OTS Talk Topputo PDF

An overview of optimal control
methods with applications

F. Topputo and P. Di Lizia!
Dinamica Srl
DINAMICA srl!
Registered office

Rembrandt,
69 !
Piazza della Repubblica, 10 Via
- 20121
- Milan (Italy)

20147 - Milano
(Italy)!
Head office

Phone
02- 40091584!
Via Morghen,
13 - +39
20158
Milan (Italy)

Phone:
+39-02-8342-2930

Fax +39
02 4816277 !
Fax: +39-02-3206-6679

e-mail: dinamica@dinamicatech.com
e-mail: dinamica@dinamicatech.com

website:
www.dinamicatech.com
website: www.dinamicatech.com
Stardust - Opening Training School!

University of Strathclyde!
Glasgow, UK!
18-22 November 2013
The Company
Founded by 5 partners in January 2008

All partners have a PhD in Aerospace Engineering
Dinamica Srl has a strong connection with Academia
More than 30 years of accumulated space experience
Dinamica Srl is located in Milano
?D7C?97
INNOVATING TECHNOLOGY
Keywords
Analysis
Design
System engineering
Mission analysis
Simulation
Control
Optimization
Optimization
Autonomy
The mission
Italian SME, founded in 2008
The mission: ... to carry on developing methods and
advanced solutions within the Space field and to
transfer their use in other industrial sectors ...
?D7C?97
Space field
Technology transfer
Industrial sector
A Tangible Example
?D7C?97
Hubble
Used to reconstruct
unmodeled structural
dynamics
Used to reduce
vibrations in large
telescope satellites
Pharmaceutical industry
System
identification
Used to reconstruct a
pharmaceutical process
dynamics
Predictive
control
Used to minimize the

energy supply
Considerable savings (10-20%) compared to standard methods

7
Optimal Control Problem (1/2)
Consider the following dynamical system:!
!
where:
is the state vector and
is the control vector
Determine the m control functions such that the following

performance index is minimized:
where the initial and final state vectors,

and
the final time
, are not necessarily fixed
, as well as
Optimal Control Problem (2/2)
In addition to the previous statements suppose that the

following constraints are imposed!
Boundary conditions at final time tf : !

, where !
Path constraints on the control variables:!

, where
Two classical solution methods:!
Indirect methods: based on reducing the optimal control

problem to a Boundary Value Problem (BVP) !
Direct methods: based on reducing the optimal control

problem to a nonlinear programming problem
Example: Low-Thrust Earth-Mars Transfer
Given the dynamics of the controlled 2 body problem:
r+u
3
r
r =
Minimize:
Subject to equality constraints:
r(t0 ) = rE (t0 ) v(t0 ) = vE (t0 )

r(tf ) = rM (tf ) v(tf ) = vM (tf )
v1
v2
rM
vM
and the inequality

constraints
max
vE
rE
Indirect Methods (1/6)
Reconsider the optimal control problem:
Given the dynamical system!
Minimize:!
Subject to:
and
Constraints are added to the performance index J by

introducing two kinds of Lagrange multipliers:!
p -dimensional vector of constants
for the final constraints!
two n- and a q -dimensional vectors of functions and

(adjoint or costate variables) for dynamics and path constraints
Augmented performance index:
The dynamics is included in the augmented performance

index as a constraint
Moreover, pertaining the costate variables for the path

inequality constraints , the generic component k must
satisfy the following relations:
Integrating by parts the term
where
yields:
and
The problem is then reduced to identify a stationary point of

This is achieved by imposing the gradient to be zero. The
optimization variables are:!
State vector
and control vector!
Lagrange multipliers and costate variables ,
Unknown components of the initial state
,!
Final state and time
and
and!

