You are on page 1of 8




Krishan Kumar1
Invertis Institute of Management Studies, Bareilly, India (

Ravendra Singh2
Department of CSIT,IET, MJP Rohilkhand University, Bareilly, India (

Zubair Khan3
Department of Computer Science, IIET, Bareilly, India (

Over the next ten years, air traffic management in the world will change dramatically,
bringing both challenges and opportunities. The biggest of those challenges is the
continuing rise in traffic volumes and will increase the movement of airbuses around the
airport due to busy runways. This air traffic congestion will increase the running cost of
flights due to unwanted movement of airbuses around the airport and it will also increase
the chances of the airbus collision. This paper considers the problem of solving conflicts
arising among several aircraft that are assumed to move in a shared airspace. Aircraft can
not get closer to each other than a given safety distance in order to avoid possible
conflicts between different airplanes. For such system of multiple aircraft, we consider
the path planning problem among given waypoints avoiding all possible conflicts. In this
paper we have proposed Adaptive Resonance Theory Map Neural networks (ART1) to
produce optimize solution. Our result shows that the average expected conflict time of
aircrafts using ARTMAP has been increased to more than 20 minutes.

Keywords- collision detection, separation distance, ARTMAP neural networks, etc, stc

1 INTRODUCTION of nearly optimal resolutions that do not violate

As air traffic keeps increasing, En-route capacity separation constraints.
Part 2 of the paper introduces the problem solver,
becomes a serious problem. Aircraft conflict
resolution and resolution monitoring, are still done its constraints and goals. Modelling is discussed in
manually by controllers. Solutions to conflicts are part 3. Part 4 introduces Adaptive Resonance Theory
techniques and the coding of the problem. Part 5
empirical and, whereas aircraft are highly automated
and optimized systems, tools provided for ATC presents different examples of resolution of very
control are very basic, even out of date. complex test problems.
In part 5 we present the complete ATC simulator,
The need for an automatic problem solver is a
serious concern when addressing the issues of free conflict detector and cluster builder used to
flight. It is still very unclear how conflicts will be benchmark the problem solver on real traffic; we also
discuss weaknesses of the system and possible
solved in free flight airspace. In this paper, we present
an optimal solution. It builds optimal resolution for improvements.
complex conflicts and also computes a large number
Today as we know that Air Traffic is going to one maneuver per aircraft should be
increase day by day and creates traffic congestion, forecasted for the next twenty minutes.
delay of flights at airport and hence loss to air 4. If possible the conflicts must be solved
industry due to increase in running cost. However, horizontally for comfort and economical
continuing rise in traffic volume can be solved by reasons, especially when aircraft is leveled.
increasing the existing capacity and will require
investment in new automated systems and 3 MODELLING
infrastructure. Improving the current systems will
In this paper we will only give resolution orders
provide a short-term solution. Artificial Intelligence in the horizontal planes. It can be extended to 3D
techniques constitute an optimized methodology conflict resolution. Optimal command theory with
effective for solving discontinuous, non-convex, non-
state constraints developed by Bryson and Ho [9]
linear, or non-analytic problems. concludes the following results by us. For a conflict
resolution involving two aircraft: at the optimum, as
long as the standard separation constraint is not
2 CONFLICT RESOLUTION achieved, aircraft fly in a straight line. When
2.1 Earlier work done for the problem achieved, aircraft starts turning, and as soon as the
Conflict resolution is a very complex separation constraint is freed aircraft fly straight
mathematical problem involving trajectory again. This result can be easily extended to the case of
optimization and constraints handling. This problem n aircrafts, when n≥2. When moving only one
has two faces: find automatically solutions to the aircraft, it can be proved [10] that trajectories are
conflicts, and find the optimal solution regarding regular, they do not include discontinuous points and
conflicts. There have been many attempts to reach the minimum increase h of the trajectory length is
these objectives, automization and optimization. given by the following equation
However, most of the time two objectives are
confused. 1
AERA 3 [3], [4], [5] considers optimum results 2+h = +
in the “Gentle-Strict” function for a two aircraft 1+
conflict, but the “ maneuver Option Manger” only 2−d2
seeks after acceptable solutions and does not focus on
the optimum.
Karim Zehgal [6] with reactive techniques for 2 + 2x + x 2
collision avoidance gives a solution to the problem of ⎛ d ⎞
automation which is robust to disturbance, but 2⎜⎜1 + h − ⎟

