Journal of Advanced Transportation, Vol. 43, No. 2, pp. 153-182
sadvanced-transportcom
‘izing Fleet
Availability. under the Presence of Flight and
Maintenance Requirements
George Kozanidis
Every aireraf, military or ivilian, must be grounded for
maintenance after it has completed a certain number of fight
hhours since its last maintenance check. In this paper, we
address the problem of deciding which available aircraft
should fly and for how long, and which grounded aircraft
should perform maintenance operations, in a group of airraft
that comprise a combat unit. The objective is to achieve
maximum availability of the unit over the plannis
We develop a multiobjective optimization model for this,
problem, and we illustrate its application and sofution on a real
life instance drawn from the Hellenic Air Foree, We also
propose two heuristic approaches for solving large seale
instances of the problem. We eonchide with a discussion that
ives insight into the behavior of the model and of the
hheurstis, based on the analysis of the results obtained.
Keywords: fleet availablity, flight and maintenance planning,
‘multiobjective mixed integer linear program, military arera.
1. Introduction
The Air Force and the commercial aitline industry have several
similarities, but also exhibit significant differences. Safety is the most
important factor in both industries; however, while maximization of
profit is naturally the overall objective in the commercial airline industry,
‘maximization of the readiness to respond to external threats is the main
objective in the Air Force. Therefore, military aireraft operational
problems should generally be treated differently than traditional
problems arising in the commercial airline industry.
In this paper, We address the problem of the joint flight and
‘maintenance planning (FMP) of military aireraft. The FMP problem
poses the question of which available aircraft should fly and for how
George Kozanidis, Systems Optimization Laboratory, Department. of
Mechanical & Industrial Engineering, University of Thessaly, Greece
Received: April 2007 Accepted: July 2156 G Kozanidis
Jong, and which grounded aireraft. should perform maintenance
operations, in a group of aireraft that comprise a combat unit. The
objective is to achieve maximum availability of the unit over the
planning horizon. The FMP problem is a very important decision making
problem in the Air Foree. Although it involves both operations and
maintenance decisions, we treat it as a unified operational problem.
The safety standards used by Air Force organizations of different
countries are often similar, due to the fact that they are usually preseribed
by the aircraft manufacturer and there are a few such manufacturers
worldwide. Each aireraft must be grounded for a routine maintenance
check as soon as it completes a certain number of flight hours since its
last maintenance check. There are also restrictions regarding calendar (as
opposed to flight) time and number of takeoffs, but these are rarely used
in practice, because the flight time restrictions usually apply sooner.
The FMP problem addressed in this paper was motivated by a case
study in the Hellenic Air Force (HAF). The HAF is primarily responsible
for Greece’s national air defense. It is split into four Divisions: Division
of Tactical Air Foree, Division of Ait Support, Division of Air Training,
and a fourth division responsible for other services. All units responsible
for air operations and missions belong to the Division of Tactical Ait
Force. Further down the organizational structure of the Division of
‘Tactical Air Force, we find the Combat Wings, which are subdivided
into squadrons.
To address the FMP problem, we formulate it as a multiobjective
mixed integer linear program, The objective functions of this model
‘maximize the unit’s readiness to respond to extemal threats. Due to the
fact that there is a large number of parameters that increase the
complexity of the problem and the time needed to reach an optimal
solution, we also propose two heuristic approaches for the solution of
large-scale instances. We explore the behavior of the model and of the
heuristics, and we illustrate their application on a real life instance drawn
from the HAF.
‘The remainder of this paper is structured as follows. In Section 2, we
describe in detail the FMP problem and we review the related literature.
In Section 3, we present the mathematical model that was developed for
solution. The application of this model on a real-life instance drawn
from the HAF is presented in Section 4. Section 5 presents two heuristic
approaches that can be utilized for solving large instances of the problem
and Section 6 concludes this work.A Maltiobjective Model 157
2. Problem Description and Background Information
‘The problem that we study arises as a routine operations
‘management problem in a typical Combat Wing of the HAF. This wing
consists of three squadrons, each of which serves as the base for several
airerafl of various types. At the beginning of each planning horizon, the
wing command issues flight requirements for each squadron and period
combination. These requirements denote the total time that the aircraft of |
the corresponding squadron should fly in each period of the planning
horizon. They are expressed as target values, from which only small
deviations are permitted. Separate requirements are issued for each
aireraft type, since different aircraft types have different light
capabilities and maintenance needs. Therefore, the model introduced in
this paper was developed for use on a specific aircraft type. Of course,
this model can be applied repeatedly until all plans have been issued, if
‘more than one aircraft type is involved.
For each specific aircraft, we define its residual flight time as the
total remaining time that this aircraft can fly until it has to undergo a
‘maintenance check. This time is also referred to as “bank time” in the
related military literature (US DoA, 2000). The residual flight time of an
aircraft is positive if and only if this aircraft is available to fly. Similarly,
wwe define the residual maintenance time of a non-available aircraft as the
total remaining time that this aircraft needs in order to complete its
maintenance check. The residual maintenance time of an aircraft is
positive if and only if this aircraft is undergoing a maintenance check
(and is therefore not available to fly). At any time, the total residual
flight time of a squadron is equal to the sum of the residual flight times
of all the aircraft that belong to this squadron. The total residual flight
time of the wing is equal to the sum of the residual flight times of its
squadrons. Clearly, there exist many possible combinations of individual
aircraft residual flight times that can result in the same total squadron or
wing residual flight time.
