Transportation Research Part C 46 (2014) 16–29

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

The time slot allocation problem under uncertain capacity

Luca Corolli a, Guglielmo Lulli b,⇑, Lewis Ntaimo c
a
Dipartimento di Ingegneria e Architettura, Università degli Studi di Trieste, Trieste, Italy
b
Dipartimento di Informatica Sistemistica e Comunicazione, Università di Milano Bicocca, Milan, Italy
c
Department of Industrial and Systems Engineering, Texas A&M University, College Station, TX 77843, USA

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents two stochastic programming models for the allocation of time slots
Received 7 September 2013 over a network of airports. The proposed models address three key issues. First, they pro-
Received in revised form 13 March 2014 vide an optimization tool to allocate time slots, which takes several operational aspects and
Accepted 9 May 2014
airline preferences into account; second, they execute the process on a network of airports;
and third they explicitly include uncertainty. To the best of our knowledge, these are the
first models for time slot allocation to consider both the stochastic nature of capacity
Keywords:
reductions and the problem’s network structure. From a practical viewpoint, the proposed
Time slot allocation
Air traffic
models provide important insights for the allocation of time slots. Specifically, they high-
Stochastic programming light the tradeoff between the schedule/request discrepancies, i.e., the time difference
Scheduling between allocated time slots and airline requests, and operational delays. Increasing
schedule/request discrepancies enables a reduction in operational delays. Moreover, the
models are computationally viable. A set of realistic test instances that consider the
scheduling of four calendar days on different European airport networks has been solved
within reasonable – for the application’s context – computation times. In one of our test
instances, we were able to reduce the sum of schedule/request discrepancies and
operational delays by up to 58%. This work provides slot coordinators with a valuable
decision making tool, and it indicates that the proposed approach is very promising and
may lead to relevant monetary savings for airlines and aircraft operators.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction

Congestion in air traffic causes very high expenses for airlines and aircraft operators every day. Although air traffic
demand is below the pre-economic crisis levels of 2007 and 2008, costs caused by flight inefficiencies in 2011 amounted
to 5.2 B€ in Europe only, as reported by EUROCONTROL’s Performance Review Unit (EUROCONTROL, 2012). Despite being
‘‘fully coordinated’’, the most congested airports in Western and Central Europe suffer persistent congestion. Full coordina-
tion of an airport means, essentially, that the number of flights scheduled at the airport per hour (or other unit of time) is not
allowed to exceed the ‘‘declared capacity’’ of the airport (deNeufville and Odoni, 2003). This picture is even gloomier if we
consider EUROCONTROL’s forecasts on the future evolution of air traffic. In fact, despite the slowdown in the air traffic mar-
ket caused by the global economic crisis, air traffic is expected to double by 2035 in the most likely scenario. This will leave
12% of traffic demand not accommodated and the 20 busiest European airports heavily congested (EUROCONTROL, 2013). In
the long term, planning the increase of current airport capacity is an absolute prerequisite to cope with the projected traffic

⇑ Corresponding author. Tel.: +39 026448 7896.

E-mail addresses: luca.corolli@gmail.com (L. Corolli), lulli@disco.unimib.it (G. Lulli), ntaimo@tamu.edu (L. Ntaimo).

http://dx.doi.org/10.1016/j.trc.2014.05.004
0968-090X/Ó 2014 Elsevier Ltd. All rights reserved.
 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29 17

demand. However, at least in the short-medium term, optimizing the use of existing capacity is necessary in order to alle-
viate congestion.
Several approaches have been proposed to alleviate congestion and resolve demand-capacity imbalances. Medium term
approaches, such as the use of slot auctions or congestion pricing, are mostly administrative or economic in nature, and try to
alleviate congestion by modifying spatial or temporal traffic patterns, see Brueckner (2002), Fan (2003) and Raffarin (2004),
among others. Short term approaches consider the operational adjustment of air traffic flows to match available capacity,
and typically span a planning horizon of less than 24 h. On short term approaches, a wide body of literature is available.
The interested reader may refer to Odoni (1987), Richetta and Odoni (1994), Dell’Olmo and Lulli (2003) and Ball et al.
(2010) to mention a few. In addition to the cited models, that are flight-based, i.e., they model each single flight, aggregate
models that consider flows of flights have been developed with the scope of reducing computational times. Among these
macroscopic models, we cite Bayen et al. (2006), Sun and Bayen (2008), Wan and Roy (2008) and Andreatta et al. (2011).
Herein, we focus on medium term initiatives aiming to allocate time slots among airlines and aircraft operators effi-
ciently. A ‘‘time slot’’, or ‘‘slot’’ for short, is the amount of capacity needed by an aircraft to land or take-off. More specifically,
it is a permission given by a coordinator for a planned operation to use the full range of airport infrastructure necessary to
arrive or depart at a Level 31 airport on a specific date and time (Worldwide Slot Guidelines, 2012).
Slot allocation is de facto handled according to the guidelines of the International Air Transport Association (IATA). Since
1974, IATA has provided the global air transport community with a single set of standards for the management of airport
slots, outlining policies, principles and procedures for slot allocation. These guidelines are the result of consultation between
airlines and airport coordinators, reflecting the proven best practice for the coordination and management of airport slots
(Worldwide Slot Guidelines, 2012). For the allocation of slots at European Union airports, IATA guidelines are enriched with
further conditions aiming to encourage the efficient use of airport capacity through the optimal allocation of slots (DotEcon,
2002, see Czerny et al. (2008) for details).
European Commission regulation 95/1993 guarantees that slot allocation is based on neutral, transparent and nondis-
criminatory rules. In this regulation process, each Member State has the duty of appointing a schedule coordinator at airport
or national level with the ultimate mission to operationalize, coordinate, supervise, and arbitrate the slot allocation process.
The fundamental principle of the slot allocation process is the grandfather right, i.e., the right of an airline to keep a slot of the
preceding equivalent season. This right is granted if and only if such a slot was used at least 80% of the time (use-it-or-lose-it
rule). However, this procedure is far from being efficient. Indeed, as reported by Airports Council International (ACI) Europe,
unsatisfied/unaccommodated demand, overbidding, late return of unwanted slots, flights operated significantly and repeat-
edly off slot time (‘‘off slot’’), and failure to operate allocated slots (the so called ‘‘no shows’’), are all factors pointing or con-
tributing to the inefficient allocation and use of an already insufficient resource (see also Zografos et al. (2012)). Several
amendments to the EC regulation have been adopted or proposed with the scope of ensuring the fullest and most flexible
use of limited capacity at congested airports. For instance, regulation (EC) 793/2004 was adopted to strengthen the coordi-
nator’s role and the monitoring of compliance, i.e., the usage of slots with respect to the allocation, to verify that airlines do
not use a slot in a significantly different way than allocated by the coordinator.
Based on the UK experience, the Commission is also considering to change the current regulation to allow for the intro-
duction of market-based mechanisms across the EU provided that safeguards to ensure transparency and undistorted com-
petition are established, including greater independence for slot coordinators. Several studies have been proposed to analyze
the introduction of market-based mechanisms in the slot allocation process. Rassenti et al. (1982) developed a sealed-bid
combinatorial auction for the allocation of time slots to competing airlines. A similar approach has been proposed by Ball
et al. (2005). Kleit and Kobayashi (1996) and Fukui (2010) examined whether slot markets have resulted in anticompetitive
activities with restricted market entry and service expansion by other carriers, especially rival carriers. Verhoef and pricing
(2010) looked into some alternative economic instruments for managing congestion at airports, notably slot sales and slot
trading. In the last few years, secondary trading has achieved a lot of attention by the research community. Some of these
studies have been commissioned by the EU Commission as a guidance on the possible market and legal impacts of the intro-
duction of such trading. For instance, Matt McDonald (2006) analyzed into details the likely effects of the introduction of
secondary slot trading into Community legislation, while Starkie (1998) examined the arguments for and against a secondary
market in slots, focusing on evidence from US airports. Furthermore, Pellegrini et al. (2012) proposed a market mechanism
for secondary trading based on budget balanced combinatorial exchange.
In this paper, we focus on the judicious assignment of time slots to aircraft operators with the purpose of obtaining an
‘‘effective’’ and ‘‘reliable’’ airline schedule. The schedule is considered ‘‘effective’’ if the assigned slots respect airline prefer-
ences. Currently, slot allocations at different airports are independent, while the necessity of coherently allocating the depar-
ture and arrival slots at origin and destination for each flight is widely recognized. Furthermore, it is necessary to consider
interdependencies among flights operated by the same airline. The approach we propose, based on mathematical program-
ming, allocates slots at all airports simultaneously, considering a true network of airports thus guaranteeing the coherence of
the final result, i.e., the airline schedule. We refer to the difference between airline requests and the schedule, i.e., the

