You are on page 1of 8

A Hybrid Ant-Based System for Gate Assignment

Problem

Camelia-M. Pintea1, Petrica C. Pop2, Camelia Chira1, and D. Dumitrescu1


1
Babes-Bolyai University, Kogalniceanu 400084, Cluj-Napoca, Romania
2
North University, V.Babes 430083, Baia-Mare, Romania
{cmpintea,cchira,ddumitr}@cs.ubbcluj.ro, pop_petrica@yahoo.com

Abstract. This paper presents an ant system coupled with a local search applied
to an over-constrained airport gate assignment problem (AGAP). In the airport
gate assignment problem we are interested in selecting and allocating aircrafts
to the gates such that the total passenger connection time is minimized. Our al-
gorithm uses pheromone trail information to perform modifications on AGAP
solutions, unlike traditional ant systems that use pheromone trail information to
construct complete solutions. The algorithm is analyzed and compared with
tabu search heuristic and Ant Colony System metaheuristic.

Keywords: ant algorithms, assignment problems, constraints, t-test.

1 Introduction

An airport is a transfer point between travel domains in which two unlike carriers,
motor vehicles and airplanes, must interact to exchange people and luggage. The main
purpose of flight-to-gate assignments is to assign aircrafts to suitable gates so that
passengers can conveniently embark and disembark. The distance a passenger has to
walk in any airport to reach various key areas including departure gates, baggage belts
and connecting flights provide for an important performance measure of the quality of
any airport.
The Airport Gate Assignment Problem (AGAP) seeks optimal flight-to-gate as-
signments to minimize the total passenger connection times. AGAP can be viewed as
an instance of the Quadratic Assignment Problem (QAP).
Braksma and Shortreed [2] have described one of the first attempts to use quantita-
tive means to minimize intra-terminal travel into a design process. Babic et al. [1]
modeled the AGAP as a 0-1 integer program and used branch-and-bound algorithm to
find solutions where transfer passengers were not considered. Yan and Chang [11]
proposed a network model that was formulated as a multi-commodity flow problem.
Network-based simulation of aircrafts at gates in airport terminals were considered by
Cheng [5,6]. Wang et al. [9] presented an optimization algorithm for solving AGAP
and as an application, they achieved in optimizing the dispatch of 334 airlines by us-
ing of 18 gates in the case of Beijing International Airport.
The most notable heuristic methods developed for AGAP are the following: tabu
searches Xu and Bailey [10], Ding et al. [7], a hybrid simulated annealing Ding et al.

E. Corchado, A. Abraham, and W. Pedrycz (Eds.): HAIS 2008, LNAI 5271, pp. 273–280, 2008.
© Springer-Verlag Berlin Heidelberg 2008
274 C.-M. Pintea et al.

[7] and a genetic algorithm Gu and Chung [8]. Ding et al. in [7] focused on the AGAP
with the objective of minimizing distance costs of the over constrained gate assign-
ment problem, minimizing the number of ungated aircraft and the total walking dis-
tances.
This paper presents a hybrid ant-local search system for the airport gate assignment
problem, denoted HAS-AGAP, which works better than known heuristics (presented
in the next section). The paper is organized as follows. The definition, a quadratic
programming formulation for the AGAP and a greedy method for minimizing ungated
flights are presented in Section 2. Our hybrid ant-local search is presented in Section
3. Then, the performance of our algorithm is analyzed by comparing with Ant Colony
System. Finally, in the last section some conclusions are drawn.

