thus, the
jp xp  κ j
vþ
number of aircraft arrivals at airport l A L within time window ∑ ∑ akv k
 ∑ ∑ arv r
jp yp ¼ 0 8 jA ND ; 8 v A U j ; ð10Þ
k A K p A Pk r A Rp A P r
w A W is restricted to Aw l .
As a result of flight cancellations the flow balance of the ( )
schedule is destroyed requiring a lower bound to be imposed on
jp xp  κ j
vþ v
∑ ∑ akv k
r j  ∑ ∑ hijp zip Z 0 8 j A ND ; 8 v A U j ;
the number of aircraft that must terminate at each overnight base. k A K p A Pk j
iAN pAP i

This lower bound is set from the solution to the planning stage and ð11Þ
is given by ∑r A R o rt . The parameter o rt describes whether the
planned flight string for aircraft r terminates at airport t within
∑ xkp þ νk ¼ 1 8 k A K n K res ; ð12Þ
the recovery period. p A Pk
The recovery of crew introduces two specific recovery policies,
namely deadheading and the use of reserve crew. Within the crew ∑ xkp r 1 8 k A K res ; ð13Þ
deadheading recovery policy, there are two different approaches p A Pk
implemented for the IRP-PR, involving the transportation of crew
as passengers to either (i) return them to their crew base immedi- ∑ yrp ¼ 1 8 r A R; ð14Þ
ately following the start of a disruption, or (ii) reposition them to p A Pr

continue the operation of disrupted duties. The first approach is

called deadheading crew back to base, which is modelled by the xkp A f0; 1g; 8 k A K; 8 p A P k ; yrp A f0; 1g; 8 r A R; 8 p A P r ; ð15Þ
variables νk that equal 1 to indicate crew k is transported directly
back to base immediately following the disruption. In the IRP-PR, zjp A f0; 1g; 8 jA N D ; p A P j ; νk A f0; 1g; 8 k A K; ð16Þ
deadheading crew back to their crew base is performed at a cost of
gDHB. The second approach permits crew to deadhead within a duty κ vj þ Z 0; κ vj  Z0 8 j A N D ; 8 v A U j : ð17Þ
to another airport to continue the operation of other disrupted
duties. This deadheading action is modelled in the IRP-PR by the The objective of the IRP-PR is to minimise the additional
variables κ vj þ that count the number of crew deadheading on flight operational cost associated with the recovery of crew and aircraft
copy jv. The dummy variables κ vj  are introduced to ensure that the and the reallocation of disrupted passengers. This involves mini-
number of crew deadheading on flight-copy jv is exactly one less mising the costs related to crew deadheading, flight delays and
than the number of crew assigned to that flight. The cost of cancellations and any work performed by crew in addition to that
deadheading one crew member on a single flight leg within a duty originally planned. Another common objective for recovery pro-
is given by gDHD. blems is to minimise the number of changes between the recovered
The use of reserve crew is a very important aspect of crew and the planned solution [27]. This is implemented in the IRP-PR by
recovery, providing additional resources to operate the perturbed subtracting a bonus from the cost of crew pairings and aircraft routes
flight schedule. Reserve crew k A K res  K are employed in the for each connection ði; jÞ used in the recovery solution that is also
recovery process to avoid the regular crew exceeding the allow- used in the planning solution.
able working hours and to minimise the number of cancelled Crew recovery in the integrated problem is described by con-
flights. The reserve crew are located at each crew base and can be straints (2)–(4), (12), and (13). The flight coverage constraints (2)
assigned flight strings that originate from their base. The cost of a ensure that every flight is either assigned to a crew group or
duty p for reserve crew k, ckp, represents the cost of any work cancelled. The term ∑v A U j κ vj þ in (2) is the total number of crew
performed including a fixed penalty to minimise their use. deadheading on flight j. The inclusion of constraint (3) ensures that
The mathematical model of the IRP-PR is given by, at most one crew group is operating flight-copy jv, i.e. all other crew
assigned to jv are deadheading, which is required to implement
min ∑ ∑ ckp xkp þ ∑ ∑ g DHD κ vj þ þ ∑ g DHB νk
k A K p A Pk j A ND v A Uj kAK crew deadheading within a duty. Cancellation is not permitted for
the carry-out activities, so each must be assigned a crew group, as
þ∑ ∑ crp yrp þ ∑ ∑ djp zjp ; ð1Þ
r A Rp A P r j A ND p A P j defined by constraints (4). Constraints (12) ensure that all originally
planned crew, K n K res , are either assigned to a recovered flight
s:t: ∑ ∑ akjp xkp  ∑ κ vj þ þ ∑ zjp ¼ 1 8 j A ND ; ð2Þ string or deadheaded back to their crew base. By contrast, reserve
k A K p A Pk v A Uj p A Pj crew are only required in exceptional circumstances, which is
indicated by the inequality in constraints (13).
jp xp  κ j þ κ vj  ¼ 1
vþ
∑ ∑ akv k
8 jA ND ; 8 v A U j ; ð3Þ The structure of aircraft recovery in the IRP-PR, given by
k A K p A Pk (5)–(8) and (14), is similar to the crew recovery detailed above.
Flight coverage is described by constraints (5) and the coverage of
∑ ∑ akjp xkp ¼ 1 8 j A N Kout ; ð4Þ carry-out activities is enforced by constraint (6). The difference
k A K p A Pk
between the flight coverage constraints of crew (2) and aircraft (5)
is that the cancellation variables are not required in the latter. This
∑ ∑ arjp yrp r 1 8 j A ND ; ð5Þ is due to the consistency of flight cancellations being enforced by
r A Rp A P r
other constraints in the model. An additional feature for the
∑ ∑ arjp yrp ¼ 1 8 jA N Rout ; ð6Þ aircraft recovery problem is constraint (7), which specifies the
r A Rp A P r minimum number of aircraft flight strings that must terminate at
each overnight airport within the recovery period. Also, the number
∑ ∑ ortp yrp Z ∑ o rt 8 t A T R; ð7Þ of arrivals permitted at each airport within a given time window is
r A Rp A P r r AR restricted by constraints (8). Finally, all aircraft that are assigned a
128 S.J Maher / Computers & Operations Research 57 (2015) 123–137

routing in the planned solution must be assigned a recovered flight to identify variables that are expected to improve the upper bound
string, which is captured by constraint (14). given by the current formulation of the RMP.
The integration of crew and aircraft in the IRP-PR is described To aid the discussion in this section the notation for the dual
by the use of short connections and consistency of flight delays. As formulation of the RMP will be provided. The dual variables for the
stated in Section 2.1, it is permissible for crew to use a short flight coverage constraints (2) and (4) for crew; and (5) and (6)
connection if that same connection is used by aircraft, which is for aircraft; are defined as αK ¼ fαKj ; 8 j A ND [ N Kout g and αR ¼
described by constraints (9). Additionally, in the integrated pro- fαRj ; 8 j A ND [ N Rout g respectively. The dual variables for the surplus
blem it is important to ensure that the amount of delay on each crew count constraints (3) are defined by η ¼ fηuj ; 8 j A N D ; 8 u A U j g.
flight is consistent between crew and aircraft, which is enforced by The dual variables ϵ ¼ fϵt ; 8 t A Tg are defined for the aircraft end-
constraints (10). of-day location constraints (7). The dual variables for the aircraft
movement constraints (8) are given by β ¼ fβl ; 8 l A L; 8 w A Wg.
w
As a variation on the conventional modelling approach, the
cancellation of flight i is indicated by ∑p A P i zip ¼ 1 in (2). The The dual variables for the short connection constraints (9), delay
contribution of this modelling approach is that if ∑p A P i zip ¼ 1, consistency constraints (10) and the passenger reallocation con-
flight i is cancelled and the booked passengers are reallocated to straints (11) are given by ρ ¼ fρij ; 8 ði; jÞ A ED g, γ ¼ fγ uj ; 8 j A N D ;
8 u A U j g and λ ¼ fλj ; 8 jA ND ; 8 u A U j g respectively. Finally, the dual
v
alternative flights, which are contained in Npost i , as described by
variables δ ¼ fδ ; 8 k A Kg and δ ¼ fδ ; 8 r A Rg are defined for the
K k R r
constraint (11). This set of constraints is in the form of big-M
constraints, only permitting the reallocation of passengers to flight crew and aircraft assignment constraints, (12), (13), and (14),
j if crew are also operating that flight. The total number of passengers respectively. The column generation subproblems for the crew
that can be reallocated to flight j is given by r j ¼ Maxcap  PaxðjÞ. To and aircraft variables are formulated as shortest path problems,
the best of the author's knowledge this modelling approach for given in Section 3.1.1. Finally, the subproblem for the cancellation
passenger recovery has not previously been considered. variables can be formulated as a bounded knapsack problem,
which is described in Section 3.1.2.