1
Level 3 or fully coordinated: airports where capacity providers have not developed sufficient infrastructure, or where governments have imposed
conditions that make it impossible to meet demand. A coordinator is appointed to allocate slots to airlines and other aircraft operators using or planning to use
the airport as a means of managing available capacity.
18 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29

allocated time slots, as ‘‘schedule/request discrepancies’’. Finally, the schedule is considered ‘‘reliable’’ if it takes the inherent
uncertainty affecting the airports’ capacity into account.
This problem is currently solved by the slot coordinator at each single coordinated airport using a First-In-First-Served
discipline, without considering connections between airports. Issues related to airport network dependencies and slot com-
plementarity are resolved off-line at the IATA Worldwide Scheduling Conference, which is held biannually. Few approaches
have been proposed for this specific problem. We cite the heuristic procedure developed by Kosters (2007), which addresses
the interrelationship between slot demand and the time difference between assigned and requested slots, in order to predict
the latter depending on slot utilization. Castelli et al. (2012) analyzed the effect of the presence of grandfather rights on total
airline costs, showing that costs are higher if they are present. The authors also considered the interdependence of slots at
different airports. The possibility to fairly redistribute time slot related costs among airlines through monetary compensa-
tions is introduced, also allowing for the removal of grandfather rights. This approach is shown not to significantly penalize
airlines that own grandfather rights. Another relevant work was developed by Zografos et al. (2012), where a model for the
single airport slot allocation problem is proposed, implementing existing EU/IATA scheduling rules and coordination proce-
dures. The authors applied their work to real data from 3 different Greek airports, showing that it is possible to reduce the
total amount of schedule/request discrepancies up to over 95%, compared to that assigned in practice without using optimi-
zation models. This is a very important result, as it shows how much it is possible to improve time slot allocation by applying
mathematical methods.
In this work we develop two stochastic programming models for the time slot allocation problem, which extend the
deterministic single airport model presented in Zografos et al. (2012) in the following directions. First, they simultaneously
allocate time slots on multiple airports explicitly considering – in addition to other operational constraints – the coherence
between the departure time slot at the airport of origin and the arrival time slot at the airport of destination of each flight.
Second, our models take into account stochastic capacity availability. This is of particular relevance since we are faced with a
strategic problem in which the uncertainty on available resources, i.e., capacity, is extremely high. Our stochastic optimiza-
tion approach provides robust solutions to the problem by balancing the immediate costs generated by schedule/request dis-
crepancies with future costs of delays that one can expect to be assigned using a proposed flight schedule on the days of
operation.
Finally, it is important to note that, as a by-product, our models are able to address an issue of great concern for slot coor-
dinators. Indeed, slot coordinators are called for fixing the declared capacity of fully coordinated airports, i.e., the capacity
that is available at each airport to accommodate the demand. Such values of capacity are currently assessed by a thorough
demand and capacity analysis whenever there are significant changes in airport infrastructure, operational policies, or
demand patterns. By using our models, airports’ capacity values can be obtained from their optimal solutions. In fact, the
total number of departures and arrivals at airports, served in periods of demand peaks, can be interpreted as values of
declared capacity (which is bounded by the nominal capacity). In this case, the declared capacity of an airport will not be
a static value but rather time-dependent, i.e., it may change period by period on each day of operation thus guaranteeing
the fullest and most flexible usage of limited capacity at congested airports.
Our paper unfolds as follows. In Section 2 we describe our stochastic programming models, providing their complete for-
mulations. Section 3 shows the computational results of our models, derived from tests on realistic instances that represent
European airport networks of different sizes. Finally, Section 4 summarizes the conclusions and future directions of our
work.

2. Stochastic programming model

In this section we present our stochastic programming models for the Time Slot Allocation Problem under Uncertainty
(TSAPU). Given the nominal capacity of each airport, i.e., the number of available time slots, and airline requests for time
slots in order to operate specific flight movements, the TSAPU allocates time slots to airlines and aircraft operators. This
results in a flight schedule that allocates time slots with the goal of accommodating airline preferences (i.e., slot requests)
by minimizing the difference between the allocated and the requested slot times, i.e., the schedule/request discrepancies.
However, on each day of operation, available capacity can be much less than the nominal one – e.g., as a consequence of
bad weather conditions – and large delays have to be assigned to flights in order to resolve demand-capacity imbalances.
Therefore, a tradeoff exists between the schedule/request discrepancies and expected delays to be assigned on the day of
operations. In fact, a very tight schedule operating at nominal capacity is likely to be subject to delays on the day of operation
when capacity is reduced. The TSAPU aims at finding a schedule that results in a good compromise for the described tradeoff.
One of the fundamental aspects of our models is the fact that they consider a network of interconnected airports, taking
downstream effects of local decisions into account. Since a movement is either the departure or the arrival of a flight taking
place at an airport, a pair of movements ðm1 ; m2 Þ is associated with each flight, where m1 is the departure movement from
the airport of origin and m2 is the arrival movement of the same flight at the destination airport. The elapse of time between
the departure time slot at the airport of origin and the arrival time slot at destination corresponds to the flight time, which is
fixed – as it corresponds to the scheduled duration of the flight – and is given as input to the models. In view of this fact,
allocating a specific time slot for the execution of the departure of a flight explicitly fixes the allocation of an arrival time
slot at the airport of destination of the flight. Decisions are therefore not merely local, as they have effects on the whole
 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29 19