completely disregards optimization. Furthermore the ⎝ 2 + 2h + h 2 − d 2 ⎠
modeling adopted implies a complete automation of
both on board and ground systems and requires speed 2
regulations which can not be handled by human pilots d (1 + d )( 2 + 2h + h 2− d + 1
and would probably be very difficult to apply to log
2 (1 + h − d )( 2 − d 2 + 1)
current aircraft engines without damaging them.
First approach to conflict resolution by genetic
algorithms was done by Alliot and Gruber [7], [8].
In this above equation d is the standard
2.2 Specification of the System separation divided by the initial distance to the
The main idea behind the solution of conflict trajectory cross, speeds are equal and normalized, the
detection is as close as possible to the current ATC trajectory angle is 85°. This equation can be
system. The solution has to handle the following generalized to different speeds and angles. Numerical
constraints: solutions shows that the length of the conflict free
1. Conflict free trajectories must respect both trajectory increases when
aircraft and pilot performances. 1. The angle of incidence between the two
2. Trajectories must take into account aircraft decreases
uncertainties in aircraft speed due to winds 2. The speed ratio gets close to 1
or turbulence. 3. Aircraft are getting closer to the conflict
3. Maneuvers orders must be given with an point
advance notice to the pilot. A maximum of
When only one of the two aircraft turns, it has space which strongly suggests that any method which
been shown that the turning point approximation requires exploring every connected component is NP.
(figure1) lengthens the optimal trajectory by less than It is important to note that this complexity is
1% if distance between the aircraft and the conflict independent of the modeling chosen. The offset
point is greater than two standard separations and the modeling seems to be very attractive, because it
angle of incidence between trajectories is greater than linearizes constraints. Nevertheless, each constraint
30 degrees. It can also be proved that that the offset multiplies by two the number of linear programs to
modelling (figure1), which moves an aircraft to put it solve. Our problem involves n(n-1)/2 constraints.
on a parallel route. The offset is thus very easy to Moreover, linearizing the minimized function,
compute, but separation constraints must be checked multiplies by 2n the number of linear programs to
during maneuvers and the complexity of the problem solve (we minimize the sum of each aircraft offset
remains. For n aircraft, 2n(n+1)/2 linear programs must which may be positive or negative). Finally, we will
be solved i.e. if n=4, 4096 linear programs must be have to solve 2n(n+1)/2 linear programs. Each one
solved . However this offset modeling is very useful involving n(n+1)/2 linear constraints. For n=4 we
to solve conflicts between overtaking aircraft. have to solve 1024 linear programs with 10
Both offset modeling and turning point modeling constraints in each program.
must be kept. For overtaking aircraft, offsets are more The model introduced above is simple enough to
efficient whereas for other conflicts, turning points be used in real time optimization program. Let us
are more efficient. However if we want to solve the consider a conflict involving n aircraft and let’s
very large conflict then we must try to start choose a time step (λ minutes for example). Let’s
maneuvers as late as possible with respect to the imagine we want to recompute all trajectories every λ
aircraft constraints. A maneuver will be determined minutes. During the optimization time, aircraft are
by: flying and must know if they change their route or
1. The maneuver starting time. not. Consequently, for each aircraft, at the beginning
2. The turning point. of the current optimization, trajectories are
3. The offset ending time determined by the previous run and can not be
4. The deviation angle changed for the next λ minutes; afterwards, an offset
can be shortened or turned into a turning point, a
turning point can become an offset, etc.

