同样为了翻译

A Decision Support Framework for Airline Flight Cancellations and Delays
Author(s): AHMAD I. Z. JARRAH, GANG YU, NIRUP KRISHNAMURTHY and ANANDA
RAKSHIT
Source: Transportation Science, Vol. 27, No. 3 (August 1993), pp. 266-280
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/25768594
Accessed: 19-07-2017 03:12 UTC
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/25768594?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Transportation Science
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
A Decision Support Framework for
Airline Flight Cancellations
and Delays
AHMAD I. Z. JARRAH and GANG YU
University of Texas at Austin, Austin, Texas 78712
NIRUP KRISHNAMURTHY and ANANDA RAKSHIT
United Airlines, Corporate Research and Development, Chicago, Illinois 60666
Aircraft shortages occasionally occur during day-to-day airline operation due to factors such
as unfavorable weather conditions, mechanical problems, and delays in the schedule of
incoming flights. Flight controllers need to respond to such shortages on a real-time basis by
delaying or cancelling flights, swapping aircraft among scheduled flights, or requesting the
usage of surplus aircraft. The choices undertaken aim at minimizing the losses incurred while
retaining an operable flight schedule. In this paper, we represent two network models for
aiding flight controllers in this complex decision environment. The models represent an
attempt at conceptualizing this important and relatively unstudied problem, and form the
basis for an evolving decision support system at United Airlines.
INTRODUCTION
T
JL he schedule of flights for an airline is built
using various techniques that consider factors such
as market demand, aircraft size and number, avail
able crew resources, maintenance requirements,
airport constraints, to name but a few. Due to its
many dimensions, the schedule building process
may span days, weeks, or months depending upon
the size of the airline. A key feature of such a
schedule is that resources are tightly coupled to it
and even minor perturbations could have a severe
impact from the standpoint of resource availability
and utilization. This is unfortunate because pertur
bations resulting from unplanned events, like air
craft shortages, occur during actual operations.
Traditionally, airlines have relied upon flight con
trollers with access to on-line operational data
to manage the day-to-day situation as it unfolds.
This entails analyzing the impact of aircraft short
ages on the whole network of flights and providing
appropriate real-time responses. These responses
usually take the form of delaying or cancel
ling flights, swapping aircraft among flights, and
requesting unused (spare) aircraf
is very complex due to the large nu
and flights that may need to be
multitude of responses that can be
the need to provide solutions in re
ous intricacies described emphas
computerized decision support tools
providing solutions with as little
airline's operation as possible. Su
to the airline industry where the
is to have the flight controllers res
based on their judgment and aided
on-line information about the in
aircraft, and crews.
The problem described has so far
attention in the published literatur
with extensive research on airline
(see [2] for a survey), and some re
maintenance scheduling (see [3, 7,
and Guberini<5[91 discuss the proble
overall passenger delays in the e
schedule perturbation. They atte
least expensive set of aircraft r
branch and bound procedure. T
266
Transportation Science
Vol. 27, No. 3, August 1993
0041-1655/93/2703-0266 $01.25
? 1993 Operations Research Society of America
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
FRAMEWORK FOR FLIGHT CANCELLATIONS AND DELAYS / 267
example with merely three aircraft but give no
further numerical results. Given the vast number
of possible routings for even moderately sized prob
lems, it is doubtful that their approach would be
practical for realistic problems. Teodorovi<5 and
Stojkovi<5[101 present a greedy heuristic for solv
ing the following goal programming problem:
given some perturbations in the flight schedule,
find the new set of aircraft routings that first mini
mizes the number of cancellations, and then min
imizes the overall passenger delays. The heuristic
algorithm processes the aircraft in sequence. For
each aircraft an attempt is made to assign as many
flights as possible; subsequently, the path with
the least amount of delays that cover the same
number of flights is found using a recursive delay
function. Again, an example is shown but no
computational results are provided.
In this paper, we start by giving an overview of
the issues involved in flight cancellations and delays
(Section 1), and by discussing a related but unpub
lished work[4] that provides a useful conceptual
starting point for our models (Section 2). Next, we
cast some of the problems faced by flight controllers
addressing aircraft shortages into minimum-cost
network flow models. The first of these models
(Section 3.1) chooses a set of flight delays that
can absorb the shortages, while the second model
(Section 3.2) chooses another set of flight cancel
lations that can achieve the same goal. These
models can form the basis for building a decision
support system to assist flight controllers in finding
good solutions in real-time. Benefits from such a
real-time system could be large; assuming 1400
flights a day, an average of 100 passengers per
affected flight, a 1% flight cancellation rate, sav
ings can be in excess of 0.5x million dollars per
year (1400 X 365 X 100 X 0.01 X x > 0.5x), where
x is the dollars per passenger that a human
machine system could save on cancellation costs
alone. Savings could be much larger if delay
costs are also included. In addition to examples,
very good results are presented for computational
tests performed at United Airlines (UA) where deci
sion support tools are under development using the
framework presented in this paper (Section 4). We
conclude by discussing the limitations of the sug
gested approaches and outlining possibilities for
future research in the area.
1. FLIGHT CANCELLATIONS AND DELAYS:
AN OVERVIEW
While the overwhelming majority of flights operate
as scheduled, aircraft shortages do occur, resulting
in flight delays or cancellations. The reason for
shortages at a certain airport at a certain point in
time are various, ranging from weather conditions
that make flying unacceptable, mechanical prob
lems that call for immediate attention, or delays in
the schedule of incoming flights. Such shortages are
often managed through cancellation of flights.
Flight controllers responsible for day-to-day opera
tions may also resort to the option of delaying some
of the flights in an attempt to avoid cancellations.
To illustrate, let us consider the following scenario:
on a particular day at 1:00 p.m., the flight con
trollers learn that a certain aircraft pi will need
immediate maintenance which will keep it inopera
ble until 4:00 p.m.. Aircraft pi was supposed to
take flight fl scheduled for 2:00 p.m., while two
other aircraft p2 and p3, arriving at 2:00 p.m. and
2:30 p.m., respectively, are scheduled to take flights
f2 and f3 at 2:30 p.m. and 3:50 p.m., respectively.
As an alternative to cancelling flights, the con
trollers may assign flight fl to aircraft p2 for a
delayed departure at, say, 2:15 p.m., flight f2
to aircraft p 3 for a delayed departure at, say,
2:50 p.m., and flight f3 to the fixed aircraft pi for
a delayed departure at 4:00 p.m. Another technique
which can be used by the controllers is that of
requesting unscheduled surplus aircraft (spares) if
such a request is deemed economically and opera
tionally attractive. These surplus aircraft may be
available either at the airports where the problem
planes are or can be ferried in from nearby airports.
Flight controllers have to make real-time deci
sions as to the set of flights that need to be can
celled or delayed because of the aircraft shortage(s).