For the maintenance needs of the wing, there exists a station,
responsible for providing service to the aircraft of the wing. This station
has certain space (also referred to as “dock space”) and time capacity
capabilities. Given the flight requirements for each squadron and period
combination, and the physical constraints that stem from the capacity of
the maintenance station, the objective is to issue a flight and maintenance
plan for each individual aircraft in each squadron of the wing, so that
Some appropriate measure of effectiveness is optimized.G Kozanidis
Maximizing the readiness to respond to external threats is the most
appropriate measure of effectiveness for this application. This readiness
is usually defined as the total number of aircraft that are available to fly:
however, this number alone gives no information about the way that the
total residual flight time of the wing is distributed across the squadrons
and the individual aircraft. With this in mind, the objective functions of
the proposed model maximize the minimum number of available aircraft
of the wing and of each squadron, respectively, and the minimum
residual flight time of the wing and of each squadron, respectively, over
all periods.
‘The HAF and many other Air Force organizations solve the FMP
problem empirically using a 2-dimensional graphical tool called the
“aircraft flowchart”, shown in Figure 1. The vertical axis of this
flowchart represents residual flight time measured in some appropriate
unit, and the horizontal axis represents the indices of the available
aircraft in nondecreasing order of their residual flight times, I being the
index of the aircraft with the smallest and V being the index of the
aircraft with the largest residual flight time, where V’is the total number
of available aircraft. Consider the line segment connecting the origin
with the point with coordinates (V, ¥), where Y is the maximum time that
an aireraft can fly between two consecutive maintenance checks, often
referred to as “phase interval” in the related military literature, This line
segment is also referred to as “the diagonal”. By mapping the available
aircraft of each squadron on an aircraft flowchart, we can visualize the
total availability of the unit.
To describe the smoothness of the distribution of the total residual
flight time among the aircraft, we use a “total deviation index”. This
index is equal to the sum of the vertical distances (deviations) of each
point mapping a single aircraft from the diagonal. The smaller this sum
is, the smoother the distribution of the total residual flight time. Ideally.
the total deviation index is equal to zero, when all points lie on the
diagonal. When issuing the individual aircraft plans, the intention is to
keep each point as close to the diagonal as possible, in order to keep the
total deviation index as small as possible, To date, this technique has
been used to issue the flight and maintenance plans of individual aireraft
in the Combat Wing that we studied.
The research literature dealing with airline operations is quite rich.
Most of the published research in this area, however, has been directed
towards problems in the commercial airline industry, which have
0, the auxiliary binary decision variable Pym inA Maltiobjective Model 165
constraint set (20) is foreed to zero value. Then, constraint set (21) forces
Gny= 10 2610 if Ygy ~ Xs Sine this implies that the residual flight time
of this aireraft drops to zero during period s. Simitarly, constraint sets
(22) and (23) ensure that an aircraft becomes available as soon as its
residual maintenance time drops to zero. If gu > 0, the auxiliary
decision variable ray: in constraint set (22) is foreed to zero value. Then,
constraint set (23) fOrCeS apyes tO LHF gins = Figs Since this implies that
the residual maintenance time of this aireraft drops to zero during petiod
Z
Vane & Yams IM, 1 =LoeeolNals T+ ey
Bene S Gadi), Hs Mly 1 =HyoslNo P= Dog (25)
Sot SX neo, 1AM, =Dy Nae = Deol 26)
Yoana, 11 =T|Mly = Dou P +1 en
fant Gaal ~ dns MAM ly lMal t= Zoo + engy
St Sts ME {Nght = Ayal 29)
Faw Sas Nolet = yon (30)
Constraint set (24) states that the residual flight time of an aireraft
can not exceed Y, and ensures that it will be zero whenever this aircraft is
not available, Similarly, constraint set (25) states that the residual
‘maintenance time of an aircraft can not exceed G, and ensures that it will
be zero whenever this aircraft is available. Constraint set (26) imposes an
‘upper bound on the maximum time that an aircraft can fly during a single
time period. Such a restriction is usually present due to technical reasons.
Constraint set (27) imposes a lower bound on the residual flight time of
each available aircraft, and constraint set (28) imposes a lower bound on
the residual maintenance time of each non-available aircraft. These
constraints are introduced to eliminate the situation in which an aireraft
has negligible residual flight or maintenance time. Constraint set 29)
ensures that the total time that an aircraft flies during a single period does
not exceed its residual flight time at the beginning of the same period
Similarly, constraint set (30) ensures that the total time that the
‘maintenance crew works on a particular aircraft during a single period166 G Kozanidis
does not exeeed the residual maintenance time of this aireraft at the
beginning of the same period.
yy, = Alyy, 8 Lye MA, ep
Yount = Yue 0 =Lye-AM|, 0 G2)
Brat = Fons LM, Nn (33)
Constraint sets (31), (32) and (33) are used to initialize the state of
the system at the first period of the planning horizon. When an aircraft
exits or enters the maintenance station at the first period of the planning
horizon, its residual flight and maintenance times are updated directly:
therefore, variables dy», and fy are never Used.
