An Optimal Fleet Assignment and Flight Scheduling

Problem for an Airline Company

Yurii A. Mezentsev1 and Igor V. Estraykh2

1
Novosibirsk State Technical University, Novosibirsk, Russia
mesyan@ yandex.ru
2
Novosibirsk State Technical University, Novosibirsk, Russia
ive7@yandex.ru

Abstract. An original problem statement and solution algorithms are presented

for an applied problem in the scheduling theory. The idea of the optimal fleet
assignment and Flight Scheduling problem considered in this paper is to find a
scheduling control method that minimizes the losses of the airline company
from aircraft schedule disruptions. The problem is NP-complete and cannot be
solved accurately for any real-life number of dimensions. An efficient paramet-
ric algorithm is proposed for finding an approximate solution of the problem.
The proposed algorithm is an extension of the schedule optimization algorithm
for a system of unrelated parallel machines with job release dates, which is
based on the makespan criterion (Cmax). A substantial example is presented of
applying the algorithm, as well as statistics of testing it on the data of a generat-
ing problem by the Cmax criterion.

Keywords: Optimal scheduling · Airline fleet assignment · Makespan criterion

·Efficient parametric algorithm

Introduction

One of the areas where optimization methods are traditionally applied in practice is
the planning of airline operations. In this case, planning involves several stages, the
most important being aircraft scheduling, fleet assignment, routing, and crew plan-
ning. A detailed review on this topic was published, e.g., by Grönkvist (2005) [1]; the
formal problem statements and approaches to solving the basic problems were dis-
cussed by Sherali, Bish, and Zhu (2006) [2] and in the numerous publications follow-
ing individual lines of applied research from among those listed above [3–22].
For instance, the problems and algorithms of airline fleet assignment modeling
(FAM) are examined in [3–7]. The main focus in [8–16] is on the aircraft routing
problem (ARP). Simultaneous solving of both problems (FAM and ARP) is consid-
ered, e.g., in [17–19]. Finally, the studies closest to the subject of the present paper [1,
5, 20–22] consider the optimal scheduling problem, including route changes and fleet
assignment.

Copyright © by the paper’s authors. Copying permitted for private and academic purposes.
In: S. Belim et al. (eds.): OPTA-SCL 2018, Omsk, Russia, published at http://ceur-ws.org
Optimal Fleet Assignment and Flight Scheduling Problem 277

Obviously, the above listed problems are closely related. All their relevant formal
representations belong to the class of intractable problems of mixed programming.
The approaches that are used to find their approximate solutions build on classical
schemes such as the Lagrangian relaxation methods, column generation, and Benders
decomposition and apply the well-known computational tools of combinatorial opti-
mization, methods of cuts, and programming in constraints [1, 2].
The core of this work is an original problem statement in the form of an optimal
scheduling problem for a system of unrelated parallel machines (aircraft) with job
release dates (flight delays), which is adapted to the airline flight scheduling problem
proposed by the authors in [23, 24], together with a special efficient parametric algo-
rithm for its approximate solution [24].

1 Conceptual and Formal Statement for the Optimal Fleet

Assignment and Flight Scheduling Problem

The input data are airline flight schedules, standard flight times for of all types of
aircraft, and standard times for ground handling and flight preparations for all types of
aircraft.
The real-time information is flight delays at any given time at all airports.
Then, conceptually, the scheduling problem consists in finding, for the flights in
the planning period, such an airline fleet assignment that will minimize the maximum
total deviation from the initial schedule for the entire fleet while satisfying all the
constraints of the initial schedule in terms of the passenger flow, number of flights,
and flight standards.
We use the following notation:
l is the airport number, l  L ;
i is the flight number, i  I l ,  I l  I , I l  I l '  , l , l ' L ;
lL
s is the type of aircraft, s  S ;
j is the tail number, j  J s ,  Js  J , J s  J s '  , s, s' S ;
sS

 i0 is the actual delay of flight i at the time of scheduling,  i0  0 , i  I l , l  L ,

0   i0 . Hereinafter,  denotes a vector, matrix, or tensor corresponding to the
context of dimension;
ti0 is the scheduled departure time of flight i , i  I l , l  L , T 0  ti0 ;
ti0   i0 is the possible actual departure time of flight i at the initial time of sched-
uling;
ti , j is the time of ground handling, preparation, and air travel of flight i of aircraft
j , T  ti , j , i  I l , l  L , j  J s , s  S .
We need to find xi , j under the constraints:
278 Yu. A. Mezentsev, I. V. Estraykh