3. Solution methods
3.1.1. Shortest path subproblem
Airline operations are very dynamic with the current state of The column generation subproblems for the crew duty and
the system changing almost every minute. In any transportation aircraft routing variables are described by
application, it is necessary for recovery problems to achieve high 8
<
quality solutions in very short runtimes. A general framework for k
c^ p ¼ min RecDutyCostðkÞ  ∑ αj akjp  ∑ ρiu jv bkiu jv p
k:
column-and-row generation is developed and applied in Maher pAP D
jAN [N K
ði ;j Þ A ED out u v

[20], demonstrating reduced runtimes for the integrated airline )

n o
recovery problem (IRP). Unfortunately, the IRP presented in [20] ηvj þ γ vj þr j λvj akv
jp  δ
k
 ∑ ∑ ; 8 k A K; ð18Þ
only applies a special case of the developed column-and-row jAN Dv A U
j

generation framework, namely the problem is formulated with

only one set of secondary variables and linking constraints. Hence, 8
<
solving the IRP does not adequately assess the potential of the r
c^ p ¼ minr ∑ ∑ crv arv  ∑ α ar  ∑ ϵ or  ∑ ∑ β vrw w

solution approach. The extension presented by the IRP-PR intro- pAP :j A ND v A U j jp jp j A ND [ NK j jp t A T t tp l A Lw A W l lp

out
duces additional variables and constraints related to the cancella-
)
tion variables that further increases the complexity of the recovery
problem. The inclusion of this set of variables and constraints þ ∑ ρ r
iu jv biu jv p þ ∑ ∑ γ v rv
j ajp 
r
δ ; 8 r A R: ð19Þ
ðiu ;jv Þ A ED j A ND v A Uj
formulates a problem to which the general column-and-row gen-
eration framework of [20] can be applied. The application of column- In problems (18) and (19), the minimisation problem is defined
and-row generation described in this section and the evaluation over the sets Pk and Pr that contain all possible flight strings for
presented in Section 4 is a contribution of this paper. crew k and aircraft r respectively. While this description indicates
k r
The column-and-row generation solution approach will be that additional variables can be identified from the sets P and P ,
presented in relation to the key components of column generation these sets represent variables included in the master problem,
and row generation. Section 3.1 will discuss the application of hence their reduced costs are at least zero. Therefore, all negative
column generation to solve the IRP-PR. This discussion is provided reduced cost variables identified by solving (18) and (19) will be
k r
assuming that the IRP-PR is solved by column generation alone. A contained in P k n P and P r n P .
focus of Section 3.1 is the column generation subproblem for the The set of all aircraft flight strings are contained in P r ; 8 r A R
cancellation variables, given in Section 3.1.2. Finally, Section 3.2 and are described as a sequence of connected flight between
will describe the application of the row generation procedure from carry-in and carry-out activities. The sets P r ; 8 r A R can be defined
the column-and-row generation framework developed in [20] to as the feasible region of a network flow problem. The connection
the IRP-PR. network used in the shortest path problem for aircraft is con-
structed with a single source node and multiple sink nodes,
D
3.1. Column generation described by N^ representing the nodes and CR representing the
arcs. In addition, maintenance is the only requirement that must
Column generation relies on the understanding that the vast be enforced in the construction of feasible flight strings. For
majority of variables are non-basic in the optimal solution. Thus, a aircraft requiring maintenance at the end of the day, all flight
restricted master problem (RMP) is initially formulated to contain strings are constructed to terminate at a maintenance station.
only a subset of all variables from the IRP-PR. This initial subset of Therefore, it is only necessary to propagate the single resource of
k r i
variables is denoted by P D P k , P DP r and P D P i for the crew, cost through the network. The cost of an aircraft flight string is
aircraft and cancellation variables respectively. The RMP is solved given by crp ¼ ∑j A ND ∑v A U j crv rv
jp ajp , which is the simple sum of the
to identify an upper bound on the original problem, which is delay costs for each flight operated by aircraft r. There are various
progressively improved with the addition of variables. The follow- standard solution algorithms available to solve this problem, many
ing sections will detail the column generation subproblems solved of which are described in Ahuja et al. [5].
 S.J Maher / Computers & Operations Research 57 (2015) 123–137 129

While a feasible flight string for crew is also a sequence of the IRP-PR, the size of the knapsack is given by the passengers on
connected flights, there are specific rules governing their construc- the cancelled flight i and the items are the seats on the alternative
tion. The crew duty rules modelled in the IRP-PR are described in flights contained in Npost i . Since flight delays are implemented
Section 2.2 and are enforced in the column generation subproblems. using flight-copies, it is possible for many identical flights to be
Each of the flight strings contained in Pk respect the origination and contained in Npost i that differ only by the departure time. The
termination location and the restriction on the maximum number treatment of this unique formulation of the column generation
of working and flying hours. As a consequence of the complex crew subproblem will be discussed in this section.
duty rules, the column generation subproblem for crew presents a Each cancellation variable defines a reallocation scheme pro-
more complex shortest path problem than the problem solved for viding alternative travel arrangements for passengers in the event
aircraft. of a flight cancellation. Ideally this modelling approach provides
The objective function for problem (18) represents the reduced an alternative travel arrangement for all disrupted passengers,
cost for crew duty variables. The cost of a crew duty is a max however this is not always possible. As such, passengers on
function related to the flying hours, fly, the total working hours, cancelled flights are partitioned into two categories, stranded
work, a minimum number of guaranteed hours, minGuar, and the and reallocated passengers. To model the cost of a flight cancella-
cost of the originally planned duty for crew k, OrigDutyCost(k). This tion dip, a cost is assigned to each of these categories. Namely, the
function is based off the crew duty cost presented by Barnhart parameter gCAN is defined as the cost of leaving a single passenger
et al. [8] and is given by stranded and gRA ij as the cost of reallocating a single passenger to