2 The Airport Gate Assignment Problem

The Airport Gate Assignment Problem (AGAP) seeks feasible flight-to-gate assign-
ments so that the total passenger walking distances and consequently connection
times are minimized. Distances in airports that are taken into account are those from
check-in to gates in the case of embarking or originating passengers, from gates to
check-out in the case of disembarking or destination passengers and from gate to gate
in the case of transfer or connecting passengers.
In the over-constrained case the number of aircrafts exceeds the number of avail-
able gates and the distance from the apron area to the terminal for aircraft assigned to
these areas is considered. Let us consider the following notations:
- N is the set of flights arriving at (and/or departing from) the airport;
- M is the set of gates available at the airport;
- n is the total number of flights;
- m is the total number of gates;
- ai is the arrival time of flight i;
- di is the departure time of flight i;
- wkl is the walking distance for passengers from gate k to gate l;
- fij is the number of passengers transferring from flight i to flight j.
Two dummy gates are used. Gate 0 represents the entrance or exit of the airport
and gate m + 1 represents the apron where flights arrive at when no gates are avail-
able. In order to model the AGAP as a quadratic program there were considered the
binary variable yik denoting that flight i is assigned to gate k if yik = 1 and yik = 0 oth-
erwise, where (0<k<m + 1).
The walking distance between gate k and the airport entrance or exit is wk0, f0i
represents the number of originating departure passengers of flight i and fi0 is the
number of the disembarking arrival passengers of flight i. The walking distance be-
tween the apron and gate k is wm+1k. The objectives refer to minimizing the number of
flights assigned to the apron and the total walking distance (the walking distance of
transfer passengers, disembarking arrival passengers and originating departure pas-
sengers). AGAP may be formulated as follows, see [7].
A Hybrid Ant-Based System for Gate Assignment Problem 275

Minimize the number of flights assigned to the apron:


n
minimize ∑y
i=1
im+1

Minimize the total walking distance:


n n m+1 m+1 n m +1 n m +1
minimize ∑∑∑∑
i =1 j =1 k =1 l =1
f ij wkl yik y jl + ∑∑ f0i w0l + ∑∑ f i0 wl0
i =1 l =1 i=1 l =1
(1)
m+1
s.t. ∑y
k =1
ik = 1, ∀ i, 1 ≤ i ≤ n

ai < di, ∀ i, 1 ≤ i ≤ n (2)

(dj − ai)(di − aj)yikyjk ≤ 0, ∀ i, j 1 ≤ i, j ≤ n, k ≠ m + 1 (3)

yik ∈{0, 1}, ∀i, 1 ≤ i ≤ n, ∀ k, 1 ≤ k ≤m + 1. (4)


Constraint (1) ensures that every flight must be assigned to one and only one gate
or assigned to the apron. Constraint (2) specifies that each flights departure time is
later than its arrival time. Constraint (3) says that two flights schedule cannot overlap
if they are assigned to the same gate. Condition (4) disallows any two flights to be
scheduled to the same gate simultaneously except if they are scheduled to the apron.
The above mentioned model is a 0-1 integer programming model with a quadratic
objective function. As in [7], the first step for the over-constrained AGAP is to mini-
mize the number of flights that need be assigned to the apron. The minimal number of
flights can be calculated by greedy algorithm described in the following. After sorting
all the flights by the departure time, flights are assigned one by one to the gates. Any
flight is assigned to an available gate with latest departure time. If there are no gates
available, the flight will be assigned to the apron.
Correctness of the greedy algorithm is proved in [7]. The greedy solution gives the
optimal number of flights that can be scheduled in gates and also helps us to get a
feasible initial solution (L+), which is used in the proposed metaheuristic.

3 Solving AGAP with an Hybrid Ant System


The idea of imitating the behaviour of the ants for finding good solutions to the com-
binatorial optimization problems was initiated by Dorigo, Maniezzo and Colorni [4].
The principle of these methods is based on the way ants search for food and find their
way back to the nest.
In the best ant system to date (Dorigo and Gambardella [3]), pheromone trails are
not only modified locally by the artificial ants during or just after the construction of a
new solution, but also globally, considering the best solution generated by all the ants
at a given iteration or even the bets solution ever constructed.
The introduced Hybrid Ant System (HAS-AGAP) uses the pheromone trails in a
non-standard way: in proceeding applications of ant systems pheromone trails were
exploited to build a completely new solution. In our paper pheromone trails are used
276 C.-M. Pintea et al.