1, if aircraft j is assigned to flight i,

xi , j   i  I l , l  L , j  J s , s  S (1)
0 otherwise,

 xi, j  1 , i  I l l  L , (2)
jJ

(constraint (2) means that only one aircraft is assigned to flight i);

bj   xi, j  b j , l  L , j  J s s  S , (3)
iI l

(constraint (3) means that aircraft with tail number j can be assigned to no less than
b j and no more than b j flights);
 i, j is a possible delay in the departure of aircraft j on flight i , i  I l , l  L ,
j  J s , s  S (  i, j can be negative, which is taken into account in constraints (5)
and (6))

 i , j  ti0   i0   ( k , j  tk , j ) xk , j , i  I k , j  J s , s  S , (4)
kI k

(constraint (4) means that the delay of aircraft j at the current step (on flight i ) is a
recursive function of the delays accumulated in the previous flights of this aircraft);

 i, j   i, j  yi, j  0 , i  I l , l  L , j  J s , s  S ; (5)

yi , j  0 , i  I l , l  L , j  J s , s  S . (6)

Constraints (5) and (6) neutralize negative delays through the compensating varia-

bles yi , j  0 ; then,  i , j  0 is the dependent variable, having the meaning of adjusted
delay between the arrival of the aircraft j and its flight i , taking into account the
required service time on the ground, and

 i, j xi, j   ti, j xi, j   , l  L , j  J s , s  S , (7)
iI l iI l

  min (8)

Relations (7) and (8) represent the minimax makespan criterion. The use of this cri-
terion helps achieve a uniform distribution of load on the fleet by minimizing the
maximum total downtime for any aircraft from the whole set of aircraft of the airline.
Another variant of constraint (7) is

 i, j xi, j   (7’)
iI l
Optimal Fleet Assignment and Flight Scheduling Problem 279

Instead of, or together with (7), one can apply an additive criterion of minimization
of the total delays

 i, j xi, j  min (9)
jJ iI

Relations (4), which mediate constraints (5) and (7), contain recursions because

any subsequent (in time) values of  i, j and  i, j depend on the previous ones.
Calculating the delays  i, j in all the i previous steps is associated with considera-
ble difficulties because, first, due to the multiplicity of the variants of their formation
with the subsequent choice of the best, and, second, the expansion of recursions and
reduction of the statement (1) - (8) to the one-stage mixed programming problem
leads to an increase in the number of Boolean variables and constraints in the problem
by a factor of I 2 , where I  sup I [23]. The structural complexity of (1)–(8) is
thereby reduced to the computational complexity of the resulting problem statement,
which remains intractable given that the initial dimension increases by a large factor.
For more detail on the expansion of recursions for a statement identical to (1)–(8)
with the formation of a one-step problem (which we call, for brevity, a direct reduc-
tion) and the subsequent formation of a simplified (relaxed) problem with two criteria
(called a bicriteria relaxation), which allows one to find close-to-optimal schedules in
terms of makespan, see [23, 24]. These works also provide experimental proof of
inefficiency of using the direct reduction. Thus, e.g., finding even an approximate
solution with no more than a six-percent deviation from the optimum for a problem
instance with 20 flights and 5 aircraft took more than 16 hours of computing time
using a 6-core processor and the latest version of the IBM ILOG CPLEX optimization
studio. In [24], one can also find the results of applying the bicriteria relaxation using
CPLEX. Below we compare the accuracy and computing time in the approach devel-
oped in our publication with the results achieved through the application of the
bicriteria relaxation (Table 9).

2 Parametric Algorithm for Finding Suboptimal Solutions

Since the problem contains recursions, DP is, most likely, the only computational
method directly applicable to solving problem (1)–(8). However, the direct applica-
tion of DP is inefficient, partly because the problem in question is NP-complete. In
attempts at finding an accurate solution of (1)–(8), DP leads to an exhaustive search
through all possible options. It is easy to calculate the number N of these options.
For example, if k is the step number and we assume in (2) that b j  0 and
b j  sup I , I  sup I , J  sup J , then, as shown below, considering that the number
of options grows in a geometric progression with DP steps, we have
 I 1 
N   J  J  2 .. Therefore, the DP method in (1)–(8) has a complexity greater
 
280 Yu. A. Mezentsev, I. V. Estraykh

than the exponential one and is not applicable in its pure form to problems with an
actual number of dimensions.
To construct an efficient approximate algorithm, we use a general DP scheme with
the sifting of locally worst options at certain DP steps. We tested this approach previ-
ously in solving optimal scheduling problems for unrelated parallel machines with job
start delays [24]; the tests showed good results in terms of accuracy and speed.
We assume that all flights i  I l , l  L , are arranged in the order of the initial
delays (the initial schedule  i ), considering the aircraft locations at the time of
0

scheduling. Then, based on the DP procedure, we determine the step number   1, I .

