1.224J/ ESD.

204J: Airline Crew

Scheduling
Outline
Crew Scheduling
Problem Definition
Costs
Set Partitioning Model and Solution
Enhanced Crew Scheduling
Link to Maintenance Routing Problem
Enhanced Model and Solution Approach
 Airline Schedule Planning

Schedule Design Select optimal set of flight legs

in a schedule

AAssign
flight specifies
aircraft types
origin,todestination,
flight legs
Fleet Assignment such thatandcontribution
departureistimemaximized

Route individual aircraft honoring

Aircraft Routing maintenance restrictions

Assign crew (pilots and/or flight

Crew Scheduling attendants) to flight legs

12/31/2003 1.224J/ESD.204J 2
 Crew Scheduling: Some
Background
This problem has been studied by
operations researchers for at least 4
decades
Most major U.S. airlines use crew
pairing optimizers for the cockpit crews
Crew costs are the airlines second largest
operating expense
Even small improvements in efficiency can
have large financial benefits
12/31/2003 1.224J/ESD.204J 3
 Airline Crew Scheduling
2-stage process:
Crew Pairing Optimization
Construct minimum cost work schedules,
called pairings, spanning several days
Bidline Generation/ Rostering
Construct monthly work schedules from the
pairings generated in the first stage
Bidlines
Individualized schedules
Objective to balance workload, maximize
number of crew requests granted, etc.
12/31/2003 1.224J/ESD.204J 4
 Some Definitions
A crewbase is the home station, or
domicile, of a crew
A crew pairing is a sequence of flights
that can be flown by a single crew:
beginning and ending at a crewbase
spanning one or more days
satisfying FAA rules and collective bargaining
agreements, such as:
maximum flying time in a day
minimum rest requirements
minimum connection time between two flights
12/31/2003 1.224J/ESD.204J 5
 Example: A Crew Pairing

LA c
b d e
Detroit
a f
Boston

12/31/2003 1.224J/ESD.204J 6
 Some More Definitions
A duty period (or duty) is a sequence of
flight legs comprising a day of work for a
crew
Alternative pairing definition: a crew pairing
is a sequence of duties separated by rests
A crew schedule is a sequence of pairings
separated by time-off, satisfying
numerous restrictions from regulatory
agencies and collective bargaining
agreements

12/31/2003 1.224J/ESD.204J 7
 Example: Duty Periods

LA
b c d e
Detroit
a f
Boston

Pairing = DP1(a,b,c) + rest + DP2(d,e) + rest + DP3(f)

a b c d e f

12/31/2003 1.224J/ESD.204J 8
 Crew Pairing Problem (CP)

Assign crews to flights such that

every flight is covered, costs are
minimized and labor rules are
satisfied:
Maximum flying time in a day
Minimum rest requirements
Minimum connection time

12/31/2003 1.224J/ESD.204J 9
 Crew Pairing Costs
Duty costs: Maximum of 3 elements:
f1*flying time cost
f2*elapsed time cost
f3*minimum guarantee

Pairing costs: Maximum of 3 elements:

f1*duty cost
f2*time-away-from-base
f3*minimum guarantee
12/31/2003 1.224J/ESD.204J 10
 Set Partitioning Model for CP:
Variable Definition and Constraints
A variable is a pairing
Binary variables: =1 if pairing is assigned
to a crew; = 0 if pairing not flown
Set partitioning constraints require each
flight to be covered exactly once
Number of possible pairings (variables)
grows exponentially with the number of
flights
12/31/2003 1.224J/ESD.204J 11
 An Example

Flights: Crew pairing solutions:

ABCDEFGH x1 => pairings 1, 4: $7
Potential pairings: x2 => pairings 2, 3: $6
A-C-D-F (y1): $1
A-B-E-F (y2): $2
C-D-G-H (y3): $4
B-E-G-H (y4): $6

12/31/2003 1.224J/ESD.204J 12
 Pairing 1

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 13
 Pairing 2

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 14
 Pairing 3

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 15
 Pairing 4

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 16
 Notation

Pk is the set of feasible pairings for fleet

family k
Fk is the set of daily flights assigned to
fleet family k
fp equals 1 if flight f is in pairing p, else 0
cp is the cost of pairing p
yp is 1 if pairing p is in the solution, else 0

12/31/2003 1.224J/ESD.204J 17
 Formulation

min c p yp
pP k

st
fp y p = 1 f F k

pP k

y p {0,1} p P k

12/31/2003 1.224J/ESD.204J 18
 Example
y1 + y2 = 1 A
y2 + y4 = 1 B
Flight y1 + y3 = 1 C
Cover y1 + y3 = 1 D
Constraints: y2 + y4 = 1 E
y1 + y2 = 1 F
y3 + y4 = 1 G
y3 + y4 = 1 H
y1 0,1 1
Binary y2 0,1 2
Pairing y3 0,1 3
Restrictions: y4 0,1 4