Nant Hts 205 1 = 1M, 1 =H ys sos|Nile t= Ga)
Ves Sime 2 OS (Mj. 11 =LssssolNile Loo + 1 G5)
Poivs ns Ge Bitar, me AM, 1 =L..1Nub 6)
ys Ayes fone BiNATY, MM =H.) slNs GB)
Finally, constraints (34), (35) and (36), (37) are the nonnegativity
and integrality constraints, respectively.
Having introduced the proposed FMP model, itis useful to illustrate
its similarities and differences with the models developed by Sgaslik
(1994) and Pippin (1998). Sgaslik (1994) develops two models for the
FMP problem which are solved separately. The first is called the Yearly
Planning Model and is utilized to provide the maintenance schedule and
the flight hour distribution. The second model is called the Short Term
Planning Model. It takes as input the maintenance schedule produced by
the Yearly Planning Model and it returns the mission assignments of the
aireraft
The main characteristic of the Yearly Planning Model is that it does
not maximize the fleet availability, but it minimizes the cost associated
With the violation of some of the problem’s constraints (e.g., those
referring to the required flight time, the maintenance capacity and the
flight time of each individual aireraft), while also maintaining a givenA Maltiobjective Model 167
lower bound on the fleet availabitity. On the other hand, the FMP model
introduced above maximizes the fleet availability, without allowing
violation of these constraints. The Short Term Planning Model goes one
step further, by assigning aircraft to specitie missions based on the
requirements of each mission and the characteristics of each aircraft,
\while in the above FMP model, the flight load is expressed in terms of
the total light time that must be flown collectively by the aireraft in each
period of the planning horizon,
Pippin (1998) adopts the aircraft flowchart methodology (US Doa,
2000) and minimizes the cost associated with the deviations of the
individual aircraft residual flight times from their diagonal-line target
values, but does not incorporate the maintenance aspect of the problem
and the apparent difficulties that this introduces. A quadratic formulation
similar to the one that we discuss in Section 5.1 is also developed, which
does not either take into consideration the constraints imposed by the
limited capacity of the maintenance station
4. Application of the Model
In this section, we illustrate the application of the model on a
problem instance drawn from the Combat Wing that we studied, with 3
squadrons, each squadron having 8 aircraft. The planning horizon is 6
monthly periods. At the beginning of the planning horizon, there are 3
aircraft at the maintenance station, one from each squadron. Table 1
shows the required flight times for each squadron and period
combination, Table 2 shows the time capacity of the maintenance station
in each period. This capacity may vary from period to period, due to
personnel vacations, holidays, etc. Table 3 shows the residual flight and
‘maintenance times of the aircraft in the first period of the planning
horizon, In Table 3 (and Table 5 introduced later), bold-style entries
denote maintenance times of non-available aircraft and plain-style entries
denote flight times of available aircraft. For example, the residual
‘maintenance times of the three non-available aircraft in the first period
are gist = 320, gizi = 190 and gss; = 130, respectively. OF course, the
‘maintenance times of an available aircraft and the flight times of a non-
available aircraft are always equal to 0, The values of the other problem
parameters are G~ 320 hours, Y= 300 hours, C= 3, Xge = 50 hours, Yip
= 0.1 hours, Gyuy = 0.1 hours, L = 0.95 and U= 1.05.168
Table 1. Required flight times in hours (S,,)
1
1 2 3 4 5 6
1 150 165 «14014031
M2 135 165145 160160165,
3 165 145135 150160150.
Table 2. Time capacity of the maintenance station in hours
B, 425 47560470 46080
‘Table 3. Initial residual flight and maintenance times in hours
nor!)
1203 4 5 6 7 8
320.248 «171 «48 ~«105 «216 292 134
106 199 77° 177-273 «39 «298-210
162 92 «13018-28829 40145
For the solution of the problem, we apply the weighted sums
approach (see Steuer, 1986). Although many other alternative approaches
exist (see Steuer, 1986 and Ehrgott, 2005), we chose this approach
because the solutions that it retums ate guaranteed to be nondominated
(Steuer, 1986). Thus, itis a good choice for illustrating the application of
the FMP model to the reader, which is our main aim in this work. The
user should keep in mind that this approach may only return very few of
the existing nondominated solutions, even if the weights are diversified.
This could make the complete solution of a multiobjective problem quite
hard.
