NATIONAL INSTITUTE OF TECHNOLOGY

KURUKSHETRA

SEMINAR REPORT
ON
A Dynamic Programming Approach for 4D Flight
Route Optimization

1
 Submitted by:
HIMANSHU ROHILLA
11610187
CO-4

Submitted To:
Prof. MANTOSH BISWAS
Department of Computer Engineering
NIT Kurukshetra

2
Abstract
The solution was aimed at optimizing flight routes so that the
overall cost depending on fuel is as low as possible. Several data
tables are taken as inputs, including aircraft positions and
destinations, weather information and other aviation related data.
The task was to produce a flight plan for each flight given a list of
(latitude, longitude, altitude, airspeed) quadruplets. The cost of
flight plan was evaluated with an open source simulator. Proposed
method produces an initial solution with the Dijkstra’s algorithm to
avoid restricted zones, and then refines it using dynamic
programming and local search techniques. Wind forecast are
extensively utilized and the plane is significantly diverted from the
great circle route if necessary. Moreover the method tries to set the
ascending and descending profiles of the flights to further decrease
the cost.

3
 INDEX
Serial Title of the section Page
Number Number
1. Introduction 5
2. Input data 8
3. Proposed solution 13
3.a Initial Route 14
3.b 2D optimization 16
3.c 1D optimization 20
3.d Implementation Details 22
4. Numerical results 23
5. Conclusion 24
6. References 25

4
Introduction
Economic growth makes air transportation more important every
day. The increasing number of flights leads to high complexity
scheduling and planning problems. Airlines are constantly seeking
ways to make the flights more efficient. From gate conflicts to
operational challenges to air traffic management, the dynamics of
a flight can change quickly and lead to costly delays.

Global airline industry fuel bill

All the information given below is from the source
Source: Industry Economics Performance - Forecast Table (IATA Economics)
Updated: 06/2018 Next Update: 12/2018
Link:
https://www.iata.org/pressroom/facts_figures/fact_sheets/Do
cuments/fact-sheet-fuel.pdf

1. The global airline industry’s fuel bill is estimated to total $149

billion in 2017 (accounting for over 21% of operating
expenses at $54.9/barrel Brent).

2. This is an increase of 10.3% over 2016’s fuel bill and is more

than 3 times 2003’s bill of $44 billion (that accounted for 14%
of operating expenses at $28.8/barrel Brent).

3. In 2018 the fuel bill is forecasted at $188 billion (accounting

for around 24% of operating expenses at $70 per barrel
Brent).

5
 4. Industry profits of $33.8 billion are forecasted for 2018,
following profits of $38.0 billion in 2017.

In a commercial airline, flight route planning is a central element of

the overall planning process. Small adjustments to the plans can
add up to substantial savings for the airline.
Input for the algorithm was obtained in several data tables,
containing current air-craft positions and destinations, weather
forecast information and other aviation related data. The input data
was given up to a “cut off” time on each test day. The task of the
contestants was to produce a flight plan for each flight, so that
average the cost of the plans is as low as possible.
The cost of the plans was evaluated with an open source simulator
and it depended on fuel consumption, delay and some other
factors. The contestants could only access forecast weather data
for creating their plans, but the final evaluation was based on live
weather data.

A flight plan was a list of (latitude, longitude, altitude, airspeed)

quadruples. This considers three space dimensions plus a time
dimensions, therefore we can call it a 4D flight plan. An example
flight plan for one plane is shown on Table I.

6
 TABLE I: Flight plan for one plane containing 5 instructions. The
altitude is given in feet the airspeed is given in knots.

Flight plans not only have to take the laws of physics into account
but also route restrictions from the air traffic control and aviation
authorities. This means that there are geographical zones in which
the planes must not fly, or there are specified altitudes for different
types of flights where they are allowed to cruise.
A good flight plan should
(1) find a short 2D path between the origin and a destination
with taking winds into account.
(2) find a proper cruising altitude and establish an efficient
altitude profile for ascent and descent.
(3) set the airspeed to keep a balance between traveling time
and fuel consumption.