network. Furthermore, each movement is requested on a set of days, to ensure that the same time slot is assigned on each
day of operation to the same movement. Therefore, assigning a time slot to a movement affects multiple days simulta-
neously. From a modeling point of view, this helps to keep the models’ dimensions at a reasonable size, as the number of
decisions to be made is independent from the number of days a movement is requested.
A second aspect that is explicitly modeled in the TSAPU is flight connectivity. Some airlines are willing to schedule – espe-
cially at their hub airports – banks of operations, i.e., a group of flights whose arrival or departure times fall within a time
window. Moreover, in many cases flights are operated by the same aircraft and crew. In this case, the departure movement
needs to be allocated a time slot that is subsequent to the time slot of the arrival movement. The minimum time separating
the arrival and departure movements of two such flights, named connected flights, is known as turnaround time. Further-
more, the TSAPU explicitly models uncertainty, which affects – predominantly – airport capacity. In the following subsection,
we show an example that highlights the benefit of incorporating uncertainty within the mathematical models.
Finally, we highlight the flexibility in several respects of the models we propose. First, they are flexible enough to accom-
modate constraints deriving from operational considerations. This could be the case of considering airport capacity as defined
by capacity envelopes representing the set of all feasible combinations of arrivals and departures that can be accommodated at
a specific airport in a given period of time, see Gilbo (1993). An obvious example involves cases in which the same runway is
used for both arrivals and departures during a particular time period: the more arrivals are assigned to the runway, the fewer
departures can be served, and viceversa. At multi-runway airports, arrival and departure capacities vary according to the run-
way configuration used and ‘‘mix’’ of operations (arrivals only, departures only, or a combination of both) which is assigned to
each of the active runways. Despite decisions related to the optimal mix of operations being of daily/operational nature, they
have a remarkable impact in setting the declared capacity of an airport. Second, they are flexible enough to be used within the
current IATA system. Although they do not implement any of the existing IATA rules for time slot allocation, the proposed
models can be easily implemented to respect the IATA guidelines. Indeed, to adhere to the fundamental principle of the IATA
allocation process, i.e., the grandfather right, the models can be solved sequentially for each single priority class, i.e., first allo-
cating requests holding ‘‘grandfather rights’’, then allocating 50% of the new capacity, including withdrawn and surrendered
slots, to requests of ‘‘new entrants’’, and finally allocating all remaining time slot requests.

2.1. Example

Before providing the complete formulations of the models, we highlight their functioning by illustrating a simple example
in which considering uncertainty on capacity availability can lead to decrease flight delays.
Consider an airport with capacity constraints defined at each hour, with a limit of 2 departure movements, 2 arrival
movements and 3 total movements per hour. Suppose that different airlines formulate the following movement requests
at this airport: (a) arrival at 8:30, (b) departure at 9:00, (c) arrival at 9:20, (d) departure at 9:50, (e) arrival at 10:20, and
(f) departure at 10:30. These requests correspond to the aggregated hour by hour demand reported in Table 1. It is easily
possible to verify that this demand respects the airport’s capacity constraints, therefore in a deterministic setting the optimal
solution allocates all requested time slots to airlines.
Let us now consider uncertainty on available capacity, by defining two equiprobable scenarios:

 Nominal capacity is respected with probability 0.5.

 Capacity is decreased by 33% (departure and arrival capacity = 1, total capacity = 2) with probability 0.5.

Considering this information and 150 wide time slots, the delay on the day of operation of flights – when flights can only
be delayed and cannot depart ahead of schedule – using the schedule defined only considering deterministic information
would be:

 With nominal capacity: 0.

 With reduced capacity: movement d is delayed to the 10:00 time slot (1 time instant) and movement f is delayed to the
11:00 time slot (2 time instants), resulting in a total of 3 instants of delay.

The average delay is therefore equal to 0  0:5 þ 3  0:5 ¼ 1:5 time instants. This delay can be reduced by considering
capacity scenarios when defining the schedule. For example, it is possible to define a schedule that is feasible for all capacity
scenarios, i.e., providing no operational delay, by simply moving departure movement b from the 9:00 to the 8:45 time slot.
This results in a delay on the schedule of 1 time instant, which is less than the average operational delay of 1.5 time instants

Table 1
Example: movement demand.

Time horizon Arrivals Departures Total

8–9 1 0 1
9–10 1 2 3
10–11 1 1 2
20 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29

on the day of operation of flights obtained by using the schedule formulated only considering deterministic information.
From this small example it is also possible to see the tradeoff between the schedule/request discrepancies (0 units consid-
ering deterministic information versus 1 unit considering stochastic information) and the delay on the day of operations (1.5
units considering deterministic information versus 0 units considering stochastic information).

2.2. Formulation

In this section, we present the two-stage stochastic programming formulation for each of the two TSAPU models. For
those who are unfamiliar with this modeling approach, stochastic programming is a framework to explicitly represent
uncertainty, modeled as a multivariate random variable (x ~ ). Some of the decisions (first-stage decisions) are made without
full information on random events. Once uncertainty is disclosed, i.e., when the realization of multivariate random variable
(x) is known, second-stage decisions (yðx; xÞ) are made (see Birge and Louveaux (1997) for details). The general two-stage
SIP model can be formulated as follows:
z ¼ Min c> x þ E½f ðx; x
~ Þ
s:t: Ax P b ð1Þ
x2X
where E denotes the mathematical expectation with respect to x
~ and f ðx; x
~ Þ is the recourse (second-stage) value function
and is given by the following mixed-integer program:

f ðx; xÞ ¼ Min qðxÞ> yðxÞ

s:t: WðxÞyðxÞ P hðxÞ  TðxÞx ð2Þ
yðxÞ 2 Y
qðxÞ; WðxÞ; TðxÞ and hðxÞ are the components of the multivariate random variable. All the matrices and vectors of appro-
priate dimensions and X and Y are given sets.
The two formulations we propose for the TSAPU are identical in the first-stage, where the program schedules flights
according to airline requests, satisfying nominal capacity at all airports. The objective is to minimize the schedule/request
discrepancies, plus the estimate of future delays that will arise by choosing such a schedule. This estimate is provided by
the second-stage recourse function, which is described in Section 2.2.2. For this recourse function we provide two alternative
formulations. The first formulation, herein called ‘‘simplified recourse’’, does not consider the downstream effect of
delays over time, providing a rounding down estimate of future delays. It has the advantage that it can be computed
considering each time instant independently of the others. The second formulation, herein called ‘‘time-linked recourse’’,
takes into account delay propagation between consecutive time instants. This formulation is more accurate in computing
future delays, but it may require longer computational time to solve the problem. In the next subsection, we describe the
first-stage subproblem, which is common to both formulations, and then we present the recourse subproblem for each
formulation.

2.2.1. Formulation of the first-stage

To formulate the problem, we assume that the time horizon is fixed and subdivided into equal size contiguous time slots.
This assumption is typical of nearly all Air Traffic Management (ATM) formulations and agrees with practice. We now
introduce the notation we use in the TSAPU formulations.

Notation
T: Set of time slots, indexed by t;
D: Set of scheduled days, indexed by d;
A: Set of airports, indexed by a;
Dda : Set of departure movements taking place at airport a 2 A on day d 2 D;
Ada : Set of arrival movements taking place at airport a 2 A on day d 2 D;
M: Set of all movements ðM ¼ [a2A;d2D Dda [ Ada Þ; indexed by m;
dm : Number of days on which movement m is requested;
W: Weight of second-stage (operational) delays compared with first-stage discrepancies;
DC a ; AC a ; TC a : Set of departure, arrival and total capacity constraints for airport a 2 A, indexed by c, each spanning over
tc time slots;
F: Set of movement pairs ðm1 ; m2 Þ corresponding to the operation of a specific flight;
tm12 : Flight time of the flight corresponding to the pair of movements ðm1 ; m2 Þ 2 F;
P: Set of movement pairs ðm1 ; m2 Þ corresponding to consecutive flights, where m1 2 M is the arrival
corresponding to departure m2 2 M;
lm12 : Minimum time difference between the two movements ðm1 ; m2 Þ 2 P;
 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29 21

tm : Requested time slot for movement m 2 M ;

ctm : Cost of allocating movement m 2 M to time slot t; with ctm ¼ jt  t m j;
dt
adt dt
a ; ba ; ca : Departure, arrival and total capacity available at airport a 2 A on day d 2 D at time t 2 T.