With this in mind, we consider strictly po jghts wi, wa, ws
and ws such that; + 12 +13 +384 = I. For scaling reasons, we multiply
the first objective with ¥/2 = 150, the second objective with [M\(¥/2) =
450, and the fourth objective with [M| = 3. This is because increasing 2,
or z; of 2, by | is on the average equivalent to increasing z, by ¥/2,A Maltiobjective Model 169
|M(2/2) and [M|, respectively. This way, we get a single criterion
problem with objective Max 150012) + 450w:25 + w3z) + 3yizu. Four
different sets of weights were considered, ic., (0.85, 0.05, 0.05, 0.05),
(0.6, 0.3, 0.05, 0.05), (0.45, 0.45, 0.085, 0.05) and (0.25, 0.25, 0.25, 0.25).
‘The order of the weight values shown above is not important, sinee every
possible combination was tested for each set. Since the number of
distinct weight combinations that we can form for each of these sets is 4,
12, 6 and 1, respectively, the total number of distinct weight
combinations that was considered is equal to 23. In general, as the
number of weight combinations considered increases, the number of
nondominated solutions that the approach will find is expected to
increase, too,
Table 4 shows the nondominated solutions that were obtained from
the solution of the problem with these 23 weight combinations. Only
three nondominated solutions were obtained, which may be partly due to
the fact that the weights are not extensively diversified and the range of
their possible values is not very large. The actual number of distinet
efficient solutions obtained was greater than three, since efficient
solutions that retumed the same objective value combinations were not
always identical. This is a main consequence of the fact that a stight
change of a given flight and maintenance plan can produce a different
solution with the same objective function values.
The ideal values (see Ehrgott, 2005) of the four criteria are 23, 7,
3380.25 and 979.5, respectively. The first solution of Table 4 gives the
ideal values for criteria 2), 2; and z,. The second solution gives higher
value for criterion 2, but lower for criterion 2s, providing this way a
tradeoff. The third solution gives the ideal values for eriteria 2, and 2,
bbut lower values for zs and 2,, providing a different tradeoff. The user
should always keep in mind that alternative nondominated solutions may
exist that give better compromise between the four objectives. Although
the optimal value of the second objective is 7 in all three solutions, the
‘minimum mumber of available aircraft in each squadron decreases if we
do not use z> as an objective function of the model, For example, the
weight combination (w1, w:, ws, v3) = (0.333, 0, 0.333, 0.333) returns the
solution (2), 23, 2. 24) = (22, 6, 3335.25, 979.5) which is dominated by
the second solution of Table 4. This supports the claim that we made in
Section 3 regarding the potential omission of one of the model's
objective functions.
In order to give a sense of the way in which the results are obtained,
wwe show in Table 5 the optimal flight and maintenance plan of the first170 G Kozanidis
airerafl of the first squadron, for the combination with 1,
= 0.25. More specifically, the first row of this table shows the planned
flight and maintenance times of this aireraft for each of the six periods,
and the second row its corresponding residual flight and maintenance
times. We see that this aireraft has residual maintenance time equal to
320 hours at the beginning of the planning horizon. After undergoing
maintenance operations for 10S and 215 hours during the first and
second petiod, respectively, it completes its maintenance service and
exits the maintenance station at the beginning of the third period, with
residual flight time equal to 300 hours. After remaining idle during the
third and the fourth period and flying for SO hours in each of the next
‘two periods, it is left with residual flight time equal to 200 hours at the
end of the planning horizon, All| decision variables take appropriate
values to ensure the correctness of the model. For example, aii = diz
O and ayys ~ ayy ~ As ~ Ais ~ a7 ~ 1, Since this aireraft is unavailable
in the first qwo periods and available in the next five ones. Additionally,
dys ~ 1, since this aireraft exits the maintenance station at the beginning
of the third period. Table 6 shows the optimal fleet availability of the
‘wing and of each individual squadron, expressed both as number of
available aircraft and as total residual flight time. Once the minimum
fleet availability over all periods has been maximized, the model can be
used to find a solution in which the total availability over all periods is
‘maximized, too.
‘Table 4. Nondominated solutions of the problem,
1 21 7 3389.25 9795
a 2 7 3335.25 979.5
3 B 1 2903 856.75
Table 5. Flight and maintenance plan of an individual aircraft
1 2 3 4 5 6 1
iulhin
vig | 320 215 = 300300 300250200A Maltiobjective Model 7
‘Table 6. Optimal fleet availability
a 7 = 7 : T
wala 3 7 , 7 ‘ ‘
Teal ae es as TT
Mat et | nim sis ues tasras ros2s tise
5. Heuristic Approaches
‘The model of the FMP problem that we introduced in Section 3 is a
mixed integer linear program. The problem instance of Section 4
contains 2628 fimetional constraints, $76 continuous and 726 binary
decision variables, not counting the decision variables whose values are
known at the first period of the planning horizon and decision variables
21, Z, 25 and z,, For the solution of the problem with each weight
combination, we used AMPL/CPLEX (see Fourer et al., 2002), version
9.1 with default values, The average time that AMPL/CPLEX needed on
a Pentium IV/2.SGHz processor to reach the optimal solution over the 23
weight combinations was approximately 1673 seconds. The minimum
time was approximately 5 seconds and the maximum time 11687
seconds,
This range of computational times for the solution of the weighted
sums problem is quite large. When the same instance was solved using
21, 2, 25 and 2, a8 the unique objective, the computational times were
approximately 1, 12, 12 and 6 seconds, respectively. Since the problem
instance is always the same and only the objective function weights
differ, this implies that certain combinations of weights and objectives
seem to make the problem harder. An interesting question for further
study is which these combinations are and what their exact influence on
the computational time is,
The above short computational study reveals that the size of the
problem alone is not indicative of the computational effort needed to
reach the optimal solution, While it may be possible to reach the optimal2 G Kozanidis
solution fast in some cases, a much larger computational effort may be
necessary in other ones. This implies that it is necessary to develop
alternative intelligent approaches in order to be able to handle such
difficulties that may arise. The two heuristic techniques that we
developed and used to solve the FMP problem are the Aireraft Flowchart
Heuristic and the Horizon Splitting Heuristic, which are described next.