Many complex factors have to be considered by the
controllers. The chosen sets of cancellations and
delays should preferably be the ones which cause
the minimum, or close to minimum, loss in direct
revenues (ticket refunds for passengers choosing
other airlines, or hotel rates for passengers choos
ing to wait overnight) and indirect costs like cus
tomer goodwill. Ideally, this revenue loss should be
analyzed for the immediately impacted flights as
well as for all the subsequently affected flights to
obtain an overall acceptable solution. The schedule
should remain operable after the cancellations and
delays; i.e., each scheduled flight should have an
aircraft. In addition, the flight controllers have to
identify how the chosen sets of cancellations and
delays will affect the crews (pilots and flight atten
dants), and interactively obtain the crews' approval
for any changes in their work schedules. Other
considerations relate to the maintenance needs of
the aircraft, for it would not be acceptable to have
cancellations or delays with consequent changes in
the flight schedule that prevent an aircraft from
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
268 / a. i. z. jarrah et al.
arriving at one of the eligible airports for receiving
scheduled maintenance. Given the above complex
amalgam of constraints and options, it would not be
reasonable to expect the flight controllers to pro
vide solutions that are globally attractive. This is
especially true because the controllers have to pro
vide solutions in real-time to the problem at hand.
2. THE SUCCESSIVE SHORTEST PATH METHOD
A successive shortest path method (SSPM), pre
sented by Gershkoff,[4] attempts to find a good set
of flight cancellations to resolve aircraft shortages;
however, it fails to consider several important fea
tures of the problem, like allowing delays and usage
of spare aircraft. We choose to discuss SSPM
because it serves as a conceptual starting point for
the later developments in the paper. We present
two models in Section 4 which are much wider in
scope and address some of the shortcomings of
SSPM.
Figure 1 can be used to describe SSPM. Time is
placed on the vertical axis, while the horizontal
axis lists various relevant stations (airports). The
rectangular nodes are sequences of flights (termed
"movement groups") during which there is one or
more aircraft on the ground. In contrast, during the
time between movement groups no aircraft are on
the ground. The idea behind movement groups is
that if a shortage exists in one of the groups, say G,
then a cancellation should necessarily take place
for one of the departures of group G, or any of the
earlier groups at station PHX (Phoenix) such
as group F; otherwise the number of departures
would exceed the number of aircraft in group G,
which is impossible. The arcs connecting the nodes
represent flights between stations, and if an arc
Fig. 1. Underlying graph for the sspm.
(flight) connects two nodes i and j, then the
arc (flight) will be called arc (flight) In
the scenario of Figure 1, an aircraft becomes
unavailable at time 12:00 at station LAX (Los
Angeles), and will become functional at time 16:00.
Hence, a set of flight cancellations should originate
from node A and terminate at station LAX at
or after time 16:00, where the recovered aircraft
will resume flying and "absorb" the cancellation.
Movement groups at a certain station are connected
by upward arcs to allow for the possibility of han
dling a shortage in a group at a certain station by
cancelling a flight in one of the earlier groups. The
cost associated with cancelling each flight is the
loss in revenue to be incurred if the flight is can
celled. Hence, the shortest path is found between
group A, and one of the several groups at LAX
after time 16:00 (like C, D, and E). The shortest
path would represent the least expensive set of
cancellations. For example, if the shortest path is
AGFKHD, then the flights (A,G), (F, K), (K, H),
and (H, D) are to be cancelled. Note that the can
celled flight, (F, K), leaves an extra aircraft to be
used in group G. Also the shortest path ends at
node D, where no further cancellations are needed,
because an extra aircraft is now available since the
shortage aircraft is now fixed and available for
usage.
Suppose, now, that there is more than one short
age (say two in our example, with one originating
at A at time 12:00 and can be recovered at time
16:00, and the second originating at time 12:00 at
F and can be recovered at time 17:00). Then, the
method arbitrarily selects one of the shortages,
say the one at A, and finds the shortest path
(AGFKHD) as described above. Next, the arcs on
the shortest path are reversed with their costs
multiplied by -1 (see Figure 2). After that, the
shortest path is found from node F to several of
the nodes at PHX after time 17:00. The backward
arcs with negative costs allow for the possibility of
"uncancelling" a cancellation, i.e., using the flight.
For example, suppose the shortest path starting at
F is FJHKI (Figure 2). This means that flight
(K,H) is not going to be cancelled after all. The
final set of cancellations would become:
(A,G), (F,K), and (K, /).
(F, J),(J,H\ and (H,D).
Hence, the shortage starting at LAX at time
12:00 ends at PHX after time 17:00, while the
shortage starting at PHX at time 12:00 ends at
LAX after time 16:00.
While this successive shortest path method is
very interesting, there is obviously no guarantee of
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
FRAMEWORK FOR FLIGHT CANCELLATIONS AND DELAYS / 269
STATION
time 1 time 2 F A
13:00 Ll%\ l\ J
14:00 * \/ 4 X n \ /U
B G X T
15:00 4 U-l / Ik
|| ik /
Recovery c Tir / t
6:00 'lir^e'l ^[Hp / I
17:00 ^y^!S^.J{ / % LjJ time 2 r-H / 4
18:00 |7l LJ M
19:00 | |^
Fig. 2. Solution for the second shortage using SSPM.
optimality, since the final solution is dependent on
the sequence of shortest paths chosen. For instance,
in our example above, if the shortest path is found
for the second shortage first, the shortest path to be
reversed for the following run cannot be FJHKI
(arc (if, K) would not exist) and a different final
solution will necessarily follow. In addition, the
model ignores many of the considerations described
in Section 2, like surplus aircraft and allowing
delays.
3. NEW NETWORK MODELS
We present two new network flow models which
will provide solutions in the form of a set of flight
delays or a set of flight cancellations, while allow
ing for swapping aircraft among flights and for
using spare aircraft. As in the SSPM, a network is
used to model the flight schedule. However we do
not use the idea of movement groups and actually
represent individual aircraft and flights as nodes in
the network. The main reason for this is to be able
to differentiate between individual flight-to-aircraft
assignments, since some of these assignments may
not be feasible due to too short or too long ground
times for some aircraft, or due to excessive delays
for some flights.
The salient features of the models are:
Multiple delays are considered.
Multiple cancellations are considered.
Swapping aircraft among flights is permitted.
Usage of spare aircraft is allowed (both those
available at the station where the problem air
craft is or those that can be ferried from other
stations).
The models assume that a disutility can be
assigned to each flight in order to reflect the value
lost if the flight is cancelled, and that the disutility
of delaying each flight is assessable. Developing
such disutility functions is by no means an easy or
exact task and is complicated by flight connectivity
considerations. However, these functions need not
capture the exact disutilities of delays and cancella
tions for the network models presented in this paper
to function properly as long as the numbers are
correct in a relative sense. The factors used to
generate the disutility of a flight delay or a can
cellation include the number of passengers on the
flight, number of passengers connecting when
the flight arrives downline, possible downline delay,
possible downline cancellations, lost crew time and
disruption of aircraft maintenance.