RecDutyCostðkÞ ¼ maxf0; maxffly; f d  work; minGuarg OrigDutyCostðkÞg; ð20Þ flight j following the cancellation of flight i. Both the costs given by
gCAN and gRA ij include a quantitative measure of passenger dissa-
where minGuar is set at 6 hours [8] and fd is a fraction which is tisfaction and loss of good will. The definition of dip for the realloca-
airline specific and is set at f d ¼ 5=8 [8] for the IRP-PR. The cost of a tion of passengers by scheme p following the cancellation of flight i is
recovered crew duty given by Eq. (20) is the cost associated with given by
any additional work performed as a result of a disruption and the  
v
respective recovery actions. In the case of reserve crew, the dip ¼ g CAN PaxðiÞ þ ∑ ∑ hijp g RA ij g
CAN
: ð21Þ
j A Npost v A Uj
original duty cost is set to zero and a fixed penalty is added to i

RecDutyCost(k). Eq. (21) consists of a fixed cost which assumes all passengers are
A benefit of column generation is that all feasible flight strings stranded and a variable cost that is dependent on the number of
contained in P k ; 8 k A K are not enumerated a priori but defined as passengers rebooked onto alternative flights.
the feasible region of a network flow problem. This conveniently The cost of a service cancellation given by Eq. (21) can be
allows for negative reduced cost paths to be identified by the applied in many different contexts. For example, in public trans-
solution to a shortest path problem. For the crew duty variables, port applications the fixed cost for cancelling a service, gCAN, can
the shortest path problem is solved on a connection network with be set to zero since all waiting passengers will travel on the next
one source node and multiple sink nodes, which is described by available service. However, due to overcrowding the variable cost,
D
N^ representing the nodes and C K [ ED representing the arcs. Since gRA
ij , can be very high. Given the similarities in the cost structure,
there are a numerous rules that must be reviewed to construct the proposed passenger recovery approach can also be applied.
feasible crew flight strings, a multi-label shortest path problem is Using the definitions of the dual variables given in Section 3.1,
used to track the resources of cost, flying and working hours. By the reduced cost of a cancellation variable for flight i is given by
tracking multiple resources through the connection network, an  
d ip ¼ g CAN PaxðiÞ  αKi þ ∑ ∑ hijp λj þ g RA
v v CAN
efficient frontier is constructed at each node by all labels that are ij  g : ð22Þ
j A N post v A Uj
propagated to there. The use of multiple labels in the shortest path i

problem allows the propagation paths that are suboptimal with It is observed that the first two terms of Eq. (22) are not
respect to cost but have a resource consumption that lies on the dependent on the reallocation of passengers to alternative flights.
efficient frontier. It is possible to adapt the algorithm used for aircraft Therefore, the objective function for the cancellation variables
to develop an efficient solution approach for the crew column column generation subproblem is defined only by the final term.
generation subproblem. The column generation subproblem for the cancellation variables
(CV-PSP) to identify passenger reallocation schemes for flight i is
General aircraft column generation subproblem given by
 
d~ ip ¼ max ∑ ∑ hijp g CAN  λj  g RA
v v
The column-and-row generation solution approach requires ij ; ð23Þ
j A N post v A Uj
the solution to the shortest path problem for aircraft to construct i

an optimal dual solution. It is observed in problem (19) that all v

dual variables, except for δ , are aircraft independent. Therefore,
R s:t: ∑ ∑ hijp r PaxðiÞ; ð24Þ
j A N post v A Uj
by setting δ ¼ δ ; 8 r A R, where δ ¼ maxr A R fδ g in (19), it is
r R R r i

possible to formulate a single, aircraft-independent shortest path v

problem, which is labelled PSPR . The solution to the PSPR is a lower hijp r r j wvij 8 jA N post
i ; 8 v A Uj; ð25Þ
bound on the minimum reduced cost flight string for all aircraft.
This problem is implemented in the row generation procedure, ∑ wvij ¼ 1 8 j A N post
i ; ð26Þ
v A Uj
which is described in Section 3.2.2.
hijp A Z þ
v
8 j A N post
i ; 8 v A Uj; ð27Þ
3.1.2. Cancellation variable subproblem
The modelling of the cancellation variables is unique in regards
wuij A f0; 1g 8 j A N post
i ; 8 v A Uj: ð28Þ
to the airline context and the column generation implementation.
While there are many application of column generation where the The problem described by CV-PSP shares many characteristics
subproblem is defined as one of the many knapsack problem with a bounded knapsack problem. The main difference between the
variants, there are few where the items to enter the knapsack are CV-PSP and a bounded knapsack problem arises from the multiple
duplicated multiple times with slightly different properties. For copies for flight j given by the set Uj. The multiple copies complicates
130 S.J Maher / Computers & Operations Research 57 (2015) 123–137

the formulation of the CV-PSP, requiring that at most one copy v for scaling, the set of constraints (11) can be restated as
each flight j is used as a reallocation option. This is enforced through
 ∑ ∑ h~ ijp zip Z 0
v
jp xp  κ j
vþ
the addition of the variables wvij and constraints (26). ∑ ∑ akv k
8 j A ND ; 8 v A U j ; ð34Þ
k A K p A Pk j
iAN pAP i
A more convenient formulation of the cancellation variable
where h~ ijp ¼ hijp =r j . Replacing constraints (11) with the scaled
v v
subproblem can be found by eliminating the need to consider
multiple copies for each flight. This is achieved for the CV-PSP by constraints (34) ensures that the coefficients of the crew duty
v0
identifying a single copy v0 for each flight j that satisfies λj ¼ variables are either zero or one in all rows of the IRP-PR. This
v0
minv A U j fλj g. Replacing ∑v A U j λj with λj and setting wij ¼ 1 for all
v u v0 replacement requires a modification to the reduced cost functions
flights j A N post
i forces the use of at most one copy for each flight. for the crew duty and cancellation variables.
The resulting problem is called the reduced pricing subproblem Additionally, this constraint modification significantly affects
(CV-PSPR), which is a bounded knapsack problem that is equiva- the structure of the column generation subproblem for the cancella-
tion variables. In particular, the scaled coefficients h~ ijp are no longer
v
lent to the CV-PSP. The CV-PSPR is given by
  integer, therefore integrality is not required for the decision variables.
d~ ip ¼ max
v0 v0
∑ hijp g CAN  λj  g RA
ij ; ð29Þ The scaled reduced cancellation variable pricing subproblem (CV-
j A Npost
i
PSPRS) is defined as
v0 v0
  
∑ hijp r PaxðiÞ; ð30Þ d~ ip ¼ max ∑ h~ ijp g CAN  g RA r j  λj ;
s:t: v0
ij ð35Þ
j A Npost
i j A Npost
i

v0
hijp rr j 8 j A N post ; ð31Þ 0
∑ h~ ijp r j r PaxðiÞ;
v
i s:t: ð36Þ
j A N post
v0 i
hijp A Z þ 8 j A N post
i : ð32Þ
h~ ijp A ½0; 1
v0
The following theorem will prove that the optimal solution to 8 j A N post
i : ð37Þ
the CV-PSPR is also optimal for the CV-PSP.
Since the CV-PSPRS is also a bounded knapsack problems formu-
Theorem 3.1. The optimal solutions to the CV-PSP and CV-PSPR are lated as a linear program, the optimal solution is found using a greedy
identical. heuristic. Similar to the CV-PSPR, the right hand side of constraints
(36) are all integer, hence only integer quantities of passengers are
Proof. Assume that an optimal solution to CV-PSP for a given flight reallocated to each available flight. The computational experiments
i has been found describing passenger reallocation scheme p with presented in Section 4 will discuss the improvements in the solution
0
an objective0 value d~ ip . The optimal solution is described by the runtimes for the IRP-PR achieved by applying the described scaling of
v
variables h ijp 40; 8 j A N post i , where v0 is the copy for flight j where the coefficients for the cancellation variables.
v0
wij ¼ 1.
v0
, such that λj0 4 minv A U j fλj0 g, and 8 j A N post
v
Let ( j0 A N post
i i n j0 ,
v0 0 3.2. Row generation
λj ¼ minv A U j fλj g. Therefore, the objective value for the optimal
v