to modify an existing solution, in the spirit of a neighbourhood search. After an artifi-


cial agent has modified a solution, taking into account only the information given by
the pheromone trail intensity, an improvement phase that consists in performing a fast
local search that takes into consideration just the objective function is applied.
Adding local search to ant systems has also been identified as very promising by
other researchers such as Dorigo and Gambardella [3] in the case of the TSP, where
they design an ant system almost as efficient as the best implementations of the Lin
and Kernighan heuristic by adding a simple 3-opt phase after the construction phase.
In HAS-AGAP each ant is associated with a problem solution that is first modified
using pheromone trail and later is improved using a local search mechanism.

3.1 Neighbourhood Search

A neighbourhood search move is the operation that maps one solution π to another
solution π1. The attractiveness of a move from π to π1 can be examined by calculating
the cost (objective function value) improvement of π1 compared to that of π. In our
neighbourhood search we used three types of neighbourhoods, which prevailed in
AGAP applications based on concepts from Xu et al. [10] and Ding et al. [7]:
- The Insert Move: move a single flight to a gate other than the one it currently as-
signs.
- The interval Exchange Move: Exchange two flights intervals in the current assign-
ment, where a flight interval consists of one or more consecutive flights in one gate.
- The Apron Exchange Move: Exchange one flight which has been assigned to the
apron with a flight that is assigned to a gate currently.
The elements of the algorithm based on ant colony system coupled with a local search
for solving the AGAP, are:
- Graph representation: the nodes of the graph are locations, and the goal of the ants is
to visit all the locations and match an activity to it, forming a bijective mapping from
activities to locations. At time t each edge (i,j) is labelled by a trail intensity τij(t).
- Initially the ants are randomly placed in the nodes of the graph.
- At each iteration an ant moves to a new node. When an ant decides which node is
the next move it does so with a probability based on the distance to that node and the
amount of trail intensity on the connecting edge. The inverse of distance to the next
node, is known as the visibility ηij.
- At each moment evaporation takes place to stop the intensity trails increasing un-
bounded. The values of rate evaporation ρ, are in the unit interval [0,1].
- A tabu list: ants memorizing locations that they have visited, so that they never visit
them again.
- A tabu list: ants memorizing the activities that have been mapped to the visited loca-
tions.
- In order to favour the selection of an edge that has a high pheromone value τ and
high visibility value η, a probability is considered. If we denote by Jki the unvisited
neighbours of node i by ant k and u ∈ Jki , then pkiu is defined as the probability of
choosing j = u as the next node if q > q0 (the current node is i) and is given by the
rule (see [4]):
A Hybrid Ant-Based System for Gate Assignment Problem 277

p =k [τ ][η ]
ij ij
β

u
∑ o∈ J ik
[τ io (t )][ηio (t )]β (5)

where β is a parameter used for tuning the relative importance of edge length in se-
lecting the next node. Consider a random variable q uniformly distributed over [0, 1]
and q0 with 0≤ q0 ≤ 1. If q ≤ q0 the next node j is chosen as:
j= arg max u∈J k {τ (t) [η (t)]β} iu iu (6)
i

- Manipulation of solutions: local search


Local search consists in applying the above mentioned neighbourhood search based
on three types of neighbourhoods with first improving strategy to a solution. We de-
note by S the solution space, V the neighbourhood structure containing the three types
of neighbourhoods and by c the cost function. The local search procedure is:

Neighbourhood search procedure


Select a starting solution s[0] from S
Repeat
Select s such that c(s)< c(s[0])
Replace s[0] by s
Until c(s)>c(s[0]) for all s from V(s[0])
s[0] is the approximation of the optimal solution.

