e-journal: Information Technology

for Economics & Management

ISSN 1643-8949

ITEM, Volume 1, No 1, 2001, Paper 10

Flight Trajectory Optimization Using Genetic Algorithm Combined

with Gradient Method
Nobuhiro YOKOYAMA: School of Aeronautics and Astronautics,
the University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo 113, JAPAN
email: [email protected]
Shinji SUZUKI: School of Aeronautics and Astronautics, the University of Tokyo,
[email protected]
a nonlinear optimization problem by descritizing the
time functions such as control or state variables using a
set of discrete variables. Sophisticated algorithms with
gradient information for nonlinear optimization problems make it easier to deal with complicated equal or
unequal constrains. However, the gradient algorithms
require the delicate selection of initial guess of the optimal solution. This paper tries to combine non-gradient
techniques with the gradient methods to solve the complicated trajectory optimization methods without accurate initial solutions.

Abstaract
This paper considers the numerical method to solve trajectory optimization problems where ecient flight trajectories are searched to minimize a performance index
with specified constrained. While a wide variety of numerical methods have been developed, the direct application of a nonlinear programming method is frequently
used in the recent applications, which transforms the
original problem into a nonlinear optimization problem.
Although these methods can be applied to complicated
applications, it is very dicult to find reasonable initial
solutions since practical problems have high nonlinearlity. This paper tries to apply the Genetic Algorithm
(GA) method. Sine the GA approach is one of the random search methods, the initial solutions are selected
at random. As shown in numerical examples, the GA
approach has slow convergence characteristics and the
meaningless fluctuation in the solutions. Therefore, the
gradient approach is utilized to refine the solutions. As
the gradient method, the authors uses the BDH method
which approximates the state and control variables using linear interpolation and the collocation method is
used to satisfy the dynamic equations. The obtained
Hessian matrix can be approximated in a Block Diagonal Form which is suitable for ecient computation.
The proposed approach is applied to the optimal accent
trajectory optimization of a spaceplane. The spaceplane
is a future space transportation vehicle and its conceptual design is investigated.

As a non-gradient optimization algorithm, a Genetic Algorithms (GA) is employed, which uses search
procedures based on the mechanics of natural genetics, which combines a Darwinian survival-of-the-fittest
strategy to eliminate unfit characteristics and uses random information exchange(3) . While the original GA
codes continuous variables into a string of bit code, the
continuous variables can be dealt with by an advanced
continuous parameter GA(4) which defines a chromosome as an array of floating point numbers. Since
the GA approach is categorized as random search algorithm, initial solutions are selected at random. However, the convergence of the search process is usually
very slow.
This paper utilizes the continuos parameter GA
in order to select the initial solutions for the BDH
method(5) which is one of the gradient methods proposed by us. The characteristics of the GA approach to
the trajectory optimization are examined for a simple
problem with an analytic solution and the limitation
of this approach will be revealed. Additionally, as a
practical complicated example, an ascent trajectory optimization problem of a spaceplane is computed by the
presented method which uses the GA approach combined with the gradient method. The space plane is a
future space transportation vehicle which takes o horizontally, reaches to a space station with a single stage,
and lands horizontally(6) . In order to perform this mission, the vehicle changes several types of engine, e.g., an
air breathing engine and a rocket engine according to
its flight phase. The ecient accent trajectory design is
strongly desired to maximize payload or minimize consumed fuel.

Introduction

Trajectory optimization problems have been widely considered in aerospace engineering fields(1) . The problem
is formulated as the calculus of variations where an objective function such as a flight time or required fuel is
minimized while satisfying initial/final conditions and
path constrains for a flying vehicle. Although some
techniques such as the dynamic programming of Bellman and the maximum principle of Pontryagin have
been developed and used, the recent development of
computer is promoting new algorithms(2) . Some of
the new methods transform the original problem into
1

Trajectory Optimization Problem

Defining the state variables x(t) and the control variables u(t), the dynamic system is defined by the state
equations as
d
x(t) = f (x, u, t)
(1)
dt
The state and control variables are optimized to minimize the following performance index
J[x, u, t]