Decision variables:


1; if movement m 2 M is allocated to time slot t
xtm ¼
0; otherwise

We highlight that the number of decision variables is not related to the number of days on which a movement is
requested. Therefore, if a single movement can be assigned to 10 different time slots, the corresponding number of decision
variables is equal to 10 even if the movement is requested on several days, as the time slot assigned is the same on each
requested day.
The first-stage formulation can now be stated as follows:
XX
Min dm ctm xtm þ W  E½f ðx; x
~ Þ ð3aÞ
m2M t2T
X
s:t: xtm ¼ 1 8m 2 M ð3bÞ
t2T
X X
xsm 6 adt
a 8a 2 A; c 2 DC a ; d 2 D; t 2 T ð3cÞ
m2Dda s2½t;tþtc Þ
X X
xsm 6 bdt
a 8a 2 A; c 2 AC a ; d 2 D; t 2 T ð3dÞ
m2Ada s2½t;tþtc Þ
X X
xsm 6 cdt
a 8a 2 A; c 2 TC a ; d 2 D; t 2 T ð3eÞ
m2Dda [Ada s2½t;tþt c Þ
tþt m
xtm1  xm2 12 ¼ 0 8ðm1 ; m2 Þ 2 F; t 2 T ð3fÞ
X X
xtm2 þ xtm1 61 8ðm1 ; m2 Þ 2 P; k 2 ½lm12 ; jTjÞ ð3gÞ
t2½0;kÞ t2½klm12 ;jTjÞ

xtm 2 f0; 1g; 8m 2 M; t 2 T ð3hÞ

where for a given outcome x of x ~ defined on a probability space ðX; A; PÞ, the so-called recourse function f ðx; xÞ is given as
defined in Section 2.2.2.
The first-stage makes a decision on the schedule of requested movements m 2 M at every airport a 2 A, taking into
account the effect that possible capacity reductions in the future (on the day of operation of a flight) may have in terms
of delays. The objective function (3a) minimizes the weighted sum of the schedule/request discrepancies and operational
delays. Constraints (3b) guarantee that a time slot is assigned to every requested movement. Constraints (3c)–(3e)
ensure that departure, arrival and total capacity constraints are satisfied at every airport a 2 A all the time. We recall
that capacity is measured on t c consecutive time periods. Constraints (3f) guarantee that the flight time for all flights
is fixed, by properly distancing departure movements from the corresponding arrivals. Note that the two decision vari-
ables referring to the departure movement at time t and to the arrival movement at time t þ t m12 of a flight will have the
same value, as they will either both be equal to 0 (flight not departing at t) or to 1 (flight departing at t, therefore arriv-
ing at t þ tm12 ). Constraints (3g) guarantee that the turnaround time for all aircraft is respected, by properly distancing all
couples of consecutive movements ðm1 ; m2 Þ 2 P. Finally, constraints (3h) impose binary restrictions on the decision vari-
ables xtm .

2.2.2. Formulation of the second-stage

We here describe the second-stage value function that measures the expected delays – assigned to flights in order to meet
airport capacity realizations – taken over all possible realizations of the random event x. In what follows, we suppose that
uncertainty only affects the capacity of airports. More specifically, the right-hand side of the constraints defining the capacity
envelopes (the c values) depends on the random variable x ~.

Decision variables:
yx
adt : Number of delayed departures from airport a 2 A on day d 2 D at time instant t 2 T
zx
adt : Number of delayed arrivals to airport a 2 A on day d 2 D at time instant t 2 T

Given a first-stage decision – which is the schedule of departures and arrivals at each airport and for each time instant –
and the realization of the random event x – which is the capacity at each airport and for each period of time – a certain
number of flights has to be delayed to balance demand and capacity. For each airport a 2 A, time instant t 2 T and day
d 2 D, the set of feasible second-stage decisions Ytad ðx; xÞ can be defined as follows:
22 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29

 x x
Ytad ðx; xÞ ¼ ðyadt ; zadt Þjyx x
adt 2 Rþ ; zadt 2 Rþ ;
X
x s
xm  adt
yadt P a ðxÞ 8c 2 DC a ;
m2Dda ;s2½t;tþtc 
X
zx
adt P xsm  bdt
a ð xÞ 8c 2 AC a ;
m2Ada ; s2½t;tþtc 
9
X =
yx x
adt þ zadt P xsm  cdt
a ð xÞ 8c 2 TC a
;
m2Dda [Ada ; s2½t;tþtc 

Note that the stochastic program has complete recourse, therefore Ytad ðx; xÞ – £ for any x and realization x. If we denote
with qa the cost for delaying a movement at airport a 2 A, the second-stage value function is given by:
X XX 
f ðx; xÞ ¼ Min qa yx x
adt þ zadt ð4aÞ
a2A d2D t2T

ðyx x t
adt ; zadt Þ 2 Yad ðx; xÞ 8a 2 A; d 2 D; t 2 T: ð4bÞ

It is important to note that the second-stage value function described above is only an estimate of the delays assigned
to movements in order to meet available capacity given slot assignment x. Indeed, this second-stage formulation
penalizes the excess of demand at each time period without taking into account the interactions between time periods,
i.e., it does not capture the propagation of delays from one time period to the following one. Delay propagation
involves having a delayed arrival (departure) movement at time instant t increase the demand of arrivals (departures)
at time instant t þ 1 by one unit. Due to this simplification in the formulation, we refer to this formulation as
‘‘simplified’’ recourse.
To overcome this issue, we consider a different formulation of the second-stage, where demand at consecutive time
instants is linked, i.e., movements not operated at time t increase the demand for movements at time t þ 1, providing a better
estimate of delays. We refer to this formulation as ‘‘time-linked’’ recourse.
X XX 
f ðx; xÞ ¼ Min qa yx x
adt þ zadt ð5aÞ
a2A d2D t2T

ðyx x x x t
adt  yad;t1 ; zadt  zad;t1 Þ 2 Yad ðx; xÞ 8a 2 A; d 2 D; t 2 T: ð5bÞ

Although the recourse value function computes the number of delayed movements resulting from a specific capacity
realization, it does not capture the effect of delay propagation on the complete network. Indeed, to compute such effect,
the second-stage formulation should keep track of each single flight. Therefore, both proposed recourse functions compute
an approximate value (lower bound) of real operational delays. Denoting with DS ; DTI the expected future delays computed
by the simplified and the time-linked recourse functions respectively, and denoting with D the exact expected future
delays, the following condition holds: DS 6 DTI 6 D. The discussed loss of precision, however, results in reduced
computational effort required to solve the instances. Furthermore, the solutions of the proposed models might be more
conservative – meaning that they might assign less schedule/request discrepancies – than the corresponding fully hedged
solutions.