This procedure examines all the possible swaps according to the considered
neighbours and immediately performs an improving swap, if one is found. The update
of the pheromone trails is done differently from the standard ant system where all the
ants update the pheromone trails with the result of their computations. The pheromone
trails are updated by taking into account only the best solution produced by the local
search. First, all the pheromone trails are weakened by setting
τij = ( 1-ρ ) · τij. (7)
A parameter ρ (0 < ρ < 1) that controls the evaporation of the pheromone trail, close to
0 implies that the pheromone trails remain active a long time, while a value of ρ close to
1 implies a high degree of evaporation. Then, the pheromone trails are reinforced by
considering only the best solution s* generated by the system so far and setting:
α
τ ij = τ ij +
* * (8)
c( s * )

4 Numerical Results
HAS-AGAP is compared with tabu search heuristic and with Ant Colony System [3].
For the comparison, a large set of problem instances is considered (see Tables 1-2),
with sizes between 100×16 (m×n) and 460×34 and used in the experiments [7]. For
small test data, (e.g. less than 25×5 as in [7]) when all the flights can be assigned to
278 C.-M. Pintea et al.

the gates, or using the apron for Tabu Search [9], ACS and HAS-AGAP found for all
problems optimal solutions, like using the Brute Force Method.
We considered as in [7] an airport with two parallel sets of terminals, where gates
are symmetrically located in the two terminals. The distance between the check-in /
check-out point to gate 1 (and gate 2) is 5 units. The distance between two adjacent
gates in one terminal (e.g., gate 1 and gate 3) is 1 unit and the distance between two
parallel gates in different terminals (e.g., gate 1 and gate 2) is 3 units.
It is assumed that the passengers can only walk horizontally or vertically. It is used
the Manhattan Distance. If one passenger wants to transfer from gate 3 to gate 2, the
walking distance is 4 units.
The arrival time ai of flight i is randomly generated in the interval [10i, 10i+7] and
the departure time di is generated in the interval [ai+60, ai+69]. There is considered, as
in real life, the total number of passengers in a flight within an interval [300,400].
The number of transfer passengers will increase if flight schedules are close, but not
too close. For each flight i, the number of disembarking passengers whose destination
is this airport fi0 and the number of embarking passengers whose origin is this airport
f0i are generated within the interval [1, 100].
The parameters chosen for the considered ant-based algorithms are: α=1, β=5,
ρ=0.01, τ0 = 0.01 and q0=0.9. The number of ants is considered the same with the
number of gates (m). The number of iterations is considered 2m ·2500. In our imple-
mentations we considered two test cases:

- Randomized large input: the test cases generated are categorized into 10 different
sizes with 10 cases in each size;
- Fully packed input: all the flights are fully packed (i.e. one flight is followed by an-
other and there are no gaps between the consecutive flights);
The computational results obtained using our algorithm HAS-AGAP, compared with
the Tabu Search (TS) [7] and Ant Colony System [3], are presented in the next tables.

Table 1. Test Set: randomized large input

Size TS-cost TS-time ACS-cost ACS-time HAS-cost HAS-time


100×16 101456 26.12 100927 5564.48 100782 5343.53
160×20 160345 64.23 160332 9247.50 158934 10762.48
220×24 215342 80.67 213987 11422.25 213243 16956.20
280×28 325412 86.24 324647 12965.20 324825 17043.87
340×32 365521 88.64 365866 13245.57 365784 18564.13
400×36 481256 125.36 480867 17453.15 480432 21130.52
460×40 498688 238.66 496334 17988.68 495487 17122.33
520×44 566325 425.66 562887 18966.22 563147 18425.40
580×48 674889 504.30 671225 20501.50 671322 19688.57
640×52 716455 702.36 710288 25457.32 709058 25036.28

We find that among these groups of test cases, HAS-AGAP performs better than TS and
ACS for AGAP and having almost the same running time in seconds as ACS, but longer
running time than TS (see the results on Tables 1-3). The computer used for numerical
tests: AMD 2600, 1.9 GHz, 512GB RAM. It is possible to obtain better values, using a
computer with higher performances or/and using better value for the parameters.
A Hybrid Ant-Based System for Gate Assignment Problem 279

