为了翻译文章

Solving the integrated airline recovery problem using

column-and-row generation

Stephen J Maher

School of Mathematics and Statistics, University of New South Wales, Sydney NSW 2052, Australia.

Abstract

Airline recovery presents very large and difficult problems requiring high quality solutions within

very short time limits. To improve computational performance, various solution approaches have

been employed, including decomposition methods and approximation techniques. There has been

increasing interest in the development of efficient and accurate solution techniques to solve an inte-

grated airline recovery problem. In this paper, an integrated airline recovery problem is developed,

integrating the schedule, crew and aircraft recovery stages, and is solved using column-and-row

generation. A general framework for column-and-row generation is presented as an extension of cur-

rent generic methods. This extension considers multiple secondary variables and linking constraints

and is proposed as an alternative solution approach to Benders’ decomposition. The application

of column-and-row generation to the integrated recovery problem demonstrates the improvement in

the solution runtimes and quality compared to a standard column generation approach. Column-

and-row generation improves solution runtimes by reducing the problem size and thereby achieving

faster execution of each LP solve. As a result of this evaluation, a number of general enhancement

techniques are identified to further reduce the runtimes of column-and-row generation. This paper

also details the integration of the row generation procedure with branch-and-price, which is used to

identify integral optimal solutions.

Key words: airline recovery, column generation, row generation

1

Introduction

Disruptions are very common in the airline industry, greatly impacting the realised operational perfor-

mance. To mitigate the effect of these disruptions, intervention by the airline is necessary to maintain

the many operational requirements of aircraft and crew. Consequently, any disruption results in a sig-

nificant increase to an airlines operational costs related to additional crew overtime and increased fuel

usage. Due to the significant associated costs, the use of efficient and accurate recovery processes is of

great importance to the airline industry.

1

1

INTRODUCTION

2

The airline recovery problem, similar to the planning problem, is a very large and complex problem

commonly broken into a number of smaller, more tractable sequential stages. These stages are broadly

categorised as schedule, aircraft, crew and passenger recovery, also defining clear boundaries for research

in this area. The complete recovery process is a sequential problem, where each stage is solved to

optimality and fixed for use in subsequent stages. Analogous to the planning problem, this typically

results in suboptimal, or even infeasible, solutions due to little interaction between the stages. The

integration of two or more stages in the planning problem has been demonstrated to provide higher

quality solutions [15, 25, 29], as such there is a similar expectation for airline recovery. The focus of this

paper is an integrated airline recovery problem, integrating the schedule, aircraft and crew recovery.

A critical constraint on the airline recovery problem is the allowable time limit to find an optimal

solution. Generally the solution to the airline recovery problem is required within minutes of a disruptive

event, prompting the development of many solution approaches to achieve this. The proposed solution

approaches to achieve improved runtime performance varies between each of the recovery stages. The

integration of multiple recovery stages presents very difficult and complex problems, which are commonly

solved with the use of decomposition techniques. Unfortunately, the decomposition techniques employed,

such as Benders’ decomposition, do not provide a guarantee of integral optimality. A contribution of

this paper is the development of a column-and-row generation framework as an exact solution approach

to solve the integrated airline recovery problem.

1.1

Airline recovery literature

The most common method employed to improve the runtimes of the aircraft recovery problem is through

the network design. There are three classes of network design that have been employed, the time-

line, time-band and connection networks. The time-line network is a very popular approach used for

the aircraft recovery problems, with Jarrah  et al.  [20] presenting an early example using this network

description. The work of [20] is extended by Cao and Kanafani [12, 13], integrating the two models

developed by [20] and implementing additional recovery policies. The use of the time-line network is

developed further by Yan and Yang [40], presenting a unique design that concisely describes the effect

of airline disruptions on the planned schedule. The time-band network is presented by Argüello [6] to

approximate the flight network by aggregating activities at airports into discrete time bands. Therefore,

the number of nodes used to define the recovery network is reduced. This network description has

been demonstrated by Bard  et al.  [8] and Eggenberg  et al.  [16] to improve the solution runtimes of the

aircraft recovery problem. Finally, Rosenberger  et al.  [31] present an aircraft recovery problem using the

classical connection network, which provides a concise description of the recovery flight schedule. Given

the potential size of this problem, the authors employ a heuristic to select a minimal number of aircraft

to include in the model for possible rerouting.

The crew recovery problem is a very complex and difficult problem having a significant impact on the

1

INTRODUCTION

3

operational cost of an airline. Similar to the aircraft recovery problem, a number of unique approaches

have been proposed to improve solution runtimes. These approaches involve reducing the problem size

by selecting only a subset of crew and flights, fixing the flight schedule by using the solution to the

aircraft recovery problem (part of the sequential recovery process), or implementing only a selection of

recovery policies.

Solving the crew recovery problem with a fixed flight schedule is first presented by Wei  et al.  [39],

attempting to repair any pairings affected by a schedule disruption. Fast solution runtimes are achieved

with the use of a branch-and-bound heuristic designed with consideration to the actions of an AOCC.

Another approach using a fixed flight schedule is the crew rescheduling problem proposed by Stojković  et

al.  [36]. In [36], the problem size is reduced by defining a recovery window to identify a set of disruptable

flights. Finally, Medard and Sawhney [24] and Nissen and Haase [28] present crew recovery problems

that are solved using heuristic approaches.

The crew recovery problem permitting the use of flight delays and cancellations is more complex

problem than the comparable fixed flight schedule problems. An example of such a problem is presented

by Stojković and Soumis [34, 35] as extensions of Stojković  et al.  [36], introducing flight delays and

constructing individual pairings for each crew member. Lettovsky  et al.  [23] present a crew recovery

problem that introduces the use of flight cancellations, extending Johnson  et al.  [21]. In [23], a number

of approaches to reduce the computational time of the recovery algorithm are proposed, including a

search scheme to identify the disruptable crew and compact storage for the generated columns. Finally,

a novel approach to the crew recovery problem is developed by Abdelghany  et al.  [1], partitioning flights

by their resource independence. The partitioning process reduces the solution runtimes by defining a

series of distinct recovery problems that are more readily solvable than the original problem.

Many solution approaches involve the approximation of the recovery problem to improve runtimes.

For the aircraft recovery problem, these approximations are made of either the network or the equipment

included in the model. Similarly, the runtimes for the crew recovery problem are reduced through the use

of heuristics or the selection of included crew. Column-and-row generation is presented in this paper as

an alternative, exact solution approach to improve runtimes. The results will show that column-and-row

generation solves the IRP in short runtimes without requiring any approximation of the affected crew

and aircraft. These results will provide a lower bound on the solution runtime reduction for practical

applications, which can be improved further using approximation techniques.

The sequential process used to solve the airline recovery problem generally results in suboptimal

solutions due to fixing the solution at each stage in the process. The PhD thesis of Lettovsky [22] is

an early attempt to develop a recovery model integrating multiple stages. This model integrates all

aspects of the recovery process; schedule, aircraft, crew and passenger recovery problems; and solved

using Benders’ decomposition. Further exploring the idea of employing Benders’ decomposition for the

integrated recovery problem, Petersen  et al.  [30] present a model that shares some of the characteristics

1

INTRODUCTION

4

of Lettovsky [22]. Petersen  et al.  [30] describe the implementation of this integrated recovery model,

providing an evaluation against a set of major disruptions. The reported runtimes are within the

specified goal of 30 minutes for the selected set of disruption scenarios. Extending the novel approach

by Abdelghany  et al.  [1], Abdelghany  et al.  [2] develop a recovery model integrating aircraft, pilots

and flight attendants. The results from experiments demonstrate significant delay reductions that are

achievable within very short runtimes.

Unfortunately the partitioning process of Abdelghany  et al.  [2] and the use of Benders’ decomposition

by Lettovsky [22] and Petersen  et al.  [30] does not guarantee integer optimality of the integrated recovery

problem. A contribution of this paper is the development of a general column-and-row generation

framework that is applied to problems commonly solved by Benders’ decomposition. Further, column-

and-row generation is an exact approach that provides a guarantee of near integer optimal solutions for

the integrated airline recovery problem.

1.2

Column-and-row generation

Column-and-row generation is a solution approach that extends standard column generation to reduce

problem complexity and improve solution runtimes. This solution approach involves the simultaneous

generation of variables and structural constraints. A key feature of column-and-row generation is the

reduction in the size of the master problem through the elimination of constraints. The problem size

reduction achieves runtime improvements through faster LP solves and quicker execution of the column

generation subproblems.

There have been a number of generic schemes developed for column-and-row generation [18, 26, 33],

however these schemes do not directly apply to the integrated recovery problem considered in this

paper. Fragioni and Gendron [18] and Sadykov and Vanderbeck [33] present schemes to solve mixed

integer programs by dynamically generating structural constraints. The solution process in these two

approaches involves identifying an upper bound from the linear relaxation and a lower bound from the

Lagrangian subproblem. A feature of the reformulations in [18] and [33], is the ability to ignore the dual

variables associated with the missing constraints without any loss of correctness in the algorithm. This

is not the case for the integrated recovery problem, requiring a row generation procedure to calculate

the dual variables of the missing constraints. The column-and-row generation approach presented here

is similar to that of Muter  et al.  [26], however the integrated recovery problem does not satisfy all

three assumptions required to apply their solution technique. As a contribution to the column-and-

row generation approach, this paper extends Muter  et al.  [26] by considering problems with multiple

secondary variables and linking constraints. This extension requires additional processes in the row

generation procedure to ensure the accurate calculation of an optimal dual solution. The framework

developed in this paper is presented algorithmically to succinctly describe the implementation.

Column-and-row generation is employed to solved various applied integer programming problems.

1

INTRODUCTION

5

Example applications of this solution approach are Zak [42] with the multi-stage cutting stock problem,

Avella  et al.  [7] solving a time-constrained routing problem and Muter  et al.  [27] to solve a robust

crew pairing problem. While these papers present implementations of column-and-row generation there

is little evaluation against a standard column generation approach. A contribution of this paper is

such an evaluation using the integrated airline recovery problem as an example. The integrated airline

recovery problem is a large-scale, real-world optimisation problem that provides an appropriate test bed

for the column-and-row generation framework. This evaluation will demonstrate the improvement in

solution runtime and quality compared to column generation and identify a number of enhancements that

contribute to the general column-and-row generation approach. The enhancement techniques identified

are a variation on the number of rows added in the row generation procedure and a row warm-up

process. Since the framework developed in this paper extends Muter  et al.  [26], the evaluation performed

demonstrates the strength of the framework by [26]

The various implementations of column-and-row generation have very little discussion regarding row

generation in the branch-and-price algorithm. For example, Zak [42] only solves the LP of the multi-stage

cutting stock problem using column-and-row generation, this is also the case in Wang and Tang [38] for

the batch machine scheduling problem. Also, Frangioni and Gendron [17, 18] solve the multicommodity

capacitated network design problem with column-and-row generation at the root node and then employ

branch-and-bound to identify integer solutions. Sadykov and Vanderbeck [33] mention the use of their

approach in a branch-and-price algorithm, however no details are provided regarding the implementation.

Finally, the framework by Muter  et al.  [26] is designed to solve large-scale linear programming problems,

as such the integration with a branch-and-price algorithm is not considered. A contribution of this paper

is a clear description of the application of row generation in a branch-and-price algorithm.

1.3

Outline of paper

This paper presents an integrated airline recovery problem, integrating the schedule, aircraft and crew

recovery stages. The integrated airline recovery problem is used in this paper as a real-world example

to demonstrate the ability of column-and-row generation to improve solution runtimes and quality.

The master problem for the integrated recovery problem is presented in Section 2, which is solved

using column generation to provide benchmark results. A general framework for the column-and-row

generation solution approach is presented in Section 3. The application of column generation and column-

and-row generation is discussed in Section 4, including the description of enhancement techniques. To

demonstrate the benefits of solving the integrated recovery problem by column-and-row generation, a

comparison with the results produced using column generation is made in both solution quality and

runtime. These results are presented in Section 5. The conclusions provided in Section 6 aim to present

the technique of column-and-row generation as an alternative solution method for integrated airline and

transportation problems.

2

INTEGRATED RECOVERY PROBLEM - IRP

6

2

Integrated Recovery Problem - IRP

The integrated recovery problem (IRP) attempts to minimise the costs associated with flight delays and

cancellations and the additional cost of crew following a schedule disruption. This problem is formulated

to integrate the schedule, aircraft and crew recovery problems. The link between the aircraft and crew

recovery problems in the IRP is provided by the flight cancellation and delay decisions and specific flights

allocated to each aircraft and crew.

The notation used to describe the IRP is introduced in Tables 1 and 2. For this problem, the set of all

crew is given by K, indexed by k, and the set of all aircraft is given by R, indexed by r. The set K also

includes all available reserve crew K ^res . As an extension on current techniques, the solution approach

for the IRP demonstrates an efficient algorithm that includes all crew and aircraft, as defined by K and

R respectively. Using all crew and aircraft allows the optimal allocation of all available resources.