12/31/2003 1.224J/ESD.204J 19
 Set Partitioning Model:
Advantages and Disadvantages
Advantages:
Easy to model complex work rules
Very few constraints
Linear objective function and constraints
Disadvantages:
Huge number of variables- number of
variables grows exponentially with the
number of flights

12/31/2003 1.224J/ESD.204J 20
 Problem Size
A typical US airline (with a hub-and-
spoke network) has millions or billions
of potential pairings
Example
150 flights 90,000 pairings
250 flights 6,000,000 pairings
Need a specialized approach to consider
problems of this size

12/31/2003 1.224J/ESD.204J 21
 Branch-and-Price: Branch-and-
Bound for Large-Scale Integer
Programs
y1 = 1 y1 = 0

y2 = 1 y2 = 0
y2 = 1 y2 = 0

y3 = 1 y3 = 0

All possible solutions at leaf nodes of tree (2n solutions,

where n is the number of variables)

12/31/2003 1.224J/ESD.204J 22
 Column Generation

Millions/Billions of Variables
constraints

Added
Initial

Never Considered

12/31/2003 1.224J/ESD.204J 23
 LP Solution: Column Generation

Step 1: Solve Restricted Master

Problem
Step 2: Solve Pricing Problem
(generate columns with negative
reduced cost)
Step 3: If columns generated, return
to Step 1; otherwise STOP
12/31/2003 1.224J/ESD.204J 24
 Network Representation

City A

City B

City C

City D

800 1200 1600 2000 800 1200 1600 2000

12/31/2003 1.224J/ESD.204J 25
 Branch-and-Bound with Too
Many Variables
Branch-and-Price
Branch-and-bound with bounding
provided by LP solutions
CP has too many variables to consider
all of them
Solve linear programming relaxation
using column generation

12/31/2003 1.224J/ESD.204J 26
 A Twist
Crew scheduling is critical to the airline
industry
Second largest operating expense
Small improvement in solution quality has
significant financial impact
For decades, researchers have worked on
finding better crew scheduling algorithms
Our approach is to instead improve
solution quality by expanding the feasible set of
solutions
12/31/2003 1.224J/ESD.204J 27
 Airline Schedule Planning

Schedule Design

Fleet Assignment

Route individual aircraft honoring

Aircraft Routing maintenance restrictions

Assign crew (pilots and/or flight

Crew Scheduling attendants) to flight legs

12/31/2003 1.224J/ESD.204J 28
 Aircraft Maintenance Routing:
Problem Definition
Given:
Flight Schedule for a single fleet
Each flight covered exactly once by fleet
Number of Aircraft by Equipment Type
Cant assign more aircraft than are available
Turn Times at each Station
FAA Maintenance Requirements
Maintenance required every 60 hours of flying
Airlines maintain aircraft every 40-45 hours of flying with the
maximum time between checks restricted to three to four
calendar days
12/31/2003 1.224J/ESD.204J 29
 Aircraft Maintenance Routing:
Objective
Find:
Feasible assignment of individual
aircraft to scheduled flights
Each flight is covered exactly once
Maintenance requirements are satisfied
Conservation of flow (balance) of
aircraft is achieved
The number of aircraft used does not
exceed the number available
12/31/2003 1.224J/ESD.204J 30
 Example: Maintenance Station in
Boston

LA
b c d e
Detroit
a f
Boston

12/31/2003 1.224J/ESD.204J 31
 String Model: Variable Definition

A string is a sequence of flights beginning

and ending at a maintenance station with
maintenance following the last flight in the
sequence
Departure time of the string is the departure
time of the first flight in the sequence
Arrival time of the string is the arrival time of
the last flight in the sequence + maintenance
time
12/31/2003 1.224J/ESD.204J 32
 String Model: Constraints
Maintenance constraints
Satisfied by variable definition
Cover constraints
Each flight must be assigned to exactly one
string
Balance constraints
Needed only at maintenance stations
Fleet size constraints
The number of assigned aircraft cannot
exceed the number of aircraft in the fleet
12/31/2003 1.224J/ESD.204J 33
 Model Solution

Complex constraints can be handled

easily
Model size
Huge number of variables
Solution approach: branch-and-price
Generate string variables only as-needed