Using classical methods, such as gradient
methods for example, becomes useless for our
problem, because of arbitrary choice of the starting
point required by these methods. Each connected
component may contain one or several local optima,
and we can easily understand that the choice of the
starting point in one of these components can not lead
by classical method to an optimum in another
component. We can thus expect only a local
optimum. Practical attempts done on LANCELOT
(Large and Nonlinear Constrained Extended
Lagrangian Optimization Techniques [12] have
4 PROBLEM DESCRIPTION confirmed this problem, and highlighted others.
Convergence is very slow, particularly when
The problem’s complexity was exposed by
introducing a speed constraint. This approach is not
Medioni, Durand and Alliot [11]. Let us consider a
efficient for real time trajectory planning.
conflict between two aircraft. We can easily prove
Genetic algorithms were used to find a problem
that the minimized function is convex, but the set of
solver and hence to find the expected time of conflict
conflict free trajectories is not. It is not even
of different aircraft [13],[14],[15]. Genetic algorithms
connected. If trajectories don’t loop, the set of
were initially developed by John Holland [16] in the
conflict free trajectories has two connected
sixties. Genetic algorithms are stochastic optimization
components. For a conflict involving n aircrafts there
techniques that mimic natural evolution.
may be 2n connected components in the free trajectory
Here we are proposing an Adaptive Resonance input vector again sends its signals to the interface
Theory Neural networks [1] based solution to find the units, which again send this as the bottom-up signal to
expected time of conflict between two aircrafts and the F2 layer, and the competition is repeated (but
later to find the solution for multiple aircrafts by without the participation of any inhibited units).
using ART1.
The process continues until either a satisfactory match
is found (a candidate is accepted) or all units are
5.1 Algorithm inhibited. The action to be taken if all units are
A discussion of the choice of parameter values
inhibited must be specified by the user.
and initial weights follows the training algorithm. The
notation we use is as follows:
5.3 Coding our problem
Coding is done by using adaptive resonance theory
n- number of components in the input vector. neural network coded in java. Each time value
m- maximum number of clusters that can be represents is coded by a positive integer. The
deviation angle can be -30,-20,-10, 0, 10, 20, 30. At
bij-bottom-up weights (from F1(b) unit Xi to F2 each new step following data is given
unit Yj). 1. Duration of the optimization
tji- top-down weights (from F2 unit Yj to F1 unit
2. Anticipation time
Xi). 3. Heading of each aircraft
ρ- vigilance parameter. 4. Speed of each aircraft
s- binary input vector (an n-tuple).
5. Uncertainty of each aircraft
x-activation vector for F1 (b) layer (binary). 6. Position of each aircraft
║x║ norm of vector x, defined as the sum of the
7. Horizontal separation in nautical miles
components xi. (whereas standard separation is 4 nm)
Other global data is required by ARTMAP e.g.
5.2 Description number of iterations
A binary input vector s is presented to the F1 (a)
layer, and the signals are sent to the corresponding X
5.4 Computing the winning node
units. These F1 (b) units then broadcast to the F2
One of the main issues is to know how to find
layer over connection pathways with bottom-up out the winning node we have a poly criteria problem
weights. Each F2 unit computes its net input, and the to solve, in fact the following criteria have to be
units compete for the right to be active. The unit with
matched together to find a single fitness value
the largest net input sets its activation to 1; all others 1. The delay due to a deviation must be as
have an activation of 0.We shall denote the inbox of small as possible.
the unit as j. This winning unit becomes the candidate
2. However, the number of aircrafts deviated
to learn the input pattern. A signal is then sent down and the total number of maneuvers must be
from F2 to F1 (b) (multiplied by the top down as low as possible.
weights). The X units (in the interface portion of the 3. The maneuver duration for an aircraft must
F1 layer) remain “on” only if they receive nonzero be as short as possible so that the aircraft is
signals from both the F1 (a) and F2 units [figure (1)]. freed as soon as possible for another
The norm of the vector x (the activation vector
for the interface portion of F1) gives the number of 4. Trajectories must handle the separation
components in which the top-down weight vector for constraints.
the wining F2 unit tj and the input vector s are both 1. Instead of considering a global fitness value that
(This quantity is sometimes is referred to as the takes into account the different length engines of the
match.) If the ratio of ║x║ and ║s║ is greater than or
trajectories and the conflicts between the aircraft, we
equal to the vigilance parameter, the weights (top keep in a n2 sized matrix F (where n is the number of
down and bottom up) for the winning cluster unit are aircraft). If i ≠ j, Fi,j measures the conflict between the
aircraft i and j. it is set to zero if no conflict occurs
However, the ratio is less than the vigilance and increases with the seriousness of the conflict. Fi,i
parameter; the candidate unit is rejected, and another measures the lengthening of aircraft i trajectory. This
candidate unit must be chosen. The current winning
fitness matrix contains much more information than
cluster becomes inhibited, so that it cannot be chosen then previous scalar global fitness and this will allow
again as a candidate on this learning trial, and the
activations of the F1 units are reset to zero. The same
Figure 1: Basic Architecture of ART1