K

is the set of all planned and reserve crew in the model, indexed by k

K ^res

is the set of all reserve crew, K ^res  ⊂ K

R

is the set of all aircraft in the model, indexed by r

P ^k , P ^r

is the set of all strings p assigned to crew k or aircraft r respectively

N

is the set of all flights j

T ^K , T ^R

is the set of crew bases/overnight airports t for crew and aircraft respectively

N ^D

is the set of all disruptable flights j, N ^D  ⊆ N

is the set of all carry-in activities j, flights and origination nodes, for crew and aircraft respectively,

N ^K

in  ,

in

N ^K

in  ⊂N∪T K  and

in  ⊂N∪T R

is the set of all carry-out activities j, flights and termination nodes, for crew and aircraft respectively,

N ^K

out ,

^out  N ^K

out  ⊂N∪T K  and

out  ⊂N∪T R

U _j

is the set of all delay copies v for flight j ∈ N

is the set of all nodes in the connection network defined by flight-copy pairs j v  , representing flights,

N

origination and termination nodes

is the set of disruptable nodes in the connection network defined by flight-copy pairs j v  , representing

N D

flights, origination and termination nodes

C ^K , C ^R

N

is the set of all connections (i u , j v  ), i u , j v  ∈

for crew and aircraft respectively

E

is the set of short connections E = C ^R \C ^K

E ^D

is the set of disruptable short connections, E ^D  = {(i u , j v  ) ∈ E|i u  ∈

N ^D  ∨ j v  ∈

N ^D }

Table 1: Sets used in the IRP.

2.1

Recovery flight schedule and connection network

A single day flight schedule is used to describe and evaluate the IRP, with the set of flights in the

schedule given by N . A critical aspect of the model is the use of a recovery window that specifies the

time allocated to return the operations back to plan. The recovery window is described as a time period

which defines the set of disruptable flights N ^D  ⊆ N . The set N ^D  contains all flights that are primarily

2

INTEGRATED RECOVERY PROBLEM - IRP

7

x ^k

p

= 1 if crew k uses string p, 0 otherwise

y ^r

p

= 1 if aircraft r uses string p, 0 otherwise

c ^k

p ,

c ^r

p

= the cost of using string p for crew k or aircraft r respectively

a ^k

a ^r

= 1 if flight j is in string p for crew k or aircraft r respectively, 0 otherwise

jp ,

jp

b ^k

b ^r

= 1 if connection (i u ,j v ) is in string p for crew k or aircraft r respectively, 0 otherwise

i u  j v  p ,

i u  j v  p

a ^kv

a ^rv

= 1 if copy v for flight j is in string p for crew k or aircraft r respectively, 0 otherwise

jp  ,

jp

o ^r

= 1 if string p, assigned to aircraft r, terminates at airport t within the recovery window, 0 otherwise

tp

z _j

= 1 if the flight j is cancelled, 0 otherwise

d _j

= the cost of cancelling flight j

κ v+j

= the number crew deadheading on flight-copy j v

κ v−j

dummy variable for counting the number of deadheading crew on flight-copy j v

ν _k

= 1 if crew k deadhead back to their crew base from the start of the disruption period, 0 otherwise

g ^DHD

= the cost of deadheading crew on one leg within a duty

g ^DHB

= the cost of deadheading crew k back to their crew base

o rt

= 1 if the planned string for aircraft r terminates at airport t within the recovery window, 0 otherwise

Table 2: Variables used in the IRP.

affected by the disruption and those that depart after the disruption occurs, but before the end of the

specified time window.

Restricting the flights included in the IRP using a recovery window requires the activities performed

by crew and aircraft before and after this window to be fixed. This introduces the concepts of carry-in

and carry-out activities. A carry-in activity is a flight or origination airport that is assigned to a crew

or aircraft directly preceding the disruption, contained in the sets N ^K

in  and

in  respectively.Acarry-

out activity is a flight or termination airport assigned to a crew or aircraft immediately after the end

of the recovery window, contained in the sets N ^K

out  and

out  respectively.Now,thecarry-inactivity

describes an origination airport when a planned crew duty or aircraft routing begins after the start of

the disruptive event. In a similar manner, if a planned crew duty or aircraft routing concludes before

the end of the recovery window, the carry-out activity is defined as the terminating airport.

To achieve an efficient solution approach for the IRP, the recovery policy of flight delays is imple-

mented using flight copies. The flight copies technique involves the multiple duplication of each flight

contained in N with each copy assigned a progressively later departure time. For each flight j ∈ N the

set of discrete copies is given by U _j  , and a flight-copy pair is described by j _v  , where v ∈ U _j  . Since the set

of all flights N is partitioned into disruptable and non-disruptable sets, the set of discrete flight copies

requires a different definition for each partition. As such, for all non-disruptable flights, j ∈ N \N ^D  ,

only one copy is included to represent the original scheduled departure, i.e. U _j  = {0}. Further, for all

disruptable flights, j ∈ N ^D  , the set of flight copies U _j  contains a copy representing the original departure

time and at least one additional copy to represent some delay on that flight. It is important to note that

the nodes representing origination and termination airports are treated as non-disruptable flights. This

2

INTEGRATED RECOVERY PROBLEM - IRP

8

definition is made for convenience in discussing carry-in and carry-out activities.

The connection network used for this model is described using the flight-copy representation for each

node in the network. The set of all nodes is represented by

N

= {j _v  |j ∈ N ∧ v ∈ U _j  }, detailing all

flight-copy pairs that exist for each disruptable and non-disruptable flight. Using the same notation,

the set of all disruptable nodes is given by

N ^D  = {j _v |j ∈ N ^D  ∧ v ∈ U _j }. The connection network for

this problem is defined by a set of nodes, representing flight-copy pairs, and a set of arcs as connections

between the nodes. A connection between two flight-copy pairs (i _u , j _v  ), i _u , j _v  ∈

N

is feasible if i) the

destination of flight i is the same as the origin of flight j and ii) departure of flight-copy j _v  occurs after

a specified amount of time following the arrival of flight-copy i _u . All feasible connections for crew are

contained in the set C ^K  and require a  minimum sit time  between the arrival of i _u  and the departure of

j _v . Similarly, all feasible connections for aircraft contained in C ^R  require a  minimum turn time  between

the connecting flights.

2.2

Aircraft routes and crew duties

The modelling approach for the IRP is based upon the string formulation introduced by Barnhart  et

al.  [9]. In the IRP, a flight string is defined as a set of connected flights from a carry-in to a carry-out

activity. Using the flight-copy notation of each node for this model, any reference to flight j without

specifying a copy v collectively states all flight-copy pairs associated with that flight. So, the parameters

a ^k

jp  anda jp  specifywhetherflightj,representinganyflight-copypairj v ,v∈U j ,

is included on string p

for crew k and aircraft r respectively. A flight string will either terminate within the recovery window,

by ending at a crew base or aircraft overnight port, or will terminate at a carry-out flight. If the flight

string assigned to an aircraft terminates within the recovery window, the parameter o ^r

tp  describesthe

terminating airport t for aircraft r. Flight cancellations are implemented as a recovery policy for the IRP,

so the flow balance of the original schedule is not maintained. To ensure enough aircraft are positioned

at each airport t to operate the schedule for the following day, the minimum number of required aircraft

must be specified. Now, the number of planned aircraft strings terminating at end-of-day locations

∑

within the recovery period is given by

r∈R  o ^t ,∀t∈T,whereo t  indicatesthataircraftrterminatesat

airport t. Therefore this expression defines the minimum number of aircraft required to terminate at

each end-of-day location t within the recovery window for the IRP.

Within the set of all aircraft connections, C ^R , it is common to have connection times less than the

minimum sit time for crew. These connections are called  short connections , and it is permissible for

crew to operate the two flights in succession, as defined by this connection, only if a single aircraft also

operates the same two flights. The set of all short connections are contained in E = C ^R \C ^K  , and

the subset of short connections that include flight-copy pairs in

N ^D  are contained in the set E ^D . If a

flight string includes two flight-copy pairs that form a short connection, the parameters b ^k

i u  j v  p  andb i u  j v  p

indicate the inclusion of connection (i _u , j _v  ) on string p for crew and aircraft respectively.

2

INTEGRATED RECOVERY PROBLEM - IRP

9

2.2.1

Legality of crew duties

There a numerous rules that dictate the construction of feasible flight strings for crew that must be

strictly adhered to in the recovery problem. A good review of the crew scheduling problem is presented

by Barnhart  et al.  [10], discussing the numerous crew rules. There are rules that are specific to the

construction of duties, pairings and schedules, however for a single day schedule, which is used for the

IRP, the most important rules that must be considered are the crew duty rules.

The origination and termination locations are critical in the construction of legal crew duties. Each

crew is employed at one of many crew bases throughout the network, so ideally a duty is constructed to

start and end at the same crew base. Unfortunately, the design of the flight schedule does not permit

all crew duties to terminate at their respective crew base, requiring an overnight stay at a permissible

airport. This rule is modelled in the IRP and if a crew duty originates from a permissible overnight

airport, it must terminate the required crew base.

The number of hours that a crew duty spans is crucial in managing the effects of fatigue. There

are two rules related to working hours modelled in the IRP, a maximum number of flying hours and

limit on the duration of the crew duty, which are set of 8 and 13 hours respectively. These are the most

important duty rules related to working hours, and are strictly adhered to in the IRP. Another important,

but complicated, rule is the 8-in-24 rule that requires crew to receive additional rest of more that 8 hours

flying is performed in a 24 hour period [10]. This rule must be considered in the construction of crew

pairings, however it may be reviewed in the recovery of crew duties. By ensuring that crew only perform

8 hours flying in a single day and given that individual crew members have planned work starting times,

the 8-in-24 rule will not be violated in most cases. In the event that this rule is violated by the solution

to the IRP, further adjustment can be made to the recovered duties at the end of the day to provide an

adequate amount of rest.

It is important to note that prior to a disruption crew may have performed part of a duty, consuming

allowable flying and working hours. The personalised schedules are respected in the recovery of crew

duties by originating each duty from a carry-in location and accounting for the working  history  prior to

the disruption. This ensures that the recovered crew duties, including the flights performed prior to the

disruption, respect the crew duty rules.

2.3

Recovery policies

The set of recovery policies implemented for the IRP include the generation of new aircraft routes and

crew duties, crew deadheading (transportation of crew as passengers), the use of reserve crew, and flight

delays and cancellations. In the column generation algorithm, feasible crew duties and aircraft routes

are generated for each crew and aircraft contained in K and R respectively. The length of delay that

is required on each flight is determined in the generation of these new flight strings by the selection of

2

INTEGRATED RECOVERY PROBLEM - IRP

10

flight-copy pairs. The parameters a ^kv

jp  anda jp  describethelengthofdelay,asspecifiedbycopyv,selected

for flight j on string p for the crew and aircraft respectively. The IRP also allows for the cancellation

of any flight that can not be covered by both crew and aircraft. Flight cancellations are defined in the

IRP through the use of the variables z _j  , which equal 1 to indicate flight j is cancelled at a cost of d _j  .

Following a disruption, it is not always possible for every crew member to operate the originally

planned flight strings as expected. Hence, the crew specific recovery policies of deadheading and the use

of reserve crew are employed. Crew deadheading transports crew as passengers to continue the operation

of disrupted flight strings. Two different types of deadheading are implemented in the IRP, deadheading

within a duty and deadheading back to base. The variables κ ^v+

j  areintroducedtocountthenumberof

crew that deadhead within a duty on flight-copy j _v  . In addition, the dummy variables κ ^v−

j  arerequired

to ensure that the number of crew deadheading on flight-copy j _v  is one less than the total number of crew

assigned to that flight-copy. The cost of crew deadheading within a duty is given by g ^DHD  . Alternatively,

the variables ν _k  indicate whether crew k deadhead to their crew base immediately following the start of

the disruption at a cost of g ^DHB  . As a result of recovery actions, it is not guaranteed that the set of

originally planned crew are able to operate the recovered schedule. Hence, reserve crew are employed to

operate crew duties from each crew base. Reserve crew are a limited, costly resource, as such their use

in the solution to the IRP is minimised by the addition of a penalty to the cost c ^k

p  foreachdutythey

perform.

