Flight Schedule Planning

with Maintenance Considerations

Julia L. Higle
Anne E. C. Johnson

Systems and Industrial Engineering

The University of Arizona
Tucson, AZ 85721

Abstract

Airline planning operations typically begin with the specification of the master flight sched-
ule. While the master schedule is typically determined based on strategic business objectives,
such as market share and competitive considerations, efficient operation of the schedule leads
to a series of challenging problems that are constrained by FAA regulations, labor agree-
ments, market forces, etc. In this paper, we propose a model that integrates a surrogate
representation of FAA regulations regarding aircraft maintenance within a master flight
schedule planning problem, and examine its impact on the resulting schedules.

This research was funded by Grant No. DMS-0400085 from the National Science Founda-
tion.
1 Introduction

Airline operations begin with a set of scheduled flights, from which costs and revenues are
generated. The development of the flight schedule involves forecasting demand for flights
between different airports, including considerations of competition, market share, and air-
plane and staffing availability. A “good” schedule maximizes profit within the constraints
of an airline’s human resources and infrastructure. It is fair to say that airline planning
problems are characterized by their large and complex nature, and give rise to some of the
most challenging of optimization problems. For example, once a flight schedule has been
settled upon, a variety of problems related to the operation of the schedule result. Aircraft
and crew must be assigned to specific flights in accordance with complicated labor regula-
tions. Aircraft must be maintained in accordance with Federal maintenance guidelines, and
ticket pricing issues must be resolved. In general, each of these problems leads to a rich and
complex area of research, such as “fleet assignment problems”, “crew scheduling”, “main-
tenance scheduling”, and “revenue management.” Illustrative works in these areas include
Hane et al. [1995], Gopalan and Talluri [1998], and McGill and Van Ryzin [1999] respectively.
Due to the complexity of these problems, they are typically considered in isolation, and only
after the flight schedule has been determined. Exceptions to this tendency include Cordeau
et al. [2001], who integrate the aircraft routing and crew scheduling problems for a single
fleet type. Other examples include Cohn and Barnhart [2003], Stojković and Soumis [2001],
and Freling et al. [2000]. Similarly, Barnhart et al. [1998] investigate a model that integrates
the Fleet Assignment Model (FAM), found in Hane et al. [1995], and a duty pairing problem
(DPP) model, which is an approximation of the crew pairing problem. Note however, that
each of these studies require the master flight schedule as input to their models.

In this paper, we consider a modelling paradigm that permits a movement toward planning
methods that permit a representation of these complicated operational problems within the
schedule planning phase. Our method involves the development of a surrogate representation
of an operational problem — a model that is relatively easy to solve, yet sufficiently repre-
sentative of the problem to reflect its impact on schedule choices. To illustrate the concept,
we develop a surrogate model of maintenance scheduling. In particular, we use the concept
of “maintenance opportunities,” which stand as surrogates for the more complicated notion
of “maintenance feasibility.” While the former concept can be determined via the solution
of a linear program (which we propose as a subproblem of a schedule planning problem),
the latter requires the solution of a complex combinatorial optimization problem designed
to ensure strict compliance with FAA maintenance regulations.

2 Maintenance Scheduling

Just as crews require periodic training, aircraft require periodic maintenance, although main-
tenance requirements are more frequent and varied than crew training. Each US-operated
Flight Schedules with Maintenance 2

aircraft must meet FAA and airline requirements for scheduled maintenance. Aircraft must
undergo four types of maintenance, as explained by Talluri [1998], from infrequent major
overhauls to minor visual inspections every few days. Major overhauls are generally accom-
modated on a rotating basis, and have the impact of an effective reduction in the number
of aircraft available at any given time. Visual inspections require one or two hours every
65 flying hours by FAA rules, although most major airlines self-impose a stricter 45-hour
maximum interval. Other types of minor maintenance include engine oil changes and other
minor tasks needed to keep an airplane in service. Our model applies to consideration of the
impact of minor maintenance on the flight schedule.