12/31/2003 1.224J/ESD.204J 34
 Airline Schedule Planning

Schedule Design

Fleet Assignment

Route individual aircraft honoring

Aircraft Routing maintenance restrictions

Assign crew (pilots and/or flight

Crew Scheduling attendants) to flight legs

12/31/2003 1.224J/ESD.204J 35
 Integrate?
Crew scheduling options are limited by
maintenance routing decisions made
earlier in the airline planning process
Solving maintenance routing and crew
scheduling simultaneously yields a large
and challenging problem
Idea is to improve crew scheduling by
incorporating relevant maintenance
routing decisions
12/31/2003 1.224J/ESD.204J 36
 Maintenance Routing and its Link to
the Crew Pairing Problem
The Maintenance Routing Problem (MR) - find
feasible routing of aircraft ensuring adequate aircraft
maintenance opportunities and flight coverage
Crews need enough time between two sequential
flights to travel through the terminal -- minimum
connect time
If both flights are covered by the same aircraft,
connection time can be reduced
A short connect is a connection that is crew-feasible
only if both flights are assigned to the same aircraft
12/31/2003 1.224J/ESD.204J 37
 Sequential Approach

Short
Maintenance Connects Flight Valid Crew
Routing Pairing
Network
Problem flown by Problem
pairings
the same
aircraft

12/31/2003 1.224J/ESD.204J 38
 Klabjan, Johnson, and Nemhauser
Solve the crew pairing problem first, including
all short connects in the crew pairing network
Given the crew solution, require all short
connects included in it to be part of the
maintenance solution, which is solved second
For good instances, this yields the optimal
solution to the integrated problem (and many
problems are good)
For bad instances, this leads to maintenance
infeasibility
12/31/2003 1.224J/ESD.204J 39
 Cordeau, Stojkovi, Soumis,
Desrosiers
Directly integrate crew and maintenance
routing models
Basic maintenance routing and crew
pairing variables and constraints, plus
linking constraints
Benders decomposition approach using a
heuristic branching strategy
For non-hub-and-spoke networks, positive
computational results
12/31/2003 1.224J/ESD.204J 40
 Our Approach

Generate different solutions to the

maintenance routing problem
Allow the crew pairing model to
choose the maintenance routing
solution with the most useful set of
short connects

12/31/2003 1.224J/ESD.204J 41
 The Example Again
Flights: Potential pairings:
ABCDEFGH A-C-D-F (y1): $1
All Possible Short A-B-E-F (y2): $2
Connects: C-D-G-H (y3): $4
A-B A-C E-G B-E-G-H (y4): $6
Crew pairing solutions:
MR solution (x1) assigns
x1 => pairings 1, 4: $7
the same aircraft to short
connects A-C and E-G x2 => pairings 2, 3: $6
MR solution (x2) assigns
the same aircraft to short
connect A-B
12/31/2003 1.224J/ESD.204J 42
 Flights

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 43
 Short Connect

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 44
 Short Connect

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 45
 Short Connect

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 46
 Maintenance Solution 1

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 47
 Maintenance Solution 2

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 48
 If MR Solution 1 (A-C, E-G) =>
Optimal: Pairings 1, 4 -- $7

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 49
 If MR Solution 2 (A-B) =>
Optimal: Pairings 2, 3 -- $6

A C D F

B E G H

12/31/2003 1.224J/ESD.204J 50
 Approach
In the sequential approach, given a maintenance
routing solution, the crew pairing problem is
solved
We allow the crew scheduler to choose from a
collection of maintenance routing solutions
Select the one containing the set of short connects that
allows the minimum cost crew pairing solution
Problem: We dont want to solve one crew pairing
problem for each maintenance routing solution
Solution: Extended Crew Pairing Model (ECP)

12/31/2003 1.224J/ESD.204J 51
The Extended Crew Pairing Model
(ECP)
Simultaneously select a cost minimizing
set of crew pairings and a corresponding
feasible maintenance routing solution
from a given set of maintenance routing
solutions
Add constraints that allow pairings with
a short connect to be selected only if the
chosen maintenance solution assigns the
same aircraft to that short connect
12/31/2003 1.224J/ESD.204J 52
 The Example Again
Flights:
ABCDEFGH Potential pairings:
A-C-D-F (y1): $1
Short Connects: A-B-E-F (y2): $2
A-B A-C E-G C-D-G-H (y3): $4
MR solution (x1) B-E-G-H (y4): $6
uses short connects Crew pairing solutions:
A-C and E-G
x1 => pairings 1, 4: $7
MR solution (x2) x2 => pairings 2, 3: $6
uses short connect
A-B
12/31/2003 1.224J/ESD.204J 53
 Matrix Representation for the
Example
x1 x2 y1 y2 y3 y4 rhs
0 0 1 1 0 0 = 1 A
0 0 0 1 0 1 = 1 B
Flights: 0 0 1 0 1 0 = 1 C
0 0 1 0 1 0 = 1 D
0 0 0 1 0 1 = 1 E
0 0 1 1 0 0 = 1 F
0 0 0 0 1 1 = 1 G
0 0 0 0 1 1 = 1 H
0 1 0 1 0 0 0 A-B
Short Connects: 1 0 1 0 0 0 0 A-C
1 0 0 0 0 1 0 E-G
Convexity: 1 1 0 0 0 0 = 1 Conv.
12/31/2003 1.224J/ESD.204J 54
 Notation