5.1 Aircraft Flowehart Heuristic (AFH)
The first heuristic procedure that we propose is based on the aircraft
flowchart discussed in Section 2. Besides the fact that it provides a
‘means for the solution of large-scale FMP instances, this heuristic is also
worthwhile studying because it enables us to evaluate the procedure
‘which is used at the moment for the production of military aireraft fight
and maintenance plans not only in the HAF, but in many other Air Force
“organizations, as Well
The flowchart is a very simple but effective technique that has been
used successfully by maintenance officers. When used properly, it
prevents unnecessary backlog of scheduled maintenance inspections
under normal conditions, it prevents a corresponding sudden surge in
requirements for aireraft parts, it allows the unit maintenance officer a
degree of control over individual aircraft hours flown, and it provides a
graphical depiction of fire scheduled maintenance requirements,
‘The intuition behind the utilization of the aircraft flowchart is
straightforward. By providing a wide range of equally spaced aircraft
residual flight times, it establishes a smooth sequence that determines the
order in which the aircraft should visit the maintenance station. This in
tum prevents bottlenecks in the maintenance station and ensures it
smooth utilization. Additionally, it ensures a fairly constant level of
aircraft availability, which is also very important.
Tn current practice, the aircraft flowchart is at best used as a
graphical device by the officer responsible for issuing the flight and
‘maintenance plans. For example, in an aviation maintenance manual of
the U.S. Army in which this flowchart is described (US Do, 2000), that
officer is simply advised to utilize the flowchart by “flying the aireraft
that are above the diagonal to get them down to the line” and “holding
the aireraft that are below the diagonal to bring them up to the line”, No
particular instructions are given on how to do this in the most effective
way. It becomes clear that this procedure is highly subjective andA Maltiobjective Model 173
dependent on numerous minor decisions made by the user. With this
motivation in mind, we outline next how this procedure can be
implemented mote systematically
The application of the AFH requires a number of minor decisions in
each period of the planning horizon. The first such decision regards the
allocation of the time capacity of the maintenance station to the grounded
aircraft. After this decision has been made, the number of maintenance
dock spaces that will become available in the next period is determined,
as well. This information is important, because the production of the
flight plans of the available aircraft depends strongly upon it. For
example, an aireraft which should normally be grounded for maintenance
in the next period, may have to be held back (fly less), if no empty dock
space is available at the maintenance station.
At each period of the planning horizon, the AFH makes the relevant
decisions by computing a “priority index” for each squadron of the wing,
Many different suggestions exist on how this index can be computed; in
our study, we used as priority index the flight load of each squadron in
period 1, divided by its total residual flight time at the beginning of the
same period. A higher index value reveals that the squadron experiences
heavier load with respect to its current availability; therefore, higher
priority should be given to the grounded aircraft that belong to this
squadron. Thus, the maintenance station gives priority to the aircraft that
belong to the squadron with the highest priority index first, and so on. In
order to free dock space, the maintenance station always works
continuously on the same aircraft until its service is completed. Of
course, if not enough time capacity exists, this service may be spread out
over more than one periods. The priority index of each squadron is
updated each time that one of its aircraft completes maintenance, to
reflect the fact that this aircraft becomes available. This is done by
adding ¥ hours to the total residual flight time of that squadron.
Besides determining the order in which the aircraft will receive
‘maintenance service, the squadron priority indices are also used to
determine the order in which the squadrons will occupy dock space that
is emptied at the maintenance station. In tum, this determines the number
of available aircraft in each squadron at the beginning of the next period.