Current methods tend to consider maintenance schedules after the flight schedule has been
determined. The reasons for this are fairly clear from a computational point of view. Sched-
uled maintenance requires knowledge of the flights assigned to each specific aircraft (called
the “tail assignment”). The immense size of an airline’s operations, involving up to 2500
flights, 200 airports and 10 distinct fleets, makes it unrealistic to expect to consider the tail
assignment and maintenance schedule while initially planning the master schedule. From a
decision modelling point of view, the reasons are equally clear. The master schedule, which
forms the nucleus of the airline’s business operations and revenue generating capacity, is the
product of a strategic planning process in which big-picture issues such as competition and
market shares are of primary importance. Maintenance, on the other hand, takes place on
an operational level. As such, it requires a closer level of attention to detail. Nonetheless,
maintenance requirements impose constraints on the operation of the master schedule, which
must be “maintenance feasible.”

Talluri [1998] and Gopalan and Talluri [1998] developed procedures for routing aircraft to
meet basic maintenance requirements. Their method heuristically finds maintenance-feasible
four-day aircraft routings in polynomial time, and can identify flights which must be shifted
to another fleet to achieve maintenance feasibility. This algorithm can be particularly use-
ful when airlines route specific airplanes late in the planning process. Alternatively, Cohn
and Barnhart [2003] show that by planning maintenance routings before crew pairings, the
solution quality of both problems improves. They state that crew pairings often produce
infeasible maintenance routings, and planning the maintenance routings first limits selected
crew pairings to those which allow for required maintenance opportunities. Throughout our
development, we assume that all minor maintenance can be performed at a set of prop-
erly equipped and staffed maintenance stations. A maintenance opportunity occurs when
an airplane spends a sufficiently long period of time at a maintenance station, whether
or not maintenance is actually performed. As a stand-in for the more complex notion of
maintenance feasibility, our basic premise is that if a schedule has sufficient maintenance
opportunities built into it, the schedule is likely to be maintainable with out excessive delay.
Flight Schedules with Maintenance 3

3 Models

We begin with a model for the identification of the master schedule. Our model involves the
determination of flights, which are modelled as flows in a dynamic network. Flights require
a given amount of time to travel from one airport to another, and airplanes must spend a
minimum amount of time at the new location upon arrival. Each airport has a finite gate
capacity, which restricts the number of aircraft that can be on the ground at any given time.
In addition, each airport has a limited number of takeoff and landing time “slots” in any
period of time. We model this as a dynamic network, in a manner that is similar to the
network description that appears in Hane et al. [1995].

Airline schedules are cyclic in nature, and are designed to repeat every c units of time.
In formulating the model, simultaneous considerations of time and space are of paramount
importance. We adopt the following notation throughout this paper.

Sets:

I denotes the set of distinct locations (airports) under consideration

T = {0, ..., tmax }, is the set of distinct times associated with an appropriate discretization
of time.
N ⊂ I × T , is the set of nodes in the dynamic network.
A ⊂ N × N is the set of arcs in the dynamic network.

Thus, each node in the dynamic network corresponds to a specific location in I and a specific
time in T . For each n ∈ N , let

i(n) denote the location to which n corresponds,

τ (n) denote the time to which n corresponds.

Let

tij denote the standard time required to travel between locations i and j;
γ(i) denote the minimum “turn” time required at location i, the time required for unloading
passengers and cargo, cleaning, fueling, inspecting, and reloading the aircraft;
cij denotes the estimated profit generated by a flight between locations i and j.

An arc (n, m) ∈ A is either a flight or time spent on the ground. Let

Flight Schedules with Maintenance 4

AF denote the set of flight arcs, where (n, m) ∈ AF implies that i(n) 6= i(m) and
τ (m) − τ (n) = ti(n)i(m) ,

AG denote the set of ground arcs, where (n, m) ∈ AG implies that i(n) = i(m) and
τ (m) − τ (n) = 1.

Finally, for each t, let

N (t) ={n ∈ N | τ (n) = t} the set of nodes that correspond to time t

A(t) ={(n, m) ∈ A | τ (n) ≤ t and τ (m) > t}, the set of arcs that are active at time t

Ni (t) ={n ∈ N (t) | i(n) = i}, the set of nodes that correspond to location i and time t.

Typically, the cycle c is one day, and tmax may be any time up to one month. In addition,
we have the following data:

ri− , (ri+ ) is the maximum number of aircraft that can takeoff (land) at location i.

gi is the maximum number of aircraft that can be on the ground at location i.

V is the number of airplanes available in the system.