The integrated recovery problem is presented in a compact formulation with variables x ^k

p  forcrew

and y ^r

p  foraircraftrepresentingfeasibleflightstrings.Thefulldescriptionofthisproblemispresented

below,

2

INTEGRATED RECOVERY PROBLEM - IRP

11

∑

∑

∑

∑

∑

∑

∑

∑

min

c ^k

g ^DHD κ ^v+

+ g ^DHB ν _k  +

c ^r

+ d _j z _j ,

(1)

p x ^p  +

j

p y ^p

k∈K p∈P ^k

j∈N ^D  v∈U j

k∈K

r∈R p∈P ^r

j∈N ^D

∑

∑

∑

s.t.

a ^k

κ ^v+

+z _j  =1

∀j ∈ N ^D  ,

(2)

jp x ^p  −

j

k∈K p∈P ^k

v∈U j

∑∑

a ^k

1

∀j ∈ N ^K

(3)

jp x ^p  =

out ,

k∈K p∈P ^k

∑∑

a ^r

jp y ^p  +z j  =

1

∀j ∈ N ^D  ,

(4)

r∈R p∈P ^r

∑∑

a ^r

jp y ^p  =

1

∀j ∈ N ^R

out ,

(5)

r∈R p∈P ^r

∑∑

∑

o ^r

o r

∀t ∈ T,

(6)

tp y ^p  ≥

t

r∈R p∈P ^r

r∈R

∑∑

a ^kv

1

∀j ∈ N ^D  , ∀v ∈ U _j  ,

(7)

jp  x ^p  −κ j+  +κ j−  =

k∈K p∈P ^k

(IRP)

∑∑

∑

∑

b ^k

b ^r

0

∀(i _u , j _v  ) ∈ E ^D  ,

(8)

i u  j v

p x ^p  −

i u  j v  p y ^p  ≤

k∈K p∈P ^k

r∈R p∈P ^r

∑∑

∑

∑

a ^kv

a ^rv

0

∀j ∈ N ^D  , ∀v ∈ U _j  ,

(9)

jp  x ^p  −κ j+  −

jp y ^p  =

k∈K p∈P ^k

r∈R p∈P ^r

∑

x ^k

p  +ν k  =

1

∀k ∈ K\K ^res ,

(10)

p∈P ^k  ∑

x ^k

1

∀k ∈ K ^res ,

(11)

p  ≤

p∈P ^k ∑

y ^r

p  ≤

1

∀r ∈ R,

(12)

p∈P ^r

x ^k

{0, 1}, ∀k ∈ K, ∀p ∈ P ^k  ,

y ^r

{0, 1}, ∀r ∈ R, ∀p ∈ P

^r ,

(13)

p  ∈

p  ∈

z _j  ∈ {0,1}, ∀j ∈ N ^D ,

ν _k  ∈ {0,1}, ∀k ∈ K,

(14)

κ ^v+

≥ 0, κ ^v−

≥ 0,

∀j ∈ N ^D  , ∀v ∈ U _j  .

(15)

j

j

The IRP is defined by equations (1)-(14) with the objective to minimise the cost of recovery for aircraft

and crew. The recovery costs include the cost of flight delays and cancellations, reserve crew, additional

crew duty costs and the cost of crew deadheading.

The coverage of flights within the recovery window by the crew and aircraft is enforced by constraints

(2) and (4). Additionally, the constraints (3) and (5) ensure that each carry-out flight is serviced by

crew and aircraft in the recovered solution. Respecting the carry-out flight coverage ensures that the

solution to the IRP positions the crew and aircraft to continue the activities as planned, following the

end of the recovery window.

This problem is solved for a single day schedule, so the aircraft are required to be positioned at

airports to maintain flow balance for the following days operations. Since all recovery actions occur

within the recovery window, the positioning of the aircraft must be considered before the conclusion of

2

INTEGRATED RECOVERY PROBLEM - IRP

12

this window. Two cases can occur in the recovery of aircraft; either i) an aircraft is assigned a carry-out

flight, allowing it to follow a planned routing to an end-of-day location, or ii) the recovered flight route

terminates within the recovery window requiring an end-of-day location to be specified. The minimum

number of aircraft required to terminate at each end-of-day location within the recovery window is

enforced through constraints (6).

The recovery policy of crew deadheading within a duty is implemented through the surplus crew count

constraints (7). This set of constraints ensure that crew deadheading is only permitted on flight-copy

pairs that are operated by at least one crew.

In the IRP, the integration between the crew and aircraft variables is described by the short connection

and delay consistency constraints, equations (8) and (9) respectively. The short connection constraints

(8) ensure that crew are only permitted to use connection (i _u , j _v  ) ∈ E ^D  only if an aircraft is also using

the same connection. The delay consistency constraints (9) ensure that the length of delay on any flight

in a feasible aircraft recovery solution is identical for the crew recovery solution, and vice versa. Since

there exists one delay consistency constraint for each flight-copy pair, this set of constraints grows very

quickly with an increase in the number of copies. It is on this set of constraints that the row generation

procedure is implemented to improve the solution runtime.

The number of crew and aircraft operating the recovered schedule is based upon the originally planned

duties and routings. Each crew that is assigned a duty from the planning stage must also be assigned

a duty in the IRP or deadheaded back to base, captured by constraints (10). This is not true for the

reserve crew since they former is not required to perform any duties during recovery, which is captured

by the inequality in constraints (11). Similar for crew, each aircraft that is assigned a flight route in the

planned solution must be assigned a flight route in recovery, which is given by (12)

It is common practice in both the sequential stage and integrated recovery problems to select a subset

of crew, aircraft and flights to reduce the problem size and improve solution runtimes. To improve the

computational performance of the IRP, the concept of a recovery window has been used to reduce

the number of flights included in the optimisation problem. While this provides an upper bound on

the optimal recovery cost, it is believed that this approach is realistic and consistent with a common

objective to quickly return operations to plan. Returning operations to plan is a consequence of the

carry-out activity coverage at the conclusion of the recovery window. Hence, the shorter the recovery

window, the quicker operations are returned to plan. To ensure that all reassignment and rerouting

options are available, the full set of crew and aircraft are used in the IRP. The selection of all crew and

aircraft for this problem demonstrates that fast solution algorithms are possible on larger data sets using

sophisticated solution techniques.

3

COLUMN-AND-ROW GENERATION

13

3

Column-and-Row Generation

This paper demonstrates the improvements in the solution runtimes and quality for the IRP achieved by

employing column-and-row generation. As discussed in Section 1.2, there have been a number of generic

column-and-row generation approaches that have been developed [18,26,33]. Unfortunately, these generic

approaches do not directly apply to the IRP, as such an alternative framework will be presented in the

following sections. The framework developed in this paper is an extension of Muter  et al.  [26] and

is presented as an alternative solution approach to Benders’ decomposition. The key contributions of

this framework are i) the extension of [26] to consider multiple sets of secondary variables and linking

constraints, ii) the application of column-and-row generation to a problem structure that is commonly

solved by Benders’ decomposition, and iii) the explicit evaluation against a standard column generation

approach, identifying various enhancement techniques.

3.1

Features of the Column-and-Row Generation Framework

The solution approach presented by Muter  et al.  [26] involves defining two restrictions on the original

problem, the restricted master problem (RMP) and the short restricted master problem (SRMP). The

RMP is constructed to contain all constraints from the original problem but with only a subset of all

possible variables. The SRMP describes a further restriction on the original problem, containing a subset

of all variables  and  constraints that form the RMP.

Both the RMP and SRMP are solved by column generation and the identical subproblems are used

in the solution process for the two problems. However, the SRMP is formed by eliminating structural

constraints from the RMP, so variable fixings in the column generation subproblems are used to restrict

the set of all possible columns. By applying the variable fixing to the column generation subproblem,

all feasible solutions to the SRMP are feasible for the RMP and the original problem.

Column-and-row generation is a solution approach proposed to solve large-scale linear programs. The

implementation of column-and-row generation in a branch-and-price framework to solve mixed-integer

programs has not been previously published. As a contribution of this paper, an efficient method for

applying column-and-row generation at each node in the branch-and-bound tree is explicitly discussed.

The critical aspects of the column-and-row generation procedure are the formulation of the RMP and

SRMP, and the method used to calculate an optimal dual solution to identify favourable rows. These

two features will be discussed in Sections 3.1.1 and 3.1.2 respectively. Finally, a general algorithm for

the row generation procedure will be presented in Section 3.1.3.

3.1.1

Formulation of the restricted problems

To provide an overview of column-and-row generation, the key features will be discussed with respect

to a generic problem. The example problem is formulated with a single set of primary variables and

3

COLUMN-AND-ROW GENERATION

14

multiple sets of secondary variables. In the problem description x is used to represent a single vector of

primary variables, and each vector of secondary variables is given by y ⁱ , i ∈ {1, 2, . . . , n}. The multiple

sets of secondary variables considered in this framework extends the generic framework presented by

Muter  et al.  [26].

This section focuses on problems solved by column generation, as such it is assumed that vectors x

and y ⁱ  contain only a subset of all possible variables from the original problem. The structure of the

original problem, and by extension the RMP, contains a set of constraints related to each x and y ⁱ , A ^x  and

A ⁱ  respectively, with a set of linking constraints A ^iL

x  andA yL  betweenxandy i .

The rows representing

the linking constraints are dynamically generated using the row generation procedure developed in this

section.

To construct the SRMP, it is necessary to redefine the variable vectors and constraint matrices used

to describe the RMP. Initially, a subset of linking constraints are eliminated from the RMP, which

involves removing rows from A ^iL

x  andA yL ,

resulting in the constraint matrices

A ^iL

x  an

A ^iL

y  respectively.

As stated previously, the elimination of rows from the RMP is coupled with the fixing of variables in

the column generation subproblem. This is required to prohibit the generation of columns with non-zero

elements in the eliminated rows. Consequently, the set of all possible variables that can be generated

for the SRMP is reduced, therefore the vectors x ⊂ x and y ⁱ  ⊂ y ⁱ  are defined. While all rows in the

matrices A ^x  and A ⁱ  are still present in the formulation of the SRMP, the restriction on the possible set

of variables requires the elimination of columns, hence the matrices

A ^x  and

A ⁱ  are defined.

The matrix representation of the RMP and SRMP is given by,

RMP

SRMP

∑

∑

min c ^x x + c ⁱ y ⁱ ,

(16)

min

c x  x +

c i  y ⁱ ,

(21)

i

i

s.t. A ^x x = b,

(17)

s.t.

A ^x x = b,

(22)

A ⁱ y ⁱ  = b ⁱ  ∀i,

(18)

A ⁱ  y ⁱ  = b ⁱ  ∀i,

(23)

A ^iL

0 ∀i,

(19)

A ^iL

A ^iL y i  = 0 ∀i,

(24)

x  x−A yL y ⁱ  =

x  x

y

x ≥ 0, y ⁱ  ≥ 0.

(20)

x ≥ 0,

y i  ≥ 0.

(25)

3.1.2

Calculation of dual solutions

To demonstrate the correctness of the column-and-row generation approach described here, an additional

problem is introduced, the RMP ^′ . This problem is formulated with all constraints from the RMP, but

only a subset of all possible variables. Thus, the formulation of the RMP ^′  is identical to the RMP at an

intermediate stage of the column generation solution process.

To describe the RMP ^′ , the matrices

A ^iL

y  areintroducedtocontainallrowseliminatedfromA yL  to

construct the SRMP. Additionally, a set of dummy variables y ⁱ  are created such that each y ∈ y ⁱ  has

3

COLUMN-AND-ROW GENERATION

15

a non-zero element in only one row of

A ^iL

This is an important condition on the construction of y ⁱ

y  .

that is exploited in Theorem 3.1.1. The dummy variables contained in y ⁱ  also have non-zero elements

in the rows of (18), thus the matrices

A ⁱ  must be introduced. Therefore, the matrix representation of

the RMP ^′  is given by,

∑

min

cx +

{c ⁱ  y ⁱ  + c ⁱ  y ⁱ },

(26)

i

s.t.

Ax = b,

(27)

A ⁱ  y ⁱ  +

A ⁱ  y ⁱ  = b ⁱ  ∀i,

(28)

(RMP ^′ )

A ^iL

A ^iL y i  = 0 ∀i,

(29)

x  x

y

−A˜ ^iL y i  = 0 ∀i,

(30)

y

x ≥ 0,

y i  ≥ 0,

y i  ≥ 0.

(31)

For each row in equations (27)-(30) there is an equivalent row in equations (17)-(19). It is assumed that

in this formulation the optimal solution to the RMP ^′  is not the optimal solution to the original problem.

Thus, additional columns with a negative reduced cost can be found by solving the column generation

subproblems for the primary and secondary variables.

By construction, an optimal primal solution to the SRMP is a feasible solution to the RMP ^′ . The

following lemma will prove that this feasible primal solution to the RMP ^′  is optimal.

Lemma 3.1.1.  The optimal primal solution to the SRMP is an optimal primal solution to the RMP ^′ .

Proof.  The constraints (30) force the variables y ⁱ  to be zero in any feasible solution of the RMP ^′ . As

such, the optimal primal solution to the RMP ^′  can be found by eliminating constraints (30) and solving

this problem with the variables y ⁱ  fixed to zero. Solving this modified form of the RMP ^′  is equivalent

to solving the SRMP.

There are two major steps in the procedure to calculate an optimal dual solution for the RMP ^′

using the solution to the SRMP. Firstly, the constraints (22)-(24) in the SRMP are identical to the

constraints (27)-(29), therefore the solutions to the related dual variables can be simply equated. The

second step involves finding the solutions for the dual variables related to the rows in (30), which are

the constraints eliminated to form the SRMP. This involves solving the column generation subproblems

for the secondary variables to accurately calculate the values of these dual variables.

Additional notation is required to describe the calculation of the dual variables for the rows eliminated

to form the SRMP. An index set Φ _i  is defined to reference each row u in the constraint matrix A ^iL

y  .