7
Input Data
This section gives more details about the input data available for
the route planning algorithms.
1. Flight Data
 Status information about flights for that the route to the
destination had to be optimized.
 Data was provided for 14 days for approximately 1000
flights on each day.
 For each flight, the departure and arrival airport, the
current position and altitude, the departure and
scheduled arrival time, fuel, delay, turbulence costs and
some other, less important, information was given.

2. Airports
 Latitude, longitude and altitude of 63 airports where the
planes had to land.
 The airports and the positions of the planes at the cut off
time of an example day is shown in Fig. 1.

Fig. 1: Airports (blue) and plane positions at the cut off time of an example day (red).

8
3. Restricted Zones
 Planes were not allowed to fly into the restricted zones
of the airspace.
 Each zone was a convex polyhedron, given with the
coordinates of the vertices and the lower and upper
altitude bounds.
 Restricted zones were the same on every day. The
number of zones was 19.

4. Turbulent Zones
 Planes got some cost penalty, if they flew the turbulent
zones of the airspace.
 Each zone was a convex polyhedron, given with the
coordinates of the vertices and the lower and upper
altitude bounds.
 Turbulent zones were different on different days.
 The number of zones varied between 5 and 9 zones per
day.

5. Weather Data
 Wind data was provided as 2D vectors given hourly at
eight altitude levels on a 451X337 geographical grid.
 There were two kinds of wind data: live wind before the
cut off time, and forecast wind from the NWS SOO
Science and Training Resource Center for the entire
flying period.
 Some other weather data was also given for the airports:
ground conditions for the arrival, such as temperature,
wind and visibility, given in hourly resolution up to the cut
off time.

9
 The polygons of the restricted and turbulent zones for an
example day, and the wind vectors on the grid of the highest
altitude level is shown in Fig. 2

Fig. 2: Restricted zones (red) turbulent zones (yellow) and

wind vectors on the highest altitude level.

6. The Simulator
 As mentioned earlier, the objective function of FQ2 was
calculated by a simulator. The simulator was written in
F#, and it contained 2300 lines.

 The simulator consists of several submodules, including

fuel consumption models for ascending, cruising and
descending, a landing, weight, atmosphere and aircraft
model and an airspeed limiter. Some of the modules
implement physical laws, others are statistical models
with parameters fit to measurement data.

10
 The simulator computes a total cost value for all flight
plans loaded into it, and finally outputs one number, the
average total cost. The structure of the cost function for
one flight can be written in the following form:

Ctotal = Cfuel + Cdelay + Coscillation + Cturbulence

The fuel cost is proportional to the burned fuel, the

delay cost is the sum of a linear and a nonlinear term,
the turbulence cost is a linear function of the elapsed
time in turbulent zones and the oscillation costs
penalize if there are too many altitude changes during
the flight.

 Simulating a flight plan means a discrete simulation

with a time step of 6 minutes, following the instructions.

A flight can take three kind of steps: ascending,

descending and cruising.

During the descending steps, the fuel consumption is a

function of the altitude of the aircraft:
Fuel descent = f1(altitude).

During the ascending steps, the fuel consumption is a

function of the altitude and weight:
Fuel ascent = f2(altitude; weight).

During the cruise steps, the fuel consumption also

depends from the airspeed:

Fuel cruise = f3(altitude; weight; airspeed).

11
One more relevant characteristic of the simulator was
that the ground speed of the aircraft was calculated as
the function of airspeed and wind:

Speed ground = g(airspeed; wind).

12
Proposed Solution
We split the problem into three main parts. First, we create an initial
flight plan for the aircraft. Second, we perform a 2D optimization
process, to set the latitude and longitude coordinates of the
waypoints. Third, we set the altitudes and the airspeed of the flight
along the route obtained from the previous step. The 2D and the
1D profile of a flight plan is visualized in Fig. 3.

Fig. 3: The 2D shape, the altitude-distance and the velocity-distance graph of an example flight.

13
1. Initial Routes

For hitting any of the restricted zones of the airspace (e.g.