Our model is designed to identify values of the following decision variables

xnm is the “flow” (i.e., the number of aircraft) on arc (n, m) ∈ A.

that maximize “profit” while maintaining an appropriate number of “maintenance opportu-

nities”, where the latter is denoted by the function fmx (x) and is defined in §4.
Flight Schedules with Maintenance 5

MSP:
X
Max ci(n)i(m) xnm (1)
(nm)∈AF
X
−
s.t. xnm ≤ ri(n) ∀ n ∈ N , (nm) ∈ AF (2)
m
X
+
xnm ≤ ri(m) ∀ m ∈ N , (nm) ∈ AF (3)
n
X
xnm ≤ gi(m) ∀ t, m ∈ N (t), (nm) ∈ AG (4)
n
X
xnm ≤ V (5)
(nm)∈A(0)
X
fmx (x) ≥ xnm (6)
(nm)∈A(0)
X
xmn ≤ xpq ∀ (pq) ∈ AG (7)
γ(i(p))
(mn)∈AF ,3 n∈{Ni(p) (τ (p)−t)}t=0
X X
xmn − xnm = 0 ∀ n ∈ N (8)
m m
xnm − xn0 m0 = 0 ∀ (nm) ∈ A, i(n) = i(n0 ), n ∈ N (t),
τ (n0 ) = τ (n) + c, t ∈ {0, . . . , max(tij )} (9)
xnm ∈ Z + ∀ (nm) ∈ A (10)

The objective function maximizes the net profit associated with the master schedule. Con-
straints (2) limit the number of flights departing a location to the takeoff capacity, while (3)
similarly limit the number of landings at each location in each period. Constraints (4) ensure
that holding capacity is not exceeded at any location, while constraint (5) prescribes that
the number of assigned airplanes does not exceed the number available, and is called a plane
count constraint. With (6), we require at least as many maintenance opportunities as the
plane count. Equations (7) require that flights stay on the ground for at least the minimum
turn times. In other words, given xpq , where (pq) ∈ AG the number of aircraft assigned
on the ground arc must include at least all of the flights that have arrived γ(i(p)) or fewer
time units before τ (p) (and are therefore in their mandatory ground hold time window). Re-
source flow is maintained by (8). Constraints (9), known as wraparound constraints, ensure
identical schedules for each cycle.

4 A Model of Maintenance Opportunities

Minor frequent maintenance requires each aircraft to pass through a maintenance station
every few days. As previously discussed, scheduling this maintenance results in a large
Flight Schedules with Maintenance 6

integer program, which is prohibitive as a component of MSP. We incorporate this mainte-

nance requirement by counting maintenance opportunities, layovers meeting or exceeding a
minimum time interval at a designated maintenance station, within the MSP. Since minor
maintenance must occur every T days (typically, T is 3-4 days), we can count the number of
maintenance opportunities in this time window, and require that the number of maintenance
opportunities is at least as large as the number of aircraft (as in (6)). Note that this does not
guarantee that each individual aircraft will be serviced, nor does it minimize maintenance
cost; it simply offers a measure of whether a schedule is “maintainable.” The maintenance
opportunity count also acts as a measure for comparison between alternate schedules to
estimate which is the most maintainable.

In the following, we assume that T > 0, and present a model from which fmx (x), the number
of maintenance opportunities associated with master schedule x, can be derived. We restrict
our attention to the case in which there is only a single type of maintenance that must be
scheduled, and comment on extensions to multiple types of maintenance at the end of the
section. We begin by constructing a “maintenance network” based on the dynamic network
used for MSP. To do this, we overlay “maintenance arcs” on the network, which correspond
to sequences of enough ground arcs to permit maintenance to take place. Figure 1 illustrates
the flight network and the resulting maintenance opportunities. Figure 1a displays the basic
MSP structure, where horizontal lines represent potential ground arcs, and diagonal lines
represent potential flights. Phoenix (PHX) appears as a maintenance station, while Salt Lake
City (SLC) and Los Angeles (LAX) do not have maintenance capability. Figure 1b depicts
a collection of flights (and ground holds) between these cities. Assuming that maintenance
requires two time periods and that at most two aircraft can undergo maintenance at any
time, Figure 1c, illustrates a reorganization of the flights that provides four maintenance
opportunities.