Φ _i

Extending this notation, the index set for the rows included in the SRMP is given by

and all eliminated

rows are contained in Φ _i \Φ i . Finally, the dual variables for each row in A ^iL

{θ ⁱ

y  isgivenbyθ i  =

u |u∈Φ i }.

This notation conveniently describes the rows which are eliminated or contained in the SRMP and the

3

COLUMN-AND-ROW GENERATION

16

dual values which must be computed. The value of θ ⁱ

u  iscalculatedfromtheminimumreducedcostĉ i

for a variable with a non-zero entry in row u of matrix

A ^iL

This is achieved by executing Algorithm 1.

y  .

Algorithm 1 Computing a feasible dual solution

1:  Assume that θ ⁱ

0 and force all feasible solutions to the column generation subproblem for the

u  =

secondary variables i to have a 1 in row u of

A ^iL

y  and0inallrowsv∈Φ i \Φ i ,v=u.

2:  Solve the column generation subproblem to identify variable ŷ that has the minimum reduced cost

ĉ i .

3:  Set θ ⁱ

u  =−ĉ i .

This calculation procedure relies on the structure of the RMP ^′  and the form of the dummy variables

that populate the rows u ∈ Φ _i \Φ i . The reasoning provided here draws upon the discussion presented

in the column-and-row generation framework by Muter  et al.  [26]. For ŷ, found by Algorithm 1, to

be eligible to enter the basis of the RMP ^′  implies that ĉ ⁱ  < 0. Since ŷ has a non-zero element in the

rows of

A ^iL

y  ,

the construction of the RMP ^′  forces ŷ = 0 upon addition of this column, resulting in a

degenerate simplex iteration. To avoid this situation, it is assumed that the minimum reduced cost for

all variables found using Algorithm 1 is at least zero. This requirement ensures that the dual solutions

that are computed for θ ⁱ  are feasible for the RMP ^′ . The following theorem will prove the feasibility of

the computed values for θ ⁱ  and that the resulting feasible dual solution is also optimal.

Theorem 3.1.1.  The dual solutions computed for  θ ⁱ  forms an optimal dual solution to the RMP ^′ .

Proof.  The first step of this proof is to show that the solutions calculated for the dual variables θ ⁱ  using

θ i

Algorithm 1 are feasible for the RMP ^′ . For this proof the variable

u  isassumedtohaveavaluethat

satisfies all dual constraints of the RMP ^′ . Additionally, the reduced cost of variable y ∈ y ⁱ  is given by

c i

y .

Algorithm 1 solves the column generation subproblem for the secondary variables i to identify ŷ that

has the minimum reduced cost ĉ ⁱ , assuming θ ⁱ

0. Comparing ŷ with the variables currently in the

u  =

RMP ^′ , there are two possible outcomes:

θ i

i) There exists a column y ∈ y ⁱ  identical to ŷ. Since y is identical to ŷ, ĉ ⁱ  = c ⁱ

In Algorithm 1,

y  −

u .

the value of θ ⁱ

hence c ⁱ

0. Therefore, setting θ ⁱ

u  issetto−ĉ i ,

y  −θ u +θ u  =

u  =−ĉ i  ensuresdualfeasibility.

ii) There exists a column y ∈ y ⁱ  that has a non-zero element in row u of

A ^iL

y  butisnotidentical

to ŷ. This implies that the variable θ ⁱ

u  existsinatleastonedualconstraint.Letsassumethatsetting

θ ⁱ

u  =−ĉ i  violatesaconstraintinthedualoftheRMP ′ .

This implies that c ⁱ

y  calculatedusingθ u  =−ĉ i

θ i

θ i

θ i

in place of

u  isnegative,i.e.

c i

y  −

u  +

θ ⁱ

u  <

0. Since ĉ ⁱ  + θ ⁱ

u  =

0, then ĉ ⁱ  > c ⁱ

y  −

u .

Now, step

θ i

2 of Algorithm 1 identifies ŷ that has the minimum reduced cost, so ĉ ⁱ  > c ⁱ

y  −

u  isacontradiction.

θ i

Therefore, c ⁱ

y  −

u  +θ u  ≥

0 must be true, hence the computed value for θ ⁱ

u  satisfiesalldualconstraints.

The first step of this proof has established that the dual solutions calculated for θ ⁱ  using Algorithm

1 are feasible for the RMP ^′ . Therefore, the calculated values for θ ⁱ  can be used as the dual solutions for

3

COLUMN-AND-ROW GENERATION

17

the constraints (30). A feasible dual solution for the RMP ^′  is then simply constructed by equating the

solutions to the dual variables representing constraints (27)-(29) to the dual solutions of the SRMP.

The second step proves that the feasible dual solution constructed for the RMP ^′  is also optimal.

Firstly, it is stated in Lemma 3.1.1 that the optimal primal solution to the SRMP is also optimal for the

RMP ^′ . In addition, the solutions to the dual variables representing constraints (27)-(29) are equated to

the dual solutions of the SRMP. Since the right hand side of the constraints represented by equations

(30) are zero, the value of the respective dual variables do not affect the optimal objective function value.

It follows that the dual objective function value for the RMP ^′  is identical to the dual objective function

value of the SRMP. Given that the dual objective function value of the SRMP is equal to the primal

objective values of the SRMP and RMP ^′ , then primal and dual objective values for the RMP ^′  are equal.

Therefore, the feasible dual solution constructed for the RMP ^′  is optimal.

3.1.3

Row generation algorithm

Using the optimal dual solution to the RMP ^′ , the row generation algorithm is executed to identify

rows to update the SRMP. This procedure involves solving the column generation subproblem for the

primary variables to find negative reduced cost columns feasible for the RMP ^′ . It is likely, due to the

eliminated constraints, that the columns identified during this procedure are infeasible for the SRMP.

Such columns are identified by displaying at least one non-zero element in the rows contained in Φ _i \Φ i .

Φ _i

If columns infeasible for the SRMP are found, then u must be added to

and the related row to the

SRMP. Consequently, the SRMP grows vertically and horizontally with the addition of rows and columns

respectively. The row generation procedure is detailed in Algorithm 2.

Algorithm 2 Row generation algorithm

Require: An optimal solution to the SRMP.

1:  Set the dual values for rows (27)-(29) to the dual solutions for the rows (22)-(24).

2:  for all secondary variable sets i do

3:

for all rows u contained in

A ^iL

y  do

4:

Execute Algorithm 1 to compute the value of θ ⁱ

u .

5:

end for

6:  end for

7:  By Theorem 3.1.1 an optimal dual solution for the RMP ^′  has been computed.

8:  Solve the column generation subproblem for the primary variables to identify variables feasible for

the RMP ^′ .

9:  if a negative reduced cost column has at least one non-zero entry in the rows of

A ^iL

y  then

10:

add the rows with non-zero entries in

A ^iL

y  to

A ^iL

y  .

11:  end if

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

18

The column-and-row generation solution approach terminates when no favourable rows are identified

by Algorithm 2. This is consistent with the termination condition of the standard column generation

approach, terminating when no columns with a negative reduced cost for the RMP are found. Since an

optimal dual solution is calculated for the RMP ^′ , the column generation subproblem for the primary

variables accurately evaluates the minimum reduced cost. Therefore, if no negative reduced cost columns

are found for the RMP ^′ , then the solution to the RMP ^′ , and the SRMP, is optimal for the original

problem.

3.2

Column-and-Row Generation Solution Algorithm

The column-and-row generation solution algorithm implemented in this paper is developed by combining

the fundamental features of the approach developed throughout Section 3.1. The first stage of the

solution algorithm involves the formulation of the SRMP, which is detailed in Section 3.1.1. Using

the solution to the SRMP, Section 3.1.2 details the calculation procedure that is required to form an

optimal dual solution for the RMP ^′ . The final step in the column-and-row generation solution algorithm,

described in Section 3.1.3, executes Algorithm 2 to identify favourable rows for the SRMP. The complete

column-and-row generation solution algorithm is given by Algorithm 3.

Algorithm 3 Column-and-row generation algorithm

1:  Eliminate columns from the original problem to form the RMP.

2:  Eliminate rows (and subsequently columns) from the RMP to form the SRMP.

3:  repeat

4:

Solve the SRMP by column generation to optimality.

5:

Use Algorithm 2 to compute the optimal dual solution to the RMP ^′  and identify any favourable

rows.

6:  until no rows are added to

A ^iL

y  inAlgorithm2

7:  The solution to the SRMP is the optimal solution to the original problems

4

Applying the Column-and-Row Generation Framework

The framework presented in Section 3 describes the features of column-and-row generation that extend

the standard column generation approach. Column generation is a critical aspect of column-and-row

generation, as such its implementation for the IRP will be discussed in Section 4.1. This will be followed

by a review of the features presented in Section 3.1, detailing the application to the IRP in Section 4.2.

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

19

4.1

Column Generation

The formulation of the IRP presents two sets of variables for which column generation is applied. These

variables are related to crew duties and aircraft routes, which are defined as flight strings. While each

of the variables types have similar structures, there are specific rules governing their generation. Hence,

two individual column generation subproblems are required in the solution process. In this section,

the column generation subproblem for each variable type is described, including the relevant solution

methods.

In the column generation procedure, a restricted master problem (RMP _IRP  ) is defined by including

only a subset of all possible columns,

P ^k  ⊆ P ^k  and

P ^r  ⊆ P ^r , and is solved to find the optimal dual

solution. The dual variables α ^K  = {α ^K , ∀j ∈ N D  ∪ N K

and α ^R  = {α ^R

are defined

j

out }

j  ,∀j∈N D  ∪

out }

for the flight coverage constraints (2)-(3) and (4)-(5), respectively. The dual variables for the aircraft

end-of-day location constraints (6) are defined by ǫ = {ǫ _t , ∀t ∈ T }. The dual variables for the surplus

crew count constraints (7) are given by η = {η ^v

j ,∀j∈N D ,∀v∈U j }.Fortheshortconnectionconstraints

(8) and the delay consistency constraints (9), the dual variables are given by ρ = {ρ _ij  , ∀(i, j) ∈ E ^D  }

and γ = {γ ^v

∈ N ^D ,∀v ∈ U _j }, respectively. Finally, the dual variables δ ^K  = {δ ^k ,∀k ∈ K} and

j ,∀j

δ ^R  = {δ ^r ,∀r ∈ R} are defined for the crew and aircraft assignment constraints, (10)-(11) and (12),

respectively. Using the set of optimal dual solutions, the column generation subproblems for crew and

aircraft are solved to identify negative reduced cost columns to add to the sets

P ^k  and

P ^r .

4.1.1

Crew Pairing Subproblem

The crew duty subproblem (PSP ^k  ) is solved as a shortest path problem with one source node and

multiple sink nodes. The general form of the PSP ^k  is given by,

{

}

∑

∑

∑

∑{

}

k

ĉ

min RecDutyCost(k) −

α _j a ^k

ρ _i

η ^v

a ^kv

−δ ^k

p  =

jp  −

u j v  b ⁱ u  j v

p −

j  +γ j

jp

p∈P ^k

j∈N ^D ∪N ^K

out

(i u  ,j v  )∈E ^D

j∈N ^D  v∈U j

(32)

The set P ^k  is defined as the feasible region to a network flow problem, as such the PSP ^k  is solved as

a resource constrained shortest path problem (RCSPP). The important features of the RCSPP used to

solve the PSP ^k  is the origination of each crew k from a unique carry-in activity and the termination

at any carry-out activity. In addition, the crew duty rules considered in this paper are the maximum

number of flying and working hours, which are set at 8 and 13 respectively. In the column generation and

column-and-row generation solution approaches the PSP ^k  is solve once for each k ∈ K in each iteration.

In the objective function of (32), the complex cost structure used for crew remuneration is denoted by

RecDutyCost(k), which defines the additional crew cost resulting from recovery actions. The recovery

crew cost RecDutyCost(k) is a max function related to the number of flying hours, f ly(k), the number

of working hours, elapse(k), a minimum number of guaranteed hours, minGuar, and the originally

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

20

planned duty cost, OrigDutyCost(k). As such, the expression for the cost of a duty in the IRP is given

by

RecDutyCost(k) = max{0, max{f ly(k), f _d  · elapse(k), minGuar} − OrigDutyCost(k)},

(33)

where minGuar is set at 6 hours [10] and f _d  is a fraction which is airline specific and is set at f _d  = 5/8 [10].

In consideration to the resource restrictions and complex cost structure, a multi-label shortest path

algorithm is implemented to solve the PSP ^k  . Label l at node i _u  stores the cost of the current shortest

path to the node, ĉ _i

u l ,

the number of flying hours, H ¹

and the total elapsed hours, H ²

Multiple

i u  l ,

i u  l .

labels are necessary to track any suboptimal paths to a node, based on the path length ĉ _i

u l ,

that have

a favourable resource consumption. While it is possible to store every path that arrives at a node, this

would be a very inefficient method to track resource consumption. As such, a dominance condition is used

to reduce the number of labels stored at each node by eliminating any suboptimal paths demonstrating

unfavourable resource consumption.