Dynamics
Constraints
The problem consists on finding the functions

,
and
by solving the differential-algebraic system:
differential
Euler-Lagrange
equations
algebraic
Note: For the sake of a more compact notation, define the Hamiltonian
The previous equations read:
The previous differential-algebraic system must be coupled to

the 2 n boundary conditions
given or
and to the p + q + 1 additional constraints
The optimal control problem is reduced to a boundary value

problem on a differential-algebraic system of equations (DAE)
Simple
Optimal
subject
to system
equationsControl
Problem
J = x2 (1)
timal control problem

with
HermiteSimpson
and certain
x 1 = 0.5x1 +method,
u
subjectistosolved
system
equations
e computation time and improving accuracy. Consider the problem of finding
(
5 2
2
nal time [0, 1] to minimize the cost
x 2 = u + x1 u +x 1 =
x 0.5x1 + u
Example 1: A Simple Optimal
Control
4 1 Problem
5 2
2
J
=
x
(1)
(68)
x
=
u
+
x
u
+
x
2
2
1
to theaboundary
conditions
sand
section,
simple optimal
control problem is solved with HermiteSimpson
4 1 method, and certai
Given the simple optimal control problem (Problem #1)
are set to speed-up the computation time and ximproving
accuracy. Consider the problem of findin
and to the boundary conditions 1 (0) = 1
(
with fixed initial
the cost
! and final time [0, 1] to minimize
x
(0)
=
0
2
x 1 = 0.5x1 + u
x1 (0) = 1
!
(69)
J = x2 (1)
(68
5of 2state and control
2
with the following
bound
box
variables
x
(0)
=
0
x 2 = u + x1 u + x1
2
4
!
ct to system equations
10 of
state
x1 (t) and
10control variables
(
with the following bound box
ons
Assignment #1
! Dynamics
init., final time
Obj.
fcn.
b. c.
1 x+
2 (t)
x 1 =10
0.5x
u 10
10 x1 (t) 10
! x1 (0) = 1
10
u(t)
10
5 2
(70)
2
10
x
(t)
10
x
=
u
+
x
u
+
x
2
x2 (0) = 0
2
1
1
4
Write the necessary conditions for10
optimality
u(t) 10and show that
of state
and the
control
variablessolution is
optimal
oxThe
the
boundary
conditions
analytical
solution
of this problem is
(
(69(
10 x1 (t) 10
(71)
x1 (0)
cosh(1
=
t) 1 is
The analytical solution
of
this
problem
(70
Optimal

10 x2 (t) 10 x1 (t) = cosh(1)
(72)
x2 (0) = 0
(
cosh(1
t) t)(73)
solution
10 u(t) 10
(tanh(1
t)
+
0.5)
cosh(1
x1 (t) =
the following bound box of state u(t)
and =
control variables
cosh(1)
(74)
cosh(1)
(tanh(1 t) + 0.5) cosh(1 t)
u(t)
=
10 x1 (t) 10
(71
his
problem
is
cosh(1)
3.1
Solution
by HermiteSimpson
method
10 x2 (t) 10
(72
cosh(1
t)
Thisx1problem
is solved by using the MATLAB function fmincon with interior-point algorithm. T
(t) =
u(t) 10
(73
3.1 Solution by10
HermiteSimpson
method
Low-Thrust Transfer to Halo Orbit (1/4)
Transfer the s/c from a given orbit (GTO raising) to a Halo orbit
around L1 of the Earth-Moon system
Dynamics:
In canonical form, the dynamics reads:
Constraints: fixed
Euler-Lagrange equations:
Performance index: minimize the quadratic functional
and
, fixed final time
Processing the last algebraic equation leads to:
which can be inserted in the differential equations

The DAE system is reduced to a ODE system
All constraints simply reduce to:
The original problem is reduced to a simple Two Point

Boundary Value Problem (TPBVP)
Solution of the TPBVP:!
Transcribe the dynamics! (x,
Evaluate the control parameters
Couple the transcribed dynamics with the constraints on
and!
Solve the resulting system with a Newton method starting from a

suitable initial condition!
(u0 , ... , uN )
manifolds that act as separatrices for the states of motion
III.
Low-Thrust Propulsion and Attainable
End-to-end Tooptimization
finite
model the controlledw/
motion
of P thrust
under the gra
3
attractions of P1 , P2 , and the low-thrust propulsion, the

differential equations are considered:
GTO-to-halo fully optimized!
Switching fcn
Control
very difficult problem!