We denote the time when aircraft j completes flight  at step  as

f , j ( , j , t , j , x , j ) , j  J s , s  S , and the conditional minimum time of completion

of all flights at steps from 1 to  as  ( , j , ti , j , xi , j ) i  1, I , j  1,


 

f , j ( , j , t , j , x , j )  max 0, [ , j x , j   1, j ( 1, j , ti, j , xi , j )]  t , j x , j , j  1, J ,
i  1,  1 . (10)

The recurrent Bellman relation for this problem is

 , j ( 1, j , ti, j , xi, j )  f , j ( , j , t , j , x , j )   1, j ( 1, j , ti, j , xi, j ), i  1,  1 , (11)
  

 ( , j , ti , j , xi , j )  max  , j ( 1, j , ti , j , xi , j ) , j  J s , s  S , i  1, .

 
(12)
j

 
To achieve the minimum makespan in (7)–(8), we should select in the last step the
 
minimum value of  I ( I , j , ti , j , xi, j ) , i.e., find.   min  I ( I , j , ti, j , xi , j ) j  J s , s  S ,
i  I l , l  L . The total number of scheduling options that we need to find to ensure
the best schedule is

2 k I  I 1 
N  J  J  ...  J  ...  J   J  J  2 (13)
 
We can sift out intermediate schedules in DP in different ways. If we discard all in-
termediate schedules at step k except the locally best one, we have a greedy algo-
rithm. If we keep all the intermediate schedules, we have an exhaustive search
through all the options. In the latter case, we have J intermediate schedule options at
2 k
step 1, J options at step 2, and J options at step k . If we look for a compromise
between accuracy and speed, then, considering that we seek to construct an efficient
algorithm, the number of intermediate schedules should be polynomially dependent
on the number of Boolean variables in (1)–(8).
Optimal Fleet Assignment and Flight Scheduling Problem 281

Let us now consider one such compromise. First, we determine the maximum
number K of the options retained at stage k for further analysis. For convenience of
description, we assume that K is a constant. For example, we assume that K  1024
k
and determine the maximum number K '  J  K . Since the number of possible
intermediate schedules increases by a factor of J at each step, we suggest sifting out
1 1 / J of the locally worst options at each step starting from k  1 .
 ln K  
 , where  is the integer part of the number.
Obviously, k : k  
 ln J 
If we calculate the total number of schedule options generated in the algorithm, we
k k
obtain J options at step 1, J options at step k , and also J options at steps from
k 1
k  1 to I . This scheme is implemented by sifting out J intermediate schedule
options at all the steps from k  1 to I . Then, the number of options that remains for
further consideration at each step beginning from k  1 is exactly J , and the total
number of intermediate schedules N ' is
2
N '  J  J  ...  J
k 1 k k


k 1



 J  ...  J   J  J  2  I  k  1  J 
k
(14)

Since k is a constant, relation (14) represents a polynomial dependence of the

complexity of the parametric DP algorithm with option-sifting on the dimension of
problem (1)–(8). In this case, k is the degree of this polynomial. For clarity, we com-
pare N with N ' , assuming k  3 , I  1000 , and J  100 .
Then, N  100  1001
  
 100 2 , N '  100 2  100 2  100  2 100 3  98009900 .
These circumstances underlie the ordinary complexity of the parametric algorithm (its
complexity is defined by the parameter k ) and the virtually infinite complexity of the
DP method.
Based on (14), we can estimate the total complexity of the parametric algorithm.
To this end, it is sufficient to determine the complexity of the step beginning from k ,
which directly depends on the number of combinations of the variables xi , j at step k .
k
If we denote this value as Pk , then, obviously, Pk  J . In fact, this means that at
each step beginning from k , the algorithm requires calculating Pk variants of con-
strains (7) for all the possible values of xk , j . In total, we have at all steps: P1  J ,
2 k
P2  J , Pl  J , l  k , I . In the above example, Pk  100 , and, considering (14),
3

we obtain a high complexity for a problem with an actual number of dimensions. We

can overcome this difficulty either by reducing k or by decomposition.
Below we describe the parametric DP algorithm with the sifting of the locally
worst intermediate options.
282 Yu. A. Mezentsev, I. V. Estraykh

Algorithm AP

1. Enter the input data ( i0 , ti , j ) , j  J s , s  S , i  I l , l  L , and the parameters k

and N ' . Assuming that 0, j ( 00 , ti , j , xi, j )  0 , determine the initial step number
 : 0 .
2.  :   1 .
3. Check the step number. If   I , proceed to point 7; otherwise, proceed to the
next point.
4. At step  , determine the sequence of the subsequent steps (rearrange the flight

list), calculate the delays  , j , generate from (10)–(12) all the feasible fleet assign-

ment options, and calculate f , j ( , j , t , j , x , j ) and the schedule lengths

 , j ( , j , ti , j , xi , j ) .
 