Once this information has been determined, the flight time of each
available aircraft for the current period, , is determined by solving a
simple nonlinear optimization problem, as described next,
‘Assume that the aircraft that will be available in period +1. are
arranged in nondecreasing order of their residual flight times. The index174 G Kozanidis
of the aireraft with the smallest residual flight time at the beginning of
period ris equal to 1. Therefore, i is the slope of the diagonal, then the
residual flight time of this aireraft at the end of period f should ideally be
equal to s(1) ~ s. Similarly, the index of the aircraft with the next largest
residual flight time at the beginning of period ¢ is equal to 2; therefore its
residual flight time at the end of period r should ideally be equal to 2s,
etc. If Vis the total number of aircraft that will be available in period #1
sis clearly equal to ¥/V. Thus, the problem of deciding the flight time of
cach available aireraft reduces to a least squares problem, in which the
total deviation index is minimized. The functional constraints of this
problem ensure that the sum of flight times of all the aireraft that belong
to squadron m lies in the interval [Z5y,, US, that the flight time of each
aircraft does not exceed Xo, and that the residual flight time of each
available aircraft at the end of period r is at least equal 0 Yn
The difference with the traditional least squares problem, solved in a
regression problem for example, is that in that case we are trying to fit an
‘unknown line to a set of known points, while here we are trying to fit a
set of unknown points to a known line. To illustrate the procedure
sketched above, assume that we have determined that J” is the number of
available aircraft that will be available in period r+1. In order to produce
the flight plans for period 1, we index these aircraft in nondecreasing
order of their residual flight times at the beginning of period 1. Let be
the index used. Then, the nonlinear optimization problem that arises is
the following:
Min 1,4. 715
SL Yiu =D Mie F
ES_ 5) )% 5 LU (8)
uoi,, f=1—F
"ver = Yon Vv
x, 20,
In the above formulation, we have replaced the indices m and n of
each aircraft by the index i, for simplicity. In the objective function, the
suum of squares of all deviations is minimized. The first set of constraints
updates the residual flight time of each aircraft at the beginning of theA Maltiobjective Model 175
next period, based on its residual flight time at the beginning of the
current period and the time that it Mies during this period. The second set
of constraints ensures that the flight requirements of the squadron are
satisfied for the current period. The next two sets of constraints ensure
that the lower and upper bounds on the flight times are not violated, and
the last set of constraints accounts for the nonnegativity of the flight
times. Note that in this formulation, y.1 and x; are decision variables,
while yy is a known problem parameter. This is a relatively simple
quadratic problem with linear constraints that can be solved rather easily
(see for example Pardalos and Kovoor, 1990). The above procedure is
repeated successively, until the aircraft maintenance and flight plans for
all periods of the planning horizon have been produced. In brief, the
AFH that we propose can be sketched as follows:
AFH
for each period of the planning horizon do
1. compute the squadron priority indices
2. produce the aircraft maintenance plans
3. determine the aircraft that will be available in the next
period
4. produce the aircraft flight plans by minimizing the total
deviation index
end.
‘The heuristic is straightforward and can be implemented quite easily:
therefore, it can be used altematively when the size of the instance under
consideration prohibits its solution using a commercial optimization
software package. The user is provided with the option of applying the
above quadratic problem once to each squadron, or once to the whole
swing, in each period of the planning horizon. Then, he can compare these
‘two solutions and choose the one that he considers better.
It should be emphasized at this point that the aim of the model
developed and the AFH do not exactly coincide, since the model
‘maximizes the fleet availability, while the heuristic minimizes the total
deviation indices. As a result, even when a solution with higher fleet
availability exists, the AFH. will prefer another one with lower deviation
index, even if this leads to lower fleet availability. Despite the existence
of such extreme peculiarities, the AFH seems to perform quite
satisfactorily in general, mainly because it prevents bottlenecks in the
‘maintenance station and ensures a smooth utilization of the maintenance
station. This is the main reason that it is currently used by many Air
Force organizations worldwide176 G Kozanidis
Testing the heuristic procedure on the above instance, We obtain the
solutions shown in Tables 7 and 8. The solution of Table 7 was obtained
‘when the quadratic problem was applied onee to the entire wing in each
period of the planning horizon (i.c., 6 times in total). ‘The solution of
‘Table 8 was obtained when the quadratic problem was applied once to
each squadron in each period of the planning horizon (ic., 18 times in
total).
Comparing the solutions of Tables 7 and 8, we observe that they are
very similar with each other and only exhibit minor differences. ‘The
objective function values (21, 22, 25, 24) that they provide are (21, 7,
3192.75, 948,72) and (21, 7, 3195.68, 948.75), respectively: thus, the
solution of Table 7 is dominated by the solution of Table 8. Of course,
the solution of Table 8 is dominated, too, by each of the first two
solutions of Table 4, Nevertheless, it is quite satisfactory, since it
provides the ideal value for eriterion z;, and values which are smaller
than the ideal by approximately 8.7 %, 5.7 % and 3.1 % for eriteria 2,2.
and 2, respectively. Compared to the second nondominated solution of
Table 6, which is the one with the most similarities, the heuristic solution
provides the same value for criterion z, and values which are smaller by
approximately 4.5%, 4.2 % and 3.1 % for criteria 2, 2 and 2,
respectively. What is even more important is that each of the two
heuristic solutions was obtained in less than a second computational time
‘on the same machine, as opposed to the approximately 690 seconds that
‘were required for the solution of this problem with AMPL/CPLEX.