(2)

while satisfying the initial and final conditions defined

as
(3)
I (x(tI )) = 0, F (x(tF )) = 0
where tI and tF are initial and final time respectively.
In many cases, during a flight path some constrains are
introduced as
G(x, u) = 0, H(x, u) 0

(4)

The initial Np population of the independent variables is prepared for the continuous parameter GA at
random. The generation is updated from the following
process:
1) Selection of parents: m + 2 parents are initially selected at random.
2) Generation of children: the gravity point of the first
m + 1 parents is determined as xp . The deviation
vectors of each parents are defined as
xp =

m+1
1 X
xi
m+1 1

(10)

d i = xi xp

(11)

The normalized orthogonal base vectors to the deviation vector di , (i = 1, ..., m) are prepared as
e1 , ..., enm

(12)

where n is the number of the individual variables

p.
Using the deviation vectors and the base vectors,
A set of children is generated in the following equation;

GA Approach

The trajectory optimization problem stated in the preceding section can be converted to a nonlinear optimization problem. The time interval from tI to tF is divided
into N elements to transform the control variables u(t)
as a function of time to a set of discrete variables. By denoting the nodal values of time by ti (i = 1, 2, ...N + 1),
control variables are represented as a set of discrete values ui = u(ti ). If the initial states x(tI ) are specified in
the initial conditions, and if the final time tF is assumed
and the descritized control variables ui are assumed,
the state variables can be calculated by an appropriate integration scheme. In this paper, the control variables between nodal values of ui are interpolated using
a third order spline function, and the 4th order RungeKutta algorithm is used as the integration scheme. This
approximation makes it possible to represent the state
variables at time node ti as a function of descretized
control variables ui and the final time tF . Therefore,
the independent variables to be optimized are defined
as
pT = [uT1 , uT2 , ..., uTN +1 , tF ]
(5)
It should be noted that unknown initial states variables
can be incorporated into the independent variables.
The performance index J, the final conditions at
tF , and path constrains are defined as a function of the
independent variables p. The trajectory optimization
problem can be transformed into the following nonlinear
optimization problem:
minimize : J(p)
subject to : F (p) = 0
G(p) = 0
H(p) 0

(6)
(7)
(8)
(9)

xc = xp

m
X
i=1

wi di

nm
X

vi Dei

(13)

i=1

where D is the distance of the orthogonal component of the vector dm+2 to the d1 , ..., dm vectors,
and wi and vi are the random variables with normal deviates. Note thet the number of chirdren is
specified as 2 Nc . The above process is called
as UNDX-m method(7) which allows the uniform
search.
3) Selection: the fitness function of each parent and
generated child is calculated. The fitness function
can be generated by using the performance index
and the constraints as
X j
F = J+
[w1 max[0, | Gj (p) | ]2
j

+w2j

max[0, Hj (p)]2 ]

(14)

where w1j , and w2j are weighting values in the

penalty method which represents the constraints,
and is a small number.
The finest chromosome is selected. Another chromosome is chosen from the roulette selection which
determines a fine chromosome at random according
to the specified probability based on the ranking.
4) Determination of the new generation: According to
the MGG (minimum generation gap) method(7) ,
the selected two chromosome are exchanged with
the two selected at random from the old parents.
Hence the new parents are generated. If the finest
solution does not converge, return to the step 2).

Gradient Approach

While the many gradient approach methods have been

presented for the trajectory optimization problem, this
paper uses the BDH (5) method that is one of the direct
collocation methods. In the direct collocation method,
not only the control variables but also the state variables
are decritized. The BDH is using linear interpolation
for this descritization. In the similar way as the GA
approach, the time is divided into N elements from the
initial time tI to the final time tF . The state and control
variables are denoted by a set of nodal values as,
x1 , u1 , ...., xN +1 , uN +1