tens of spirals!
thrust saturation
Ty TO
@!3 Tx
@!MINGOTTI,
3
% ;
y" % 2x_ #
%
x" $ 2y_ #
@x
m
@y
m
B. Optimal Control Problem Denition
@!3 Tz
T
_ # $ the guidance law
z" #control
% aims
; at nding
m
The optimal
Isp g0
@z
m
! 2 #t ; t $, that minimizes
0
q!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Z !t
2
2
2 f T!!"
where T # Tx % TyJ %
% Tz is thed! present thrust m
t0 Ispacecraft
sp g0
Continuous variations of the
mass m are t
account through the last of Eqs. (7). This causes a singular
Itwhen
is easymto!
verify
through
the
last ofsolution
Eqs.singularities
(11) that J is given
the prob
Optimal
0 beside
the
well-known
mass;
is, J % m0 & mf , where m0 and mf are the initial an
of Pthat
3 with P1 or P2 (r1;2 ! 0). The guidance la
spacecraft masses. The thrust magnitude must not exc
!Tx !t"; Ty !t"; Tz !t"", t 2 't0 ; tf (, in Eqs. (7) is not given; r
maximum
threshold given by technological constraints. T
found through an optimal control step where objective fun
imposed along the whole transfer through
boundary conditions are specied (see Sec. IV). However,
construct a rst-guess solution,
the prole of T ove
T!t" ' Tmax
prescribed at this stage. Using tangential thrust, attainable
dened
in the
fashion asavailable
reachablethrust.
sets dened
in [3
where
Tmax
is same
the maximum
In additio
!x0constraints
; t0 ; t" be the
of system
(7)impacts
at time wt
Let #T!$"
following
path
are ow
imposed
to avoid
Earth
and moon
guidance
law along
T!$",the$ transfer:
2 't0 ; t(, and starting from !x0
_ 0 ; y_ 0 ; z_0 ; m0 ". p
The
generic point of a !
!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
px!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
0 # !x0 ; y0 ; z0 ; x
y2 ( z2 >
!x ( " & 1"2 ( y2 ( z2
!x ( ""2 (
low-thrust
trajectory
is RE ;
J = x2 (1)

with
HermiteSimpson method, and certain
subjectistosolved
system
equations
x 1 = 0.5x1 + u
Assignment #2
Example 1: A Simple Optimal Control Problem
5
J = x2 (1)
x 2 = u2 + x1 u(68)
+ x21
s section, a Given
simple optimal
control
method, and certai
4
Problem
#1 problem is solved with HermiteSimpson
are set to speed-up the computation time and improving accuracy. Consider the problem of findin
to the boundary conditions
! and
with fixed initial
and final time [0, 1] to minimize the cost
x1 (0) = 1
!x 1 = 0.5x1 + u
(69)
J
=
x
(1)
(68
5 2
2
2
x2 (0) = 0
!x 2 = u + x1 u + 4 x1
ons
! with the following bound box of state and control variables
init., final time

Obj.1 +
fcn.
b. c.
! Dynamics
x 1 = 0.5x
u
!
10 x1 (t) 10
x1 (0) = 1
!
5 2
(70)
2
!
10
x
(t)
10
x
=
u
+
x
u
+
x
2
x
(0)
=
0
2
1
1
2
!
4
!
10 u(t) 10
x
of
state
and
control
variables
o the boundary conditions
Solve the TPBVP associated
10 x1 (t) 10
(71)
x
(0)
=
1
1
The
Matlab
built-in
bvp4c,
bvp5c
analytical solution of this problem is
10 x2 (t) 10
(72)
x2 (0) = 0
10Sixth-order
cosh(1 t)
u(t) 10 method bvp_h6
(73)
x1 (t) =
the following bound
box of state
and control variables cosh(1)
(74)
available
at http://www.astrodynamics.eu/Astrodynamics.eu/Software.html
(69
(a) discrete solution and analytical time history of x1
0.9
0.8
0.7
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
1a
x1
0.9
(b) discrete solution and analytical time history of u
cosh(1 t)
x1 (t) =
3.1 Solution
cosh(1)
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
(c) error of x1 and u located at discrete points
0
(tanh(1
t) + 0.5) cosh(1 t)
5
u(t)
=
10 x1 (t) 10
10
cosh(1)
15
10 x2 (t) 10 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
u(t) 10
by10
HermiteSimpson
method
(d)
reintegrate dynamics with discrete control solution
r))
his problem is
log10(abs(error))
(70
ua
x1 error
u error
0.9
(71
(72
(73
Indirect Methods: Remarks
Main Difficulties:!
Deriving Euler-Lagrange equations and transversality