Edwards Air Force Base), the simulator gives a penalty of
100,000$ per plane, and in some cases the shortest (great
circle) path hits some of the zones. The first step of our
method is to create initial solutions avoiding these incidents.
Since the restricted zones are convex polyhedrons, we can
translate this problem into a simple geometry question: how
can we connect two points on a plane (current position,
destination) if we have to avoid a set of convex polygons? Of
course, to create good initial routes, we would like to find the
shortest path. This problem can be solved with Dijkstra’s well-
known algorithm.
The Dijkstra’s algorithm computes the shortest path from one
point to all others in an edge-weighted graph with non-
negative weights. In our case, the points of the graph are the
current position, the destination airport and the vertices of the
restricted zones. The weights are the distances between the
points. With the 2D paths given by the algorithm we can
guarantee that none of the planes will hit a restricted zone.
Creating an initial altitude and airspeed profile also, we are
already able to make a reasonable initial plan for all flights. In
Fig. 4, the Dijkstra’s path for an example plane is shown for
that the great circle path would hit two of the restricted zones.

14
Fig. 4: An initial 2D path with two additional waypoints created with Dijkstra’s algorithm.

15
2. 2D Optimization

In the cost model, the burned fuel is a function of the airspeed,

but the ground speed is the sum of the velocity relative to the
air and the wind vector. This means that for some flights a
wind-optimal path may be far from the great circle. Taking
advantages on the wind can significantly reduce the fuel cost,
and since the fuel consumption is directly proportional to the
elapsed time, it can also decrease the delay costs.
Turbulence cost can be also reduced in these cases. In Fig.
5, a 2D profile of a flight is shown, where the wind-optimal
path is far from the great circle and significantly reduces the
fuel, delay and turbulence costs.

Fig. 5: A wind optimal 2D path far from the great circle can
significantly reduce the fuel, delay and turbulence costs.

For the flight above, the blue route reduces the total cost with
nearly 15 percent. The core part of our solution was to explore

16
the airspace to discover such routes with a dynamic
programming-based approach.

The algorithm creates a grid in the airspace. It divides the

initial path into N equal parts with waypoints, and determined
n points perpendicular to the initial path after d miles in both
directions. We call these 2n+1 points “levels”, and we have N
levels if we call the current position level 0, and the destination
airport level N. We denote the grid points with Pij, where
1<=i<= N-1 and -n<=j<=n, and P00 is the current, and PN0 is
the destination point. N, n and d are parameters of the
algorithm. In Fig. 6, an example grid is shown with parameters
N= 6, n = 2, d = 50.

Fig. 6: Grid points for dynamic programming with parameters N = 6, n = 2, d = 50.

17
The original objective function cannot be evaluated for partial
routes, as the delay cost is not defined. Therefore, we created
one by omitting the delay term from the formula of the total
cost.

We recursively determined the cheapest route from the

current position to every point of the grid. Each point had three
labels:

Costij: The lowest possible cost to reach Pij from P00.

Previj: The index of the previous point at level i-1 in the best
route.

Propij: The status of the aircraft after reaching Pij in the best
route.

During the algorithm for all grid points pij we simulate the
routes from all the points pk in the previous level, with starting
properties Propk and register the best one (to make the
notation shorter, we use the single index k instead of (i 1; k)
for the points in the previous level). Apart from level 1 this is
the minimum of 2n+1 possible routes. At the end of the
procedure we obtain a new route that is not worse than the
initial one since the initial path was also present among the
ones we considered. The pseudo-code of our dynamic
programming-based route optimizer is shown on the next
page.

18
>Procedure to construct wind-optimal routes with dynamic
programming

1: function DYNAMIC 2D OPTIMIZATION (N, n, d)

2: Cost00 := 0
3: Prop00 := (initial position, initial weight, initial time)
4: for i = 1, N do
5: for j = -n, n do:

>Simulate path to Pij from all the points Pk of level i-1

and store the arrival properties

6: Costij,k := total cost across point Pk

7: Propij,k := (position, weight, time)
>Find the best path to Pij

8: Costij := mink Cij,k

9: Previj := k
10: Propij = Propij,k
11: end for
12: end for
13: return Optimized path
14: end function

After the dynamic programming phase, we further enhanced

the 2D route with local search procedures. This means that
we perturbed the waypoints of the routes in several directions
and kept the perturbation if it reduced the cost. We call this
part of our algorithm 2D refinement.