(15)

and the time interval of the element is given

ti = ti+1 ti

(16)

where X k is the vector X at kth iteration step, indicates the gradient of a function associated with the variables X, and dk is the update vector (= X k+1 X k ).
Vector G0 is composed from all the equal constraints.
Matrix B in Eq. (26) denotes the approximation of the
Hessian matrix H for the following Lagrange function
L:
(29)
L = J(X) + T G0 (X) + T H(X)
H = 2 L

where and are called the Lagrange variables

vectors. The SQP formulation assumes the initial
B to be a unit matrix and modifies the matrix B
through iteration steps by using gradient information. The Broyden-Fletcher-Goldfarb-Shanno (BFGS)
formulation(8) is used to update the matrix B in the
following way:

The variables X i (i = 1, 2, ..., N + 1) composed of

the discretized state and control variables and the time
interval are defined as
X i = [xTi , uTi , ti ]T

(17)

pk = X k+1 X k

xi + f (xi , ui )

ti
2

ti+1 Ki
]=0
2 Ki+1
Ki+1 ti Ki ti+1 = 0

H =

(21)
(22)

F (X) = 0
H(X) 0

(23)
(24)

5.1

0
0

2X1 X1 L
0

0
..

.
2XN +1 XN+1 L

(34)

Numerical Examples
Rocket Thrust Direction Control

(27)
(28)

where x and y are the inertia coordinates and u and v

are the particles velocity.
Lets consider the transfer of the particle to a specified height h with the zero vertical velocity in a fixed

subject to
=

(26)

(25)

Note that the Eq. (21) indicates the continuity condition of states at the mid point in time element ti .
In order to solve the nonlinear programming
problem, a sequential quadratic programming (SQP)(8)
method is employed. The SQP tries to solve the following quadratic programming problem in a sequential
manner:

G0 (X k ) + G0 (X k )dk
H 0 (X k ) + H 0 (X k )dk

(32)
T

To check the validity of the GA approach, a planar powered flight was investigated. A point mass m is guided
by controlling the direction angle of a constant thrust
with magnitude ma on a planar inertial coordinate system as shown in Fig.1.
The state equations are

0 0 1 0
0

0 0 0 1
d
0
x +
x=
(35)

0 0 0 0
a cos

dt

0 0 0 0
a sin

where Ki is a constant defined as,

1
minimize J(X k )dk + (dk )T Bdk
2

k+1

Since the Hessian matrix has a diagonal form, each

block in the matrix B is updated by BFGS formulation independently. The fundamental characteristics of
the BDH methods were examined in Reference (10).

[xi+1 + f (xi , ui )

Ki = ti /(tF tI )

k+1

) L(X , , )
(33)
The Hessian matrix for Eqs. (19)-(24) is derived

subject to
(20)

k+1

as

(19)

G(X) = 0

q = L(X

(18)

minimize J(X)

(31)

where
k

Using this approximation, the trajectory optimization problem is transformed into a nonlinear programming problem defined as,

B k pk (pk )T B k
q k (q k )T

k
T
k
(q ) p
(pk )T B k pk

B k+1 = B k +

The independent variables X are obtained as,

X = [X T1 , ..., X TN +1 ]T

(30)

x = [x, y u, v]T

(36)

given time T so as to maximize the final horizontal velocity. The initial and final conditions and performance
index for this problem are

0 1 0 0
h
F =
x(T )
=0
(37)
0 0 0 1
0
J = v(T )

(39)

Figure 2 indicates the transition of the finest performance index during the GA process. Figures 3,4 and
5 are the horizontal velocity, the vertical velocity, and
the trajectory, respectively. Those values are compared
with analytical solutions. The numerical solutions show
fluctuation around the analytical ones. This is remarkable in Fig. 6 showing the control variables, i.e., the
time history of the thrust direction. These results indicate that the GA approach has a great advantage in
the selection of initial solutions since those are selected
at random but has a disadvantage in the convergence
characteristics. It should be noted that the fluctuation in the GA method can be reduced if the number
of initial population increases and the huge number of
generations is allowed.
Figures 7 and 8 show the optimal solutions obtained from the BDH approach in which the initial solutions are obtained from the GA approach. These results
indicate the good accuracy of the BDH method.