5. Check N , i.e., the number of options of  , j ( , j , ti , j , xi , j ) at step  . If   k

, i.e., N  N ' , then proceed to point 2; otherwise, proceed to the next point.
k 1
6. Sift out J of all the options generated at point 4 with the largest schedule

lengths  , j ( , j , ti , j , xi , j ) . Return to point 2.
7. Choose schedule options with the minimum length. Construct the final schedules
using the inverse DP procedure.

A note on the AP algorithm regarding estimates for the delays  i, j :

At each step of the algorithm, one needs to estimate the delays  i, j for aircraft that

have not yet arrived at the airport of departure. The estimates for  i, j are found by
solving the subproblems of finding all the shortest paths in a graph composed of the
possible connections between the airports. In general, one should find these estimates
at each step because the delays can vary from step to step, depending on the previous
local fleet assignments.

3 Illustrative Example

Below is an illustrative fragment demonstrating the application of the proposed al-

gorithm for constructing a close-to-initial flight schedule and fleet assignment for
three aircraft (designated by their tail numbers 1, 2, and 3) of two different types.
The input data on the flights and time costs are given in Table 1.
In Table 1, the Time column contains ti , j ; the Location column shows the pres-
ence or absence of aircraft at the airport of departure at each step of scheduling; and
the initial schedule corresponds to  i0 . The Tail Number column shows the numbers
Optimal Fleet Assignment and Flight Scheduling Problem 283

j . The data in Table 1 are sorted by  i0 , taking into account the aircraft location. The
flights are numbered in the same order.
The algorithm parameters are k  1 , ( N ' 3 ), J  3 , and I  11 .
Since k  1 , we can reduce the number of dimensions (the eliminated options are
highlighted by filling).

After the third step, both the aircraft locations and the current delays  i, j change,
necessitating a new tail assignment sequence, i.e., a change in the sequence of the
algorithm steps (see point 4 in AP ).

Table 1. Flight Data

Departure Destination Aircraft data Type and initial

Flight airport airport Tail
Type Time Number location schedule
1 1 2 1т 5 1 1т 1
2 2 3 1т 3 2 1т 2
3 2 1 2т 4 3 2т 1
4 3 1 2т 2 1
5 1 3 1т 2 2
6 3 2 1т 4 2
7 1 2 2т 4 4
8 1 3 2т 2 4
9 3 1 1т 3 4
10 2 1 1т 5 4
11 2 3 1т 3 5

Table 2. Algorithm: Step One

 1 f1, j 1, j  f1, j  1  max 1, j 
j
,
Flight 1 1 0 0 (1+5,0,0) 1  max{6;0;0}=6
0 1 0 (0,~,0)  
1  max 0, 1,2 ,0   1,2 ,
0 0 1 (0,0 ,~) 1  max 0,0,   

Table 3. Algorithm: Step Two

 2 x2,1 x2, 2 x2,3 f 2, j 2, j  f 2, j  1, j ,  2  max  2, j 
j
Flight 1 + 1 0 1 0 0 (6+3,0,0)  2  max{6+3;0+0;0+0}=9
Flight 2 1 0 0 1 0 (0, 2+3, 0)  2  max{6+0;0+5, 0+0}=6
1 0 0 0 1 (0, 0 , ~)  2  max 6  0,0,   
0 1 1 0 0 (13+3,0,0)  2  max{13+3;0+0;0+0}=16
0 1 0 1 0 (0, 2+3, 0)  2  max{13+0;0+5, 0+0}=13
0 1 0 0 1 (0, 0 , ~)  2  max 13  0,0,   
284 Yu. A. Mezentsev, I. V. Estraykh

The change in the tail assignment is given in Table 4, whose rows are arranged in

the order of increasing  i, j , considering the current aircraft locations.

Table 4. Ordered Data

Departure Destination Aircraft data Type, Initial

Flight airport airport Type Time Number location schedule
1 1 2 1т 5 1 1т 1
2 2 3 1т 3 2 1т 2
3 2 1 2т 4 3 2т 1
4 3 1 2т 2 1
5 1 3 1т 2 2
6 3 2 1т 4 2
7 1 2 2т 4 4
8 1 3 2т 2 4
9 3 1 1т 3 4
10 2 1 1т 5 4
11 2 3 1т 3 5