In addition to the notation established for MSP,

tmx is the time required to perform scheduled maintenance

I mx is the set of locations where scheduled maintenance can be performed;

Amx is the set of maintenance arcs, (nm) ∈ Amx if i(m) = i(n), i(n) ∈ I mx , and τ (m) −
τ (n) ≥ tmx ;

Amx (t) ={(n, m) ∈ Amx | τ (n) ≤ t and τ (m) > t}, the set of maintenance arcs active at
time t;

hi is the number of crews at maintenance station i;

T is the maintenance cycle time, typically three to four days.

With xnm as the flight schedule proposed by MSP, we have the following decision variables:
Flight Schedules with Maintenance 7

SLC

PHX
(MX)

LAX

a) Dynamic Network Model

SLC

PHX
(MX)

LAX

b) Scheduled Flights

SLC

PHX
(MX)

LAX
c) Flights Scheduled, with four Maintenance Opportunities

Figure 1. Network Structure. Time moves from left to right.

Flight Schedules with Maintenance 8

ynm is the “flow” on maintenance arc (nm) ∈ Amx

znm is the “flow” on non-maintenance arc (nm) ∈ AG , i(n) ∈ I mx

We define the number of maintenance opportunities for a given master schedule, x, as

X
MX: fmx (x) = Max ynm (11)
(nm)∈Amx
X
s.t. y`m ≤ hi(n) ∀ n ∈ N mx (12)
(`m)∈Amx (τ (n))
X
y`m + znp = xnp ∀ n, (np) ∈ AG , i(n) ∈ I mx (13)
(`m)∈Amx (τ (n))

ynm ≥ 0 ∀ (nm) ∈ Amx (14)

znm ≥ 0 ∀ (nm) ∈ AG (15)

Constraints (12) ensure that the number of maintenance opportunities does not exceed the
number of available maintenance crews at any place and time. Equations (13) assign the
flow on ground arcs at maintenance stations in the master schedule to maintenance or non-
maintenance arcs.

Our presentation of the MSP-MX model highlights the potential to solve this problem using a
variation of Benders’ Decomposition. Indeed, this is readily accomplished. In cases where the
flight network is only moderately sized, the maintenance constraints can simply be directly
incorporated into the MSP without resorting to a decomposed presentation. In §7 we discuss
extensions of the model, many of which favor solution by decomposition. We note that the
use of the surrogate measure “maintenance opportunities” in place of the requirement of
maintenance-feasibility eases the computational requirements of such integrated decision
making. Nonetheless, it simultaneously raises questions regarding the suitability of the
surrogate. Ideally, a large number of maintenance opportunities would indicate a readily
maintainable schedule, while a small number would indicate a schedule with the potential
for maintenance difficulties. In other words, additional effort must be expended in order to
determine the validity of the approximation.

5 Validation of the Surrogate Model

To test and demonstrate the validity of the surrogate model, we undertake the more involved
task of actually testing the maintenance feasibility of the master schedule identified via the
solution of the integrated model (MSP). This is accomplished via a “string model” similar
Flight Schedules with Maintenance 9

to those described in Barnhart et al. [1998]. A string is a series of flights that are assigned
to a single aircraft. Although string models are not the most efficient method for solving
the maintenance scheduling model, they are relatively simple and provide a useful basis
for comparison with the approximation results. The heuristic method proposed by Talluri
[1998] is more efficient computationally, but is more difficult to implement. Our goal in
this section is to explore the validity of the output of our surrogate model rather than to
efficiently identify an operationally viable maintenance schedule. For this reason, we validate
the surrogate representation with the more easily implemented, but time consuming, string
model.

Our validation model begins with a collection of binary variables,

fnms is 1 if flow on arc (nm) ∈ A is assigned to string s (i.e., aircraft s); 0 otherwise

φns is 1 if a maintenance arc begins at node n on string s; 0 otherwise

σs is 1 if string s contains at least one maintenance arc; 0 otherwise

P
In this manner, s σs denotes the number of aircraft that utilize at least one maintenance arc.
If this number is equal to the number of aircraft needed to fly the schedule, then the schedule
is “maintenance feasible.” If the maximum number of strings that contain maintenance arcs
is less than the number of aircraft, then the schedule is not maintenance feasible.

To represent potential maintenance arcs and minimum ground holds, we introduce two sets.
These sets are primarily notational aids for expressing the model. For each n such that
i(n) ∈ I mx , let

Amx
n ={(`, m) ∈ Amx | i(`) = i(n), andτ (`) = τ (n)}, the set of maintenance arcs that
originate with node n.