solution to the CV-PSP can be written as

    The full column-and-row generation solution approach is
0 v0 v0
d~ ip ¼ ∑ h ijp g CAN  λj  g RA
v0 v0
ij þ h ij0 p g CAN  λj0  g RA
ij0 : ð33Þ described by Algorithm 3 in Maher [20]. The following sections
j A N post
i
nj0 will describe the application of the key steps of Algorithm 3 to the
v0 v0 IRP-PR. These steps include, (i) the formulation of the restricted
¼ 0. Let λj0 ¼ minv A U j fλj0 g and
v″ v
Since h ij0 p 4 0 ) ∑v A U j nv0 h ij0 p problem, (ii) computing an optimal dual solution, and (iii) the
0 0

since r j0 does not depend on v A U j0 another feasible solution to CV- addition of rows to update the restricted problem.
v v″ 0
PSP is given by setting h^ ijp ¼ h ijp ; 8 jA Npost ; 8 v A U j with h^ ij0 p ¼ h ij0 p
v v
i
v0 3.2.1. Formulation of the restricted problems
and h^ ij0 p ¼ 0. Modelling flight delays using flight copies is convenient regard-
v0
Now λj0 o λj0 , so the objective value for this new feasible solution
v″
ing the solution algorithms for the column generation subpro-
~d 00 is greater than d~ 0 . Hence, d~ 0 is not the objective value for the
ip ip ip blem; however, an unfortunate consequence is the large number
optimal solution to the CV-PSP, which is a contradiction to the of related constraints in the RMP. This is a result of the number of
original assumption. the delay consistency (10) and passenger reallocation constraints
Therefore, the optimal solution to the CV-PSP exists only when (11) being directly proportional to the number of flight copies.
0
λvj ¼ minv A U j fλvj g; 8 j A N post
i , which is the objective value given by The short restricted master problem (SRMP) is defined as a
the solution to the CV-PSPR. So the optimal solutions to the CV- further restriction of the RMP by eliminating a set of structural
PSP and CV-PSPR are identical. □ constraints. The SRMP is formulated for the IRP-PR to including
The CV-PSPR describes a bounded knapsack problem, hence a only a subset of all possible rows from the delay consistency (10)
greedy heuristic solves the linear relaxation to optimality. Addi- and passenger reallocation constraints (11) that form the RMP.
tionally, the right hand side of constraints (30) and (31) are Since the optimal integer solution to the IRP-PR selects exactly one
integer, so the solution to the greedy heuristic also provides the delay copy v0 for each flight j, it would be ideal to formulate the
integer optimal solution. SRMP with U j ¼ fv0 g; 8 j A N D . While it is not possible to identify
v0 ; 8 j A N D in the construction of the IRP-PR, the SRMP is formed
with the set U j D U j . The set U j provides an initial guess at the
Scaling of the passenger reallocation constraint coefficients optimal formulation of the IRP-PR that is progressively updated by
a row generation procedure. The elimination of rows to form the
The passenger reallocation constraints (11) can be likened to SRMP results in a problem that can be solved by column genera-
big-M constraints, which are well known for negatively affecting tion using the methods described in Section 3.1.
the ability to identify integer solutions. A method identified to The sets U j are initialised for each flight j by including the copy
improve the computational performance of the IRP-PR is to scale that represents the scheduled departure time, i.e. U j ¼ f0g. How-
the passenger reallocation constraints (11) by rj. By applying this ever, it is likely due to the delays associated with the disruption
 S.J Maher / Computers & Operations Research 57 (2015) 123–137 131

scenarios that for some flights j0 there exists no feasible connection from the implementation of these techniques, as such they are also
containing the flight-copy j00 . In these situations, an additional employed in the solution process of the IRP-PR. In the following
copy v0 is included in the set U j , such that j0v0 is the flight-copy pair section an additional enhancement technique is described that
that provides at least one feasible connection and the smallest uses properties of the row generation framework to improve the
amount of delay for j0 . convergence of the branch-and-price algorithm.
A key feature of column-and-row generation is the fixing of
variables in the column generation subproblems as a result of 3.3. Variable fixing
eliminating constraints. The variable fixings in the column gen-
eration subproblems ensure that all generated flight strings and The feasible regions of the network flow problems PSPk and
reallocation schemes have zero coefficients in the eliminated rows. PSPr conveniently describe the variables contained in Pk and Pr
The elimination of rows and the related variable fixings reduce the respectively. Such a network flow problem is formulated with arc-
size of the column generation master and subproblems respec- based variables wkiu jv p for crew (wriu jv p for aircraft) that equal 1 to
tively, having a positive effect on the solution runtimes. In indicate that connection ðiu ; jv Þ is used on flight string p. The
addition, it permits the property that the solution to the SRMP is elimination of rows to form the SRMP is coupled with the fixing of
an upper bound on the RMP and the IRP-PR. arc-based variables wkiu jv p ¼ 0 (wriu jv p ¼ 0) in the network flow pro-
blems, where v A U j n U j . This variable fixing can also be described as
the elimination of connections from the networks used to solve the
3.2.2. Row generation algorithm shortest path problems for crew and aircraft. As a result, the runtimes
The row generation procedure is described by Algorithm 2 in required to solve the column generation subproblems for the SRMP
Maher [20]. This algorithm involves two major stages, the calcula- are much shorter than that for the RMP.
tion of the optimal dual solution and identifying favourable rows The row generation procedure described in Section 3.2.2 involves
to include in the SRMP. the addition of rows to the SRMP and the unfixing of variables in the
In Maher [20], the RMP0 is constructed to contain all rows from column generation subproblems. This causes the size of the connec-
the RMP and all variables from the SRMP. In addition, a set of tion networks used for each shortest path problem to increase, having
artificial variables are introduced to populate the rows in the RMP0 a negative effect on solution runtimes. A variable fixing heuristic is
that are eliminated to form the SRMP. The construction of the proposed in this section that reduces the size of the column generation
RMP0 is critical in the row generation procedure, facilitating the subproblems using the optimal solution to the SRMP at the root node
calculation of an optimal dual solution using the solution to the SRMP. of the branch-and-bound tree. This heuristic assumes that the flight-
The rows contained in the RMP0 are partitioned into two groups, copy pairs used in the optimal solution of the LP relaxation for the
those common and uncommon with the SRMP. For the rows SRMP have a high probability of occurring in the integer optimal
common between the RMP0 and SRMP, the dual solutions can be solution. Given the high level of degeneracy in airline planning and
simply equated. A procedure is then executed to calculate the dual recovery problems, it is also assumed that the flight-copy pairs
solutions for the rows uncommon between the two problems. included on variables that have a reduced cost of zero may be used
The dual variables for the rows related to U j n U j for the two in the integer optimal solution. More formally, the set I ¼ fjv jc kp ¼
secondary variables in the RMP0 are given by γ 0 ¼ fγ vj ; 8 j A N D ;
0
k
0
0 4 akv
jp ¼ 1; 8 k A K; 8 p A P g defines the set of flight-copy pairs that
0
8 v0 A U j n U j g and λ ¼ fλj ; 8 j A N ; 8 v0 A U j n U j g. Since the dual
v D
are expected to be included on basic variable flights strings. This
0
variables contained in γ 0 and λ are related to the rows not identifies a subset of flight-copy pairs that can be used to eliminate
included in the SRMP, the solutions are calculated using the PSPR rows from the SRMP and reduce the size of the column generation
and CV-PSPRS respectively in step 2 of Algorithm 1 [20]. Similar to subproblems. From I, the set of connections included in the column
[20], the PSPR is used as a problem specific enhancement to reduce generation subproblems is defined as C ¼ fðiu ; jv Þjiu A I 4 jv A Ig, hence
the time required for the row generation procedure. Unfortunately, the variable fixing, wkiu jv ¼ 0; wriu jv ¼ 0; 8 ðiu ; jv Þ2
= C can be applied.
no such general column generation subproblem can be formed for This heuristic uses information from the optimal solution to the
the cancellation variables, requiring the CV-PSPRS to be solved LP relaxation of the IRP-PR as a proxy for the integer optimal
v0 0
once for each i A N j to calculate λj A λ . solution. As such, it is not guaranteed that the solutions providing
Favourable rows for the SRMP are identified by evaluating the the best lower and upper bounds for the IRP-PR will contain only
reduced costs of primary variables feasible for the RMP0 . For the the flight-copy pairs included in I. Consequently, it is possible that
IRP-PR, the primary variables are the crew duty variables, as such the integer optimal solution that is found using this heuristic will
problem (18) is solved in step 8 of Algorithm 2 [20]. To identify overestimate the true optimal solution. To provide a meaningful
crew duty variables feasible for the RMP0 , problem (18) is solved comparison of the solution quality by column generation and
with all possible delay copies given by Uj. If any negative reduced column-and-row generation, the optimality gap in this paper is
cost variables are found, the flight-copy pairs contained on the calculated using a lower bound set at the LP solution of the
related flight strings indicate the rows that must be included in the root node.
SRMP. In addition to adding rows to the SRMP, the related variable
fixings in the column generation subproblems are relaxed, increas-
ing the size of the connection network. If no negative reduced cost 4. Computational results
variables are found, the current solution for the SRMP is the
optimal solution to the IRP-PR. In an effort to improve the solution There are two distinct aims of the computational results presented
runtimes of the row generation procedure, a greater number of in this section, (i) demonstrate the benefits of considering passengers
rows to add to the SRMP are identified by solving problem (18) in the integrated recovery process, and (ii) provide an analysis of the
once for each k A K. This is a problem specific enhancement general column-and-row generation framework presented in Maher
technique that is also employed in Maher [20]. [20]. The modelling of passenger recovery in the IRP-PR involves the
A number of enhancement techniques for column-and-row reallocation of passengers to alternative flights when flight cancella-
generation are developed and evaluated in Maher [20]. In parti- tions are required as part of the recovery process. The benefits from
cular, a method to increase the number of rows added to the SRMP considering this passenger recovery approach is assessed against the
and a row warm-up procedure are discussed. The results pre- IRP from [20], with a comparison made in regards to runtime, cost and
sented by [20] demonstrate an improved runtime performance disruption statistics. At present, the only evaluation of the general
132 S.J Maher / Computers & Operations Research 57 (2015) 123–137