Pk is the set of feasible pairings for fleet

family k
Fk is the set of flights assigned to fleet
family k
Tk is the set of short connects for the
flights assigned to fleet family k
Sk is the set of feasible MR solutions for
the flights assigned to fleet family k
12/31/2003 1.224J/ESD.204J 55
 Notation, cont.

fp is 1 if flight f is included in pairing p,

else 0
ts is 1 if MR solution s includes short
connect t, else 0
tp is 1 if short connect t is contained in
pairing p, else 0
cp is the cost of pairing p

12/31/2003 1.224J/ESD.204J 56
 Notation, cont.

xs is a binary decision variable with

value 1 if MR solution s is chosen, else
0

yp is a binary decision variable with

value 1 if pairing p is chosen, else 0
12/31/2003 1.224J/ESD.204J 57
 ECP Formulation

min c p yp
pP k

st
Flights: fp y p = 1 f F k
pP k

Short Connects:
ts xs tp y p 0 t T k
sS k pP k
Convexity:
x s =1
sS k

xs , y p {0,1} s, p
12/31/2003 1.224J/ESD.204J 58
 ECP Enhancements
By exploiting dominance relationships, can
dramatically reduce the number of MR columns
considered in finding an optimal ECP solution
MR1 containing short connects AB, CD, GH
dominates MR2 containing short connect AB
Do not need to include MR2 in ECP
Theoretical bounds and computational
observations
Example: 61 flights => >> 25,000 MR solutions =>
4 non-dominated MR solutions (bounded by 35)
Can find these 4 non-dominated MR solutions by
solving 4 MR problems
12/31/2003 1.224J/ESD.204J 59
 ECP Enhancements, cont.
Proof: Can relax the integrality of MR columns
and still achieve integer solutions:
Same number of binary variables as original CP
min c p y p min c p y p
p p

st st
y
p: f p
p =1 f y
p: f p
p =1 f

z y
s:ts
s
p:tP
p 0 t z y
s:ts
s
p:tP
p 0 t

z
s
s =1 z
s
s =1

z s {0,1} s zs 0 s
y p {0,1} p y p {0,1} p

LP relaxation of ECP is tighter than LP relaxation

of a basic integrated approach
12/31/2003 1.224J/ESD.204J 60
 Computational Experiment
Problem A:
Lower bound: 31,396.10
ECP with 16 MR columns: 31,396.10
Optimality gap: 0%

Problem B:
Lower bound: 25,076.60
ECP with 20 MR columns: 25,498.60
Optimality gap: 1.7%
12/31/2003 1.224J/ESD.204J 61
Airline Crew Scheduling Successes
Excess crew costs in the planning
process has been driven to 0-3%
AA was 8-10% 15 yrs ago: now 0-2%
Each 1% is worth about $10
million/yr
1997 had 9,000 pilots costing $1.2 billion
Larger schedules and complex rules

12/31/2003 1.224J/ESD.204J 62
 Crew Pre-Month Planning
Ellis Johnson, Georgia Tech

Vacation Scheduling

Crew Pairing Optimization Initial Training Scheduling

Regular/Reserve Bidline Generation

Time

Bidding and Conflict Resolution

Recurrent Training Scheduling

Supplemental Regular Lines Flight Instructor Scheduling

Month of Operation

11/04/2001 Operations Research / Delta Technology Inc. 63

 Reserve Demand
Ellis Johnson, Georgia Tech

Net Reserve Demand

Pre-Month Planning Irregular Operations

(1) Vacation Conflict (1) Weather Disruptions

(2) Initial Training Conflict (2) Aircraft Maintenance
(3) Transition Conflict (3) Sick Leave
(4) Recurrent Training Conflict

Could be up to 25% Open Time Trips Higher Reserve

total trips built Availability Desired

Voluntary Flying Premium Flying Reserves

Could cover 10% of the open trip flying

11/04/2001 Operations Research / Delta Technology Inc. 64

 Crew: The Rest of The Story
Ellis Johnson, Georgia Tech

Manpower planning, conflicts, over-

time flying, and reserves
In US airlines as high as 30% of the
pilots may be on reserve bid lines
Actual flying is about 50% of usual
Of that flying, more than half is to
cover conflicts and as little as 1/3 is to
cover disruptions
12/31/2003 1.224J/ESD.204J 65
 Conclusions
Crew scheduling is critical to airline
profitability
Making maintenance routing decisions
independently increases costs
A model that fully integrates MR and CP can
be inflexible and difficult to solve
ECP exploits that only some maintenance
routing information is relevant and uses
dominance to reduce the size of the problem
More work to be done especially post-
pairing optimization
12/31/2003 1.224J/ESD.204J 66