To illustrate, consider the case where a flight ft
incurs a delay d{ at station st. The cost of the delay
dt can be related to the cost of the immediately
following downline delay di+1 of flight fi + 1 at
downline station si+1 using, for example, the fol
lowing recursive function:
DC(dj,/,i,?i) = afC(dj,/i,?i) + cc(/;+1>?l+1)
+ DC(di+1,fi+1,si+1)
where,
DC(di9 fif st) = cost of a dt minute delay of flight
fi at station st.
MC(di9 fi9 Sj) = loss of revenue for passengers
leaving for a flight with another airline + illwill
costs at station st due to d{ minute delay of
flight ft.
CC(fi9 Sj) = cost of missed connections on flight
fi at downline station st + connecting passenger
illwill at st.
One of several stopping rules can be used to
terminate the recursive function, for example, end
of a flying day. United Airlines has been able to
quantify the elements of the disutility function with
acceptable confidence through analysis of past data.
We do not present, in this paper, a detailed analytic
approach for developing these functions.
We choose the flows in our models to be short
ages rather than aircraft. While this may sound
unusual, it is a very natural choice for airline man
agers and flight controllers who are accustomed to
thinking in terms of routing shortages through the
already established flight schedules. This mode of
thinking arises from the fact that it is easier to
track the impact of several shortages on the flight
schedule than to evaluate the possible routings
of the hundreds of aircraft involved. At any rate,
equivalent network formulations are attainable
with aircraft rather than shortage flows.
Finally, the models are meant to be run indepen
dently for each of the fleets of an airline in order to
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
270 / A. I. Z. JARRAH ET AL.
insure compatibility between the flights and the
aircraft undertaking them.
3.1. The Delay Model
This model solves the problem of aircraft short
ages at a station by delaying flights until the
shortage aircraft is fixed. It allows aircraft swap
ping among flights as well as the use of spare
aircraft available at the station or ferried from
other stations. The model is a pure minimum-cost
network with arcs bounded by a flow of unity.
Figure 3 depicts an example for the underlying
network model for station LAX. Here we are
assuming that the current time is 12:00 noon and
that the user has just learned that a certain air
craft will not be available at time 13:15 as planned,
but will need maintenance at the station and will
be available once again at time 16:30. The vertical
axis is a time axis depicting the hours of the day.
The nodes on the left represent aircraft placed at
the points of time when these aircraft are ready to
fly. For example, node 1 represents an aircraft that
can be made ready to fly at noon. The time at which
an aircraft can fly is usually determined by its
arrival time from its previous flight plus a certain
amount of turnaround time which is needed for
refueling, pre-flight checks, loading and unloading,
etc. The nodes on the right represent scheduled
departures of flights. For example node 2' repre
sents a flight departure scheduled for time 13:00.
An arc connecting a node on the left to another on
the right represents an original flight-to-aircraft
assignment at the station. For example, the flight
of node 1' is originally assigned to the aircraft of
node 1. In what follows, an aircraft represented by
a node n will be referred to simply as aircraft n,
and a flight departure represented by a node n' will
be referred to as flight n'.
Time:_Station: LAX_
12:00 I iTT
14:00
17:00
16:00
15:00
13:00
18:00
Aircraft Right
Nodes Nodes
Fig. 3. Structure of the network for the delay model.
The fact that there will b
13:15 is represented by a sup
(with supply presented by >
connected using backward
towards the left) to each of t
than the node of the aircraft
take the flight and other
aircraft associated with the s
node 3' is connected to the a
arcs: (3', 1), (3', 2), (3', 4), (3',
Similar arcs emanate from th
the various aircraft nodes, b
in order to avoid crowding th
flow of one occurs on a
the flight at the tail of the
aircraft at its head. If a
upward, then no delay is invo
departure is later than the
ready to fly, and, hence, th
is that of swapping aircraf
example, if the flow on arc (3
3' will be taken up by the ai
cost involved. The swapping of
4' to flight 3' would, howev
associated with such proced
flight gate for flight 3' from
3 is parked to the gate wher
informing the crews and pass
etc. On the other hand, back
downward involve actual del
the departures. For exampl
(3', 7) is one, then the depart
to be delayed to 4:00 p.m. at
will be available. Arcs that
(3', 7), have associated costs th
the delays.
A recovery node, R, is placed at time 16:30 with a
demand (represented by O >) of one or less to
indicate that the repaired aircraft can be used any
time after time 16:30. Several flights beyond the
repair time (16:30) should be considered in order to
have a bigger pool of flights that can use the recov
ered aircraft; here, we choose to consider the three
flights 6', 5', and 7'. Each of the flight nodes, V
through 7', is connected to the recovery node to
indicate that the repaired aircraft can be used for
any of these flights. The cost on these arcs are,
again, either swap costs if the arcs point upward, or
delay costs, if the arcs point downward. Node SI
represents a surplus aircraft which can either be
available at the station or be ferried from other
stations. The position of the node indicates the time
at which the aircraft will be available to fly from
the station. Arcs connect the surplus node to the
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
FRAMEWORK FOR FLIGHT CANCELLATIONS AND DELAYS / 271
Time:_Station: LAX_
12:00 10^*^
15 00 6o-^sksT% ^^^^6
1600 7?^'-^r^^ R
18:00
Fig. 4. One possible solution for the delay model.
various flight nodes and the cost on these arcs
include both the cost of securing the surplus air
craft and of any delay or swap costs involved. Any
other available surplus aircraft can be modelled in
the exact same way.
The model is to be solved as a minimum-cost
network.151 The solution would be a sequence of arcs
that starts at a supply node and terminates at
a recovery node or a surplus node. For example,
Figure 4 shows the solution to be: 33'66'77'#. This
means that flight 3' will be delayed to be picked up
by aircraft 6, flight 6' will be picked up by aircraft
7, and flight 7' will use the recovered aircraft.
Hence, this solution involves only one delay. Figure
5, shows another solution, 33'44'S1, which involves
no delays and makes use of the surplus aircraft.
Here, flight 3' is picked up by aircraft 4, while
flight 4' uses the surplus aircraft.
A mathematical formulation of the delay model
will be given next. In what follows, we define
the candidate flights for an aircraft as those most
suited for reassignment to the aircraft given the
flights' departure times and the aircraft availabil
ity time. While this set can conceivably include all
the remaining flights (other than the originally
assigned flight) of the day which do not violate the
maintenance and crew restrictions, some of these
flights do not represent interesting choices because
of the associated excessive departure delays or
aircraft layover time. Exclusion of such flights
would reduce the input data collection and proces
sing needed for the model. Similarly, we define the
candidate aircraft for a flight as those most suited
for undertaking the flight given the aircraft's avail
ability times and the flights' departure times. In
addition, the following terms are defined prior to
the statement of the model:
Time: Station: lax
18:00
17:00
14:00
16:00
15:00
13:00
12:00
1 P*CK^
Fig. 5. A second possible solution for the delay model.
A = set of aircraft originally scheduled for
flights.
a = index for aircraft cgA.
F= set of flights considered.
f= index for flight feF.
S = set of surplus aircraft considered.
s = index for surplus aircraft seS.