AG n ={(`, m) ∈ AG | i(`) = i(n), τ (n) ≤ τ (`) < τ (n) + γ i (n)}., the set of ground arcs
necessary to satisfy the minimum ground hold constraint for flights landing at n.

The set Amxn represents the arcs that may be used to schedule maintenance at location
i(n) ∈ I mx beginning at time t(n). Similarly, the set AG n represents the set of ground arcs
that are associated with the minimum ground hold constraint for flights landing at location
i(n) at time t(n).

Given a master schedule x, a more precise representation of a maintenance schedule may be

Flight Schedules with Maintenance 10

obtained via the solution to

X
max σs (16)
s
X
s.t. φns ≥ σs ∀s (17)
n∈I mx
X
f`ms ≥ φns ∀ n ∈ N mx , s (18)
(`m)∈Amx
n
X X
φns ≤ hi ∀ t, ∀ i ∈ I mx (19)
s n:i(n)=i, τ (n)≤t<τ (n)+tmx
X X
fmns = fnms ∀n ∈ N,s (20)
m m
X
fnms = xnm ∀ (nm) ∈ A (21)
s
X X
fmns ≥ f`s ∀n ∈ N,s (22)
m:(mn)∈AF (`p)∈AG n

σs , φns , fnms ∈ [0, 1] ∀ n, m, s (23)

Constraints (17) and (18) drive variables σs to indicate whether string s contains a mainte-
nance arc or not. Since σs is binary, the objective is increased by 1 if one or more maintenance
arcs is assigned to a string, so that assigning more than one maintenance arc to a string does
not result in any additional benefit. In (19), flow on maintenance arcs at any one place and
time is limited to the number of available crews, as in (12). Flow is conserved by (20). These
constraints also associate individual flight legs with strings. Equations (21) require that each
unit of flow on an arc is assigned to a string, and (22) ensure that landing flights enter a
ground hold or maintenance arc. In a maintainable P flight schedule, every string will include
at least one maintenance opportunity, so that s σs equals the total number of aircraft.

The string model follows the same maintenance time requirement as the approximation
model. The solution obtained from the approximation model indicates whether the number
of maintenance opportunities in the schedule is at least as great as the number of aircraft.
Consequently, the string model cannot find more maintenance arcs than the surrogate model.
If the surrogate model indicates that not enough maintenance arcs exist for the schedule to
be maintainable, it is impossible for the string model to show a contradictory result. When
the surrogate model indicates maintenance feasibility, it is not immediately clear that the
schedule is necessarily maintainable. We will test this empirically.

6 Experiment

Using the string model, we now have a basis from which to test the validity of our surrogate
representation of maintenance. Within our experimental setting, we used “small” networks
Flight Schedules with Maintenance 11

containing five airports. Each of the problems that we used was constructed using randomly
generated data, as described below.

To generate problem data, we begin with networks of 5 airports over 144 thirty-minute time
periods, or three days, with a maximum of 15 aircraft in each schedule. For each network, we
used a pseudo-random number generator to produce grid coordinate pairs for each city. The
coordinates were uniformly distributed on the interval [100,3000] in the east-west direction
and [100,1500] in the north-south direction. From the coordinates, we calculated Euclidean
distances between the cities, and linearly transformed these distances into matrices of travel
times. We then randomly generated profit per flight and demand per flight ({cij } and
{dij }, respectively) using uniform distributions, and selected airport capacities and number
of maintenance crews at each airport. Initially, each airport was identified as a maintenance
station with probability 0.2. Maintenance stations are then designated sequentially, although
this designation is adjusted to ensure that at least one, and no more than two, airports are
designated as maintenance stations. Within each maintenance station, there are 2, 3, or 4
maintenance crews with probabilities 0.1, 0.2, and 0.7, respectively.

Flight schedules were generated for each network. Maintenance feasibility of these schedules
was evaluated using both the surrogate model and the string model. The results were then
compared for consistency. Our goal is to test the predictive capability of the surrogate model
for a variety of master schedules on these various networks. If the string model finds a prob-
lem to have fewer maintainable strings than the number of aircraft needed to fly the schedule,
then we say that the schedule is not maintainable. Similarly, if the surrogate model finds
fewer maintenance opportunities than aircraft, the schedule is identified as unmaintainable.
For each of the networks generated, we compared the indications of maintenance feasibility
obtained from the surrogate model to those obtained from the string model.