2.3. Second-stage formulation example

We now show a practical example to provide a better understanding of the two second-stage formulations and highlight
the differences between them. Consider a (first-stage) schedule x and a capacity realization x. The proposed schedule cor-
dt
responds to a specific demand dac for departures and adt ac for arrivals at each considered airport a 2 A on each scheduled day
d 2 D and at each time instant t 2 T. We also consider capacity defined at each time instant, having effect only on the same
time instant, i.e., the cdt
ac ðxÞ values define the maximum demand that can be accommodated at each time instant t. In the
example illustrated in Table 2, we can observe how the objective values of the simplified and time-linked recourse formu-
lations can be significantly different.
dt
In Table 2, the first row (Time) provides the different time instants considered. The second row (dac ) shows the instant
departure demand at the considered airport a on the considered day d for the departure capacity constraint c. Similarly,

Table 2
Second-stage delay calculation example.

Time 1 2 3 4 5 6
dt
dac 5 6 7 6 6 5
dt
cac 6 6 6 6 6 6
Simpl. delay 0 0 1 0 0 0
T-link. delay 0 0 1 1 1 0
 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29 23

the third row (cdt

ac ) depicts available capacity at each time instant. Finally, rows 4 and 5 provide the amount of delay esti-
mated by each recourse formulation. In this table, we can observe that at time t ¼ 3 capacity is exceeded by 1 departure
movement. Demand matches capacity at the next 2 time instants, therefore this departure delay has a downstream effect,
causing delays at time instants 4 and 5. This fact is correctly captured by the time-linked recourse formulation, which will
have an objective value for this part of the problem equal to 3. The simplified recourse formulation, on the other hand, only
detects on how many time instants capacity is exceeded, without considering the downstream effect of delays, thus provid-
ing an objective value equal to 1.

3. Computational experience

In this section, we present the computational experience with the mathematical models presented in Section 2. We
considered four realistically generated problem instances, representing slot requests over different networks of intercon-
nected airports. Each problem instance considers four different days. For each instance, we evaluated two different delay
weight factors, i.e., W ¼ 1 and W ¼ 4, as decisions may be very different depending on the relative weight between one
unit of schedule/request discrepancy and one unit of second-stage delay. Each day has been subdivided into 5 min time
instants. We consider nominal airport capacity and possible capacity realizations at each airport for all capacity
constraints, i.e., departure, arrival and total movement constraints. Capacity constraints all have a 1 h duration, e.g., a
total capacity constraint of 20 movements at airport a and at the time instant corresponding to 10:00 imposes a
limitation of 20 total movements (arrivals and departures) between 10:00 and 11:00. Capacity reduction scenarios were
generated from historical weather conditions at the different considered airports. We consider every airport and each
Q
simulated day independently, and thus the number of possible scenarios for each test instance is given by a2A XjDj a . Here,
Xa is the number of possible capacity realizations at airport a 2 A and jDj is the number of scheduled days. As mentioned
above, we are not considering correlations among capacity scenarios at different airports; also because the airports
included in our instances are quite spread out on a large geographical area. In case of correlations among capacity
realizations at different airports - if evidenced by mean of a statistical analysis on historical data -, this would produce
a twofold effect. On one side, uncertainty should be represented by a smaller number of scenarios thus reducing the
computational effort to solve the stochastic program. On the other side, there might be a reinforcement of the delay
propagation phenomenon.
We here list all airports included in our test instances and provide the International Civil Aviation Organization (ICAO)
code for each of them. The description of each test instance is preceded by an abbreviated name which we use in the rest
of the paper.

 5europe: Hamburg (EDDH), London Gatwick (EGKK), Copenhagen (EKCH), Lisbon (LPPT), Basel-Mulhouse-Freiburg (LFSB)
 6europe: Hamburg (EDDH), Edinburgh (EGPH), Liverpool (EGGP), Warsaw (EPWA), Barcelona (LEBL), Porto (LPPR)
 france: Paris Charles de Gaulle (LFPG), Toulouse (LFBO), Lyon (LFLL), Basel-Mulhouse-Freiburg (LFSB)
 large: Bergamo (LIME), Beauvais Tillé (LFOB), London Gatwick (EGKK), Tarbes-Lourdes (LFBT), Rhodes International
(LGRP), Menorca (LEMH), Porto (LPPR), East Midlands (EGNX), Pisa (LIRP), Ibiza (LEIB)

In Table 3 we provide details on the size of each test instance. The figures reported in this table are valid for both recourse
formulations and for both weight factors. Column 1 (Instance) provides the name of each test instance. Columns 2 (Flights)
and 3 (Mov.s) provide the number of flights and the number of movements considered for each test instance, respectively.
We recall that each flight, even those operated on several days, correspond to a single pair of movements. Columns 4 (1st
Cols) and 5 (1st Rows) provide the size of first-stage problems, while columns 6 (2nd Cols) and 7 (2nd Rows) provide the
size of (second-stage) subproblems. Subproblems have the same size with the simplified and time-linked recourse formu-
lations. Finally, column 8 (# Scen.) provides the total number of possible scenarios for each test instance, calculated with
the previously described formula.
Furthermore, in Table 4 we provide some insight on the level of congestion presented by each test instance. In column 2
we provide the number of average movement requests per airport that exceed available ‘‘instant capacity’’. By instant capac-
ity we refer to the result of the subdivision of total capacity, which is defined for a set of contiguous time instants (in our
case, the set spans over a 1 h horizon), over single time instants (that are 5 min wide). In column 3 we provide the average
number of time instants per airport at which demand is greater or equal to the instant capacity.

Table 3
Description of test instances.

Instance Flights Mov.s 1st Cols 1st Rows 2nd Cols 2nd Rows # Scen.
5europe 9267 6402 81,057 15,716 11,520 10,939 5:5  1011
6europe 8458 5872 74,057 13,917 13,824 10,119 6:9  1013
france 8474 5288 68,010 11,209 9281 6366 6:5  1010
large 6520 5702 73,068 25,594 23,074 20,423 1:4  1017
24 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29

Table 4
Test instance congestion.

Instance Exceeding demand Congested instants

Per airport Per airport
5europe 203.6 196.0
6europe 140.2 99.2
france 74.3 77.8
large 375.0 189.2