Table 2. Test Set: fully packed input

Size TS-cost TS-time ACS-cost ACS-time HAS-cost HAS-time


100×16 221485 24.80 219432 4214.42 218865 4205.26
140×18 350426 32.96 347885 6210.24 346299 5833.45
180×20 522830 44.25 519887 14335.70 519758 13742.48
220×22 647239 71.12 648335 17567.80 647232 16956.20
260×24 871246 86.44 870843 18499.25 870916 18823.15
300×26 1015284 89.20 1013244 20430.18 1013428 20104.28
340×28 1143906 105.15 11344952 27211.26 1144823 26521.30
380×30 1434850 143.42 1434695 31218.36 1434507 31126.22
420×32 1572625 175.03 1572439 38846.52 1572483 36228.18
460×34 1951522 239.11 1949927 42185.63 1949912 40815.25

Table 3. Results of t-test

t-test Test Set: Test Set:


results randomized large input fully packed input
ACS-HAS TS-ACS TS-HAS ACS-HAS TS-ACS TS-HAS
t 2.31 2.81 3.23 1.87 2.32 2.73
probability 0.046 0.020 0.010 0.095 0.045 0.023

The t-test is performed [11]. The results of a paired t-test performed are shown in
Table 3. The degrees of freedom are 9 in all considered cases. The probability, assum-
ing the null hypothesis, is less than 0.05, proving the significant statistic difference
between all considered algorithms from Table 1, test set. The difference between
ACS and HAS for test set - fully packed input is not significant.
For test set - fully packed input, the probability, assuming the null hypothesis, is less
than 0.05, proving the significant statistic difference between TS–ACS and TS–HAS.

5 Conclusions
Nowadays results in the application of metaheuristic algorithms for providing good
solutions to difficult combinatorial optimization problems seem to indicate that com-
bining local optimization with good metaheuristics can give powerful optimization
algorithms. In this paper we present an ant colony system coupled with a simple local
search applied to the over-constrained airport gate assignment problem. Comparison
with tabu search heuristic and ant colony system is showing good results for the hy-
brid algorithm.

References
1. Babic, O., Teodorovic, D., Tosic, V.: Aircraft stand assignment to minimize walking.
Journal of Transportation Engineering 110, 55–66 (1984)
2. Braaksma, J., Shortreed, J.: Improving airport gate usage with critical path method.
Transportation Engineering J. of ASCE 97, 187–203 (1971)
280 C.-M. Pintea et al.

3. Dorigo, M., Gambardella, L.M.: Ant colony system: A cooperative learning approach to
the traveling salesman problem. IEEE Trans. on Evol. Comp. 1, 53–66 (1997)
4. Dorigo, M., Maniezzo, V., Colorni, A.: The ant system: Optimization by a colony of
cooperative agents. IEEE Trans. on Systems, Man, and Cybernetics- Part B 26, 29–41
(1991)
5. Cheng, Y.: Network-based simulation of aircraft at gates in airport terminals. J. of
Transportation Engineering, 188–196 (1998)
6. Cheng, Y.: A rule-based reactive model for the simulation of aircraft on airport gates.
Knowledge-based Systems 10, 225–236 (1998)
7. Ding, H., Lim, A., Rodrigues, B., Zhu, Y.: The over-constrained airport gate assignment
problem. Computers and Oper. Research 32, 1867–1880 (2005)
8. Gu, Y., Chung, C.A.: Genetic algorithm approach to aircraft gate reassignment problem. J.
of Transportation Engineering 125(5), 384–389 (1999)
9. Wang, Z., Shang, H., Ning, X.: Assignment algorithm for airport gate and its application.
J. of Nanjing University of Aeronautics and Astronautics 39(6), 819–823 (2007)
10. Xu, J., Baile, G.: The airport gate assignment problem: Mathematical model and a tabu
search algorithm. In: 34th Annual Hawaii Int. Conf. on System Sciences, vol. 3 (2001)
11. Yan, S., Chang, C.-M.: A network model for gate assignment. J. of Advanced
Transportation 32(2), 176–189 (1998)
12. http://www.physics.csbsju.edu/stats/t-test.html

You might also like