conditions for the problem at hand!
Nonlinearity of the dynamics!
Lack of a plain physical meaning of Lagrange multipliers
Solution of the DAE system itself!

Solution of the boundary value problem on the DAE
system!
difficulty at identifying good first guesses for Lagrange

multipliers (primer vector theory)
Optimal Control Problem
Given a dynamical system:!
Determine u(t) which minimize the performance index:
and satisfy the constraints:
Two classical solution methods:!
Indirect methods: based on reducing the optimal control

problem to a Boundary Value Problem (BVP) !
Direct methods: based on reducing the optimal control

Nonlinear Programming Problem
Generally constrained optimization problem
Given a function!
f (x) = f (x1 , x2 , ..., xv )
Minimize:!
Subject to K equality constraints:!
f (x)
ck (x) = 0, k = 1, ..., K
and J inequality constraints:!
gj (x)
!
where
can exceed
0, j = 1, ..., J
(K
v)
Unconstrained Optimization (1/6)
Minimize: f (x)
The necessary condition for the identification of the
optimum is:
x
f =0
The optimization problem in the v variables x is reduced

to the solution of a system of v nonlinear equations!
Note: given the Hessian of f , Hf , a sufficient condition is:
x Hf x > 0,
The solution can be found using the Newton method
Newton Method (1/3)
Consider the problem:
F (x) = 0
The Newton method is an iterative method based on a

linearization of F around the current iterate!
1. Select an initial guess! x

2. Consider the first order approximation of! F
F (x) F (
x) + F (
x) (x
x
) = 0
3. Find the correction:!
x = (x
x
) =
[F (
x)]
F (
x)
4. Update current iterate and repeat from 2 until convergence
Newton Method (2/3)
Graphical interpretation:
F
x
x2 x1
Since it is based on a first

order approximation of F
the method is local
Different first guesses might
lead to different solutions
x0
x
x
x1,1 x0,1 x0,3 x0,2 x1,2
Newton Method (3/3)
Classical methods to stabilize the iteration!
Line Search:!
Instead of updating the current iterate using!
xnew = x +
Reduce the step size using a parameter
xnew = x +
!
where
is chosen such that!
:!
||F (
xnew )||
||F (
x)||
Trust region:!
The direction of the computed
is slightly modified
Solve:
f =0
Newton algorithm:!
Select an initial guess! x
While stopping criterion is not satisfied!
Find the corrections

the linear system!
!
where
Hf
x to the current solution by solving

x=
x f
! f =
Hf is the Hessian of f : H
Update the current solution:
x+
2
x

Important note:!
Consider the following optimization problem!
Minimize the quadratic form:!
!
!
1
xT Hf x +
2
T
f
x
x
Necessary optimality conditions:!
Hf
x+
f =0
which can be written as:
Hf
x=
x f
Finding the corrections x , i.e. the search direction, in the

original optimization problem is equivalent to minimizing the
previous quadratic form
Given the function to be minimized, f (x) , each iteration

of the Newton method is equivalent to:!
Approximate f around the current solution

quadratic form!
Find the offset,

quadratic form!
Use
x with a
x , to the zero-gradient point of the
as a correction in the original optimization problem