3.1. Experimental results

We solved all test instances using both simplified and time-linked recourse formulations on a computer with Intel Xeon
X5670 quad core CPU at 2.93 GHz with 16 GB of RAM. To solve the test instances we implemented the Sample Average
Approximation (SAA) method, which is an exterior sampling method. In SAA, a sample is defined from the set of scenarios
and a corresponding approximation of the recourse function is determined. The approximate objective, called sample average
approximation, of the recourse function is determined, and optimization is performed without further sampling. For details
on the SAA methodology, the interested reader may refer to chapter 6 of Ruszczynski and Shapiro (2003). The SAA was coded
in C++ language using the CPLEX 12.1 library. To compute the lower bound, we performed 20 SAA replications for each test
instance. We chose a different sample with 100 scenarios at each replication. The SAA optimality gap was set to 5% for all test
instances.
We report the computational results for the simplified recourse formulation in Table 5, while Table 6 describes the results
for the time-linked formulation. The two tables have the following structure. In column 1 (Instance) we provide the name of
each test instance. Column 2 (Weight) indicates the weight factor W used in the definition of the objective function. Column
3 (Lower Bound) indicates the 95% confidence interval for the lower bound obtained by the 20 SAA replications. The solution
out of the 20 SAA replications whose value is the closest to the average value computed by the SAA is then used to compute
the upper bound value. This solution was evaluated over 20 different samples of size 1000, resulting in the 95% upper bound
confidence interval indicated in column 4 (Upper Bound). Column 5 (1st Stage) shows the first-stage objective value, i.e., the
schedule/request discrepancies, of the solution used to compute the upper bound. Finally, columns 6 (Lower Bound t (h)) and
7 (Upper Bound t (h)) provide the computation times, in hours, for the lower and upper bound, respectively.
The upper bound computation times reflect the level of difficulty of solving the subproblems. Comparing the values in the
two tables, we can observe that it is equally difficult to solve subproblems with simplified or time-linked recourse. However,
simplified recourse subproblems can easily be decomposed into smaller independent problems, one for each time instant,
airport and constraint, while subdivision by time instants is not possible using the time-linked recourse formulation. Com-
putation times for subproblems can be improved by implementing parallel algorithms, with the simplified recourse formu-
lation providing a much higher level of parallelism, therefore being potentially much faster than the time-linked recourse
formulation. Lower bound computation times, on the other hand, can differ greatly between simplified and time-linked
recourse due to the different cuts provided by subproblems. Averaging over our test instance set, calculating the lower bound
is faster using the simplified recourse formulation. With weight factor 1, the lower bound is calculated in 5.9 h and 10.1 h for
the simplified and time-linked recourse formulations, respectively. Weight factor 4 drastically increases computation times,
showing that the tradeoff between the first- and second-stage is much more difficult. The lower bound in this case is com-
puted in 27.5 h and 34.2 h on average using the two different formulations.
In order to provide more insights on the stochastic solutions, in Tables 7 and 8 we report an estimation of the cost, i.e.,
sum of schedule/request discrepancies and future delays, of the ‘‘nominal solution’’ and the ‘‘no discrepancy solution’’. With
nominal solution we indicate the first-stage solution that only considers nominal capacity at airports, i.e., that minimizes the
first-stage cost and does not consider future delays. With no discrepancy solution, on the other hand, we indicate the
solution that satisfies all airline requests. This solution has no first-stage cost, but can expect higher future delays due to
not considering capacity constraints at the strategic phase at all. Tables 7 and 8 have the following structure. Column 1
(Instance) and column 2 (weight) identify the instance name and weight factor applied. Columns 3 (Nominal Conf. Int.)

Table 5
Simplified recourse results.

Instance Weight Lower Upper 1st Lower Upper

bound bound stage bound t (h) bound t (h)
5europe 1 [812.728, 814.234] [807.621, 813.613] 696 4.1 0.4
5europe 4 [974.149, 983.871] [962.152, 985.844] 832 32.4 0.4
6europe 1 [94.475, 97.384] [86.503, 97.178] 0 3.6 0.4
6europe 4 [376.046, 387.833] [340.59, 383.326] 6 7.7 0.4
france 1 [97.649, 98.111] [98.199, 99.543] 9 1.7 0.2
france 4 [133.815, 135.376] [166.893, 168.499] 123 30.3 0.2
large 1 [1,647.72, 1,650.73] [1,649.42, 1,660.78] 1516 14.2 1.4
large 4 [2,017.18, 2,029.93] [2,011.27, 2,055.79] 1572 39.4 1.3
 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29 25

Table 6
Time-linked recourse results.

Instance Weight Lower Upper 1st Lower Upper

bound bound stage bound t (h) bound t (h)
5europe 1 [823.527, 825.855] [836.301, 842.435] 710 8.1 0.4
5europe 4 [1,019.02, 1,032.14] [1,031.71, 1,055.71] 810 34.1 0.4
6europe 1 [200.94, 216.614] [199.081, 230.874] 0 3.5 0.4
6europe 4 [792.541, 845.915] [788.957, 890.763] 98 16.1 0.4
france 1 [97.727, 98.108] [98.168, 99.364] 9 1.8 0.2
france 4 [133.544, 135.434] [153.326, 154.502] 120 30.6 0.2
large 1 [1,829.23, 1,843.111] [1,849.81, 1,900.6] 1531 27.0 1.4
large 4 [2,765.82, 2,818.93] [2,802.6, 3,003.28] 1620 55.8 1.4

Table 7
Simplified recourse results for nominal and no discrepancy solutions.

Instance Weight Nominal conf. int. No discr. conf. int. Nominal 1st stage
5europe 1 [812.358, 818.479] [713.561, 717.134] 696
5europe 4 [1161.43, 1185.91] [2,854.25, 2,868.53] 696
6europe 1 [87.645, 97.028] [89.566, 99.2] 0
6europe 4 [350.58, 388.112] [358.264, 396.8] 0
france 1 [97.711, 99.177] [96.52, 98.037] 8
france 4 [366.845, 372.707] [386.078, 392.146] 8
large 1 [1,648.37, 1,660.81] [1,224.84, 1,236.54] 1,516
large 4 [2,045.48, 2,095.22] [4,899.35, 4,946.17] 1,516

Table 8
Time-linked recourse results for nominal and no discrepancy solutions.

Instance Weight Nominal conf. int. No discr. conf. int. Nominal 1st stage
5europe 1 [822.326, 829.727] [1,359.6, 1,365.24] 696
5europe 4 [1,201.3, 1,230.91] [5,438.41, 5,460.97] 696
6europe 1 [187.363, 236.635] [197.873, 236.979] 0
6europe 4 [749.454, 945.458] [791.491, 947.917] 0
france 1 [97.748, 98.892] [96.955, 98.393] 8
france 4 [366.991, 371.569] [387.822, 393.57] 8
large 1 [1.817.16, 1,860.78] [2,523.55, 2,564.57] 1516
large 4 [2,720.62, 2,895.12] [10,094.2, 10,258.3] 1516

and 4 (No Discr. Conf. Int.) provide the 95% confidence interval of the solution computed using the nominal and no
discrepancy solution, respectively. For the first confidence interval, the values represent the cost of the first- and
second-stage costs. For the no discrepancy confidence interval, on the other hand, the cost is all in the second-stage, since
the discrepancy is null. Finally, column 5 (Nominal 1st Stage) provides the first-stage value of the nominal solution.

3.2. Results interpretation

First of all, we briefly comment the weight factors used (W ¼ 1 and W ¼ 4). With weight factor W ¼ 1, we expect to
obtain results equivalent to those of the nominal solution. In fact, if strategic discrepancies and operational delays have
the same weight, it may be preferable to postpone delay decisions until the uncertainty is resolved, i.e., to assign delays
on the day of operations. Weight factor W ¼ 4, on the other hand, gives an incentive to assign schedule/request discrepancies
aimed at saving operational delays. From the computational results it is possible to analyze the tradeoff between schedule/
request discrepancies and (expected) operational delays. Comparing column 3 of Table 7 with columns 4 and 5 of Table 5 we
may evaluate the possible benefit of using the stochastic programming approach with simplified recourse. The comparison of
the stochastic solution is performed with respect to the deterministic counterparts, i.e., the nominal solution (column 4) and
the no discrepancy solution (column 5). Similarly, the comparison of column 3 of Table 8 with columns 4 and 5 of Table 6
allows to evaluate the benefit of the stochastic programming approach with time-linked recourse. A more detailed compar-
ison between the solutions is given in Table 9 and 10 for the simplified and time-linked recourse formulations, respectively.
Column 3 (1st Add.) reports the additional delay, i.e., the schedule/request discrepancy, assigned by the stochastic solution
with respect to the nominal solution. By paying this ‘‘price’’, we have milder demand-capacity operational imbalances, with a
reduction of expected operational delays, as reported in column 4 (2nd Save) of the two tables. Column 5 (Global gain)
reports the global gain measured as the percentage of the objective value that arises from using the schedule proposed
26 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29

Table 9
Delay tradeoff for simplified recourse instances.