R = set of recovered aircraft considered.
r= index for recovered aircraft r e R.
</>(a)= flight originally assigned to aircraft a.
a(f)= aircraft originally scheduled to undertake
flight f. (Due to the one to one mapping of
flights to aircraft, we have a = <?-1).
Fa = subset of F consisting of candidate flights
considered for aircraft a. If a is a shortage
aircraft, Fa is set to empty.
Af = subset of A consisting of candidate air
craft considered for flight f.
Fs = subset of F consisting of candidate flights
considered for surplus aircraft s.
Sf = subset of S consisting of candidate sur
plus aircraft considered for flight f.
Fr = subset of F consisting of candidate flights
considered for recovered aircraft r.
Rf = subset of R consisting of candidate recov
ered aircraft considered for flight f.
cfa = the delay and/or swap costs involved in
reassigning flight f to aircraft a.
cfs = the total ferrying and delay and/or swap
costs involved in reassigning flight f to
surplus aircraft s.
cfr = the delay and/or swap costs involved
in reassigning flight f to recovered
aircraft r.
qa= - 1 if there exists a shortage involving
aircraft a; 0, otherwise.
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
272 / A. I. Z. JARRAH ET AL.
The decision variables involved are:
(1 if the original assignment of
flight <f>(a) to aircraft a
is discarded
(or the original assignment
of flight f to aircraft
a(f) is discarded)
otherwise
Xfa =
Xfs =
Xfr ~~
,0
if the flight f is reassigned
to aircraft a
otherwise
if flight f is assigned to
surplus aircraft s
otherwise
if flight f is assigned to
recovered aircraft r
otherwise
The model can be expressed as:
mmLf&F(LaGAfcfaxfa + Ls Esfcfsxfs
+ ^r<ERf CfrXfr)
subject to:
^f^Fa\{<f>(a)} Xfa ~ y<Ka),a
VcgA
= Fa\{<Ka))
^a^Af\{a(f)} Xfa + ?
< 1 VsgS
Sf xfs + ? reRf
*fepa *fs
?/* e Fr Xfr < 1 VrGi?
^(a), a > ^V, > */a > X/s > */>
e [0,1] Va,/\s,r
(1)
(2)
(3)
(4)
(5)
(6)
The objective function (1) sums over all the flights
the costs involved in rescheduling flights among the
initially scheduled aircraft (term 1), in assigning
flights to surplus aircraft (term 2), and in assign
ing flights to recovered aircraft (term 3). These
costs include delay and/or swap costs in addition to
possible ferrying costs for surplus aircraft. Equation
2 has a right-hand side of -1 when there exists a
shortage involving aircraft a; otherwise, the right
hand side of the equation is zero. In the first case,
the set Fa would be empty, since no flights can be
assigned to the shortage aircraft a, and the equa
tion would simply enforce y^a\a to be equal to
one thus deleting the assignment of flight (f>(a)
to aircraft a. In the case where there is no shortage
involving aircraft a, the equation simply forces the
sum of assignments of flights to aircraft a (term 1)
to equal 1 only if flight </>(a) is no longer to be
assigned to aircraft a (i.e., when j^(a)>a is 1). Simi
larly, Equation 3 enforces the sum of reassign
ments of flight f to the various aircraft (left-hand
side) to be equal to 1 only if flight f is no longer to
be assigned to aircraft a(f) (i.e., when right-hand
side term is 1). Equation 4 ensures that at most one
flight gets reassigned to each surplus aircraft s,
and Equation 5 does the same for each recovered
aircraft r. Finally, (6) enforces the flow to be
bounded by 1. Of course, since the model is a pure
network the decision variables will assume values
of exactly 0 or 1 as desired.
3.2. The Cancellation Model
This minimum-cost network model solves the
problem of aircraft shortages by providing an opti
mal solution consisting of a set of flight cancella
tions. All the flows on the arcs of the network are
restricted to be less than or equal to one. The model
can handle multiple cancellations, and makes use
of aircraft swapping and surplus planes.
Figure 6 illustrates the structure of the proposed
network. The nodes within each station are defined
in the same way as for the Delay model. Unit
supplies at nodes 3 and 9, indicate a shortage of
one aircraft at each of the nodes. Node Rl repre
sents the time at which the problem aircraft at
LAX becomes available once again and has a
demand less than or equal to one thus allowing, but
not requiring, the use of the recovered aircraft.
Because no delays are allowed in this model, only
flights 6', 5', and 7' have connecting arcs into Rl
since these flights can use the repaired aircraft
without incurring any delays. Similarly, R2 repre
sents the time at which the problem aircraft at
PHX becomes operational and can receive arcs from
flight 14' and any other later flights. The arcs
connecting nodes across stations represent actual
transfers across stations of aircraft performing
flights. For example, the arc (4', 21) means that the
aircraft performing flight 4' will be physically
transferred to JFK where it will be ready to fly
again at time 16:30. The costs on such a "transfer"
arc represents the revenue which would be lost if
the flight represented by the tail of the arc were to
be cancelled. In order to avoid congesting the fig
ure, only three such transfer arcs are drawn. The
nodes SI and S2, represent two surplus aircraft
available in the system. The demand at each of
these nodes is less than or equal to one thus allow
ing but not requiring the use of these aircraft.
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
FRAMEWORK FOR FLIGHT CANCELLATIONS AND DELAYS / 273
Station: LAX PHX JFK
Time:
13:00
18:00
16:00
14:00
17:00
15:00
12:00
Fig. 6. Structure of the underlying network for the cancella
tion model.
Furthermore, an arc would connect a flight to one
of these nodes only if it is possible to have
the corresponding surplus aircraft available for the
flight without incurring delays. For example, if an
arc connects 19' to S2 then the surplus aircraft S2
can be made available for flight 19'. The cost on
such arcs would be the cost incurred in making the
surplus aircraft available for the desired flights.
Note that the cancellation flow can now cancel
either into a recovery node where a shortage air
craft is recovered, or into a surplus aircraft node.
The only backward arcs allowed at each station
are those that involve no delays. For example, the
only backward arc allowed out of node 9' would be
arc (9', 8). Backward arcs for the other flights are
not shown in the figure to avoid congestion. Each of
these arcs is assigned a cost associated with swap
ping (see Section 3.1 above for more explanation).
It should be noted that an alternative model can
be obtained if the aircraft and flight nodes are
combined at all the stations, and aircraft supply
nodes are defined only for aircraft that are unable
to undertake their scheduled flights on time. In this
network, an arc connects a flight node i to a flight
node j, if the aircraft released after the completion
of flight j can be used for flight i without incurring
a delay and if the waiting time for the released
aircraft is not deemed excessive. This would reduce
the number of needed nodes in the model; however
it leads to a complication. Several unit flows can
now arrive at and leave from each of the flight
nodes which is not acceptable since each flight can
provide at most one aircraft for swapping. This
situation will not occur when both aircraft and
flight nodes are used (as in Figure 6), since the
forward arcs connecting the two sets of nodes have
an upper bound of one.