Table 1 contains a summary of these comparisons for five different sets of problem data.
MSP was used to generate flight schedules, with and without the constraint on maintenance
opportunities, (6). We use “a” to indicate an MSP solution based solely on profit (i.e.,
with (6) omitted), and “b” to indicate the profit maximizing solution with (6) included.
Consequently, in the first data set, profit-only considerations yield six maintenance opportu-
nities, and results in six maintainable flight strings for 15 airplanes. Adding the constraint
on maintenance opportunities results in an abundance of maintenance opportunities (120
opportunities for 15 airplanes), which the string model confirms as maintainable. In all
cases, the surrogate and the string model reached the same conclusion regarding mainte-
nance feasibility. The results of this experiment indicate that the surrogate model is useful
for determining whether a schedule is maintainable.

It is interesting to note that in each case, the MSP objective value was not affected by the
inclusion of the constraint on maintenance opportunities. This indicates that the algorithm
explored alternate optimal solutions as the Benders’ cuts shaped the master schedule towards
a maintenance-feasible one.

As expected, the solution times of the two maintenance models differed sharply. The time
Flight Schedules with Maintenance 12

Problem Surrogate Model String Model

Number # Planes MX opportunities # maintained
1a 15 6 6
1b 15 120 15
2a 15 12 9-12
2b 14 36 14
3a 15 6 6
3b 15 84 15
4a 15 12 0
4b 15 30 15
5a 15 6 6
5b 15 30 15

Table 1: Results of Model Validation Experiment.

required to solve the surrogate model is essentially negligible, while the string model requires
much more time to solve, even for these small networks. In some cases (e.g., 2a), the string
model failed to identify an optimal solution after several hours (although it indicated a
lower bound of 9 maintainable strings and an upper bound of 12). When employing a
solution technique that requires the subproblems to be solved many times, such as Benders’
decomposition, these time savings are extremely helpful in finding an overall optimal schedule
within a “reasonable” amount of time.

To demonstrate scalability, we applied the master schedule planning process to much larger
networks, with 20, 25, and 30 airports and 144 time periods. The resulting problems are
too large to subject the schedule to an investigation of maintenance feasibility via our string
model. However, in each of the problems that we generated and solved, we found that the
solutions obtained using the MSP with the constraint on maintenance included sufficient
opportunities for the aircraft to be maintained, whereas the initial solutions for the same
problems were generally not maintainable schedules. As in the smaller problems, the objec-
tive values were not affected by the inclusion of the maintenance constraint. Table 2 shows
the results of solving a few large problems with and without maintenance considerations in
the model.

Problem Initial # MX Final # MX

Number # Airports # Planes opportunities opportunities
6 20 60 0 66
7 25 75 6 92
8 30 90 60 112
9 30 90 0 112
10 30 90 18 108
11 30 90 24 109

Table 2: Demonstration of Scalability.

Flight Schedules with Maintenance 13

7 Extensions

The discussion above explores the validity of a surrogate model in approximating main-
tenance scheduling, and thus to rapidly judge a schedule for maintenance feasibility. For
simplicity, the MX model did not consider several factors that impact real aircraft mainte-
nance scheduling. Some of these considerations may be explored as extensions of our model.

“Safety factor.” In our model, we required that the number of maintenance opportunities
be at least as great as the number of aircraft needed to fly the schedule. Since the number of
aircraft available may vary, especially for larger networks, and since delays and other schedule
disruptions may cause deviations from the planned schedule, we may want to include a
multiplier on V in the model. For example, we may want the schedule to contain at least
αP> 1 maintenance opportunities per plane, so the right-hand side of constraint (6) becomes
α (nm)∈A(0) xnm . Inclusion of this safety factor may decrease the profit of the schedule, so
the tradeoffs between profit and reduced risk of loss of maintenance feasibility should be
analyzed.