obj
x
x0
x1
x2 ...
f (x1 , x2 ) =
1 2
(x1 + x22 )
2
example1.m
f (x1 , x2 ) = 5 x1 + e
(x1 +5)
+ x22
example2.m
example3.m
Assignment #3
Numerically re-compute the three unconstrained
optimizations in the previous slide; i.e.,
!
!
!
1 2
2
!
f
(x
,
x
)
=
(x
+
x
1
2
1
2)
!
2
!
!
!
(x1 +5)
2
!
f
(x
,
x
)
=
5
x
+
e
+
x
1
2
1
2
!
!
!
!
!
Advice
Use Matlab built-in fminunc
Code a SQP solver
Earth-Mars 2-impulse Transfer (1/3)
Optimization variables: departure date t0 and time of flight ttof
Compute the positions of the starting and arrival planets

through the ephemerides evaluation:
(rE , vE ) = eph(t0 , Earth), (rM , vM ) = eph(t0 + ttof , M ars)
Solve the Lamberts problem to evaluate the escape

velocity v1 and the arrival one v2
Objective function:
V =
V1 +
V2
Minimize: f (x) =
Necessary conditions for optimality:

V
V
=0
=0
ttof
t0
In a generic iteration, given the current estimate

x = {t0 , ttof }, evaluate the corrections x = { t0 ,
x=
Hf
x f
where Hf is:
2
2t
0
2
V
ttof t0
V (x) =
t0 ttof
2
2t
tof
V (t0 , ttof )
x
f =0
ttof }:
Search space:
t0 [0, 1460] M JD2000 = 4 years
ttof [100, 600] day
objective function
exampleEM.m
Equality Constrained Optimization (1/5)
Minimize:!
f (x)
Subject to:
ck (x) = 0, k = 1, ..., K
(K
The classical approach to the solution of the previous

problem is based on the method of Lagrange multipliers
Method of Lagrange multipliers:
v)
Introduce the Lagrange function:
L(x, ) = f (x)
where L is a function of the
Lagrange multipliers
c(x)
variables
and the K
The necessary conditions for the identification of the

optimum are:
x L(x, ) = x f (x) CT (x)

L(x, ) = c(x) = 0
where
=0
C(x) is the Jacobian of c(x)
The constrained optimization problem in the v variables

x has been reduced to the solution of a system of v + K
equations in the v + K variables (x, )
Solution by Newton method

Algorithm:!
Select an initial guess!(x,
While stopping criterion is not satisfied!
Find the corrections ( x,

solving the linear system!
! HL
! C
C
0
) to the current solution by

=
x f
c
Karush-KuhnTucker (KKT)
where ! HL = 2xx f
k=1
Update the current solution:
2
k xx ck
(x, )
(x +
x, +

Important note:!
Consider the following optimization problem:!
Minimize the quadratic form:!
!
!
1
xT HL x + (
2
Subject to the linear constraints:
C x=
T
f
)
x
x
Use the approach of Lagrange multipliers!
Lagrange function:
1
xT HL x + (x f )T x
2
(C x + c)
Necessary optimality conditions:
HL x + x f
C x+c=0
which can be written as:
HL
C
C
0
CT
x f
c
=0
Karush-KuhnTucker (KKT)
Finding the corrections ( x,

) , i.e. the search direction,
in the original optimization problem using the KKT system is
equivalent to minimizing the previous quadratic form
Inequality Constrained Optimization

Minimize:!
f (x)
Subject to:
gj (x)
0, j = 1, ..., J
A possible approach!
Interior Point Method:
The inequality constraints are added to the objective function

as a penalty term
f(x) = f (x) +
j e
gj (x)
j=1
Solution is forced to move into the set of feasible solutions by

means of the barrier function e gj
Problem Set #3 (With Corrections)
Assignment #4
1. Consider the following NLP problem:
Solve the NLP problem(*)
!!
!!
!!
!!
min3 f (x) := x21 + x1 x2 + 2x22 6x1 2x2 12x3
xIR
s.t. g1 (x) := 2x21 + x22 15