Instance Weight vs Nominal vs No discrepancy

1st Add. 2nd Saved Global gain (%) 1st Add. 2nd Saved Pct. (%)
5europe 1 0 4.8 0.6 696 600.7 84.0
5europe 4 136 335.7 17.0 832 2,719.4 95.0
6europe 1 0 0.5 0.5 0 2.5 2.7
6europe 4 6 13.4 2.0 6 21.6 5.7
france 1 1 0.6 0.4 9 7.4 7.6
france 4 115 317.1 54.6 123 344.4 88.5
large 1 0 0.5 0.0 1516 1091.6 88.7
large 4 56 92.8 1.8 1572 4461.2 90.6

Table 10
Delay tradeoff for time-linked recourse instances.

Instance Weight vs Nominal vs No discrepancy

1st Add. 2nd Saved Global gain (%) 1st Add. 2nd Saved Pct. (%)
5europe 1 14 0.7 1.6 710 1,233.1 90.5
5europe 4 114 286.4 14.2 810 5,216.0 95.7
6europe 1 0 3.0 1.4 0 2.4 1.1
6europe 4 98 105.6 0.9 98 122.8 14.2
france 1 1 0.6 0.5 9 7.9 8.1
france 4 112 327.4 58.3 120 356.8 91.3
large 1 15 21.2 2.0 1531 2199.9 86.5
large 4 104 8.9 3.4 1620 8893.3 87.4

by our model. Notice that with a moderate assignment of additional schedule/request discrepancies, we may obtain a
significant reduction of expected operational delays. Furthermore, we compare the stochastic solutions with the no
discrepancy solutions in columns 6, 7, and 8. This analysis is of great concern for the US. In fact, with few exceptions, at
US airports all flights are scheduled at their preferred time, leading to cases – not uncommon – in which scheduled demand
exceeds even nominal capacity, thus generating persistent phenomena of airport congestion. Column 6 (1st Add) reports the
total schedule/request discrepancies assigned by the stochastic solutions. Column 7 (2nd Saved) provides expected saving in
terms of operational delays obtained by applying the stochastic solution with respect to the no discrepancy solution. These
savings are then provided in percentage terms in column 8 (Pct.), with respect to the operational delays expected by using
the no discrepancy solution.
The global gain that can be achieved with the discussed tradeoff differs greatly from test instance to test instance. First,
we highlight that imposing a weight W ¼ 1 provides no benefit, as expected. The global gain is always close to 0%, and the
small gains (or losses) reported are only given by the stochasticity of the results that are computed considering different
samples. Using weight factor W ¼ 4 the results greatly differ. Instances 5europe and france report a global gain of 17:0%
and 54:6% with the simplified formulation, and a gain of 14:2% and 58:3% with the time-linked formulation. Furthermore,
from the analysis of the no discrepancy solution, we can see that, unless the schedule/request discrepancies are small, i.e.,
the stochastic solution is close to the no discrepancy solution, not considering capacity constraints at all at the strategic
phase results in huge amounts of delay on the day of operations. This fact proves that time slot allocation is a needed stra-
tegic operation for the management of air traffic at airports subject to demand-capacity imbalances. In general, Tables 9,10
constitute a good analysis instrument for evaluating our results, indicating different possible actions from case to case.
Finally, we analyze the distribution of time slots for the four instances which reported the largest global gain, i.e., sim-
plified and time-linked recourse 5europe (see Fig. 1) and france (see Fig. 2) instances with W ¼ 4. The figures show the dis-
tribution over time of departure and arrival time slots. The values refer to the sum of discrepancies over all airports in the
considered networks. Notice that the total amount of discrepancies for departures is equal to the total amount of discrep-
ancies for arrivals. The two discrepancies, however, have different distributions in time. We show the distribution of the
schedule/request discrepancies of the stochastic solution with the gray line, and the distribution of the discrepancies of
the nominal solution with the black line. The vertical axis represents the average number of departure or arrival time slots
per day of discrepancy between the requests and the schedule. The horizontal axis represents the hours of the days. The pro-
file of the distribution of the schedule/request discrepancies for the stochastic solution is similar to that of the nominal solu-
tion for instance 5europe. Instance france, on the other hand, only has an average daily schedule/request discrepancy of 2
units with the nominal solution. The discrepancies are assigned at 17:00 for departures, and 18:00 for arrivals. The stochastic
solution proposes discrepancies that spread between 14:00 and 20:00 for departures, and between 17:00 and 22:00 for
arrivals. These figures indicate a link between the time at which discrepancies are proposed by the stochastic and the nom-
inal solution. While there is a link between the stochastic and the nominal solution in both cases, the differences between
them generate significantly different amounts of future delay. This shows the importance of evaluating the future effect of a
proposed allocation of time slots.
 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29 27

Fig. 1. Discrepancy distribution for 5europe.

Fig. 2. Discrepancy distribution for france.

4. Conclusions

In this paper, we have presented two stochastic programming models for the TSAPU, with the goal to determine robust
flight schedules, i.e., schedules that consider both airline requests for time slots and future delay on the day of operation that
can be expected by using a specific schedule. Today, slot allocation practice suffers several inefficiencies. Slot coordinators
have to make complex decisions with limited support from rule-driven slot management applications. These decisions are
28 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29