Multiple forms of maintenance. Our investigation assumed a single type of aircraft

maintenance. In reality, different types of maintenance must be performed at different inter-
vals, and possibly in different locations. Major maintenance, such as annual inspections, may
be modeled by reducing V , reflecting the fact that a given number of aircraft are routinely
removed from service for an extended period of time. With varying forms of frequent main-
tenance required, we may use multiple sets of maintenance arcs, ie., {Amx i }, and constrain
each type separately.

Random maintenance times. When mechanics discover problems on an aircraft during

a short inspection, repair times can easily exceed tmx . To capture scheduled maintenance
events that exceed the predicted time, we may reformulate MX as a stochastic linear program
with randomness in the number of time periods required for membership of nm in set Amx .
Since the number of potential scenarios for extended maintenance can be large, our solution
technique would need to incorporate stochastic programming solution techniques, which
typically favor the decomposed presentation of the integrated model.

Unscheduled maintenance.Unscheduled maintenance occurs when an airplane “breaks”

during a trip and requires immediate attention to remain in service after landing. These
problems range from burned-out lights to engines damaged by birds, and generally must be
fixed at the airport where the plane landed. Sometimes technicians or parts must be flown
to non-maintenance stations to repair the plane, which takes time and exacerbates schedule
disruption. While scheduled maintenance is largely predictable, unscheduled maintenance
occurs randomly. While our focus in this paper is on scheduled maintenance, the impact
of schedule disruptions on the operation of the flight schedule is examined in a forthcoming
paper.
Flight Schedules with Maintenance 14

8 Conclusions

As discussed, airline schedules are very large, complex problems which currently are not
typically developed in a manner that explicitly recognizes the impact of operational problems
on the profitability of the schedule. Approximations of these related operational problems
can allow the development of a schedule based on a more realistic representation of its
ultimate profitability. Our validation experiment demonstrates the viability of one such
approximation model. In particular, it appears that the approximation model is a useful
surrogate for the exact model, and can be used for indicating maintenance feasibility of
flight schedules.

The data and models used in this demonstration are simplified. For example, the models
do not account for common requirements such as overnight layovers. Similarly, the profit-
per-flight data depend only on the origin and destination of the flight, ignoring the influence
of time of day, crew costs, winds, and myriad other factors. Many of these items can be
incorporated in a more comprehensive version of MSP. Nevertheless, the concept of the
surrogate model as demonstrated is still effective.

The surrogate model for maintenance is clearly useful in identifying feasible schedules. The
model may also be used for long-term planning decisions, such as how many maintenance
stations to operate and where to locate them. These types of problems can be analyzed by
adding relative cost coefficients for the maintenance stations to the objective function, and
by changing the maintenance crew availability data to reflect each option.

Future research in this investigation will explore surrogate models for other subproblems in
airline planning, such as crew scheduling and revenue management. Our goal is to develop
a set of surrogate models that can be used in concert to find an optimal schedule in an
integrated manner.

Acknowledgement
This research was funded by Grant No. DMS-0400085 from the National Science Foundation.
Flight Schedules with Maintenance 15

References
Barnhart, C., F. Lu, and R. Shenoi (1998). Integrated airline schedule planning. pp. 384–403.

Cohn, A. M. and C. Barnhart (2003). Improving crew scheduling by incorporating key

maintenance routing decisions. Operations Research 51(3), 387–396.

Cordeau, J.-F., G. Stojković, F. Soumis, and J. Desrosiers (2001). Benders decomposition for
simultaneous aircraft routing and crew scheduling. Transportation Science 35(4), 375–388.

Freling, R., D. Huisman, and A. P. M. Wagelmans (2000). Models and algorithms for
integration of vehicle and crew scheduling. Technical Report ERS-2000-14-LIS, Erasmus
Research Institute of Management.

Gopalan, R. and K. Talluri (1998). The aircraft maintenance routing problem. Operations
Research 46, 260–271.

Hane, C., C. Barnhart, E. Johnson, R. Marsten, and G. Nemhauser (1995). The fleet as-
signment problem: solving a large-scale integer program. Mathematical Programming 70,
211–232.

McGill, J. I. and G. J. Van Ryzin (1999). Revenue management: Research overview and
prospects. Transportation Science 33(2), 233–256.

Stojković, M. and F. Soumis (2001). An optimization model for the simultaneous operational
flight and pilot scheduling problem. Management Science 47(9), 1290–1305.

Talluri, K. T. (1998). The four-day aircraft maintenance routing problem. Transportation

Science 32(1).