Table 5. Step Four

 4 x2,1 x2, 2 x4,1 x4, 2 f 4, j  4, j  f 4, j  3, j ,  4  max  4, j 
j
Flights 1 0 1 0 (4,0,0) 4  max{6+3+0+4; 0+0+0+0; 0+0+5+0} =13
1,2,3,6 1 0 0 1 (0,4,0) 4  max{6+3+0+0; 0+0+0+40; 0+0+5+0} =9
0 1 1 0 (4,0,0) 4  max{6+0+0+4; 0+5+0+0; 0+0+5+0} =10
0 1 0 1 (0,4,0) 4  max{6+0+0+0; 0+5+0+4; 0+0+5+0}=9
0 1 1 0 (4,0,0) 4  max{13+0+0+4; 0+5+0+0;0+0+5+0} =17
0 1 0 1 (0,4,0) 4  max{13+0+0+0;0+5+0+4; 0+0+5+0} =13

The calculations at the fourth step are given in Table 5.

Consistent with the parameter k  1 , we cross out the locally worst current sched-
ule options (for  4  13 ).
The results of the calculations at the final eleventh step are summarized in Table 6.
The tail assignment options excluded from consideration are highlighted by filling.
The square brackets in the sums of the aircraft handling times highlight the options
when the aircraft with the corresponding tail numbers are absent at the airport of de-
parture. In this case, the aircraft needs to be delivered from another nearby airport.
The delivery time is indicated inside the square brackets.
Thus, we have obtained four tail assignment options and flight schedules, equiva-
lent in terms of criterion (7)–(8) (   17 for all the options).
Tables 7 and 8 show all the best aircraft assignment options and the corresponding
flight schedules. The Start columns show the departure times, and the End columns
show a total of the arrival time plus the time on ground handling and preparations for
the next flight.
Optimal Fleet Assignment and Flight Scheduling Problem 285

Table 6. Final Step

  11 x8,1 x8, 2 x11,1 x11, 2 f11, j 11, j  f11, j  10 , j  ,   11  max 11, j 
Flights: 11  max{6+0+0+0+0+0+3+[2+2]+0+0+3; j

1 0 1 0 (3,0,0) 0+5+0+4+0+5+0+0+0+0+0;
1,2,3,6, 0+0+5+0+4+0+0+0+[4+2]+2+0} =17
7,10,11, 11  max{6+0+0+0+0+5+0+2+0+0+3;
1 0 0 1 (3,0,0) 0+5+0+4+0+0+3+0+0+0+0;
5,8,4,9 0+0+5+0+4+0+0+0+[4+2]+2+0} =17
11  max{6+0+0+0+0+0+3+0+0+0+3;
1 0 0 0 (3,0,0) 0+5+0+4+0+5+0+2+0+0+0;
0+0+5+0+4+0+0+0+[4+2]+2+0} =17
11  max{6+0+0+0+0+0+3+[2+2]+0+0+0;
0 1 1 0 (0, 3, 0) 0+5+0+4+0+5+0+0+0+0+[3+2];
0+0+5+0+4+0+0+0+[4+2]+2+0}=19
11  max{6+0+0+0+0+5+0+2+0+0+0;
0 1 0 1 (0, 3, 0) 0+5+0+4+0+0+3+0+0+0+3;
0+0+5+0+4+0+0+0+[4+2]+2+0} =17
11  max{6+0+0+0+0+0+3+0+0+0+0;
0 1 0 0 (0, 3, 0) 0+5+0+4+0+5+0+2+0+0+3,
0+0+5+0+4+0+0+0+[4+2]+2+0} =19

Table 7. Shortest Schedule: Options 1 and 2

i= j=1 j=2 j=3 Start End j=1 j=2 j=3 Start End
1 1 0 0 1 6 1 0 0 1 6
2 0 1 0 2 5 0 1 0 2 5
3 0 0 1 1 5 0 0 1 1 5
4 0 1 0 5 9 0 1 0 5 9
5 0 0 1 5 9 0 0 1 5 9
6 0 1 0 9 14 1 0 0 6 11
7 1 0 0 6 9 0 1 0 9 12
8 1 0 0 11 13 1 0 0 11 13
9 0 0 1 13 15 0 0 1 13 15
10 0 0 1 15 17 0 0 1 15 17
11 1 0 0 13 16 1 0 0 13 16

This example clearly demonstrates the universal applicability of the AP algorithm.