column-and-row generation framework presented in Maher [20] window is documented in Table 2. It is important to note that
involves solving a simplified model with a single set of secondary while the number of included flights is reduced, no approximation
variables and linking constraints. The IRP-PR is a more complex is made to reduce the number of affected crew and aircraft.
problem, as such the computational results will provide a more in- Flight delays and cancellations are two of the most effective
depth analysis of the framework developed in [20]. actions that are available to airlines in the recovery from disruptive
The measure of solution quality is based upon the optimality events. Flight delays have been modelled using the common flight
gap achieved at the termination of the maximum allowable runtime. copies technique, with 7 copies used for the IRP and IRP-PR to
This measure is important for the scenarios that fail to solve within discretise a maximum delay of 180 min. Attempting to provide a
this runtime, requiring the best found integer solution to be imple- true cost of delays, the estimate detailed in the EUROCONTROL
mented. In Section 3.3, a variable fixing heuristic is introduced which report by Cook and Tanner [14] has been used in these experiments.
potentially overestimates the lower bound of the IRP-PR. As stated, a In [14] the cost of delaying a full aircraft for a minute is estimated at
meaningful comparison between the column generation and column- $100 AUD.
and-row generation solution approaches is provided by calculating the The cost of cancelling a flight in the IRP-PR accounts for the
optimality gap as the relative difference between the LP solution at number of passengers reallocated to operating flights. If a passen-
the root node and the best found integer solution. Any reference to the ger is rebooked onto another flight operated by the airline, the
optimality gap hereafter will represent the gap calculated in this full revenue is retained minus any delay costs associated with
manner. waiting for a later departure time. It is common that the number
of available flying seats is less than the number of passengers
4.1. Description of data and disruption scenarios requiring reallocation, as such some passengers are not provided
alternative arrangements resulting in the full loss of ticket revenue
The data and disruption scenarios that are used for this problem for the airline.
are identical to those related to the point-to-point flight schedule The IRP solves a recovery problem integrating aircraft and crew
presented in Maher [20]. The motivation for using the same point-to- to fit within the commonly applied sequential recovery process.
point schedule is to provide consistency in the analysis of the two As part of this process, passenger recovery is undertaken after all
different recovery models. Therefore, only a brief description of the preceding stages have been solved. To directly compare the impact
of considering passenger reallocation in the integrated recovery
key features will be provided in this section.
The IRP and IRP-PR are evaluated using a single day flight problem, as in the IRP-PR, this reallocation process is undertaken a
posteriori using the solution to the IRP. This is achieved by employ-
schedule that contains 262 flights transporting 28,492 passengers
who are travelling on 48 aircraft operated by 79 crew groups. The ing a greedy approach to rebook passengers onto the next available
alternative flights, which are defined as having the same O-D pair
flight schedule services 20 airports, of which 12 are overnight
bases for aircraft and 4 are crew bases. Of the 20 airports, the and departing within a maximum allowable delay window.
This model is implemented in C þ þ calling SCIP 3.0.1 [4] to
majority of flights originate or terminate at only 2. A set of 16
disruption scenarios are generated for the evaluation of the IRP solve the integer program using CPLEX 12.4 as the linear program-
and IRP-PR, each representing a closure at one of these two major ming solver.
airports starting at 6 am, 7 am, 8 am or 9 am and lasting for a total
duration of 3 or 5 hours. Airport closure scenarios are selected for 4.2. Improvement in recovery achieved with passenger reallocation
evaluation purposes due to the significant associated recovery
costs. Throughout this section the scenarios will be referred to by The analysis of the IRP and IRP-PR aims to identify the effect
an identifying number as detailed in Table 1. modelling passenger reallocation has on the recovery costs, the
A full set of recovery policies, including the generation of new recovery actions taken and the solution runtime. The results are
aircraft routes and crew duties, flight delays and cancellations, produced by solving the two recovery models against the set of 16
crew deadheading and the use of reserve crew are implemented scenarios detailed in Table 1. The optimal solution to each of the
for the IRP and IRP-PR. The recovery actions are permissible models is identified when the relative difference between the
directly after the disruption occurs until the end of a selected primal and dual bounds is less than 5%.
recovery window. For these experiments the recovery window is
set at 6 hours, commencing after the affected airport is reopened. 4.2.1. Recovery costs
While the use of a recovery window approximates the complete The motivation for considering the alternative passenger recov-
recovery problem, this approach is consistent with the objective to ery model given by the IRP-PR is to reduce the recovery costs incurred
return operations back to plan as quickly as possible. The number while accommodating disrupted passengers. The relative difference in
of flights that depart within the closure period and recovery the recovery costs between the IRP and IRP-PR is given in Fig. 1. This
figure demonstrates that a significant reduction in the operating costs
of an airline can be realised by employing this alternative modelling
Table 1
technique. On average, using the IRP-PR to solve the recovery problem
Scenario numbers used in the presentation of results. The bracketed values indicate
two different closure durations (3 h, 5 h). reduces the operational costs of the airline by 22.98%.
The largest relative improvement in recovery costs is 42.89%,
Start time 6 am 7 am 8 am 9 am which is achieved in scenario 4. Comparing the solution for this
scenario given by the two alternative recovery models, there is a
Airport 1 (0, 8) (2, 10) (4, 12) (6, 14)
Airport 2 (1, 9) (3, 11) (5, 13) (7, 15)
marked difference in the number of flight cancellations. Specifi-
cally the IRP cancels 5 flights compared to 10 for the IRP-PR. Now,

Table 2
The number of disruptable flights for each scenario.