The mathematical statement for the cancella
tion model is similar to that of the delay model
except that, now, no flight-to-aircraft reassignments
involving delays will be considered. The set A now
includes the set of aircraft indices at all the sta
tions. Each flight f has a corresponding aircraft
index r(f). Each flight f releases its aircraft r(f)
for further scheduling. Similarly, an aircraft with
index a is said to be released for scheduling after
performing the flight T~1(a). To illustrate, in Figure
6 flight 4' releases the aircraft with index 21. The
following definitions are needed in addition to the
definitions used in the delay model:
r(f) = the aircraft released for scheduling
by flight f.
r_1(a) = the flight that releases aircraft
a for scheduling.
cf = the cost of cancelling flight f.
zf (or zT-i(a))
1 if flight f (or r_1(a))
is cancelled
{0 otherwise
The model is described as:
minEfeF(EaG Af cfaxfa + Tt8GSfcfsxfs
~^^r^Rf Cfr Xfr + CfZf)
subject to:
^feFa\{<Ka)} Xfa ~ y<f>(a),a + Zr~\a) -
= qa VaEA
^o? Af\{a(f)) Xfa~^^s^Sf X fs + e R f Xfr+Zf ^
LfeFtxf8<l VsgS (10)
EfeFr*/r<l Vrefl (11)
ya,Ma)> yf,a(f)> Zf> ZT-\a)> Xfa> Xfs> Xfr
e [0,1] Va,/*,s,r
The objective function (7) now has an additional
term over objective function (1) for capturing can
cellation costs. The interpretation of Equation (8)
identical to that of Equation (2) except that now
shortage can be initiated either due to an extern
cause (by having the right-hand side equal to -
or by cancelling an incoming flight (i.e., if zT-i(a)
1). Similarly Equation (9) is identical to Equatio
(3) except that now if the right-hand side is
(indicating that the original assignment of flight
to cf)(f) is discarded) the situation can be handled
either by a swap (when xfa is 1), a surplus o
recovered aircraft (when xfs or xfr is 1, respe
(7)
(8)
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
274 / a. i. z. jarrah et al.
tively), or a flight cancellation (when zf is 1).
4. COMPUTATIONAL EXPERIENCE
In this section, we will present the results of the
experiments conducted for the delay and cancella
tion models on representative sample data from
UA. These models are general minimum-cost net
works which involve multiple sources and sinks.
However, we can combine the source nodes to one
master source node using arcs with flows bounded
by one. The supply at the master node would
be equal to the total number of shortages impac
ting the system. Similarly, all the recovery and
surplus nodes can be made to feed into one master
sink node using arcs with flows bounded by one,
and the demand at the master sink node would be
set to the total supply into the system. Based on
this transformation, we implemented the Busacker
Gowen's dual algorithm11' 8] for the minimum-cost
flow network problem in which the shortest path is
solved repeatedly to achieve the necessary flow.
The shortest path procedure was modified to gener
ate simple paths, since the costs on the edges could
be negative and negative cycles could exist.[6] The
algorithm was coded in C and implemented on the
DEC3100 workstation.
4.1. Results for Delay Model
Three stations, Chicago, San Francisco and
Denver, were chosen for experimentation. These
stations have a large number of departures per day
and meaningful problems could be studied.
Inputs to run the model are:
the flight data for the station in question.
the disutility of delay for the outgoing flights at
that station.
The flight data give us information about the
incoming flights and the outgoing flights. A typical
flight turn at, say Chicago, is represented as fol
lows:
734 1433, GEG B 248 1514 ICT
This means that flight 734 lands in Chicago at
2:33 p.m. from Spokane (GEG) and turns to flight
248 leaving Chicago at 3:14 p.m. for Wichita (ICT).
The aircraft used here is an equipment 737 type B.
Tables I, II, and III show the results of the runs
for Chicago, San Francisco and Denver respec
tively. Column 1 represents the equipment types
for which delay has occurred. Swaps can be effected
within equipment type only. Columns 2 and 3 spec
ify the time bank or interval within which the
model is run. Only aircraft which either arrive
during this time period or are already on the ground
are considered for swaps. Column 4 indicates the
number of delayed aircraft. This includes aircraft
that were actually delayed and also those air
craft which are at the station but cannot leave due
to mechanical problems. Column 5 indicates the
number of spare aircraft available for swaps. These
include actual spare aircraft if any, as well as
aircraft which arrive at the station and leave
the next day (overnight layovers). Column 6 speci
TABLE I
Experimental Results for the Delay Model at CHICAGO
Total Delay Total Delay Disut. Disut. Run
Time Bank No. Delayed No. of No. of No. of w/o Model w/Model w/o Model w/Model Time
Equip Begin End /Mech. Spare Swaps Delays (min) (min) ($) ($) (sec)
B,Q,N 500 900 2/0 0 4 2 70 321 3800 290 r_
B,Q,N 500 900 2/1 0 6 3 146 80 4504 931
B,Q,N 500 1100 10/0 0 15 10 416 317 19,379 4632
B,J,P 1600 2359 28/2 33 35 5 503 238 57,289 12,503 30
B, J,P 1600 2359 28/2 33 35 5 503 208 57,289 11,132 28
TABLE II
Experimental Results for the Delay Model at SAN FRANCISCO
Total Delay Total Delay Disut. Disut. Run
Time Bank No. Delayed No. of No. of No. of w/o Model w/Model w/o Model w/Mo
Equip Begin End /Mech. Spare Swaps Delays (min) (min) ($) ($) (sec)
B,E 800 1100 3/0 0 4 2 127 114 10,622 8847 *~
B,E,N 800 1100 5/0 0 10 3 187 167 15,308 3242 *
B, J, F 1500 2359 9/1 13 9 4 245 365 17,472 5410 *
B, J, F 1500 2359 13/2 13 17 7 423 488 22,602 6834
B,J,F,P 1500 2359 20/2 13 25 7 550 556 45,707 7071 16
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
FRAMEWORK FOR FLIGHT CANCELLATIONS AND DELAYS / 275
TABLE III
Experimental Results for the Delay Model at DENVER
Total Delay Total Delay Disut. Disut. Run
Time Bank No. Delayed No. of No. of No. of w/o Model w/Model w/o Model w/Model Time
Equip Begin End /Mech. Spare Swaps Delays (min) (min) ($) ($) (sec)
E, K, B 730 1100 2/0 0 6 1 121 90 4830 521 r~
E,K,B 900 1200 4/1 0 9 2 112 60 12,186 1812
E,K,D 730 1200 7/3 0 13 7 244 324 17,326 4965
B,E,D 1500 2359 9/1 17 10 3 181 207 12,830 2868
B,P,L 1500 2359 13/2 17 18 5 542 455 41,837 9030 *
All runs made on DEC3100 workstation; * indicates CPU less than 10 seconds.
fies the number of swaps suggested by the model.
Column 7 indicates the number of delays in the
solution. Columns 8 and 9 give the total delay of all
flights in minutes before and after running the
model. The total delay before using the model is the
cumulative delay of all flights if the aircraft took
the original turns as specified in the flight data.