It is also suitable for solving both the FAM problem and the mixed FAM + ARP
problem. The same feature allows finding for the target optimal scheduling problem a
k -best solution that minimizes the airline company losses from disruptions in the
initial aircraft flight schedules.
286 Yu. A. Mezentsev, I. V. Estraykh

Table 8. Shortest Schedule: Options 3 and 4

i= j=1 j=2 j=3 Start End j=1 j=2 j=3 Start End
1 1 0 0 1 6 1 0 0 1 6
2 0 1 0 2 5 0 1 0 2 5
3 0 0 1 1 5 0 0 1 1 5
4 0 1 0 5 9 0 1 0 5 9
5 0 0 1 5 9 0 0 1 5 9
6 0 1 0 9 14 1 0 0 6 11
7 1 0 0 6 9 0 1 0 9 12
8 0 1 0 14 16 1 0 0 11 13
9 0 0 1 13 15 0 0 1 13 15
10 0 0 1 15 17 0 0 1 15 17
11 1 0 0 9 12 0 1 0 12 15

4 Exchange-Based Improving Algorithm

The solution obtained by the parametric AP algorithm can be improved using a

procedure based on an exchange of flights between aircraft. Below we describe this
algorithm.
At the first step, we select an aircraft with the maximum total flight time and try
reassigning one of its flights to another aircraft. If the exchange reduces the value of
the objective function, we repeat the process; otherwise, we move to the next flight.
At the second step, we select an aircraft with the maximum total flight time and
consistently review the flights assigned to this aircraft. In each case, we search for
flights that this aircraft makes in less time yet assigned to another aircraft. Then, the
flights are exchanged between the aircraft. If the exchange reduces the value of the
objective function, we repeat the second step; otherwise, we cancel the exchange and
search for another flight suitable for reassignment. We denote the general exchange-
based algorithm, as well as the one implemented at step two, as AС . The application
of the AС algorithm at step two can be detailed as follows:

Algorithm AС
1. Select aircraft m with the maximum total flight time m   .
2. Assume i : 1 .
3. If i : m , then go to point 14.
4. Assume l : 1 .
5. If xm,l  0 , then go to point 12.
6. Assume j : 1 .
7. If j  l or xi , j  0 or tm, j  tm,l or ti ,l  ti, j  m  i , go to point 10.
8. Assume xi , j  0 , xm, j  1 , xi ,l  1 , xm,l  0 .
Optimal Fleet Assignment and Flight Scheduling Problem 287

9. Calculate the value of the objective function. If it has decreased, go to point 1;

otherwise, assume xi , j  1 , xm, j  0 , xi ,l  0 , xm,l  1 .
10. Assume j : j  1 .
11. If j  J , go to point 7.
12. Assume l : l  1 .
13. If l  J , go to point 5.
14. Assume i : i  1 .
15. If i  I , go to point 3; otherwise, stop the algorithm.
Let us estimate the complexity of the AС algorithm, which searches for flight ex-
change options between aircraft for any finite schedule obtained by the AP algo-
rithm. Since the former conducts an exhaustive search among all aircraft and all
flights and calculates the value of the objective function for every exchange option,
the complexity of the exchange-based algorithm for one schedule is O I  J  . But
2

 
since the AP algorithm generates J schedules at the final step, the total upper-bound

estimate for the complexity of AС is O I  J  .

3

 

5 Results of Testing the Algorithms

A software implementation of the parametric AP and exchange-based AС algo-

rithms allowed us to investigate their properties for instances with a close-to-actual
number of dimensions. The algorithms were tested on the data of the optimal schedul-
ing problem for a system of unrelated parallel machine with job start delays [23].
Table 9 contains the results of testing the instances of problem (1)–(8) with the use of
the above-mentioned means of solving the bicriteria relaxation of problem (1)–(8)
[23] and the AP and AС algorithms. All the tests have the same number of dimen-
sions. The number of flights is I  100 , and the number of aircraft is J  5 . The
algorithms AP and AС were applied with two values of the parameter: k  4 and
k 5.
In Table 9, tda and da are, respectively, the solution time (hh:mm:ss) and the
value of the efficiency criterion, which were obtained by applying the basic algorithm
based on bicriteria relaxation and IBM ILOG CPLEX [24].The values tdp , dp , pdp ,
and  dp are, respectively, the solution time (in seconds); the value of the criterion;
and the relative and absolute worsening (improvement at a negative value) of the
criterion achieved by the AP algorithm, compared with the basic algorithm. The
corresponding cdp , pdp
c
, and cdp values were obtained by the AС algorithm.
288 Yu. A. Mezentsev, I. V. Estraykh

In general, there is an evident absolute gain in speed due to the efficiency of the
AP algorithm and its combination with AС . Moreover, the solutions obtained show
almost complete superiority over the basic algorithm in terms of closeness to the op-
timal solutions.