Definition 4.1.1.  (Dominance Condition)

Given two labels at node  i _u , ( ĉ _i

and ( ĉ _i

u 1 ,

i u  1 ,

i u  1 )

u 2 ,

i u  2 ,

i u  2 ),thatarenotequal.Label1

dominates label 2 if

H ¹

ĉ i u  1  ≤

ĉ i u  2 ,

i u  1  ≤

i u  2  and

i u  1  ≤

i u  2 .

Using Definition 4.1.1, the dominance of any new label arriving at a node is evaluated against all

currently stored labels. There are three possible outcomes from the comparison between the new label

and all currently stored labels. Firstly, if the new label dominates any stored label, the dominated labels

are removed from the node. Second, if the new label is dominated by any stored label, then the new

label is discarded. Finally, if no dominance is established between the new label and the stored labels,

then the new label is added to the list of labels stored at that node. At the sink node, the label that

achieves the lowest cost is selected and the resulting path is the minimum reduced cost path.

The connection network described in Section 2 forms an acyclic directed graph. Given this network

structure, all the nodes can be listed in a topological order, where node i _u  is ordered before node j _v  if

∃(i _u , j _v  ) ∈ C ^K  [4]. Using a topological ordering, the shortest path problem can be solved in O(ml) time in

the worst case, where m is the number of arcs in the acyclic directed graph and l is the maximum number

of labels. A pulling algorithm is implemented, which solves the shortest path problem by “pulling” labels

from previously processed nodes. Such a pulling algorithm for solving the shortest path problem on an

acyclic directed graph is presented in Ahuja  et al.  [4]. This algorithm can be easily adapted to solve the

PSP ^k  with multiple labels.

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

21

4.1.2

Aircraft Routing Subproblem

The column generation subproblem for the aircraft routing variables solves a shortest path problem from

an origination location to one of the permissible termination locations. The general form of the column

generation subproblem for the aircraft routing variables (PSP ^r  ) is given by,

{

}

∑

∑

∑

∑

∑

ĉ r

min

c ^r

α _j a ^r

ǫ _t o ^r

ρ _i

γ ^v

−δ ^k

(34)

p  =

p  −

jp  −

tp  +

u j v  b ⁱ u  j v  p  +

j a ^jp

p∈P ^r

j∈N ^D ∪N ^K

t∈T

j∈N ^D  v∈U j

out

(i u  ,j v  )∈E ^D

Similar to P ^k  , the set P ^r  is defined as the feasible region of a network flow problem. However, unlike

the PSP ^k  , the PSP ^r  does not consider any resources in addition to cost, therefore the column generation

subproblem is solved as a standard shortest path problem. The key features of the PSP ^r  is that each

flight string must originate from a unique carry-in activity and may terminate at any permissible carry-

out activity or overnight airport. In addition, if an aircraft is planned to receive maintenance at the end

of the day, the termination locations will ensure that this requirement is met. In the column generation

and column-and-row generation solution approaches the PSP ^r  is solve once for each r ∈ R in each

iteration.

The PSP ^r  describes a shortest path problem for which a large number of solution algorithms are

available. Similar to the connection network for crew, the network for aircraft is an acyclic directed

graph. As such, the nodes can be listed in a topological order and an efficient pulling algorithm presented

in Ahuja  et al.  [4] is implemented to solve the PSP ^r  .

In a given iteration of the column generation algorithm, the most negative reduced cost for all aircraft,

ĉ R

p  ,

can be found by solving the PSP ^r  for each aircraft r and setting ĉ ^R

p

= min _r∈R {c ^r

p }.However,

all connection costs and dual variables, except for δ ^R  = {δ ^r  , ∀r ∈ R}, included in (34) are aircraft

independent. Therefore, by setting δ ^r  = δ ^R , ∀r ∈ R, where δ ^R  = max _r∈R {δ ^r  } in (34), it is possible to

find a lower bound on ĉ ^R

p  ,

labelled as c ^R

p  ,

by solving the aircraft routing shortest path algorithm only

once. The aircraft routing subproblem to be solved only once will be labelled PSP ^R  and will be used as

part of the row generation procedure described in Section 4.2.

4.2

Row Generation

An important feature of the IRP is the use of a  full  set of recovery policies, which includes flight

delays. There are a number of different methods that are available to implement flight delays, such

as time windows [34] and discrete flight copies [37], each with relative strengths regarding the problem

formulation and solution methods. The technique of flight copies has been selected to model delays as

a result of its simplicity in implementation for column generation and to fit within the column-and-row

generation framework.

By implementing flight delays using flight copies, a critical consideration of the integrated problem is

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

22

to ensure that the crew duty and the aircraft routing solutions use the same copy (delay) for each flight.

The delay consistency constraints, equation (9), capture this, at the expense of adding a large number

of constraints to the RMP _IRP  . Since the optimal variables have non-zero coefficients in only a small

subset of the delay consistency constraints, many rows related to these constraints are not required in

the RMP _IRP  .

The implementation of delay copies in the IRP provides alternative flight departure times given by a

uniform discretisation of a maximum allowable delay. While this is a popular method of implementation

that has been employed by Yan and Young [41], Thengvall  et al.  [37] and Andersson and Varbrand [5],

Bratu and Barnhart [11] state that a number of copies may be dominated by shorter delay options.

Further, Petersen  et al.  [30] suggests the modelling of flight delays by an event-driven approach, linking

delays to activities related to each flight. This reduces the size of the recovery problem by only including

the delays for each flight that provide feasible connections. While uniform delay options are implemented

for the IRP, column-and-row generation provides an optimisation approach to select the most important

delay options, significantly reducing the number delay consistency constraints (9). While it is possible to

implement the recovery flight network reductions as described by [11] and [30], this would simply result

in the further enhancement of the column-and-row generation approach.

Comparing the IRP with the RMP in Section 3.1.1, it is clear that the delay consistency constraints

(9) describe linking constraints similar to (19). These constraints provide the link between the primary

and secondary sets of variables, which are given by the crew duty and aircraft routing variables in

the IRP respectively. While the RMP in Section 3.1.1 describes a problem with multiple secondary

variables, the IRP is a special case of this problem class with the aircraft routing variables as the only

set of secondary variables.

The implementation of the column-and-row generation algorithm, Algorithm 3 and a description of

each feature of this algorithm with respect to the IRP will be provided in this section. As a contribution of

this paper, the column-and-row generation solution approach developed by Muter  et al.  [26] is evaluated

against column generation to identify any potential enhancement techniques. A description of the

techniques identified by this evaluation will be provided throughout this section.

4.2.1

Formulation of the restricted problems

The column-and-row generation framework requires the formulation of a RMP _IRP  and SRMP _IRP  as

restrictions on the original problem. The formulation of the RMP _IRP  is provided in Section 4.1, including

only a subset of all variables. The SRMP _IRP  is a further restriction on the RMP _IRP  , initialised with

all rows related to constraints (2)-(8) and only a subset of rows for the delay consistency constraints

U _j

(9) as defined by v ∈

U _j ,∀j ∈ N ^D . The set

is initially populated with one copy for most flights j,

U _j

which is generally the copy representing the scheduled departure time, i.e.

= {0}. However, as a

result of flight delays caused by the initial disruption it is possible that no feasible connection containing

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

23

the flight-copy pair j ₀  exists. In these situations, the set

U _j

= {0, v ^′ } is defined for flight j, where v ^′

represents the copy with the earliest departure time that provides at least one feasible connection for

flight j.

The elimination of rows to form the SRMP _IRP  is coupled with the fixing of variables in the column

generation subproblems. This variable fixing restricts the use of specific flight-copy pairs related to the

rows eliminated to form the SRMP _IRP  . Thus, flight strings can only be constructed using the flight-copy

pairs j _v  , ∀j ∈ N ^D  , ∀v ∈

U _j . Since the set of columns feasible for the SRMP _IRP  is restricted, the solution

to the SRMP _IRP  provides an upper bound on the optimal solution of the IRP.

4.2.2

Row generation algorithm

The calculation of the optimal dual solution to the RMP ^′  is a fundamental part of the row generation

procedure. By solving the SRMP _IRP  to optimality using column generation, Theorem 3.1.1 states

that the optimal dual solution to the RMP ^′  can be calculated using the solution to the SRMP _IRP

and Algorithm 1. The dual solutions that must be calculated are related to the rows eliminated to

form the SRMP _IRP  , which are given by γ ^′  = {γ v ′ , ∀j ∈ N D  , ∀v ′  ∈ U j  \U j  }. The solutions to each of the

j

variables contained in γ ^′  are found by executing Algorithm 1, solving the PSP ^R  as the column generation

subproblem in step 2. The use of the PSP ^R  in this algorithm is a problem specific enhancement technique

that reduces the runtimes required to calculate the solutions to the dual variables for all eliminated rows.

The second part of the row generation procedure uses the optimal dual solution to the RMP ^′  to

identify favourable rows to add to the SRMP _IRP  . This process is described by steps 8-11 of Algorithm

2, which involves solving the column generation subproblem for the primary variables to find columns

feasible for the RMP ^′ . Since the primary variables for the IRP are the crew duty variables, the PSP ^k  is

solved as the pricing subproblem in step 8 of this algorithm. The crew duty variables generated by this

subproblem describe individual schedules for each crew k, hence a larger number of favourable rows are

identified by solving the PSP ^k  once for each k ∈ K. This is a natural modification of the row generation

procedure which is necessary to achieve an efficient solution approach for the IRP.

The dual variable calculation of the row generation procedure is a feature of column-and-row gen-

eration that is identified to be very computationally expensive. Given the set of flight-copy pairs,

∪ _j∈N D U _j \U j , the PSP ^R  must be solved once for each flight-copy pair contained in this set. Even with

the most efficient shortest path algorithm, the large number of executions required to calculate all dual

variables can have a significant negative impact on the solution runtimes. Consequently, the number of

times that Algorithm 2 is executed will affect the overall performance of the column-and-row generation

solution process. One approach to address this runtime issue is to vary the number of rows that are

added in each call to the row generation procedure. It has been observed that by adding too few rows

at each execution requires more calls to the row generation algorithm. Similarly, adding too many rows

has the effect of increasing the size of the SRMP _IRP  too rapidly. A successful approach involves adding

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

24

more rows to the SRMP _IRP  based upon the value of the calculated dual variables. This is achieved for

the IRP by adding a row for every flight-copy pair with a positive dual variable value, as calculated by

Algorithm 1. The rows identified by this procedure is in addition to those identified in Algorithm 2. The

ideal number of rows to add at each iteration is difficult to determine. However, this approach described

here improves the solution runtimes compared to the standard row generation procedure.

4.2.3

Row generation warm-up

The subset of rows initially included in the SRMP _IRP  greatly affects the efficiency of the column-and-

row generation solution process. In Section 4.2.1, the initialisation of the SRMP _IRP  involves selecting a

single delay copy for each flight, which is na¨ıvely set to the scheduled departure time. Ideally, the only

rows included in the SRMP _IRP  should represent the amount of delay for each flight that is required in

the optimal solution of the IRP. Unfortunately the optimisation problem to identify the optimal set of

delay options is analogous to the original recovery problem, hence an alternative technique is required.

An approach implemented for the IRP uses information from the standard column generation ap-

proach to provide a warm-start for column-and-row generation. This approach involves formulating the

RMP _IRP  with all rows from the original problem but only the columns contained in the initial formu-

lation of the SRMP _IRP  . The RMP _IRP  is then solved by column generation and in each iteration of the

solution algorithm, flight strings are constructed with no restriction on the permissible flight-copy pairs.

The initial set of delay copies for the SRMP _IRP  is updated by reviewing each generated flight string p

and if j _v  ∈ p, then the delay copy v is added to the set

U _j . After n iterations of the column generation

solution process, the SRMP _IRP  is formed to contain only the delay consistency constraints (9) described

by the sets

U _j ,∀j ∈ N ^D  and all columns in the current formulation of the RMP _IRP  .

A key feature of this approach is that no additional development work is required and the compu-

tational time is equivalent to that of the standard column generation approach. During this warm-up

period the runtime of the two solution approaches is identical, therefore the expected runtime improve-

ments are observed by applying column-and-row generation in the succeeding iterations. By retaining

the initial columns added during this process, the column-and-row generation approach is provided with

a warm-start for the set of columns and rows. The use of the warm-up process is a contribution of this

paper to the column-and-row generation solution approach.

The runtime improvements achieved by this approach demonstrate the importance of an intelligent

selection of rows in the initial formulation of the SRMP _IRP  . This is expected, since this observation is

similar to the well known relationship between the initial set of columns and the efficiency of the standard

column generation approach. A complicating factor of applying a warm-up period for column-and-row

generation is the additional parameter required to specify the number of column generation iterations

executed. The value of this parameter has been observed through experiments to greatly affect the

efficacy of this approach with an acceptable runtime improvement for the IRP achieved with n = 20

4

APPLYING THE COLUMN-AND-ROW GENERATION FRAMEWORK

25

iterations.

4.3

Branching Rules

Integral optimality is achieved for the IRP by employing the technique of branch-and-price. This problem

includes many different variable types and thus a set of problem specific branching rules have been

designed for each. The first of the branching rules described here for the IRP is a pure variable branching

rule, and the last two are derived from the Ryan/Foster branching technique [32]. A description of each

rule is provided below in the order of their assigned priority.