Columns 10 and 11 give the disutility for the same
two situations.
The following remarks pertain to the results of
the runs:
The problem situations tested included inci
dental delays which consist of a few every day
delays, and mass delays as a result of incle
ment weather at a station. The model generated
effective implementable solutions in reasonable
time and is definitely amenable to real-time
implementation.
In some instances, the time of total delay
in the proposed solution was greater than if
the delayed flights took their scheduled turns.
Upon examining the solution in more detail
it was clear that in these cases a single flight
was delayed for a very long time because this
flight had a relatively flat disutility versus time
curve. For example in Table II, run 3 had a
delay of 245 minutes initially (using originally
scheduled turns) and the proposed solution had
a delay of 365 minutes. Table IV shows the
scheduled turns for the flights for this problem
situation. Table V shows the revised turns pro
posed by the model. The model recommends
flight 1222 to Spokane (GEG) be delayed for
270 minutes which constitutes 78% of the total
delay in the proposed solution. One can observe
that, the flight had very little or no delay cost.
From another viewpoint, we may say that this
flight is a candidate for cancellation since not
too many passengers will wait for 270 minutes.
Runs 4 and 5 in Table I represent the mass
delay situation due to inclement weather. A
total of 30 planes have delays, 28 of them due
to weather and 2 due to mechanical problems.
The number of spares is indicated as 33 all of
which are in overnight layovers. In run 5 an
actual spare was added to the model which
resulted in a disutility reduction of 1471. The
solutions obtained were very attractive for both
runs as evident from the dramatic reduction in
disutility.
We also tested the model with stations which
had under 20 departures per day. Overall, we
concluded that the delay model proved effective
in stations which had a high volume of flights.
Stations with very few flights have very little
operational flexibility for the model to be very
effective although it finds obvious solutions.
4.2. Results for the Cancellation Model
For testing the model, three scenarios were
considered by dividing the country into regions:
the eastern region, the central-west region and the
entire country as one region. The experiments were
conducted for one UA sub-fleet, the 737Bs which
are aircraft with a seating capacity of 128 and a
range of about 4 hours of non-stop flying.
Tables VI, VII and VIII show the results of the
runs for the cancellation model. It clearly validates
the efficiency and applicability of the model for
implementation in a real-time decision support sys
tem. Note that it is often the case that the num
ber of cancellations exceed the number of aircraft
shortages in the system. This is due to the fact that
each of these aircraft are typically required to per
form multiple flights during each day. Columns 4
through 7 display the number of cancellations, the
number of swaps, the disutility, and the time taken
to generate the solution.
The following remarks pertain to the results of
the runs:
The aircraft shortages included those which
became available later on during the day (like a
delayed arrival) as well as real shortages in
which the aircraft is just not available for the
rest of the day. In the former case, the model
attempts to avoid cancellations by performing
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
276 / A. I. Z. JARRAH ET AL.
TABLE IV
Initial Schedule Turns for Run 3 Table II
Incoming
Flight No.
Arrival
Time
Arriving
from
Equip
Type
Outgoing
Flight No.
Departure
Time Destination
1712
1759
1112
0535
1114
1037
0250
1194
1294
1121
0675
1042
0973
1770
1242
1714
0897
1116
1769
1716
1118
1170
1499
1718
1127
1071
0346
1228
1283
0571
1634
1178
1439
1068
1186
1473
0961
1120
1073
0347
1122
0175
1033
1176
1722
1171
1435
1260
1124
0517
1132
1234
1177
1026
1274
1016
1537
0439
0019
0093
1126
1128
1781
0027
1449
1427
1422
1516
1519
1519
1523
1525
1526
1527
1530
1530
1531
1534
1535
1549
1615
1619
1622
1647
1715
1745
1746
1749
1749
1750
1751
1751
1753
1753
1755
1756
1758
1802
1802
1805
1805
1819
1826
1907
1920
1921
1926
1940
1947
2002
2014
2015
2019
2022
2025
2026
2028
2030
2040
2040
2043
2045
2052
2112
2121
2217
2225
2320
LAX
SEA
LAX
ORD
LAX
PDX
SAN
ONT
LAS
SEA
GEG
MRY
MCI
BUR
PHX
LAX
DEN
LAX
SEA
LAX
LAX
LAS
MFR
LAX
EUG
PDX
BUR
SBA
SLC
MSY
LGB
SNA
BOI
ONT
MRY
PHX
GEG
LAX
SEA
DEN
LAX
ORD
SEA
BUR
LAX
MFR
EUG
LGB
LAX
ORD
SBA
SNA
SEA
LAS
ONT
BUR
SAN
PHX
EWR
BOS
LAX
LAX
SEA
IAD
N
B
B
B
B
N
B
D
B
N
E
N
D
B
E
N
B
N
K
C
J
E
B
N
N
F
B
N
J
B
B
F
B
E
E
E
K
B
B
F
J
E
B
B
J
N
B
B
N
M
D
B
B
B
E
B
E
E
M
F
J
N
B
F
1719
1514
1758
535
1247
1222
250
1194
1063
1121
675
1042
514
1770
818
1721
567
1123
1772
1723
1125
1458
1499
1718
1127
1593
995
1725
682
1275
1074
1178
1439
490
1186
1473
1502
1778
1073
38
1131
1777
1033
1173
1421
1101
1476
582
1245
318
1273
1234
1177
1026
1051
1701
1750
882
24
16
1752
1703
1257
92
1530
1610
1530
1615
1601
1600
1605
1605
1610
1600
1605
1610
1615
1630
1630
1630
1700
1700
1730
1730
1800
1845
1835
1830
1900
2055
1835
1830
1830
1835
1845
1840
1840
1840
1840
1845
0720
2110
1905
2210
2100
2100
2100
2055
0905
0700
2115
2115
2100
0630
0850
2110
2140
2125
0915
0730
0630
0830
2215
2210
0730
0830
0645
0800
LAX
EUG
SEA
BUR
SAN
GEG
ONT
PDX
LGB
LAX
MRY
MFR
DEN
SEA
ORD
LAX
SNA
LAX
SEA
LAX
LAX
PHX
LAS
SEA
LAX
SNA
BUR
LAX
SLC
LGB
GEG
PDX
SAN
DEN
BOI
ONT
PDX
SEA
SBA
EWR
LAX
PHX
BUR
SBA
SAN
LAX
EUG
DEN
ONT
DEN
SBA
MFR
SAN
MRY
ONT
LAX
SEA
DEN
JFK
IAD
SEA
LAX
BUR
BOS
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
FRAMEWORK FOR FLIGHT CANCELLATIONS AND DELAYS / 277
TABLE V
Turns Proposed by the Delay Model for Run 3 of Table II
Delay Action Incoming Arrival Arriving Equip Outgoing Departure