us to define more deterministic crossover and fitness does not take into account the delays induced
mutation operators[17]. by maneuvers.
At each time step t, we compute Ct,i,j as the
difference of the standard separation and the distance 5.4 Training Algorithm
between the segments i and j describing aircraft i and The training algorithm an ART1 net is presented
j position at time t. these values are added and give a next. A discussion of the role of the parameters and
measure of the conflict between i and j. So the fitness an appropriate choice of initial weights follows.
matrix is computed as follows: Step0. Initialize parameters:
L > 1, 0 < ρ ≤ 1
totaltime Initialize weights:
F i, j = ∑C t ,i , j 0 < bij(0) < L / L- 1 + n , tji(0) = 1
t =0 Step1. While stopping condition is false, do
Steps 2-13
It is obvious that the fitness matrix is Step2. For each training input, do steps 3-12.
symmetrical. A triangular matrix can also be used. Step3. Set activation of all F2 units to zero.
We can now define scalar fitness as follow: Set activations of F1(a) units to input
vector s.
1 Step4. Compute the norm of s:
∃(i, j ), i ≠ j , Fi , j ≠ 0 ⇒ F = ║s║= ∑ si
2 + ∑ Fi , j Step5. Send input signal from F1(a) to the F1(b)
layer: Xi = si.
This fitness function guarantees that if a calculated Step6. For each F2 node that is not inhibited:
value of winning matrix of winning node is larger If yj ≠ -1, then
than 1/2 , no conflict occurs. If a conflict remains the Yj = ∑ bij*xi.
Step7. While reset is true, do step 8-11. Now the algorithm works as follows in our case:
Step8. Find J(winning node) such that yJ ≥ yj for all
Initialize parameters:
nodes j.
L =50
If yj = -1, then all nodes are inhibited and
ρ=0.8 (high vigilance)
this pattern can not be clustered.
Initialize weights:
Step9. Recompute activation x of F1 (b):
bij(0) =0.2
xi = si*tji.
tji(0) = 1
Step10. Compute the norm of vector x:
We have taken 9 aircrafts in sky, but here we
║x║ = ∑ xi.
have taken only two aircrafts. Each aircraft is
Step11. Test for reset:
presented as an input pattern one by one in a sequence
If ║x║ / ║s║ < ρ, then yj = -1 (inhibited
and the top down weight matrix is updated. The flight
node J) (and continue executing Step 7 again)
sequence is defined by the following table1:
If ║x║ / ║s║ ≥ ρ,

Then proceed to Step 12.

Binary Input pattern Aircraft Number
Step12. Update the weights for node j (fast learning):
bij(new) = L*xi / L-1 + ║x║ , tji (new) = xi. 0001 S1
Step13. Test for stopping condition. 0010 S2
0011 S3
5.5 Parameters Used
0100 S4
n: number of components in the input vector. Used 0101 S5
as a sequence of flights.
0110 S6
m: maximum number of clusters that can be formed. 0111 S7
Used as a runway assignment.
1000 S8
bij: bottom-up weights (from F1(b) unit Xi to F2 unit
1001 S9
Yj).Used to store different clusters values. Permissible
range is given by
0 < bij(0) < L / (L – 1 + n) sample value 1 /( 1 + n). Table 1: Input Sequence Pattern as per Aircraft
tji: top-down weights (from F2 unit Yj to F1 unit
Xi).Used to store runways assignment for different
After the completion of one sequence pattern
different conflict time are calculated for that pattern.
ρ - vigilance parameter.(For deciding the learning That sequence pattern to be selected is on the basis of
node). conflict time that is calculated and is then compared
with the other value of conflict time and the
s - binary input vector (an n-tuple). Input array to
maximum one is selected.
store different input values.
x - activation vector for F1 (b) layer (binary). 6 CONCLUSION AND RESULTS
Output array to decide the learning node.
In the first application, we consider the
║x║ - norm of vector x, defined as the sum of the conflict between two aircraft described in figure2.The
components xi. anticipation is set to 2 minutes, aircraft speed (400
knots) are known with an error of 5 percent and
delay[] -array to store different delay values.
trajectory is forecast for 20 minutes. In this case, if
Resul[] -hash table to store different sequences delay aircraft1 speed is 380 knots and aircraft2 speed is 420
obtained. knots, because of uncertainty, at last no conflict
occurs the standard separation is 4nm).
etc[] -to store expected time of conflict of different Our main objective of finding conflict time
flights. greater than ½ between two aircrafts and hence
stc[] -to store schedule time of conflict. minimizing conflict by increasing the expected time
of conflict in minutes (table 2,figure 3) is successfully
Full algorithm is coded by using core java. achieved. The use of Adaptive Resonance Theory
(ART1) instead of Genetic Algorithm has drastically
reduced the conflict time and hence collision. Results
so obtained are more accurate and optimum. Generations Fitness value
5 0
10 0.296556555
15 0.665555555
This technique can be further implemented 20 0.554444444
to aircraft collision of more than two aircrafts control
25 0.554444444
of ATC to provide the optimum air traffic control at
the busiest airport. Also the first module, which we 30 0.554444444
implemented with this problem, can be combined 35 0.554444444
with second module by using adaptive resonance 40 0.554444444
theory (ART2) to get the best air traffic simulation 45 0.554444444
tool. Further work will concentrate in refining the 50 0.554444444
modeling and the global criteria to optimize, taking
into account for example take-off sequencing needs of
approach sectors or priority levels for slotted Table 2: Delay according to our proposed ART1