A cancellation variable branching rule is implemented for the IRP to force the decision of either

covering or cancelling a specific flight. Upon identifying flight j ^′  with the most fractional cancellation

variable, z _j ′  , branches are created by enforcing z _j ′

= 0 on the left branch and z _j ′  = 1 on the right

branch. The described rule is very simple and fast, and is designed to eliminate fractional cancellation

variables early in the branch-and-bound tree.

A very effective branching technique for airline optimisation problems formulated in a set partitioning

framework is branching on follow-on’s. The implementation of follow-on branching for the IRP is similar

to that presented in Froyland  et al.  [19]. The aim of this branching rule is to select the most fractional

pair of connected flights, flights occurring in succession on a flight string, for either the crew duty or

aircraft routing variables. In the implementation of this rule, the multiple copies for each flight are

ignored and the connection (i, j) identifies all connections (i _u , j _v  ) ∈ C ^K  ∪ C ^R , ∀u ∈ U _i , ∀v ∈ U _j  . The set

of fractional variables for crew k is defined as P ^k

= {p ∈ P ^k  |x ^k

f

p  ∈Z},andsimilarlythesetoffractional

variables for aircraft r is defined as, P ^r

= {p ∈ P ^r  |y ^r

∈ Z}. The fractionality of a connection (i, j) is

f

p

calculated by,

{∑

∑

∑ ∑

}

frac ^K

fOn (i,j)=min

x ^k

p ,1−

x ^k

p

,

and

k∈K p∈P ^k

f

k∈K p∈P ^k

f

|(i,j)∈p

|(i,j)∈p

{∑

∑

∑ ∑

}

(35)

frac ^R

fOn (i,j)=min

y ^r

p ,1−

y ^r

p

,

r∈R p∈P ^r

f

r∈R p∈P ^r

f

|(i,j)∈p

|(i,j)∈p

for crew and aircraft variables respectively. The connection with the greatest fractionality for either crew

or aircraft is identified as (i ^∗ , j ^∗ ) and is selected as the branching candidate. The candidate variable

type, crew or aircraft, that this branching applies to is also identified by this selection. The branching

is performed by enforcing the use of connection (i ^∗ , j ^∗ ) on a string that contains either flight i ^∗  or j ^∗  on

the left branch for the identified variable type. On the right branch, flight j ^∗  must not directly follow

flight i ^∗  on any string for the identified variable type. However, on the right branch, flights i ^∗  and j ^∗

are not precluded from existing on the same string, provided that the connections (i ^∗ , k) and (l, j ^∗ ) are

used, where k = j ^∗  and l = i ^∗ .

An alternative branching rule is developed for the IRP that examines the allocation of specific flights

to individual crew and aircraft. This branching rule selects a crew group k or aircraft r and enforces or

5

COMPUTATIONAL RESULTS

26

disallows the use of an identified flight-copy. The fractionality of a variable identifier/flight-copy pair,

(k, i _u ) and (r, i _u ), is calculated by,

{∑

∑

}

frac ^K

x ^k

x ^k

,

and

flt (k,i u )=min

p ,1−

p

p∈P ^k

f

p∈P ^k

f

|i u  ∈p

|i u  ∈p

{∑

}

(36)

∑

frac ^R

y ^r

y ^r

,

flt (r,i u )=min

p ,1−

p

p∈P ^r

f

p∈P ^r

f

|i u  ∈p

|i u  ∈p

for the crew and aircraft variable types respectively. Branching is performed on the variable identifier/flight-

copy pair that has the greatest fractionality, as described by the equations (36). On the left branch, all

variables associated with the identifier, k ^∗  or r ^∗ , must contain the flight-copy i ^∗

u ^∗  intheflightstring.

On the right branch all flight strings for the variables associated with the identifier, k ^∗  or r ^∗ , must not

contain the flight-copy i ^∗

u ^∗  .

As a contribution to the column-and-row generation solution approach, the row generation procedure

is integrated into the branch-and-price framework. Since the branch-and-price algorithm is executed with

a subset of all rows contained in the IRP, without allowing the addition of rows throughout this process,

any identified lower bounds potentially overestimate the true bound. To avoid this inaccuracy in the

solution process, the row generation algorithm is called only at nodes where the column generation

procedure concludes with a lower bound greater than the current best bound. By executing the row

generation algorithm in these selected situations ensures that the optimal solution is found with branch-

and-price and avoids unnecessary executions of this time costly procedure.

5

Computational Results

The computational results demonstrate the benefit of using column-and-row generation (CRG) to solve

the IRP compared to a standard column generation approach (Colgen). The following discussion pro-

vides a comparison between these two approaches based upon  computational performance . For this

analysis computational performance is defined as the runtime required to solve the IRP and the final

solution quality as measured by the use of recovery policies. These results demonstrate column-and-

row generation as a superior alternative to column generation for solving integrated airline optimisation

problems for the given flight schedules.

5.1

Description of data and disruption scenarios

The performance of the IRP is evaluated on two different flight schedules that are designed for a point-to-

point and hub-and-spoke carriers, labelled as F262-A20 and F441-A36 respectively. The point-to-point

schedule consists of 262 flights, operated by 48 aircraft of a single fleet type and 79 crew groups. This

schedule services 20 airports; 12 of the airports are overnight bases for aircraft and 4 are crew bases.

5

COMPUTATIONAL RESULTS

27

The hub-and-spoke schedule consists of 441 flights, operated by 123 aircraft of a single fleet type and

182 crew groups. This schedule services 36 airports; 6 of the airports are hubs, 32 are overnight bases

for aircraft and 3 are crew bases. The original duties and routings for both schedules are generated by

solving an integrated airline planning problem with an objective of minimising crew costs and the total

number of aircraft operating the schedule.

A set of 16 disruption scenarios are generated as test cases for the IRP. The scenarios describe

airport closures at two major airports in each network, occurring in the morning at 6am, 7am, 8am

and 9am for either 3 or 5 hours. The numbers used to reference each scenario are provided in Table

3. An airport closure imposes a delay on all flights that are scheduled to arrive at or depart from the

affected airport until the end of the closure period. In these experiments a recovery window of 6 hours is

used, representing the total time allowed to return operations back to plan. The recovery window starts

from the reopening of the affected airport, thereby the set of disruptable flights N ^D  includes all flights

departing within a 9 or 11 hour window from the start of the closure. Within the recovery window, the

IRP implements a full set of recovery policies including flight delays and cancellations, crew deadheading

and the generation of new crew duties and aircraft routes.

Scenario Start Time

6am

7am

8am

9am

Scenario Number

(0,8), (1,9)

(2,10), (3,11)

(4,12), (5,13)

(6,14), (7,15)

Table 3: Scenario numbers.

( x , y ) indicates scenario  x  has a closure of 3 hours and scenario  y  has a

closure of 5 hours. Bold represents the scenarios related to airport two.

A common approach to improve the runtime of airline recovery problems is to use only a subset

of crew, aircraft and flights by only including those identified as disruptable. These results aim to

demonstrate the runtime improvements achieved by implementing column-and-row generation alone.

Hence, no approximation of the set of disruptable crew and aircraft is made. It is expected that with

approximations of the dataset in practical applications of the algorithm, further runtime improvements

will be observed. Now, a recovery window is used to identify the set of flights N ^D  to include in the

IRP and the size of this set for each scenario is documented in Table 4. While the use of a recovery

window is an approximation of the full recovery problem, it is consistent with the common objective to

return operations to plan as quickly as possible. The size of the recovery window dictates the allowable

recovery time and at the conclusion of the window no further recovery actions can be taken.

Scenario

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

F262-A20

150

151

149

150

147

145

150

149

182

183

185

186

184

182

184

183

F441-A36

265

261

246

254

251

251

250

251

315

311

304

310

305

305

292

293

Table 4: The number of disruptable flights for each scenario.

Airlines incur significant realised and unrealised costs due to flight delays and cancellations during

the recovery process. These costs are modelled in the IRP to quantitatively define the effects of the

5

COMPUTATIONAL RESULTS

28

disruption on the airline and passengers. The data provided for this problem includes the number of

passengers booked on each flight, which is used in the calculation of the delay and cancellation costs.

The cost of flight delays has been estimated from the EUROCONTROL report by Cook and Tanner [14],

where it is stated that the average cost of delay for a full aircraft is e81 per minute. For convenience,

this value is converted into Australian dollars, so the cost of a full aircraft delayed for a minute is $100

AUD.

Flight delays are implemented in the IRP using the technique of flight copies, as described in Section

2. The maximum allowable delay on any flight is set at 180 minutes, and 7 flight copies have been

used to divide this delay into discrete blocks. Therefore, the minimum possible delay on any flight is

30 minutes, with each subsequent flight copy departure occurring at 30 minute intervals. Since flight

delays are discretised with the use of flight copies, the resulting recovery costs are an overestimate

of the best possible solution. This occurs because there potentially exist shorter, feasible connections

within the 30 minute delay window that could provide an improved recovery solution. It is possible

to increase the number of flight copies to improve the solution quality, however the number of delay

consistency constraints (9) is dependent on the chosen number of copies. Providing a greater granularity

of delays with more flight copies results in a much larger column generation master problem and a

larger connection network for the pricing subproblem, degrading the computational performance. The

results will demonstrate that by using column-and-row generation the improvement in the computational

performance over a standard column generation approach is still achieved as the problem size grows with

an increased number of copies.

Quantitatively defining the cost of flight cancellations is difficult due to the indirect costs related

passenger dissatisfaction. In the event of a flight cancellation, passengers are either i) rebooked onto

another flight operated by the airline, ii) rebooked onto another flight operated by a different airline, or

iii) provided a refund and some compensation and must rebook their own flight. Case iii) is the most

uncommon and results in the greatest passenger dissatisfaction compared to cases i) and ii). However,

in all situations it is difficult to estimate the proportion of passengers that are  lost  from potential future

bookings with the airline. In these experiments, it is assumed that only the ticket revenue is lost and

passengers are not deterred from booking with the airline in the future. The calculation of the total

lost revenue for each flight assumes an average ticket price of $350 multiplied by the number of booked

passengers. The IRP is solved assuming that the passengers are not rebooked by the airline onto any

flights, resulting in the loss of the total expected revenue from the cancelled flight.

This model is implemented in C++ by calling SCIP 3.0.1 [3] to solve the integer program using

CPLEX 12.4 as the linear programming solver.

5

COMPUTATIONAL RESULTS

29

5.2

Comparison of solution runtimes

It is of high importance for the practical application of any recovery algorithm that a solution can

be found in short runtimes. Figure 1 compares the runtime required to solve the IRP for the two

different flight schedules when using the solution approaches of column generation and column-and-row

generation. To demonstrate the appropriate use of this model in practical applications, a maximum

runtime of 1200 seconds (20 minutes) is applied. This maximum runtime is selected to be within the

runtime of 30 minutes set by Petersen  et al.  [30] for their evaluation of an integrated airline recovery

problem.

Figure 1 shows that in the vast majority of experiments, the optimal solution is found with runtimes

much less than 1200 seconds. On average, column-and-row generation solves the IRP in 427 seconds for

the F262-A20 schedule and 400 seconds for the F441-A32 schedule. In addition, the results presented in

Figure 1 show that there is no scenario solved by column-and-row generation that exceeds the maximum

allowable runtime. Therefore, it is possible to reduce the maximum allowable runtime without affecting

the optimality of the IRP solved by column-and-row generation.

Runtime of the IRP using column generation and column-and-row generation

F262-A20 schedule

F441-A32 schedule

Solution Process

Approach

14

Solving LP

Colgen

Column Generation

CRG

Time Cutoff - 1200sec

Row Generation

Time Cutoff - 1200sec

12

10

8

6

4

2

0

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Scenario Number

Scenario Number

Figure 1: The runtime to solve the IRP for each scenario with a maximum of 1200 seconds. This figure

compares the solution approaches of column generation (bars) and column-and-row generation (bars

with hatching).

The results presented in Figure 1 demonstrate that the solution to the IRP using column-and-row

generation is achieved much faster than when column generation is used for all but three cases. Two

of the cases are for the F262-A20 schedule (scenarios 1 and 9) and the other is for the F441-A32

schedule (scenario 7). For the F262-A20 schedule, the average relative improvement between the two

solution methods is 27.12%, with a range of -84.14% (scenario 1) to 127.13% (scenario 14). This level of

improvement is also observed for the F441-A32 schedule, with an average improvement in the solution

runtimes of 25.18%. However, the range of the results for the F441-A32 schedule is more restricted

with a minimum improvement of -15.91% (scenario 7) and a maximum of 74.97% (scenario 8). These

improvements demonstrate a significant benefit from using column-and-row generation to solve the IRP.

A key feature of column-and-row generation is the smaller set of rows used in the formulation of the

SRMP _IRP  compared to the RMP _IRP  . There is a well known direct relationship between the number

of constraints in a problem and the expected time required to solve the linear programming relaxation.