(min) Code* Flight No. Time From Type Flight No. Time Destination
0 1712
0 1759
0 1112
0 0535
0 1114
0 S 1037
0 0250
0 1194
0 1294
0 1121
0 0675
0 1042
0 0973
0 1770
0 1242
0 1714
0 0897
0 1116
0 1769
0 1716
0 S 1118
15 D 1170
0 S 1499
0 1718
0 1127
0 1071
0 0346
0 1228
0 S 1283
0 0571
0 1634
0 1178
0 1439
0 1068
65 SD 1186
0 S 1473
15 SD 0961
270 SD 1120
0 1073
0 0347
0 S 1122
0 0175
0 1033
0 1176
0 1722
0 1171
0 1435
0 1260
0 1124
0 0517
0 1132
0 1234
0 1177
0 1026
0 1274
0 1016
0 1537
0 0439
0 0019
0 0093
0 1126
0 1128
0 1781
0 0027
1449 LAX N
1427 SEA B
1422 LAX B
1516 ORD B
1519 LAX B
1519 PDX N
1523 SAN B
1525 ONT D
1526 LAS B
1527 SEA N
1530 GEG E
1530 MRY N
1531 MCI D
1534 BUR B
1535 PHX E
1549 LAX N
1615 DEN B
1619 LAX N
1622 SEA K
1647 LAX C
1800 LAX J
1830 LAS E
1900 MFR B
1749 LAX N
1749 EUG N
1750 PDX F
1751 BUR B
1751 SBA N
1810 SLC J
1753 MSY B
1755 LGB B
1756 SNA F
1758 BOI B
1802 ONT E
1835 MRY E
1905 PHX E
1830 GEG K
2000 LAX B
1826 SEA B
1907 DEN F
2130 LAX J
1921 ORD E
1926 SEA B
1940 BUR B
1947 LAX J
2002 MFR N
2014 EUG B
2015 LGB B
2019 LAX N
2022 ORD M
2025 SBA D
2026 SNA B
2028 SEA B
2030 LAS B
2040 ONT E
2040 BUR B
2043 SAN E
2045 PHX E
2052 EWR M
2112 BOS F
2121 LAX J
2217 LAX N
2225 SEA B
2320 IAD F
1719 1530 LAX
1514 1610 EUG
1758 1530 SEA
535 1615 BUR
1247 1601 SAN
1499 1835 LAS
250 1605 ONT
1194 1605 PDX
1063 1610 LGB
1121 1600 LAX
675 1605 MRY
1042 1610 MFR
514 1615 DEN
1770 1630 SEA
818 1630 ORD
1721 1630 LAX
567 1700 SNA
1123 1700 LAX
1772 1730 SEA
1723 1730 LAX
682 1830 SLC
1458 1900 PHX
1778 2110 SEA
1718 1830 SEA
1127 1900 LAX
1593 2055 SNA
995 1835 BUR
1725 1830 LAX
1186 1840 BOI
1275 1835 LGB
1074 1845 GEG
1178 1840 PDX
1439 1840 SAN
490 1840 DEN
1125 1905 LAX
1131 2100 LAX
1473 1900 ONT
1222 2030 GEG
1073 1905 SBA
38 2210 EWR
1502 0720 PDX
1777 2100 PHX
1033 2100 BUR
1173 2055 SBA
1421 0905 SAN
1101 0700 LAX
1476 2115 EUG
582 2115 DEN
1245 2100 ONT
318 0630 DEN
1273 0850 SBA
1234 2110 MFR
1177 2140 SAN
1026 2125 MRY
1051 0915 ONT
1701 0730 LAX
1750 0630 SEA
882 0830 DEN
24 2215 JFK
16 2210 IAD
1752 0730 SEA
1703 0830 LAX
1257 0645 BUR
92 0800 BOS
* Indicates whether outgoing flight has been swapped (S), delayed (D) or both (SD).
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
278 / a. i. z. jarrah et al.
table vi
Results for the Cancellation Model for EASTERN Region
Total
Start Aircraft No. No. of Disutility Total CPU
Equip. Time Shortage Cancelled Swaps ($1000) (sec)
737s 500 1 2 3 72 SM
737s 500 2 4 5 258 3.09
737s 1200 3 2 6 54 2.50
737s 500 4 4 9 111 3.20
737s 500 5 6 11 187 3.33
table vii
Results for the Cancellation Model for CENTRAL WEST Region
Total
Start Aircraft No. No. of Disutility Total CPU
Equip Time Shortage Cancelled Swaps ($1000) (sec)
737s 500 1 2 7 50 490
737s 500 2 5 7 127 5.11
737s 1200 3 7 10 177 5.20
737s 500 4 9 17 266 5.33
737s 500 5 14 18 381 5.45
table viii
Results for the Cancellation Model for Entire Airline
Total
Start Aircraft No. No. of Disutility Total CPU
Equip Time Shortage Cancelled Swaps ($1000) (sec)
737s 500 I 2 3 51 ioo
737s 500 2 4 6 61 4.22
737s 500 3 7 10 158 4.33
737s 500 4 10 10 185 4.46
737s 000 5 12 18 222 6.50
swaps whereas in the latter case it is forced to
cancel flights.
The flight turns information used contained the
turns for a 24-hour period and it was sufficient
to balance cancellations for the 737s because of
the large size and high flight frequency of the
fleet per day. Balanced cancellation sequences
or loops can be found within 24 hours of flying.
For fleets which have a high average flying
time per flight, turns over 2 or 3 days may be
required to balance cancellations.
5. EXTENSIONS, LIMITATIONS AND RESEARCH
PROSPECTS
The authors view the models and discussions pre
sented in this paper as an initial effort in address
ing the problem of flight cancellations and delays in
the airline industry. The insights gained from the
analysis formed the basis for an evolving decision
support system at UA. However, many complex
issues exist besides and beyond the immediate
decisions of swapping aircraft and delaying or can
celling flights. It is important to discuss and under
stand these issues and to identify possibilities for
future research for improved approaches. Through
out the discussion, one should bear in mind that
the problem at hand is a real-time one, and, hence,
future related efforts should focus on good and
quick solutions for realistic models.
5.1. Crew Considerations
Crew scheduling is the important problem of
assigning flight crews to tours of duty (called bids)
extending several weeks (see Introduction). The
problem is complicated by a host of constraints
that define a feasible bid, like the maximum
allowed flight time, and time away from home,
maximum allowed layover time between assign
ments, etc. Cancelling or delaying flights often
necessitates changes in the scheduled crew assign
ments. Because of the enormous complexity involved
in combining the crew scheduling problem with
the cancellations/delays problem, we suggest an
approach similar to what is done in practice: once a
solution is identified, the resultant extensions or
modifications in the tours of duty are negotiated
with the affected crews to obtain their approval. If
a solution that involves an extension in the duty
period of a certain crew gets rejected by that crew,
then the backward arc causing the unacceptable
extension is deleted, and the model is rerun for an
alternative solution which is acceptable to the crew.
For example, if choosing arc (3', 7) (see Figure 3)
results in an unacceptable extension of the duty
period of the crew for flight 3', arc (3', 7) is then
removed for rerunning the model. Similarly if a
cancellation results in unacceptable changes in the
schedule of the affected crew, the corresponding
cancellation arc is deleted in the next search for a
feasible solution.