made locally, despite having a relevant network effect, as assigning a departure time slot at some airport needs to correspond
to the assignment of a specific arrival time slot at another airport. Local decisions are currently coordinated with biannual
conferences, which constitute a big overhead in the decision making process.
With our models, we propose a decision support tool for slot coordinators, as they can use it to assign time slots according
to airline requests and the future impact of the schedule. We propose two alternative formulations, which consider multiple
airports at the same time, thus taking into account the complex network interdependency that is not considered when
schedules are proposed locally.
Our contribution to literature is in providing two models that, to the best of our knowledge, are the first on time slot allo-
cation to consider multiple airports being scheduled as a whole network and the uncertainty on future capacity realizations.
The proposed models do not implement any of the existing IATA rules for time slot allocation (e.g., we do not consider dif-
ferent priority rules for flights) and are based on optimization criteria only. It is however possible to easily align these models
with the existing IATA slot allocation procedures by applying the models sequentially to each priority class, i.e., first allocat-
ing requests holding ‘‘grandfather rights’’, then allocating 50% of the new capacity, including withdrawn and surrendered
slots, to requests of ‘‘new entrants’’, and finally allocating all remaining time slot requests. This would lead to much faster
solutions for each subproblem than the times reported in Section 3, as each of them would be of smaller dimensions.
The practical impact of our models is that they can be a tool for proposing to the slot coordinator a tradeoff between the
schedule/request discrepancies and future delays; i.e., by paying a ‘‘cost’’ given by an increased distance of the schedule from
that requested by airlines, it is possible to obtain a reduction of delays that are expected to occur on the day of operation of
the flights. As highlighted in Section 3, the use of the proposed approach can be beneficial. In one of the test instances, it was
possible to reduce total delay costs, i.e., the sum of the schedule/request discrepancies and operational delays, by over 58%
using the time-linked recourse formulation, which is our most precise formulation. Also notice that another possible inter-
pretation of our results is determining the optimal capacity levels at each considered airport that minimizes the sum of sche-
dule/request discrepancies and operational delays.
The computation times obtained are viable for the application, as they are consistent with the time constraints associated
with the current decision making interaction between slot coordinators and airlines, which is a process spanning over sev-
eral days. Good results and computational viability together show that the proposed approach is very promising and
deserves further investigation, as it may lead to relevant monetary benefits for airlines and all other stakeholders. Future
research will focus on expanding these promising results to wider networks of coordinated airports, considering a larger
scheduling horizon. Both factors are critical as they will make the problem of larger scale, however the discussed subdivision
of the problem into smaller subproblems considering ‘‘grandfather rights’’ from IATA rules will help mitigate this increase in
the size of the problem to solve. Furthermore, an appropriate definition of the weight factor adopted in the objective function
should be studied. Such a definition needs to compare the costs of definition of a schedule at the strategic and operational
phases. Operational delays, in fact, present additional crew costs, as well as passenger-related costs, resulting in higher costs
compared with the strategic ones. Different airlines may desire the application of different weight factors to their flights,
possibly distinguishing flight by flight, in order to represent the real operation costs associated with each flight. However,
allowing for weight factors differentiated by airline or flight – possibly based on information provided by airlines – would
raise equity issues, as well as the possibility of ‘‘gaming’’ from airlines by misrepresenting their information. All of these fac-
tors go beyond the scope of this paper, but need to be addressed for a practical implementation of the proposed
methodology.

Acknowledgements

The authors are grateful to Saravanan Venkatachalam for his help in the development, execution and analysis of the com-
putational experiments.

References

Andreatta, G., Dell’Olmo, P., Lulli, G., 2011. An aggregate stochastic programming model for air traffic flow management. Euro. J. Operat. Res. 215 (3), 697–
704.
Ball, M.O., Donohue, G.L., Hoffman, K., 2005. Auctions for the safe, efficient and equitable allocation of airspace system resources. In: Crampton, P., Shoham,
Y., Steinberg, R., (Eds.), Combinatorial Auctions. MIT Press (chapter 20).
Ball, M.O., Hoffman, R., Mukherjee, A., 2010. Ground delay program planning under uncertainty based on the ration-by-distance principle. Trans. Sci. 44 (1),
1–14.
Bayen, A., Raffard, R., Tomlin, C., 2006. Adjoint-based control of a new Eulerian network model of air traffic flow. IEEE Trans. Cont. Syst. Technol. 14 (5), 804–
818.
Birge, J.R., Louveaux, F., 1997. Introduction to stochastic programming. Springer-Verlag.
Brueckner, J.K., 2002. Internalisation of airport congestion. J. Air Trans. Manage. 8 (3), 141–147.
Castelli, L., Pellegrini, P., Pesenti, R., 2012. Airport slot allocation in Europe: economic efficiency and fairness. Int. J. Rev. Manage. 6 (1/2), 28–44.
Czerny, A.I., Forsyth, P., Gillen, D., Niemeier, H.-M., (Eds.), 2008. Airport Slots: International Experiences and Options for Reform. Ashgate.
Dell’Olmo, P., Lulli, G., 2003. A dynamic programming approach for the airport capacity allocation problem. IMA J. Manage. Math. 14, 235–249.
de Neufville, R., Odoni, A.R., 2003. Airport Systems: Planning, Design and Management. McGraw-Hill.
EUROCONTROL. Performance Review Report on European Air Traffic Management Performance in 2011 <http://www.eurocontrol.int/documents/
performance-review-report-european-air-traffic-management-performance-2011> (May).
EUROCONTROL, 2013. Challenges of Growth Study.
Fan, T.P., 2003. Airport Demand Management and Congestion Pricing. Ph.D. thesis, Massachusetts Institute of Technology, Cambridge, MA.
 L. Corolli et al. / Transportation Research Part C 46 (2014) 16–29 29

Fukui, H., 2010. An empirical analysis of slot trading in the unites states. Trans. Res. Part B 44 (3), 330–357.
Gilbo, E.P., 1993. Airport capacity: representation, estimation optimization. IEEE Trans. Cont. Syst. Technol. 1, 144–154.
Kleit, A., Kobayashi, B., 1996. Market failure or market efficiency? Evidence on airport slot usage. Res. Trans. Econ. 4 (1), 1–32.
Kosters, D., 2007. Airport scheduling performance – an approach to evaluate the airport scheduling process by using scheduled delays as quality criterion.
In: Annual World Conference on the Proceedings of the Air Transport Research Society (ATRS), vol. 6.
Odoni, A., 1987. The flow management problem in air traffic control. In: Odoni, A., Szego, G. (Eds.), Flow Control of Congested Networks. Springer-Verlag, pp.
269–288.
Pellegrini, P., Castelli, L., Pesenti, R., 2012. Secondary trading of airport slots as a combinatorial exchange. Trans. Res. Part E 48, 1009–1022.
Raffarin, M., 2004. Congestion in european airspace: a pricing solution? J. Trans. Econ. Pol. 38, 109–126.
Rassenti, S.J., Smith, V.L., Bulfin, R.L., 1982. A combinatorial auction mechanism for airport time slot allocation. Bell J. Econ. 4 (2), 402–417.
Richetta, O., Odoni, A., 1994. Dynamic solution to the ground-holding problem in air traffic control. Trans. Res. Part A, Policy Pract. 28, 167–185.
Ruszczynski, A., Shapiro, A., 2003. Stochastic programming. Handbooks in Operations Research and Management Science, vol. 10. Elsevier.
Starkie, D., 1998. Allocating airport slots: a role for the market? J. Air Trans. Manage. 4 (2), 111–116.
Study on the Impact of the Introduction of Secondary Trading at Community Airports, 2006. Technical Report, Mott MacDonald <http://ec.europa.eu/
transport/modes/air/studies/doc/airports/airports.zip> (November).
Sun, D., Bayen, A., 2008. A multicommodity Eulerian–Lagrangian large capacity cell transmission model for enroute traffic. AIAA J. Guid., Cont. Dynam. 31
(3), 616–628, 200.
Verhoef, E., pricing, Congestion, 2010. slot sales and slot trading in aviation. Trans. Res. Part B 44 (3), 320–329.
Wan, Y., Roy, S., 2008. A scalable methodology for evaluating and designing coordinated air-traffic flow management strategies under uncertainty. IEEE
Trans. Intell. Trans. Syst. 9 (4), 644–656.
Worldwide Slot Guidelines, second ed. IATA, January 2012.
Zografos, K.G., Salouras, Y., Madas, M.A., 2012. Dealing with the efficient allocation of scarce resources at congested airports. Trans. Res. Part C 21 (1), 244–
256, 4.