Scenario 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

D 150 151 149 150 147 145 150 149 182 183 185 186 184 182 184 183
jN j
 S.J Maher / Computers & Operations Research 57 (2015) 123–137 133

Fig. 1. The relative difference between the recovery costs of the IRP (x) and IRP-PR (y). The value of the bars is given by (x  y)/x.

this increase in the number of cancelled flights achieves a smaller This result has a direct positive impact on passenger satisfaction by
recovery cost for the solution to the IRP-PR since all except ensuring that a greater number of passengers arrive at their destina-
27 passengers are provided alternative travel arrangements. By tion within a reasonable time frame. The average on-time perfor-
contrast, the greedy reallocation approach applied using the mance of the airline is very similar using the solutions to the IRP and
solution to the IRP leaves 380 passengers stranded by the airline. the IRP-PR. The major difference between the two models is improved
The number of passengers stranded following a flight cancellation flow of passengers through the network achieved by solving the IRP-
is greater than the total number of seats on a single flight. This PR. On average, the solution to the IRP-PR disrupts 4376 passengers
demonstrates the magnitude of the disruption experienced by compared to 4988 for the IRP, providing a decrease of 612 passengers.
passengers as a result of solving the IRP. The ability of the IRP-PR to Given that the solution to the IRP-PR provides a smaller recovery cost
provide alternative travel arrangements for the majority of passengers than the IRP, the decrease in the number of disrupted passengers
on cancelled flight is common across all scenarios. Therefore, the demonstrates that this cost saving is not at the expense of passengers.
consideration of passenger reallocation allows the strategic cancella-
tion of flights to reduce operational costs of the airline.
4.2.3. Solution runtimes
The additional variables and constraints related to passenger
4.2.2. Cancellation and delay information
reallocation decisions in the IRP-PR increases the problem com-
The consideration of passengers in the recovery process aims to
plexity compared to the IRP. As a result, it is expected that longer
reduce the impact of cancellations and delays on passenger travel
runtimes are required to identify the optimal integer solution for
arrangements. In particular, directly modelling reallocation options
the IRP-PR. This is evident in Fig. 3 where the IRP is solved to
following flight cancellations ensures that passengers are routed to
optimality faster than the IRP-PR for 11 of the presented scenarios.
their destination with minimal delay. The reduction in recovery
The main difference between the solution process for each model
costs, presented in Section 4.2.1, benefits the airline, however this
is that the IRP-PR generally requires a greater number of branches
does not necessarily imply improved passenger satisfaction. For the
to converge to the integer optimal solution. Across all scenarios,
analysis presented in this section it is assumed that providing
the IRP-PR requires 66.25 nodes on average compared to 8.56 for
passengers with alternative travel arrangements directly improves
the IRP. This is a significant difference in the solution process that
passenger satisfaction.
has a great effect on the algorithm runtimes.
Fig. 2 presents the total number of disrupted passengers and
The slow convergence displayed by the IRP-PR can be explained
the effect of the disruption on their travel arrangements. The bars
further by analysing the optimality gap of the two models at the
are divided into three different groups, (i) the reallocated group are
root node, which presented in Table 3. This table shows that 4 of
the passengers on cancelled flights that are rebooked onto alter-
the scenarios are solved to optimality by the LP solution at the root
native operating flights, (ii) the delayed group includes all passen-
node for the IRP, where this does not occur in any scenario for the
gers that are booked on flights that depart later than scheduled,
IRP-PR. In addition, the optimality gap for the IRP-PR is signifi-
and (iii) the cancelled group are the passengers for which no
cantly larger than that for the IRP in all but one scenario, indicating
alternative travel arrangements are provided. For all passengers in
the potential requirement of a greater number of branches in the
the cancelled group, the tickets are refunded and the passengers
branch-and-price algorithm. This is to be expected since the passen-
must rebook themselves onto alternative flights. The passengers
ger reallocation constraints (11) are in the form of big-M constraints,
that are placed in this group are the most affected by the
which commonly display difficulties in identifying integer solutions.
disruption, resulting in poor passenger satisfaction and the poten-
The runtime comparison between the IRP and the IRP-PR
tial loss of future bookings.
demonstrate that the increased complexity introduced by the
The passenger reallocation approach developed for the IRP-PR
passenger reallocation decisions increases the solution runtime.
increases the number of cancelled flights, however this is com-
While this increase in solution runtimes is not ideal for an airline
bined with an increase in the number of passengers reallocated
recovery problem, the difference is not prohibitively large. Fig. 3
onto alternative flights. Since reallocation options are directly
demonstrates that the vast majority of scenarios for the IRP-PR are
modelled, flight delay decisions are made to ensure an adequate
solved within 1200 sec (20 min), which is within an acceptable
number of seats are available for the passengers on cancelled flights. It
runtime for practical use of the algorithm. Comparing Figs. 1 and 3,
is clear in Fig. 2 that by solving the IRP-PR more passengers are
it can be concluded that the potential gains from included
provided with alternative travel arrangements compared to the IRP,
passenger reallocation outweigh the resulting increase in runtime.
greatly reducing the number of cancelled passengers in each scenario.
134 S.J Maher / Computers & Operations Research 57 (2015) 123–137

Fig. 2. The number of reallocated, delayed and cancelled passengers from the IRP (bars) and IRP-PR (bars with hatching) solutions.

Fig. 3. The runtimes to solve the IRP (bars) and IRP-PR (stars) using column-and-row generation for each scenario with a maximum of 2700 sec.

Table 3
The optimality gap at the root node from solving the IRP and the IRP-PR.