5

COMPUTATIONAL RESULTS

30

Both solution approaches include a column generation process, which involves solving the LP relaxation

of the RMP _IRP  , and SRMP _IRP  , each time a set of columns is added. Since the RMP _IRP  and SRMP _IRP

are formulated as very similar problems, with the latter containing less constraints, solving the LP for

the SRMP _IRP  requires significantly less simplex iterations resulting in faster execution times.

Figure 1 provides a breakdown of the solution runtimes into the processes of LP solve, column

generation and row generation. It is clear from this figure that the reduction in the time spent solving

the LP is a major component of the solution runtime improvement for the F262-A20 schedule. Using

scenario 12 for the F262-A20 schedule as an example, solving the LP of the SRMP _IRP  requires a total

of 111.7 seconds for column-and-row generation, where column generation requires 180.22 seconds to

solve the LP of the RMP _IRP  . The total solution runtime improvement for scenario 12 is 109.76 seconds,

of which a significant proportion can be attributed to the reduced execution time for solving the LP

relaxation. This demonstrates that column-and-row generation provides a significant improvement in a

solution process which is integral to both approaches.

Another significant improvement in solution runtimes is observed in the time required during the

column generation procedure. Since the reduced set of rows results in a smaller connection network,

the column generation subproblems are solved much quicker in the column-and-row generation solution

approach. It is observed that the time required for each call to the column generation subproblem is on

average 0.2857 and 0.6754 seconds quicker for column-and-row generation compared to column generation

in the F262-A20 and F441-A32 schedule results respectively. The magnitude of this improvement can

be explained using scenario 4 for the F262-A20 schedule as an example, where the column generation

subproblems are called 311 times for both solution approaches. For this scenario, column-and-row

generation requires 0.1369 seconds less per call, which results in a 42.57 second improvement in the

solution runtimes. This result is also observed in scenario 1 for the F441-A32 schedule, where the

total execution time of the column generation subproblem is 57.37 seconds less with the column-and-

row generation approach. The reduction in column generation subproblems execution times contributes

54.6% of the total runtime reduction. This reduction in runtimes for the column generation subproblems

is achievable as a result of eliminating rows to form the SRMP _IRP  . It is clear from Figure 1 that the

required additional process of row generation does not greatly contribute to the runtimes of the column-

and-row generation approach. Therefore, there is a significant runtime advantage in solving the LP

relaxation and the column generation subproblems from applying column-and-row generation.

5.2.1

Effects of problem size

The number of rows initially removed from the RMP _IRP  to form the SRMP _IRP  is directly proportional

to the number of flight copies used in the model. An increase in the number of flight copies impacts the

two competing factors affecting the runtime of the column-and-row generation approach, i.e. the smaller

problem size and the row generation algorithm. With a greater number of flight copies, the SRMP _IRP

5

COMPUTATIONAL RESULTS

31

initially defines a problem much smaller than the related RMP _IRP  , potentially providing a considerable

speed up in the runtime required for each LP solve. However, the more flight copies that are removed may

require additional executions of the row generation procedure to identify the optimal set of rows, having

a negative effect on the runtime of the column generation subproblems. Figure 2 displays the relative

difference in runtimes between the column generation and column-and-row generation approaches by

varying the number of flight copies for the F262-A20 schedule. The results demonstrate that across

all experiments performed, column-and-row generation outperforms the column generation approach in

most cases, with an average improvement of 27.07%.

Runtime comparison between column generation and column-and-row generation

Avg Diff (27.07%)

100

CRG - better runtime performance

50

0

-50

Colgen -

better runtime performance

Number of Copies

5

7

10

13

19

-100

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Scenario Number

Figure 2: The relative difference in runtimes between the column generation (x) and column-and-row

generation (y) approaches with different sets of copies for the F262-A20 schedule. The values in the

figure are calculated by (x − y)/ min{x, y} with a maximum reported improvement capped at 100%.

Note: maximum runtime of 1 hour was used for these results.

Comparing the relative difference between the runtimes of column generation and column-and-row

generation in Figure 2 demonstrates a better average performance of the latter approach. While this is

true for the average case, there are many individual experiments where the reverse result is observed.

This generally occurs when the column-and-row generation approach requires more branching to identify

the optimal integer solution. For example, scenario 1 formulated with 7 flight copies is solved significantly

faster with column generation due to the column-and-row generation approach requiring 29 more nodes

in the branch-and-bound tree. While the LP of the root node is solve much faster by column-and-row

generation (327 seconds compared to 439 seconds), the branch-and-price algorithm has more difficultly

converging to integrality for the SRMP _IRP  . The structure of the SRMP _IRP  appears to affect the efficacy

of the branching rules and the performance of the primal heuristics. This is a common observation

regarding the implementation of column-and-row generation within the branch-and-price framework.

The length of the recovery window is another feature of the IRP that affects the problem size by

directly impacting the number of flights included in the set N ^D  . As the length of the recovery window

increases, the number of flights that depart within that period also increases. A larger set of disruptable

5

COMPUTATIONAL RESULTS

32

Comparison of recovery window lengths for the IRP

F262-A20 schedule

F441-A32 schedule

16

Solution Approach

Recovery Window

14

Colgen

CRG

6 hours

8 hours

End of day

12

10

8

6

4

2

0

0

200

400

600

800

1000

1200

0

200

400

600

800

1000

1200

Runtime (seconds)

Runtime (seconds)

Figure 3: Time required to solve scenarios with different recovery window lengths (maximum runtime of

1200 seconds and 7 flight copies per flight). The recovery windows are set at 6 hours, 8 hours and until

the end of the day.

flights N ^D  has two main effects on the IRP, i) an increase in the number of constraints in both the

RMP _IRP  and SRMP _IRP  , and ii) an increase in the size of the connection network used in the column

generation subproblem.

Figure 3 presents the time required to solve all of the scenarios used in the experiments for both

flight schedules with different recovery window lengths, 6 hours, 8 hours and until the end of the day.

The results presented for the F262-A20 schedule demonstrate that the runtimes required to solve all

scenarios using a window of 6 hours is significantly shorter than when the other two window lengths are

used. This indicates that the increase in the size of N ^D  has a great effect on the solution runtime, which

is very evident when the recovery window is extended from 6 to 8 hours. This result is also observed for

the F441-A32 schedule, however the separation of the frontiers is not as pronounced.

It is observed in Figure 3 that the increased problem complexity from extending the recovery window

has a more pronounced effect on the runtime of the column generation approach. This is evident from

the significant decrease in the number of scenarios that column generation solves to optimality within

the runtime of 1200 seconds for both flight schedules. Also, the separation of the frontiers produced by

the two solution approaches using an 8 hour window demonstrates the greater degradation of runtime

performance from column generation.

Comparing the results presented in Figures 1 and 3, it is clear that the improvement in runtimes

is more pronounced for the F262-A20 schedule. In addition, the number of scenarios solved for the

F262-A20 schedule degrades much quicker with an increase in the recovery window length compared to

the F441-A32 schedule. This is an interesting result which can be explained by the percentage increase

in the number of included flights. By increasing the recovery window length from 6 hours to the end

of the day, the average percentage increase in flights for the F262-A20 schedule is 38.45% compared to

27.13% for the F441-A32 schedule. In addition, the maximum percentage increase in the number of

included flights for the F262-A20 schedule is 73.97% compared to 52.11% for the F441-A32 schedule.

Since the increase in recovery window lengths results in a comparatively larger increase in the number

of included flights, it is expected that a greater degradation of the solution runtimes will be observed

5

COMPUTATIONAL RESULTS

33

for the F262-A20 schedule.

5.2.2

Analysing enhancement techniques

A contribution of this paper is the explicit evaluation of the column-and-row generation approach against

column generation and the development of enhancement techniques. While the techniques introduced

in Section 4.2 are presented in relation to the IRP, they may be applied in any implementation of

column-and-row generation. The most important enhancement techniques discussed are the number of

rows added during the row generation procedure and the row warm-up technique used to provide an

initial formulation of the SRMP _IRP  with a meaningful set of delay options. Using the F262-A20 schedule,

Figure 4 presents the relative difference in solution runtimes for the column-and-row generation approach

with different enhancements compared to the standard column generation approach.

Runtime comparison of column-and-row generation enhancement techniques

150

Enhancement

No Enhancement (Avg: 11.91%)

Warm-up Only (Avg: 1.54%)

Additional Rows Only (Avg: 8.36%)

All Enhancements (Avg: 27.01%)

100

CRG -

better runtime performance

50

0

-50

Colgen - better

runtime

performance

-100

-150

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Scenario Number

Figure 4: The relative difference in the solution runtimes from applying different enhancement techniques

for the F262-A20 schedule. The values in the figure are calculated by (x − y)/ min{x, y}.

The strength of the column-and-row generation approach to improve upon solution runtimes of

column generation is evident in Figure 4. This is demonstrated by all implementations of column-and-

row generation achieving an improvement in the solution runtimes compared to the standard column

generation approach. In particular, column-and-row generation implemented without any enhancement

improves upon the solution runtimes of column generation by 11.91%. This result also outperforms the

implementations of column-and-row generation using the additional rows or warm-up enhancements in

isolation. This can be explained by the nature of the two enhancement approaches. The row warm-up

technique slows the execution of the initial iterations of the column generation stage by permitting the use

of all flight-copies in the subproblem. This coupled with only a small set of rows added in each iteration

of the row generation procedure does not provide much reduction in the solution runtimes. By constrast,

the additional rows enhancement rapidly increases the size of the SRMP _IRP  . As a result, very little

decrease in the LP solving time is observed and on average more time is required per iteration to solve the

column generation subproblem. Finally, the use of no enhancements achieves a smaller optimality gap

5

COMPUTATIONAL RESULTS

34

at the root node compared to the two individual enhancement approaches in most scenarios, requiring

less nodes to find the integer optimal solution.

While each enhancement implemented individually degrades the performance of the column-and-row

generation approach, significant runtime improvements are observed by their combined use. This is

demonstrated in Figure 4, implying that a strong relationship exists between the two enhancement ap-

proaches. In this figure, column-and-row generation using all enhancements demonstrates runtimes that

are 27.01% shorter on average when compared to the standard column generation approach. Comparing

this result to the standard implementation of column-and-row generation, the use of all enhancements

provides a runtime improvement of 15.88%. This is exhibits a significant runtime improvement from the

use of enhancement techniques in the implementation of column-and-row generation.

The results presented in this section demonstrate that the greatest difference in the efficiency of the

solution approaches is the convergence to the integer optimal solution. In each of the figures presented

above, the largest variations in solution runtimes is commonly caused by an increased number of nodes

in the branch-and-bound tree. This effect is observed in a number of results presented in Figure 4,

where the use of enhancement techniques greatly reduces the runtime required to solve the root node

with ineffective branching eroding any gains. For example, the implementation of column-and-row

generation with all enhancements outperforms column generation in the solving time for the root node,

with scenario 1 and 9 being solved 110 and 153 seconds faster respectively. These two scenarios display

the greatest improvement in the root node solving time but the integer optimal solution is found quicker

by column generation. While column-and-row generation achieves an improvement in solution runtimes,

these results demonstrate that greater reductions can be achieved through more effective branching

techniques.

5.3

Analysis of recovery statistics

The composition of the recovery policies in the solution to the IRP greatly affects passenger satisfaction

and the acceptability to the operations control centre. The main recovery policies that are implemented

for the IRP are flight delays and cancellations and the crew specific policies of deadheading and the use

of reserve crew. Figures 5 and 6 demonstrate the use of each of these recovery policies in the solutions

achieved by column generation and column-and-row generation for the two considered flight schedules.

Since both the column generation and column-and-row generation solution approaches are solved to

within an optimality gap of 1% for most scenarios, very little difference between the solutions is observed.

This is demonstrated in Figures 5 and 6 with only subtle variations in the use of the reported recovery

policies. The largest difference is observed for Scenario 8 in Figure 6, which is a consequence of column

generation being unable to solve the IRP to integral optimality within the maximum runtime. It is also

important to note that scenario 0 in Figure 5 is not solved to optimality by column generation, therefore

the recovery solution is expected to present a greater use of recovery policies compared to generating

5

COMPUTATIONAL RESULTS

35

Recovery solution statistics for the F262-A20 dataset

24

24

Flights

Deadheaded crew

22

Colgen

CRG

Reserve

crew

22

Delayed

Delayed

20

Cancelled

Cancelled

20

18

18

16

16

14

14

12

12

10

10

8

8

6

6

4

4

2

2

0

0

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Scenario Number

Figure 5: Comparison of recovery solution statistics from the column generation and column-and-row

generation solutions using the F262-A20 schedule. The left axis is related to the flight statistics and the

right axis is related to crew specific recovery policies.

recovered crew duties and aircraft routes.