Table 7: Fleet availability of the solution obtained using the AFH
(application of the quadratic problem to the wing)
poe oa
Tet 3A Maltiobjective Model 7
Table 8: Fleet availability of the solution obtained using the AFH
(application of the quadratic problem to each squadron)
om anras 99S sis 18672 04a
S368 5200548926 SSROL TH SERTOS—sIOS
As depicted in Tables 7 and 8, the fleet availability of the two
heuristic solutions additionally exhibits low variability, since it does not
‘fluctuate significantly from period to period. In general, the AFH seems
to produce solutions with satisfactory fleet availability. It is likely that
the quality of these solutions can be further improved by considering
different rules for the various decisions involved. It should be noted
though, that the evidence provided in this work is not sufficient to judge
this quality at the moment, This is an open area for future research with
lots of potential
5.2 Horizon Splitting Heuristic (HSH)
‘The second heuristic that we propose for the solution of large
instances is less sophisticated but more straightforward. It makes use of
the simple idea of splitting the initial planning horizon into several
consecutive ones, and solving an FMP subproblem for each of them. The
ending state of the system for each smaller horizon becomes the
beginning state for the next one, and so on. The smaller horizons do not
necessarily need to have equal lengths.
The quality of the solution obtained this way is expected to be
inferior to the one obtained when the problem is solved up front for the
entire planning horizon, On the other hand, the total computational time178, G Kozanidis
needed in order to reach a solution is expected to deerease, especially as
the length of the smaller horizons decreases. This is mainly because the
computational time needed for the solution of the problem is expected to
gTow faster as the size of the problem increases, as some initial
experiments suggest.
For the problem with weight combination w; = w3 = ws = wi = 0.25,
\we applied this simple technique by splitting the original six-period FMP
problem into two three-period subproblems. The solution obtained after
the model was applied twice (once to each subproblem) is displayed in
Table 9. As already mentioned, the time needed for the solution of the
six-period problem was approximately 690 seconds, while the total time
needed for the solution of the two three-period ones was approximately 5
seconds,
‘Table 9: Fleet availability of the solution obtained using the HSH
7 z ¥ at
Read Ws ae aE
fiat : "
It becomes clear from the results shown in Table 9 that the HSH
exhibits a myopic behavior. The residual flight time availability of the
first three periods of the planning horizon is significantly higher than that
of the next three ones. This is a consequence of the fact that the heuristic
focuses on optimizing fleet availability for the subhorizon first, but when
the model is applied to the next subhorizon, the system’s residual flight
time availability drops drastically. In general, the number of periods thatA Maltiobjective Model 179
Will be considered in each smaller horizon seems to have a strong effect
‘on the quality of the solution that will be obtained.
The objective function values that the solution of Table 9 provides
are (23, 7, 2603, 714.25). Thus, the HSH provides the ideal values for
criteria’ z) and z:, and values which are smaller than the ideal by
approximately 23.2 % and 27 % for criteria z: and zy, respectively. OF
course, this solution is dominated by the third solution of Table 4.
Compared to that solution, the heuristic solution provides the same
values for criteria 2, and 2, and values which are smaller by
approximately 10.3 % and 16.6 % for criteria 2: and z,, respectively.
Compared to the two solutions provided by the AFH, the solution
provided by the HSH gives the same value for 22, a value which is larger
by approximately 9.5 % for criterion 2; (which is ideal, too), and values
which are smaller by approximately 18.5 % and 24.7 % for eriteria z5 and
2, Tespectively. Of course, the evidence provided in this work is again
not sufficient to conclude which of the two heuristics produces solutions
of higher quality. Further experimentation is required in order to answer
this, In summary, the HSH provides an alternative way of obtaining an
acceptable solution when the size of the problem instance prohibits its
solution using a commercial optimization software package. Therefore, it
can also be considered altematively for the solution of large-scale FMP
instances,
‘An interesting observation that arises fiom the myopie behavior of
the HSH is that since FMP is an on-going problem that is repeatedly
solved in successive horizons, the length of the horizons for which the
‘wing command issues the flight requirements has a strong impact on the
long term availability of the unit. As the number of periods over which
the command issues the flight requirements increases, the fleet
availability of the obtained solutions is expected to increase, too. given,
of course, that these requirements are kept the same.
Another interesting observation stemming from the fact that the
problem must be solved repeatedly in successive horizons is that such a
myopic behavior will always be present, independently of what the exact
length of each horizon is. Therefore, it seems also very important to try
to find the maximum fleet availability that can be Kept constant over
time. This can be achieved by also minimizing the variability of the
obtained fleet availability between different time periods, besides
maximizing its value. Of course, this has the prerequisite that the flight
requirements issued by the wing command will not fluctuate
significantly trom horizon to horizon180 G Kozanidis
6. Conclusions and Future Research
In this work, we introduced a multiobjective mixed integer model for
‘maximizing a combat unit's fleet availability subject to flight and
‘maintenance requirements. Maximizing a unit’s availability increases its
readiness to respond fo external threats. The model was developed for
use by the HAP and incorporates the various aspects of the problem
under consideration. Certain model parameters seem to have a strong
influence on its computational requirements. Therefore, the decision of
how exactly the model should be utilized in order to solve a specific
FMP instance must be made carefully.