5.2. Aircraft Maintenance Considerations
The scheduled assignment of aircraft to flights is
done in a way which insures that each of the air
craft receives all its various types of maintenance
checks which are required after certain prespeci
fied numbers of flying hours. This, in itself, is
a complex problem, which requires extensive effort
(see Introduction). When aircraft are resched
uled (due to delays, cancellations, or swaps) it is
important to ensure that all affected aircraft will
still receive their scheduled maintenance. To check
if this condition is met, the network of flights can
be searched exhaustively to see if the aircraft
undertaking the flight can reach an appropriate
maintenance airport in time. So again, the approach
is to attempt and recapture feasibility after the
decision to cancel/ delay flights or swap aircraft is
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
FRAMEWORK FOR FLIGHT CANCELLATIONS AND DELAYS / 279
made. In some severe situations, the problem can
not be resolved satisfactorily without the usage of
surplus aircraft beyond those suggested by the
models.
5.3. Combining Cancellations and Delays
Flight controllers typically attempt to remedy the
problem of aircraft shortages first by considering
the possible usage of delays. If needed delays are
deemed excessive, the controllers would consider
cancelling flights to absorb the shortage. The mod
els we presented in Sections 3.1 and 3.2 are shown
to be effective in aiding the controllers in these
efforts. In addition, it would be desirable to investi
gate the possibility of finding even better solutions
through combining delays and cancellations. This
can be modeled using the network in Figure 6
except that we now allow both upward and down
ward backward arcs. For example, flight 3' would
now have backward arcs emanating from it to nodes
1, 2, 4, 5, 6 and 7. Similar backward arcs emanate
from the various flight nodes at a given station to
all the surplus and recovery nodes at the station.
The above model, if solved as minimum-cost net
work, will not, necessarily, give a correct solution,
because whenever a delay arc is chosen, the posi
tion of some of the network nodes will be impacted
and the costs on some of the backward arcs will no
longer be valid. To illustrate, suppose the optimal
solution to the network of Figure 6, once solved as
a minimum-cost network, involves a flow of one on
the backward arc (2', 6). This means that the
departure of flight 2' is to be delayed to 15:15,
the time at which aircraft 6 is ready to fly. This in
turn means that node 18 will have to be shifted
downward to reflect the subsequent delay in the
arrival of flight 2', and this will impact the delay
costs on the backward arcs which connect into node
18 at JFK. This problem represents an opportunity
for future research for approaches that can capture
this delay/cancellation interaction. The salient fea
ture of any such approach would be the treatment
of the departure time of the flights as a variable
that is updated depending on the delays chosen in
the solution. The authors are currently involved
in developing and testing a model along these lines.
5.4. Multicommodity Approaches
The exposition in this paper is based on the
assumption that shortage problems involving air
craft of a certain fleet are to be resolved using
cancellations, delays, and swaps within that same
fleet. This was done to insure compatibility of
the aircraft with the asssigned flight segments.
Another approach is to consider all the fleets simul
taneously while allowing assignment of flights to
compatible fleets only, which results in a multicom
modity formulation. The obvious shortcoming of
such an approach is the added complexity, while
the advantage is the opportunity of obtaining bet
ter solutions because of the increased number of
aircraft to choose from.
Another relevant aspect is the concept of hub and
spokes which is prevalent in the modern airline
industry. Here, a bank of incoming flights feed
passengers from various airports (spokes) into a
major airport (hub), where the passengers redis
tribute among the next bank of departing flights.
The system has proven advantages from both oper
ational and revenue considerations. However, the
system makes the development of disutility func
tions for delays and cancellations a difficult task
and the assessment of the impact on connecting
passengers only approximate. As an alternative,
one can visualize an additional commodity repre
senting the flow of the passengers in the system,
and relate the commodity to the aircraft commodi
ties and flows. Needless to say, the resultant
problem is rather complex.
In conclusion, we have presented two network
models which can assist in choosing which flights to
delay or cancel in the event of unexpected short
ages of aircraft due to situations that may arise
during the operation of an airline. Computational
experiments reveal the amenability of the models
to real-time interactive usage. A decision support
system is currently under development at United
Airlines based on this study. We have also identi
fied various avenues for further research in this
important airline scheduling problem.
ACKNOWLEDGMENTS
We would like to thank S. G. Regulinski of United
Airlines for his helpful comments. In addition we
would like to thank the Editor for his insightful
remarks, and the anonymous referees for their sug
gestions regarding the paper's conceptual develop
ment and the model formulations.
REFERENCES
1. R. G. Busacker and P. G. Gowen, "A Procedure for
Determining a Family of Minimum-Cost Flow Pat
terns," Operations Research Office Technical Report
15, Johns Hopkins University, 1961.
2. M. M. Etschmaier and D. F. X. Mathaisel, "Airline
Scheduling: An Overview," Trans. Sci. 2, 127-138
(1985).
3. T. A. Feo and J. F. Bard, "Flight Scheduling and
Maintenance Base Planning," Mgmt. Sci. 35, 12,
1415-1432 (1989).
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms
280 / A. I. Z. JARRAH ET AL.
4. I. Gershkoff, "Aircraft Shortage Evaluator," pre
sented at the ORSA/TIMS Joint National Meeting,
St. Louis, MO (October), 1987.
5. F. Glover, D. Klingman and N. V. Phillips, Network
Models in Optimization and Their Applications in
Practice, John Wiley & Sons, New York, 1992.
6. R. V. Helgason and J. L. Kennington, Algorithms
for Network Programming, John Wiley & Sons, New
York, 1980.
7. O. Holst and B. Sorensen, "Combined Scheduling
and Maintenance Planning for an Aircraft Fleet,"
Operational Research '84, 735-747 (1984).
8. M. M. Syslo, N. Deo and J. S. Kowalik, Discrete
Optimization Algorithms, Prentice Hall, Englewood
Cliffs, NJ, 1983.
9. D. TEODOROVltf AND S. Guberini<5, "Optimal Dispatch
ing Strategy on an Airline Network after a Schedule
Perturbation," Eur. J. Opnl. Res. 15, 178-182 (1984).
10. D. Teodorovi<3 and G. Stojkovi6, "Model for Oper
ational Daily Airline Scheduling," Trans. Plan.
Technol. 14, 273-285 (1990).
11. G. Yu and A. I. Z. Jarrah, "A Mathematical Program
ming Model for Airline Maintenance Scheduling,"
Department of Management Science and Information
Systems Working Paper 89/90-3-3, University of
Texas at Austin, 1990. Also, presented at the ORSA/
TIMS Joint National Meeting, Philadelphia, PA
(October), 1990.
(Received, July 1991; revisions received March 1992, October
1992; accepted January 1993)
This content downloaded from 155.69.24.171 on Wed, 19 Jul 2017 03:12:54 UTC
All use subject to http://about.jstor.org/terms