Figure 2: Two aircraft conflict



fitness value





0 10 20 30 40 50 60

Figure 3: Results according to our proposed ART1

[15] N. Durand, J.M. Alliot, and J. Noailles. Collision
REFERENCES avoidance using neural networks learned by genetic
algorithms. In Ninth International Conference on
[1] Gail A.Carpenter, Stephen Grossberg, Adaptive
Industrial and Engineering Applications of Artificial
Resonance Theory MAP.
Intelligence and Expert Systems, Fukuoka, 1996.
[2] Yang Yuying, Shi Xizhi, Li Guoyi, Rules from
[16] J.H Holland. Adaptation in Natural and Artificial
fuzzy adaptive resonance theory map, Progress in
Systems. University of Michigan press, 1975.
Natural Science May 1999, Vol.9 No.5, p.382.
[17] N. Durand, O. Chansou, and J.M. Alliot. An
[3] W.P.Niedrighaus, I.Frowlow, J.C.Corbin, a.h.
optimizing conflict solver for atc. ATC Quarterly,
Gisch, N.J.Taber, and f.H. Leiber. Automated enroute
Air Traffic Control Algorithmic Specifications: Flight
Plan conflict probe. Technical report, FAA, 1983.
[4] W.P.Niedrighaus, A mathematical formulation
for planning automated aircraft separation for
AERA3. technical report, FAA, 1989. DOT/FAA/DS-
[5] W.P.Niedrighaus, A mathematical formulation for
planning automated aircraft separation for AERA3.
technical report, FAA, 1989. DOT/FAA/DS-89/20.
[6] Karim zehgal. Vers Une theory de la coordinated
d’actions. Applications a’la navigation aerienne.
Ph.D. thesis, University Paris VI, 1994.
[7] Daniel Delahaye, Jean-Marc Alliot, Marc
Schenauer, and Jean-Loup Farges. Genetic algorithms
for partitioning air space. In proceedings of the Tenth
Confrence on Artificial Intelligence and Application.
IEEE, 1994.
[8] Jean-Marc Alliot, Herbe Gruber, and Marc
Schoenauer. Using genetic algorithms for solving
ATC conflicts. Inproceedings of the ninth Conference
on Artificial Intelligence and Application. IEEE,
[9] Bryson and Ho. Applied Optimal Control.
Hemisphere publishing Corporotation, New York
[10] Nocolas Durand. Conflict free trajectory
modeling for enroute control. Technical report,
ENAC/CENA, January 1994.
[11] Frédéric Médioni. Algorithmes génétiques et
programmation linéaire appliqués a la resolution de
conflits aériens. Master's thesis, Ecole Nationale de
l'Aviation Civile (ENAC), 1994.
[12] A.R. Conn, Nick Gould, and Ph. L. Toint. A
comprehensive description of LANCELOT.
Technical report, IBM T.J. Watson research center,
1992. Report 91/10.
[13] N. Durand, O. Chansou, and J.M. Alliot. An
optimizing conflict solver for atc. ATC Quarterly,
[14]Nicolas Durand, Nicolas Alech, Jean-Marc Alliot,
and Marc Schoenauer. Genetic algorithms for optimal
air traffic conflict resolution. In Proceedings of the
Second Singapore Conference on Intelligent Systems.
SPICIS, 1994.