19
3. 1D Optimization

After we created 2D paths, the next step is to optimize the

altitudes (ascending, cruising and descending part of the
flight) and the speed profile. Since the oscillations (altitude
changes) are also highly penalized -100$ per extra oscillation.

we divided every flight into three parts: to an ascending, a

cruising and a descending part. Analyzing the cost models for
the different flight profiles we found that the most critical part
of the planning is to determine the starting point of the
descending phase. Since the 2D profile is fixed, this point can
be given with its distance from the destination airport.

Our initial flight profile was the following: if the flights were far
enough from the destination they ascended to the maximum
altitude level, cruised, and started descending. As we
described it in the Simulator subsection, the airspeed
instruction affects only in the cruising phase, so there was no
reason the decrease the airspeed from the maximum in the
ascending and descending phases. In some cases, it is worth
to decrease the airspeed during the cruising part: if the plane
is on time, which means the delay cost would be still zero even
with a lower cruise speed, it is possible to reduce the fuel
consumption by decreasing the speed.
This means that the parametrization of the 1D profile can be
done with two variables: the first one is the descending
distance and the second one is the cruise speed. For this 1D
optimization part we implemented an exhaustive search
method to optimize these two constants. For some of the
flights (an example for the optimized 1D profiles plotted on

20
Fig. 7) this improvement spared another 10 percent of the total
cost:

Fig. 7: Optimized altitude-distance and groundspeed-distance profiles for a plane arriving

on time.

During the 1D optimization process, we found an extra

possibility to reduce the cost. Due the artifacts of the
simulator, it was worthwhile to insert extra waypoints into the
ascending and descending phases of the plans just to
increase the density of waypoints in these phases. Of course,
this 1D refinement procedure is effective only in the FQ2
simulator, it would have no impact in real life.

21
4. Implementation Details

Bash scripts was used for automation, and Python for data
processing. FQ2 simulator was rewritten and dynamic
programming-based approach was implemented in MATLAB.
The flights on each day were divided into 4 parts, and the
optimization process was run on 14x4 = 56 cores. The
hardware used during the competition was a 64 core Linux
server with 1 terabyte memory.

22
Numerical Results

In this section, we compare the efficiency of different parts of our

solution. Table II shows objective function values for the solutions
created after each step.

TABLE II: Reduction of the average cost after different steps of our algorithm.

23
Conclusion
The main elements of our route optimization method are the
Dijkstra’s algorithm, dynamic programming, local and exhaustive
search procedures. Our method is able to produce reasonable
initial plans in a short time. To improve the initial solution the most
effective and time-consuming part was the dynamic programming
part. Most of our methods can be useful for real life flight route
optimization, since the simulator used in FQ2 was quite realistic in
many aspects.

24
References

1. “GE Flight Quest 1,” https://www.gequest.com/c/flight.

2. “GE Flight Quest 2,” https://www.gequest.com/c/flight2-final.
3. S. Altus, “Effective flight plans can help airlines economize,”
2009,
http://www.boeing.com/commercial/aeromagazine/articles/qt
r 03 09/article 08 1.html.
4. K. Bousson and P. Machado, “4d flight trajectory optimization
based on pseudospectral methods,” Engineering and
Technology, vol. 4, pp. 471–477, 2010.
5. D. Merle, “Flight path optimization for an airplane,” Master’s
thesis, Norwegian University of Science and Technology,
2011.
6. E. M. Atkins, “Flight Plan Management with George Jetson as
Pilot,” in AAAI Spring Symposium, 2004.
7. E. Dijkstra, “A note on two problems in connexion with
graphs,” Numerische Mathematik, vol. 1, pp. 269–271, 1959.

25