Table 9. Comparative Characteristics of the Algorithms

No t da da tdp dp pdp  dp cdp pdp
c
cdp tdp dp pdp  dp cdp pdp
c
cdp
.
hh:mm:ss
I  100 , J  5 , AС J 5,k  5 AС
1 00:07:18 389 17 368 -5.40 -21 364 -6.43 -25 435 363 -6.68 -26 359 -7.71 -30
2 00:15:29 335 18 337 0.60 2 335 0.00 0 436 331 -1.19 -4 329 -1.79 -6
3 00:09:21 401 17 382 -4.74 -19 382 -4.74 -19 438 375 -6.48 -26 375 -6.48 -26
4 00:01:42 356 18 370 3.93 14 366 2.81 10 437 365 2.53 9 361 1.40 5
5 00:12:11 334 17 337 0.90 3 333 -0.30 -1 435 337 0.90 3 337 0.90 3
6 04:43:14 363 17 377 3.86 14 369 1.65 6 434 375 3.31 12 361 -0.55 -2
7 00:11:56 403 18 404 0.25 1 404 0.25 1 434 404 0.25 1 401 -0.50 -2
8 01:34:02 395 17 398 0.76 3 397 0.51 2 433 386 -2.28 -9 386 -2.28 -9
9 00:03:33 364 17 372 2.20 8 372 2.20 8 434 371 1.92 7 371 1.92 7
10 00:15:37 395 17 392 -0.76 -3 366 -7.34 -29 434 389 -1.52 -6 376 -4.81 -19
Average 0.16 0.2 -1.14 -4.7 -0.93 -3.9 -1.99 -7.9

Tables 10 and 11 show the algorithm testing statistics to estimate the solution times
for instances of the optimal airline fleet assignment and flight scheduling problem
with an actual number of dimensions.

Table 10. Tests: I  100 , J  10 , k  3 Table 11. Tests: I  100 , J  30 , k  2

Test tdp dp cdp Test tdp dp cdp

No. No.
1 325 163 154 1 18 106 106
2 322 126 122 2 18 105 105
3 322 145 145 3 18 107 107
4 338 139 136 4 18 106 106
5 329 157 152 5 17 108 108
6 322 134 132 6 18 114 114
7 324 147 147 7 18 105 104
8 322 135 132 8 18 108 106
9 323 153 153 9 18 106 106
10 326 137 137 10 19 105 105

Concerning the accuracy (i.e., closeness of the flight assignments and aircraft
schedules generated by the AP and AС algorithms to the optimal ones) estimates, we
note the following points. There are no a priori accuracy estimates for AP and for the
combination AP + AС , but there are a posteriori ones at small dimensions, which are
as follows [24]: in approximately 82% of cases, an accurate solution was obtained in
the generated tests. In the other cases, the deviation from the optimum was no more
Optimal Fleet Assignment and Flight Scheduling Problem 289

than 6%. This conclusion was derived from a comparison of the AP + AС testing
results with the solutions of the same tests in CPLEX by the expansion of recursions
and the direct reduction of problem (1)–(8) in milp.
The dimensions of the tests (100 flights and 30 aircraft) and the solution time offer
hope that the designed toolkit would be efficient in solving real-life problems of air-
line planning.

Conclusions

The results obtained demonstrate the efficiency of the proposed approaches in

solving real-life problems of air transportation planning, including aircraft fleet as-
signment, routing, and, if necessary, flight scheduling. Thus, our approach would
make a promising a contribution to the planning practice of an airline company of any
size. A posteriori estimates for the accuracy and speed of the algorithms lead us to
conclude that the developed toolkit has evident advantages over its analogs.
The testing confirms experimentally, in terms of computing time, the efficiency of
the AP + AС pair. Noteworthy is the insignificant contribution of the AС algorithm
to the total complexity. The use of this algorithm adds no more than a second to the
total computing time tdp for the tests in Tables 9–11.

References
1. Grönkvist, M.: The Tail Assignment Problem. PhD thesis, Chalmers University of Tech-
nology and Göteborg University, Göteborg, Sweden (2005)
2. Sherali, H.D., Bish, E.K., Zhu, X.: Airline fleet assignment concepts, models, and algo-
rithms. European Journal of Operational Research 172, 1-30 (2006)
3. Hane, C. A., Barnhart, C., Johnson, E. L., Marsten, R. E., Nemhauser, G. L., Sigismon-
di, G.: The fleet assignment problem: solving a large-scale integer program. Mathematical
Programming 70, 211-232 (1995)
4. Gu, Z., Johnson, E. L., Nemhauser, G. L., Wang, Y.: Some Properties of the Fleet As-
signment Problem. Technical report, School of Industrial & Systems Engineering, Georgia
Institute of Technology, Atlanta, GA, USA (1993)
5. Subramanian, R., Scheff, Jr. R. P., Quillinan, J. D., Wiper, D. S., Marsten, R. E. Coldstart:
Fleet Assignment at Delta Air Lines. Interfaces 24(1), 104-120 (1994)
6. Ozdemir, Y., Basligil, H., Nalbant, K.G.: Optimization of fleet assignment: a case study in
Turkey. An International Journal of Optimization and Control: Theories & Applications
2(1), 59-71 (2012)
7. Blegur, F.M.A., Bakhtiar, T., Aman, A.: Scenarios for fleet assignment: a case study at Li-
on Air. IOSR Journal of Mathematics 10(5) Ver. I, 64-68 (2014)
8. Kabbani, N. M., Patty, B. W.: Aircraft routing at American Airlines. In: Proceedings of the
Thirty-Second Annual Symposium of AGIFORS (1992)
9. Clarke, L. W., Johnson, E. L., Nemhauser, G. L., Zhu, Z.: The aircraft rotation problem.
Annals of Operations Research 69, 33-46 (1997)
290 Yu. A. Mezentsev, I. V. Estraykh