5.2

take o weight m0
wing area Sw
ATR intake area
SCR intake area
RE thrust TRE
RE Isp

(38)

Dividing the time scale into N = 20 elements, the independent variables to be optimized are represented as a
set of i .
In this calculation, the constant values are a = 2,
h = 20, and T = 10, respectively. The following values
are selected in the GA method:
Np = 100, m = 4, Nc = 100

Table 1: Numerical Data

Space Plane Accent Trajectory Optimization

As a practical and complicated trajectory optimization

problem, an accent trajectory of a spaceplane (Fig. 9)
is investigated. The space plane is planned as a future space transportation vehicle which takes o horizontally, accents to the space station directly, and descends to the earth to land horizontally. In order to
perform this mission in a single stage, the plane is designed to equip with several types of engines to be selected according to the flight environment. A typical
accent flight path is shown in Fig. 10 where the space
plane rises above 90 km by using an air-turboramjet
engine ATR, a scramjet engine SCR, and a rocket engine RE sequentially. The RE is cut-o and the vehicle
zooms up to 400 km in the elliptic orbit. At the apogee
in the elliptic orbit, the vehicle is put on a circular orbit
as shown in Fig.11.
As shown in Fig. 12, the state variables are the
altitude h (= r R0 ), velocity v, flight-path angle and
weight m. The control variable is defined as the angle

300 [ton]
336.8 [m2 ]
17.96 [m2 ]
17.96 [m2 ]
233.6 [ton]
450 [sec]

of attack . Note that the engine holds the max thrust.

The state equations are defined as
dh
dt
dv
dt
d
dt
dm
dt

= v sin

(40)

T cos D
(41)
+ (2 r 2 ) sin
m
r

2r
T sin + L
v
=
+( 2 +
) cos
mv
r
vr
v
+2
(42)
T
(43)
=
Isp g0

where and g0 (= /R20 ) are the angular velocity of

the earth and the gravity constant at the earth surface.
Additionally, D and L are the lift and the drag. T and
Is p are the engine thrust and the specific impulse which
are determined according to the Mach number M as
M 6
6 M Ms
Ms M

:
:

ATR
SCR

(44)
(45)

RE

(46)

The aerodynamic characteristics of the spaceplane

are determined from Reference (6) and the ATR and
SCR engine characteristics are obtained from Reference
(10). The other data used are listed in the following
table.
Initial conditions are specified to consider take-o
conditions at time tI = 0 as
h(0) = 0 [km], v(0) = 150 [m/s]
(0) = 0 [deg], m(0) = 300 [ton]
d
(0) 0
dt

(47)
(48)
(49)

The Mach number Ms which indicates the switching

timing from SCR engine to RE engine is specified as
Ms 12

(50)

The terminal conditions at t = tF when the RE is cuto are specified as

h(tF )

(tF )
Ha [h(tF ), v(tF ), (tF )] =

90 [km]

(51)

0 [deg]
400 [km]

(52)
(53)

where Ha is the function which defines the apogee altitude on the elliptic orbit. Several path constraints must

be considered in the following items:

altitude :
dynamic pressure :
angle ofattack :
heating rate :

Reference

h 0 [km]
(54)
1 2
(55)
q = v 100 [kPa]
2
20 [deg]
(56)
8 3.07

Q = 4.4575 10
v
2

load factor :

300 [kW/m ]
L cos + D sin
4
mg0

(57)

v[h(tF ), v(tF ), (tF )]

)
g0 Isp

(59)

where v is the velocity change required to put the

veihcle on the circular orbit.
As the same way in the previous example, the GA
approach is applied to find initial solutions for the BDH
method. In the GA calculation, the following data are
used:
Np = 500, m = 10, Nc = 100
(60)
Figures 13-16 show the comparison between the
GA results after 10,000th generation and the BDH results in which the time scale is divided into N = 80
elements. Note that the initial solutions in the BDH
are obtained from the GA method. Figure 13, 14, and
15 are the time histories of the velocity, the altitude and
the control variables, respectively. It should be noted
that the BDH can refine the solutions. The final mass
increases from 75.93 ton to 80.34 ton and the terminal
time tF is reduced from 1387.8 sec to 835.9 sec. Figure
16 shows the v-h diagram which shows the relation the
velocity and the altitude. This clearly indicates the refinement of the trajectory obtained by the BDH which
strictly satisfies the constraints of the dynamic pressure
and the heating rate.