In order to be able to solve large instances of the FMP problem, we
also proposed two heuristic approaches, we illustrated their application
on a real-life instance, and we presented a discussion that provides
insight into their behavior. The first heuristic tries to ensure a smooth.
utilization of the maintenance station and to prevent bottlenecks, by
establishing a smooth sequence of aircraft grounded for service, The
solutions produced by this heuristic seem to be satisfactory and their fleet
availability seems to exhibit low variability, which is also very important
for the FMP problem
‘The second heuristic that was developed for the FMP problem
exhibits a myopic behavior due to the fact that, by construction, it
focuses on consecutive subhorizons and not on the entire planning
horizon of a given problem. Nevertheless, the solutions that this heuristic
produces for practical applications seem to be acceptable, too. What is
even more important is that the computational requirements of both
heuristics seem to be considerably smaller than those of the FMP model.
A conclusion that should be pointed out based on the findings of this
work is that knowing in advance the flight requirements of the wing
command for a greater number of periods can increase the fleet
availability of the unit. Another interesting conclusion is that, besides
‘maximizing the fleet availability level, it seems also very important to try
to minimize its variability, in order to ensure that this level remains
relatively constant over time,
In summary, the FMP model and the two heuristics provide a set of
tools that can be used collectively in order to address this important
problem. The final decision on how to utilize these tools in order to get
the most effective solution relies upon the user and depends on many
parameters, such as the desired compromise between computational time
and solution qualityA Maltiobjective Model Ist
The proposed model can be easily extended to include additional
aspects of the problem that may arise in different situations, such as the
rotation of the aircraft that are used to satisfy’ the flight requirements, the
incorporation of special maintenance checks and parts removal, the
observance of the daily schedule of the pilots, etc, A stochastic model
that incorporates the uncertainty that some of the problem’s parameters
might exhibit can also be developed and studied, based on the results of
this work. We believe that future research should also be directed
towards the development and extensive testing of efficient heuristics
stich as the ones proposed in Section 5, in order to make the solution of
large instances more tractable.
In practice, a significant number of military applications are
addressed with simple decision rules which are not based on analytical
methodologies. We believe that future research should also be directed
towards developing analytical tools for such problems, due to the
importance of the related issues and the serious impact that the involved
decisions can have,
Acknowledgements
George Kozanidis is a recipient of a research grant from the State
Scholarships Foundation of Greece, which partially supported this work,
This support is greatly appreciated
References
Arguello, M.F., Bard, LF. and Yu, G., “Models and Methods for
Managing Airline Irregular Operations Aircraft Routing”,
Operations Research in the Airline Industry, Ed: Yu, G., pp: 1-45,
Kluwer Academic Publishers, Boston, MA, 1997,
Barnhart, C., Belobaba, P, and Odoni, A.R., “Applications of Operations
Research in the Air Transport Industry”, Transportation Science, 37-
4, pp: 368-391, 2008.
Ehrgott, M., Multicriteria Optimization, Springer, Berlin, 2005.
Fourer, R., Gay, D.M. and Kernighan, B.W., AMPL: A Modeling
Language for Mathematical Programming, Duxbury Press, 2002,
Gopalan, R. and Talluri, K-T., “Mathematical Models in Airline
Schedule Planning: A Survey”, Annals of Operations Research, 76,182 G Kozanidis
pp: 155-185, 1998.
Jardine, A.K.S. and Hassounah, ML, “An Optimal Vehicle-Fleet
Inspection Schedule”, The Journal of the Operational Research
Society, 41-9, pp: 791-799, 1990.
Kurokawa, T. and Takeshita K., “Air Transportation Planning using
Neural Networks as an Example of the Transportation Squadron in
the Japan Air Self-Defense Force”, Systems and Computers in Japan.
35-12, pp: 1223-1232, 2004,
Pardalos, P.M. and Kovoor, N., “An Algorithm for a Singly Constrained
Class of Quadratic Programs Subject to Upper and Lower
Constraints”, Mathematical Programming, 46, pp: 321-328, 1990,
Pippin, B.W., “Allocating Flight Hours to Army Helicopters”, MSe
‘Thesis, Naval Postgraduate School, Monterey, CA, USA, 1998,
Radosavijevic, Z. and Babic, O., “Assigning Fighter Plane Formations to
Enemy Aircraft using Fuzzy Logie”, Transportation Planning and
Technology, 23-4, pp: 353-368, 2000.
Seaslik, A., “Planning German Army Helicopter Maintenance and
Mission Assignment”, MSc Thesis, Naval Postgraduate School,
Monterey, CA, USA, 1994.
Steuer, R-E., Multiple Criteria Optimization: Theory, Computation and
Application, Wiley, New York, 1986.
US. DoA, United States Department of the Army, Field Manual No. 3-
04.500: Army Aviation Maintenance (Appendix D: Maintenance
Management Tools). Washington,
00/index. html
‘Yeung, T.G., Cassady C.R., and Pohl, E.A., “Mission Assignment and
Maintenance Sched for Multi-State Systems”, Military
Operations Research, 12-1, pp: 19-34, 2007.