Scenario 0 1 2 3 4 5 6 7

IRP 2083.49% 5305.47% 1825.2% 2957.86% 1735.34% 0.0% 0.0% 2276.56%

IRP-PR 3558.8% 6199.04% 4462.38% 5407.34% 4241.88% 3071.69% 3969.75% 4976.35%

Scenario 8 9 10 11 12 13 14 15

IRP 872.49% 0.0% 1201.56% 1238.49% 1469.05% 0.0% 3354.78% 3547.17%

IRP-PR 1392.51% 1916.24% 1654.11% 2195.46% 2459.03% 1516.11% 5755.05% 4772.27%

4.3. Evaluation of modelling approaches aim to improve the efficiency of the planned solution by permit-
ting shorter ground times for crew. Thus, it is common for
The IRP-PR incorporates a number of modelling approaches to scheduling decisions to be made by considering such “short
improve robustness and maintain the consistency of the recovered connections”, especially at smaller airports in the network. For
solution. In particular, short connection constraints force crew and such airports it may be necessary to include a subset of constraints
aircraft to operate the same connections, increasing the permis- (9) in the recovery problem formulation. To avoid causing infea-
sible set of connections, and a bonus parameter is introduced, sibilities by omitting the short connection constraints, those
which reduces the number of changes between the planned and required as a result of scheduling decisions are still included in
recovered solutions. Both of these approaches, while beneficial in the problem formulation.
an operations context, can result in suboptimal recovered solu- Fig. 1 presents the relative difference in the recovery costs from
tions. This section demonstrates the effect that implementing each solving the IRP-PR with and without the short connection con-
of these approaches has on the optimal recovered solution. straints. Since the formulation of the IRP-PR given by (1)–(17)
includes the short connection constraints, this is used as the base
case. The results in Fig. 4 demonstrate significant increases in the
4.3.1. Effect of short connection constraints optimal recovery costs as a result of omitting the short connection
This section evaluates the impact of omitting the short con- constraints from the problem formulation. This is not surprising,
nection constraints from the IRP-PR. Short connection constraints since the omission of the short connection constraints reduces the
 S.J Maher / Computers & Operations Research 57 (2015) 123–137 135

Fig. 4. The relative difference between the standard implementation of the IRP-PR (with short connection constraints) (x) and the IRP-PR without short connection
constraints (y). The value of the bars is given by (y  x)/y.

IRP-PR recovery cost comparison using different bonus values

Bonus Values
12 1 10 100 1000
Relative difference in recovery cost (%)

10

-2

-4
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Scenario Number

Fig. 5. The relative difference in the recovery costs between solving the IRP-PR with a bonus of 0.01 (x) and bonuses of 1, 10, 100, and 1000 (y). The values in the figure are
calculated by (y  x)/x. For improved presentation the maximum relative difference is capped at 12%.

number of available connections for crew in the recovered solu- introducing a bonus parameter, however small, to reduce the
tion. So, including short connection constraints in the IRP-PR number of changes that are made during recovery is beneficial.
improves the robustness of the resulting solution and reduces The results presented in Section 4 solve the IRP-PR using a
the overall costs. bonus parameter of 0.01. This small bonus is used to prevent the
The largest relative difference in recovery costs is given by generation of a completely new set of pairings and routes, but still
scenario 10, which is due to the model failing to identify a good permitting sufficient flexibility to achieve a recovered solution that
optimality gap in the allocated runtime. Additionally, it is observed is close to optimal. Fig. 5 presents the relative difference in the
for the larger scenarios (8 – 15) that omitting the short connection recovery costs between the base case and that from applying
constraints from the IRP-PR results in an increase to the solution bonuses greater than 0.01.
runtimes. In this set of scenarios, the time limit is reached for four The results presented in Fig. 5 demonstrate much variation
and there is only one where the runtime is shorter than for the with the use of different parameters across the set of scenarios.
standard implementation of the IRP-PR. This suggests that the However, there appears to be a general trend where solving the
inclusion of the short connection constraints forms a better LP IRP-PR with larger bonus parameters, i.e. 100 and 1000, results in a
relaxation, reducing the time required to identify the optimal recovery cost increase. This is to be expected, since the use of
integer solution. bonus parameters reduces the flexibility of the recovery problem.
There are two different effects resulting from this reduced flex-
ibility, (i) the construction of suboptimal crew pairings and aircraft
4.3.2. Effect of the bonus parameter routes, and (ii) an increase in the number of delayed and cancelled
Solving the IRP-PR identifies a set of crew pairings and aircraft flights. The first of these effects is observed in scenario 1, where
routes to operate on the perturbed flight schedule. As a result of the cost of crew increases substantially as a result of increases in
the schedule perturbations, the generated pairings and routes may the bonus parameter. This indicates that the recovered solution for
be vastly different to those constructed in the planning stage. This scenario 1 forces the crew to work longer hours than originally
situation, while providing the optimal solution, may present planned. Since the crew costs increase, little change in the number
difficulties for the operations controllers in the practical imple- of delayed and cancelled flights are observed. Alternatively,
mentation the resulting solution. Hence, it is assumed that scenario 6 demonstrates a decrease in the crew costs as the bonus
136 S.J Maher / Computers & Operations Research 57 (2015) 123–137

parameter increases but an increase in the number of delay PR when solved with column-and-row generation. These result are
minutes. This increase in delay minutes is a consequence of the encouraging since it supports the results presented in Maher [20].
model attempting to construct crew pairings and aircraft routes This suggests that the general column-and-row generation frame-
without much deviation from the planned solution. Hence, the work is also very effective on problems with multiple secondary
main recovery policy available in this situation is flight delays. variables and linking constraints.
The runtime results in Fig. 6 indicate that higher quality solutions
are achievable using column-and-row generation compared to column
4.4. Evaluation of column-and-row generation
generation with shorter runtimes. The runtimes presented in Fig. 6
show that column-and-row generation outperforms column genera-
A general column-and-row generation framework to solve
tion in 12 of the presented experiments. This provides a relative
problems with multiple secondary variables and linking con-
improvement in solution runtimes of 24.35%, significantly improving
straints is developed and evaluated in Maher[20]. However, the
upon the results given by Maher [20]. It is clear through the explicit
IRP presented in [20] is a special case of problems to which
evaluation of the solution runtimes, column-and-row generation
the column-and-row generation framework can be applied. Since
significantly outperforms a standard column generation approach.
the IRP-PR is modelled with two sets of secondary variables and
linking constraints, this paper presents an evaluation of the
column-and-row generation framework against a more general
4.4.2. Runtime enhancement techniques
problem formulation. The following sections provide a comparison
A number of enhancements techniques designed to improve
between standard column generation and column-and-row gen-
the solution runtime and quality of column-and-row generation
eration and the assessment of enhancement techniques.
are proposed in Section 3. In particular, a variable fixing heuristic
and the reformulation of the cancellation variables have been
4.4.1. Comparison between column generation and column-and-row developed specifically for the IRP-PR. However, it is possible to
generation adapt the variable fixing heuristic to other implementations of the
The runtime required to solve the IRP-PR using column gen- column-and-row generation framework.
eration and column-and-row generation is presented Fig. 6. This The variable fixing heuristic restricts the allowable delay copies
figure demonstrates a reduction in solution runtimes for the IRP- used to construct columns, therefore reducing the feasible region

Fig. 6. The runtimes to solve the IRP-PR for each scenario with a maximum of 2700 sec. This figure compares the solution approaches of column-and-row generation (bars)
and column generation (stars).