10. Cordeau, J. F., Stojkovi´c, G., Soumis, F., Desrosiers, J.: Benders decomposition for sim-
ultaneous aircraft routing and crew scheduling. Transportation Science 35(4), 55-76 (2001)
11. Barnhart, C., Boland, N.L., Clarke, L.W., Johnson, E.L., Nemhauser, G.L., Shenoi, R.G.
Flight string models for aircraft fleeting and routing. Transportation Science 32(3), 208-
220 (1998)
12. Elf, M., J¨unger, M., Kaibel, V.: Rotation planning for the continental service of a Europe-
an Airline. In: W. Jager and H.-J. Krebs, editors, Mathematics – Key Technologies for the
Future. Joint Projects between Universities and Industry. pp. 675-689, Springer (2003)
13. Sarac, A., Batta, R., Rump, C. M.: A Branch-and-Price Approach for Operational Aircraft
Maintenance Routing. Working Paper, Department of Industrial Engineering, University
of Buffalo, Buffalo, NY, USA, ( 2003)
14. Gopalan, R., Talluri, K. T.: The aircraft maintenance routing problem. Operations Re-
search, 46(2), 260-271 (1998)
15. Ahuja, R. K., Goodstein, J., Mukherjee, A., Orlin, J. B., Sharma, D.: A Very Large-Scale
Neighborhood Search Algorithm for the Combined Through and Fleet Assignment Model.
Working Paper 4388-01, MIT Sloan School of Management, Cambridge, USA (2001)
16. Birbil, S.I., Frenk, J.B.G., Gromicho, J.A.S., Zhang, S.: A Network Airline Revenue Man-
agement Framework Based on Decomposition by Origins and Destinations. Submitted
manuscript (2015)
17. Yan, S., Tseng, C.H.: A passenger demand model for airline flight scheduling and fleet
routing. Computers & Operations Research 29, 1559-1581 (2002)
18. Barnhart, C., Boland, N.L., Clarke, L.W., Johnson, E.L., Nemhauser, G.L., Shenoi, R.G.:
Flight string models for aircraft fleeting and routing. Transportation Science 32(3), 208-
220 (1998)
19. Sandhu, R., Klabjan, D.: Integrated Airline Planning. Working paper, submitted for publi-
cation, Department of Mechanical and Industrial Engineering, University of Illinois at Ur-
bana-Campaign, Urbana, IL, USA (2004)
20. Lettovsky, L.: Airline Operations Recovery: An Optimization Approach. PhD thesis,
School of Industrial & Systems Engineering, Georgia Institute of Technology, Atlanta,
GA, USA (1997)
21. Kohl, N., Larsen, A., Larsen, J., Ross, A., Tiourine, S.: Airline Disruption Management -
Perspectives, Experiences and Outlook. Research and Technology Report CRTR-0407,
Carmen Systems AB, Gothenburg, Sweden (2004)
22. Rosenberger, J. M.: Topics in Airline Operations. PhD thesis, School of Industrial & Sys-
tems Engineering, Georgia Institute of Technology, Atlanta, GA, USA (2002)
23. Avdeenko, T. V., Mesentsev, Y. A.: Efficient approaches to scheduling for unrelated paral-
lel machines with release dates. IFAC-Papers Online (IFAC Proceedings Volumes).
49(12), 1743-1748 (2016)
24. Avdeenko, T. V., Mezentsev, Y. A., Estraikh, I.V.: Heuristic approach to unrelated parallel
machines scheduling under availability and resource constraints. IFAC-PapersOnline,
50(1), 13096-13101 (2017)
25. Mezentsev, Y. A.: Binary cut-and-branch method for solving linear programming prob-
lems with boolean variables. In: Proc. DOOR 2016, Vladivostok, Russia, September 19-
23, CEUR-WS 1623. Pp. 72-85 (2016)