(2) Hargraves, C. R. and Paris, S. W., Direct Trajectory Optimization Using Nonlinear Programming
and Collocation, Journal of Guidance, Control,
and Dynamics, Vol. 10, No. 4 (1987), pp.338-342.

(58)

The objective function to be minimized is the final

mass at the circular orbit and is given as.
J = m(tF )exp(

(1) Bryson, A. E., Jr., and Ho, Y. C., Applied Optimal Control, Blaisdell Publishing Company, MA,
(1969).

Conclusion

The numerical method for the flight trajectory optimization problem is investigated in this paper. The
new approach is combined the Genetic Algorithm (GA)
method and the BDH method which is one of the gradient approach. The GA approach is used to find initial
solutions for the BDH method. While the GA approach
can select initial solution at random, it is dicult to obtain the smooth converged solution. On the other hand,
the BDH method has the good convergence characteristics if the good initial solutions are specified. The
proposing approach utilizes the good characteristic in
each method. The simple example which has an analytic solution is used to check the validity of the method,
and the complicated example of the accent flight trajectory optimization of a spaceplane is demonstrated to
show the applicability of the proposing approach.

(3) Holland, J., Adaptation and Natural and Artificial

System , Univ. of Michigan Press, Ann Arbor, MI,
1975.
(4) Wright, A. Genetic Algorithms for Real Parameter Optimization, in G.J.E. Rawlins (Ed.), Foundations of Genetic Algorithms, San Matero, CA:
Morgan Kaufmann, pp. 205-218.
(5) Tsuchiya, T., and Suzuki, S., Spaceplane Trajectory Optimization with Vehicle Size Analysis,
14th IFAC Symposium on Automatic Control in
Aerospace, 1998, pp.444-449.
(6) Nomura, S., Hozumi, K., Kawamoto, I., and
Miyamoto, Y., Experimental Studies on Aerodynamic Characteristics of SSTO Vehicle at Subsonic to Hypersonic Speeds, The 16th International Symposium on Space Technology and Science, (1988)pp.1547-1554.
(7) Kita, H., Ono, H., and Kobayashi, S, A New Generation Alternation Method of Genetic Algorithms
and Its Assessment (in Japanese), Journal of Artificial Inteligence, Vol. 12, No. 5 (1997), pp.734744.
(8) Ibaragi, H. and Fukushima, M., FORTRAN77
Optimization Programming, Iwanami Shoten,
Tokyo, (1991)
(9) Tsuchiya, T., and Suzuki, S., Computational
method of Optimal Control Problem Using Mathematical Programming (2nd Report), Introduction
of Block Hessian Method (in Japanese), Journal of
Japan Society for Aeronautical and Space Sciences,
Vol. 46, No.536 (1998), pp.497-503.
(10) Kato, K., Spaceplane (in Japanese), University
of Tokyo, Tokyo, (1989).

Figure 1: Coordinate system

Figure 5: Flight trajectory

Figure 2: History of fitness function vs generation

Figure 6: Time history of control

Figure 3: Time history of horizontal velocity

Figure 7: Flight trajectory

Figure 4: Time History of vertical velocity

Figure 8: Time history of control

Figure 11: Orbit transfer of spaceplane

Figure 9: Spaceplane

Figure 12: State and control variables

Figure 10: Spaceplane accent trajectory

Figure 13: Time history of spaceplane altitude

Figure 14: Time history of spaceplane velocity

Figure 15: Time history of spaceplane control

Figure 16: H-V diagram of spaceplane accent trajectory