Fig. 7. Time required to solve scenarios by employing different enhancement techniques, specifically the variable fixing heuristic and/or the cancellation variable scaling
enhancements. A maximum runtime of 2700 s (45 min) is applied.
 S.J Maher / Computers & Operations Research 57 (2015) 123–137 137

of the IRP-PR. This technique attempts to reduce the runtimes of Future work on this problem involves integrating the passenger
the subproblem and LP solves in each iteration of the column reallocation approach into a broader passenger recovery scheme
generation algorithm. The modelling of the cancellation variables in an attempt to reduce the impact of flight delays. There is an
in this paper is a unique approach for considering passenger realloca- expectation that further reductions in recovery costs and improve-
tion for the integrated recovery problem. Section 3.1.2 details the ments in passenger satisfaction can be achieved with the greater
scaling of the coefficients for the crew duty and cancellation variables consideration of passenger flows in airline recovery problems.
in the reallocation constraints (11). Since the reallocation constraints
(11) are in the form of big-M constraints, this scaling to form
constraints (34) is proposed to alleviate the common issues related Acknowledgements
to identifying integer solutions.
Fig. 7 details the time required to solve the scenarios with the SJM is supported by the Australian Research Council Centre of
IRP-PR, implementing the variable fixing heuristic and/or the Excellence for Mathematics and Statistics of Complex Systems
cancellation variable scaling enhancement techniques. It is inter- (MASCOS) and an Australian Postgraduate Award.
esting to note that all except one scenario is solved to optimality
within the maximum runtime of 2700 sec (45 min). The best References
runtime performance of the IRP-PR is given when the heuristic is
used in isolation or both enhancements are used. As expected, the [1] Challenge ROADEF 2009: disruption management for commercial aviation.
URL: http://challenge.roadef.org/2009/en/. Access date: 3rd September 2014.
best overall runtime result is achieved when both of the enhance-
[2] Abdelghany A, Ekollu G, Narasimhan R, Abdelghany K. A proactive crew recovery
ment techniques are implemented. decision support tool for commercial airlines during irregular operations. Ann
The enhancement techniques attempt to form the SRMP that is Oper Res 2004;127(23):309–31.
more easily solvable by branch-and-price. Comparing the frontiers [3] Abdelghany KF, Abdelghany AF, Ekollu G. An integrated decision support tool
for airlines schedule recovery during irregular operations. Eur J Oper Res
created by each of the enhancement techniques, the greatest 2008;185(2):825–48.
improvement is observed for scenarios requiring greater than [4] Achterberg T. SCIP: solving constraint integer programs. Math Program
500 s runtime. Now, the average time to solve the root node when Comput 2009;1(1):1–41.
[5] Ahuja R, Magnanti T, Orlin J. Network flows: theory, algorithms, and applica-
no enhancements are used is 173.25 s, with a maximum of 242 s. tions. Upper Saddle River, NJ, USA: Prentice Hall; 1993.
By comparison, the average time to solve the root node when all [6] Bard J, Yu G, Argüello M. Optimizing aircraft routings in response to ground-
enhancements are employed is 164.53 s. This suggests that the ings and delays. IIE Trans 2001;33(10):931–47.
[7] Barnhart C, Boland NL, Clarke LW, Johnson EL, Nemhauser GL, Shenoi RG. Flight
enhancements aid in improving the convergence to the integral string models for aircraft fleeting and routing. Transp Sci 1998;32(3):208–20.
optimal solution for the larger scenarios. [8] Barnhart C, Cohn A, Johnson E, Klabjan D, Nemhauser G, Vance P. Airline crew
scheduling. International series in operations research and management
science. In: Hall RW, editor. Handbook of Transportation Science, 56. USA:
5. Conclusions Springer; 2003. p. 517–60.
[9] Bisaillon S, Cordeau J-F, Laporte G, Pasin F. A large neighbourhood search
heuristic for the aircraft and passenger recovery problem. 4OR 2011;9(2):139–57.
This paper presents a novel approach for considering passen- [10] Bratu S, Barnhart C. Flight operations recovery: new approaches considering
gers in an integrated airline recovery problem through the model- passenger recovery. J Sched 2006;9(3):279–98.
[11] Cao J, Kanafani A. Real-time decision support for integration of airline flight
ling of cancellation variables. The cancellation variables have been cancellations and delays Part I: mathematical formulation. Transp Plann
modelled as knapsack variables to describe the possible reallocation Technol 1997;20(3):183–99.
options for passengers in the event of a flight cancellation. This [12] Cao J, Kanafani A. Real-time decision support for integration of airline flight
cancellations and delays Part II: algorithm and computational experiments.
modelling approach provides a simple and effective method to Transp Plann Technol 1997;20(3):201–17.
consider passengers in recovery. To achieve fast solution runtimes [13] Clausen J, Larsen A, Larsen J, Rezanova NJ. Disruption management in the
for the IRP-PR, the column-and-row generation solution approach is airline industry-concepts, models and methods. Comput Oper Res 2010;37
(5):809–21.
applied. The computational experiments presented in this chapter [14] Cook A, Tanner G. European airline delay cost reference values. URL: 〈http://
provide an extensive evaluation of the general column-and-row www.eurocontrol.int/documents/european-airline-delay-cost-reference-va
generation framework developed by Maher [20]. lues〉; 2011.
[15] Eggenberg N, Salani M, Bierlaire M. Constraint-specific recovery network for
The passenger reallocation approach developed for the IRP-PR is
solving airline recovery problems. Comput Oper Res 2010;37(6):1014–26.
demonstrated through experiments to greatly reduce the recovery [16] Jafari N, Zegordi SH. Simultaneous recovery model for aircraft and passengers.
costs achieved by the IRP. Further, the solution to the IRP-PR achieves J Frankl Inst 2011;348(7):1638–55.
a greater passenger flow through the network as a result of strategic [17] Jarrah AIZ, Yu G, Krishnamurthy N, Rakshit A. A decision support framework
for airline flight cancellations and delays. Transp Sci 1993;27(3):266–80.
delay decisions. The benefits achieved by solving the IRP-PR are [18] Lettovsky L. Airline operations recovery: an optimization approach [Ph.D. thesis].
realised by the airline and passengers through a reduction in costs Georgia Institute of Technology; 1997.
and the magnitude of disruption. [19] Lettovsky L, Johnson E, Nemhauser G. Airline crew recovery. Transp Sci
2000;34(4):337–48.
The computational experiments present a comparison between [20] Maher SJ. Solving the integrated recovery problem using column-and-row
column generation and column-and-row generation, demonstrat- generation. Transp Sci; 2014. Accepted for publication.
ing significant runtime improvements achieved by the latter. In [21] McCarty L. Preemptive rerouting of airline passengers under uncertain delays
[Ph.D. thesis]. The University of Michigan; 2012.
addition, a number of enhancements have been proposed and the [22] Petersen J, Sölveling G, Johnson E, Clarke J, Shebalov S. An optimization
results present a reduction in the solution runtime of the IRP-PR approach to airline integrated recovery. Transp Sci 2012;46(4):482–500.
by improving the convergence to the integer optimal solution. The [23] Rosenberger JM, Johnson EL, Nemhauser GL. Rerouting aircraft for airline
recovery. Transp Sci 2003;37(4):408–21.
variable fixing heuristic is developed using characteristics of the [24] Stojković M, Soumis F. An optimization model for the simultaneous opera-
column-and-row generation solution approach, as such it is tional flight and pilot scheduling problem. Manag Sci 2001;47(9):1290–305.
possible to implement such a heuristic for similar problems where [25] Stojković M, Soumis F. The operational flight and multi-crew scheduling
problem. Yugosl J Oper Res 2005;15(1):25–48.
column-and-row generation is applied.
[26] Stojković M, Soumis F, Desrosiers J. The operational airline crew scheduling
Passenger recovery in the IRP-PR is a simple and novel approach problem. Transp Sci 1998;32(3):232–45.
that is designed to reduce passenger dissatisfaction resulting from [27] Thengvall BG, Bard JF, Yu G. Balancing user preferences for aircraft schedule
flight cancellations. While this modelling approach reduces the recovery during irregular operations. IIE Trans 2000;32(3):181–93.
[28] Wei G, Yu G, Song M. Optimization model and algorithm for crew manage-
number of cancelled and disrupted passengers, it does not have a ment during airline irregular operations. J Combinator Optim 1997;1
significant effect on the average delay experienced per passenger. (3):305–21.