The most significant difference in the solutions presented in Figure 5 is given by the number of

deadheaded crew. The column generation solution requires more deadheading crew than the column-

and-row generation solution in four of the scenarios examined compared to two in the reverse. A feature

of the solution approaches affecting the composition of recovery policies is related to the construction

of flight strings for the SRMP _IRP  and RMP _IRP  . Since the SRMP _IRP  contains less rows than the

RMP _IRP  throughout the solution process, there are many columns that are initially not permissible in

the column-and-row generation approach. This results in feasible integer solutions with minimal delay,

hence requiring the use of alternative recovery policies, being added to the solution pool in the column-

and-row generation approach. Consequently, the different sets of columns and feasible solutions found

during the solving process can impact upon the composition of recovery actions in the optimal solution.

Recovery solution statistics for the F441-A32 dataset

100

20

Flights

Deadheaded crew

90

Colgen

CRG

Reserve

crew

18

Delayed

Delayed

Cancelled

Cancelled

80

16

70

14

60

12

50

10

40

8

30

6

20

4

10

2

0

0

0

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

Scenario Number

Figure 6: Comparison of recovery solution statistics from the column generation and column-and-row

generation solutions using the F441-A32 schedule. The left axis is related to the flight statistics and the

right axis is related to crew specific recovery policies.

Figure 6 greatly emphasises the similarities between the solutions achieved by column generation and

6

CONCLUSIONS

36

column-and-row generation. This little difference is the result of the IRP being solved to a 0% optimality

gap at the root node in the majority of experiments for the F441-A32 schedule. Both column generation

and column-and-row generation are exact solution approaches, so the results presented in Figure 6 are

not surprising. The most interesting result is in Scenario 8, where the recovery statistics for the column

generation result exhibits a large percentage of cancelled flights. Since column-and-row generation solves

the IRP faster than standard column generation on average, the former solution approach is expected

to provide a better solution quality in shorter runtimes.

Finally, reserve crew are a valuable resource for an airline, however there is a limitation to their

use across a full scheduling period. It is important to manage the number of reserve crew that are

employed to operate a disrupted schedule to ensure their availability for subsequent days. Figures 5 and

6 demonstrate that the modelling approach used for the IRP results in the very little use of reserve crew

during recovery. This is an important result that allows the practical management of the reserve crew

resource. Additionally, it is trivial to modify the model to limit the number of reserve crew that are

available for each disruption.

6

Conclusions

This paper presents an alternative approach to solving the integrated airline recovery problem by imple-

menting column-and-row generation. The integrated recovery problem is a very large and computation-

ally difficult problem for which there have been many attempts to develop efficient solution methods. A

general framework for column-and-row generation has been developed as an alternative exact solution

method to Benders’ decomposition and approximation techniques. The use of column-and-row genera-

tion to solve the IRP has been demonstrated here as a superior approach in regards to solution runtime

and quality compared to column generation.

The column-and-row generation framework developed in this paper is an extension of Muter  et

al.  [26]. The extensions of this framework are the consideration of multiple secondary variables and

linking constraints, providing a wider applicability of the solution approach. Further extending the

analysis of the framework by Muter  et al.  [26], an explicit evaluation of column-and-row generation

against a standard column generation approach is performed. As a result of this evaluation, a number of

enhancement techniques have been developed that may be applied to general column-and-row generation

implementations. In particular, it is demonstrated that the addition of extra rows during the row

generation procedure and the use of a warm-up period achieves the greatest improvement for the column-

and-row generation approach when implemented in combination. The evaluation performed in this paper

demonstrates the ability of column-and-row generation to improve upon the solution runtimes achieved

by column generation. Finally, this paper details the integration of column-and-row generation and

branch-and-price, which has not previously been published.

REFERENCES

37

A motivation for applying column-and-row generation to solve the IRP is to reduce solution runtimes.

The results in Section 5 demonstrate that across the majority of the experiments, column-and-row gen-

eration outperforms column generation in solution runtime and quality. The improvement in runtimes

is observed through a decrease in the time required for each LP solve in the column-and-row generation

procedure. The number of flight copies and the length of the recovery window has a direct effect on

the size of the IRP, impacting the solution runtimes. The results demonstrate that as the problem

size increases, column-and-row generation still achieves significant improvements over a standard col-

umn generation approach. The improvements achieved through the use of column-and-row generation

demonstrate a practical solution approach for the integrated recovery problem.

Column-and-row generation provides a direct solution approach for the IRP that achieves near op-

timal solutions within the desired time-frame. This is a significant improvement on alternative solution

approaches where integral optimality is not guaranteed, such as Benders’ decomposition. Further, at

each iteration of the column-and-row generation approach the optimal solution to the SRMP provides

an upper bound on the original problem. This is a significant advantage of this approach, permitting

the early termination of the solution algorithm.

There are two directions for future work on the integrated recovery problem solved using column-and-

row generation. Firstly, the integrated recovery problem presented here integrates the schedule, aircraft

and crew recovery problems, however more detailed solutions for the complete recovery problem can be

achieved through further integration. The integration of passenger considerations in the IRP will aid in

reducing the impact of recovery on passengers. Secondly, while the use of column-and-row generation

improves upon the solution quality and runtime when compared to the standard column generation

approach, it is likely that additional enhancements can ameliorate these even further. Research into the

selection of rows during the row generation procedure is expected to aid in reducing the runtime for the

column-and-row generation solution approach.

Acknowledgements

SJM would like to thank Gary Froyland for discussing concepts presented in this paper and providing

insight to the problem.

References

[1] A. Abdelghany, G. Ekollu, R. Narasimhan, and K. Abdelghany. A proactive crew recovery decision

support tool for commercial airlines during irregular operations.  Annals of Operations Research ,

127(23):309-331, 2004.

REFERENCES

38

[2]

K. F. Abdelghany, A. F. Abdelghany, and G. Ekollu. An integrated decision support tool for

airlines schedule recovery during irregular operations.  European Journal of Operational Research ,

185(2):825-848, 2008.

[3]

T. Achterberg. SCIP: solving constraint integer programs.  Mathematical Programming Computa-

tion , 1(1):1-41, 2009.

[4]

R. Ahuja, T. Magnanti, and J. Orlin.  Network flows: theory, algorithms, and applications . Prentice

Hall, 1993.

[5]

T. Andersson and P. Varbrand. The flight perturbation problem.  Transportation Planning and

Technology , 27(2):91-117, 2004.

[6]

M. Argüello.  Framework for exact solutions and heuristics for approximate solutions to airlines’

irregular operations control aircraft routing problem . PhD thesis, University of Texas at Austin,

1997.

[7]

P. Avella, B. D’Auria, and S. Salerno. A LP-based heuristic for a time-constrained routing problem.

European Journal of Operational Research , 173(1):120-124, 2006.

[8]

J. Bard, G. Yu, and M. Argüello. Optimizing aircraft routings in response to groundings and delays.

IIE Transactions , 33(10):931-947, 2001.

[9]

C. Barnhart, N. L. Boland, L. W. Clarke, E. L. Johnson, G. L. Nemhauser, and R. G. Shenoi. Flight

string models for aircraft fleeting and routing.  Transportation Science , 32(3):208-220, 1998.

[10]

C. Barnhart, A. Cohn, E. Johnson, D. Klabjan, G. Nemhauser, and P. Vance. Airline crew schedul-

ing. In R. W. Hall, editor,  Handbook of Transportation Science , volume 56 of  International Series

in Operations Research and Management Science , pages 517-560. Springer US, 2003.

[11]

S. Bratu and C. Barnhart. Flight operations recovery: new approaches considering passenger

recovery.  Journal of Scheduling , 9(3):279-298, 2006.

[12]

J. Cao and A. Kanafani. Real-time decision support for integration of airline flight cancellations and

delays part I: mathematical formulation.  Transportation Planning and Technology , 20(3):183-199,

1997.

[13]

J. Cao and A. Kanafani. Real-time decision support for integration of airline flight cancellations and

delays part II: algorithm and computational experiments.  Transportation Planning and Technology ,

20(3):201-217, 1997.

[14]

A. Cook and G. Tanner.

European airline delay cost reference values.

URL:

http://www.eurocontrol.int/documents/european-airline-delay-cost-reference-values  , 2011.

REFERENCES

39

[15]

J.-F. Cordeau, G. Stojković, F. Soumis, and J. Desrosiers. Benders’ decomposition for simultaneous

aircraft routing and crew scheduling.  Transportation Science , 35(4):375-388, 2001.

[16]

N. Eggenberg, M. Bierlaire, and M. Salani. A column generation algorithm for disrupted airline

schedules. Technical report, Ecole Polytechnique Federale de Lausanne, 2007.

[17]

A. Frangioni and B. Gendron. 0-1 reformulations of the multicommodity capacitated network design

problem.  Discrete Applied Mathematics , 157(6):1229-1241, 2009.

[18]

A. Frangioni and B. Gendron. A stabilized structured Dantzig-Wolfe decomposition method.  Math-

ematical Programming , 140(1):45-76, 2013.

[19]

G. Froyland, S. J. Maher, and C.-L. Wu. The recoverable robust tail assignment problem.  Trans-

portation Science . To appear.

[20]

A. I. Z. Jarrah, G. Yu, N. Krishnamurthy, and A. Rakshit. A decision support framework for airline

flight cancellations and delays.  Transportation Science , 27(3):266-280, 1993.

[21]

E. Johnson, L. Lettovsky, G. Nemhauser, R. Pandit, and S. Querido. Final report to Northwest

Airlines on the crew recovery problem. Technical report, Georgia Institute of Technology, 1994.

[22]

L. Lettovsky.  Airline operations recovery: an optimization approach . PhD thesis, Georgia Institute

of Technology, 1997.

[23]

L. Lettovsky, E. Johnson, and G. Nemhauser. Airline crew recovery.  Transportation Science ,

34(4):337-348, 2000.

[24]

C. P. Medard and N. Sawhney. Airline crew scheduling from planning to operations.  European

Journal of Operational Research , 183(3):1013-1027, 2007.

[25]

A. Mercier, J. Cordeau, and F. Soumis. A computational study of Benders’ decomposition for

the integrated aircraft routing and crew scheduling problem.  Computers & Operations Research ,

32(6):1451-1476, 2005.

[26]

İ. Muter, ¸.

İ. Birbil, and K. Bülbül. Simultaneous column-and-row generation for large-scale linear

programs with column-dependent-rows.  Mathematical Programming . To appear.

[27]

İ. Muter, ¸.

İ. Birbil, K. Bülbül, G. ¸ahin, H. Yenigün, D. Tas, and D. Tüzün. Solving a robust airline

crew pairing problem with column generation.  Computers & Operations Research , 40(3):815-830,

2013.

[28]

R. Nissen and K. Haase. Duty-period-based network model for crew rescheduling in European

airlines.  Journal of Scheduling , 9(3):255-278, 2006.

REFERENCES

40

[29]

N. Papadakos. Integrated airline scheduling.  Computers & Operations Research , 36(1):176-195,

2009.

[30]

J. Petersen, G. Sölveling, E. Johnson, J. Clarke, and S. Shebalov. An optimization approach to

airline integrated recovery.  Transportation Science , 46(4):482-500, 2012.

[31]

J. M. Rosenberger, E. L. Johnson, and G. L. Nemhauser. Rerouting aircraft for airline recovery.

Transportation Science , 37(4):408-421, 2003.

[32]

D. M. Ryan and B. A. Foster. An integer programming approach to scheduling. In A. Wren, editor,

Computer Scheduling of Public Transport: Urban Passenger Vehicle and Crew Scheduling , pages

269-280. North-Holland Publishing Company, Amsterdam, 1981.

[33]

R. Sadykov and F. Vanderbeck. Column generation for extended formulations.  EURO Journal on

Computational Optimization , 1(1-2):81-115, 2013.

[34]

M. Stojković and F. Soumis. An optimization model for the simultaneous operational flight and

pilot scheduling problem.  Management Science , 47(9):1290-1305, 2001.

[35]

M. Stojković and F. Soumis. The operational flight and multi-crew scheduling problem.  Yugoslav

Journal of Operations Research , 15(1):25-48, 2005.

[36]

M. Stojković, F. Soumis, and J. Desrosiers. The operational airline crew scheduling problem.

Transportation Science , 32(3):232-245, 1998.

[37]

B. G. Thengvall, J. F. Bard, and G. Yu. Balancing user preferences for aircraft schedule recovery

during irregular operations.  IIE Transactions , 32(3):181-193, 2000.

[38]

G. Wang and L. Tang. A row-and-column generation method to a batch machine scheduling problem.

In  Proceedings of the Ninth International Symposium on Operations Research and Its Applications

(ISORA-10) , pages 301-308, 2010.

[39]

G. Wei, G. Yu, and M. Song. Optimization model and algorithm for crew management during

airline irregular operations.  Journal of Combinatorial Optimization , 1(3):305-321, 1997.

[40]

S. Yan and D.-H. Yang. A decision support framework for handling schedule perturbation.  Trans-

portation Research Part B: Methodological , 30(6):405-419, 1996.

[41]

S. Yan and H.-F. Young. A decision support framework for multi-fleet routing and multi-stop flight

scheduling.  Transportation Research Part A: Policy and Practice , 30(5):379-398, 1996.

[42]

E. J. Zak. Row and column generation technique for a multistage cutting stock problem.  Computers

& Operations Research , 29(9):1143-1156, 2002.