g2 (x) := x1 2x2 x3 3
x1 , x2 , x3 0.
Using Matlab built-in fmincon

Use SQP algorithm, quasi-Newton
update, and line search
(a) Find an optimal solution x to this problem using the function fmincon in MATLA
Set Toolbox:
solution tol, function tol, constraint tol to 1e-7
Make
that to
thex0medium
scale
Specify
initialsure
guess
= (1, 1,
1) SQP algorithm, with Quasi-Newton update
the selected solver in fmincon;
Make sure g2 is treated as linear constraint by fmincon
Set the solution point tolerance, function tolerance and constraint tolerance in
"T
Solvew/o
providing gradient of obj !fcn and constraints;
then re-do
Specify the initial guess as x0 = 1 1 1
;
by providing
analytic gradients
Make sure that the inequality constraint g2 is treated as a linear constraint b
Repeat
optimization
from
x0; doapproximation
we find the ofsame
Solve
the NLP by
usinga adifferent
finite-dierence
the gradients of th
constraints
first.
Then, resolve
the problem by providing explicit expressio
optimal and
solution
found
previously?
Why?
(*)
of the objective function and constraints (ask fmincon to check for the gradi
the optimization in this latter case);
Make sure that the solver terminated successfully in each case.
B. Chachuat, Nonlinear and Dynamic Optimization, EPFL
Direct Methods
Direct methods are based on reducing the optimal control

The core of the reduction of the optimal control problem to

a nonlinear programming problem is:!
The parameterization of all continuous variables!

The transcription of the differential equations describing the
dynamics, into a finite set of equality constraints
Classical transcription methods:!
Collocation !
Multiple Shooting
The original optimal control problem is solved within the accuracy

of the parameterization and the transcription method used
Parameterization
The parameterization is based on the discretization of the

continuous variables on a mesh, typically settled up on the
time domain!
Discretize the time domain as:!
t0 = t1 < t2 < ... < tN = tf
Discretize the states and the controls over the previous mesh
by defining xk = x(tk ) and! uk = u(tk )
x(t)
u(t)
{x1 , x2 , ..., xN }
{u1 , u2 , ..., uN }
Consequently, a new vector of variables can be defined:
X = {tf , x1 , u1 , ..., xN , uN }
Transcription: Collocation (1/2)
Collocation methods are based on the transcription of the

differential equations into a finite set of defects constraints
using a numerical integration scheme
Simplest case: Eulers scheme!
But
x i,e
ti )
xi
, then:
xi+1 = xi + h f(xi , ui , ti )
x i+1
Solution is approximated
using a linear expansion
i ) (ti+1
xi+1 = xi + x(t
i) h
= xi + x(t
ti
which can be written in terms of defects constraints:
h i = xi+1
xi
h f(xi , ui , ti ) = 0
t i+1
Transcription: Collocation (2/2)
Other numerical integration schemes can be applied!
Runge-Kutta schemes:
h i = xi+1
xi
j fij
hi
=0
j=1
The optimal control problem has been parameterized:!
x(t)
and
Minimize:
Dynamics:
u(t)
ciao
X = {tf , x1 , u1 , ..., xN , uN }
ciao
ciao
Minimize:
J(X)
Subject to: h (X) = 0
Nonlinear Programming Problem
Transcription: Multiple Shooting (1/2)
Time domain in discretized:
t0 = t1 < t2 < ... < tN = tf
Within each time interval, splines are used to model the control
profile u(t)
each time interval contains M 1 subintervals,
where M is the number of points defining the splines
On a generic node, the vector of variables will be:

Xi = {xi , u1i , ..., uM
i
xi
xi
RK integration
u2i
u1i 1
ti
u...
i 1
xci
xi
RK integration
u1i
u...
i
1
uM
i 1
ti
xi+1
u i
2
iu
uM
i
ti+1
Transcription: Multiple Shooting (2/2)
Within a generic time interval, the splines are used to map the
u(t)
1
discrete values {u1i , ..., uM
into
continuous
functions
}
i
c
Numerical integration can be used to compute xi+1
The dynamics is transcribed into a set of defects constraints:
hi = xci
xi = 0
The vector of variables for the nonlinear programming problem is:
X = {tf , X1 , ..., XN }
xi
xi
RK integration
u2i
u1i 1
ti
u...
i 1
xci
xi
RK integration
u1i
u...
i
1
uM
i 1
ti
xi+1
u i
2
iu
uM
i
ti+1

with
HermiteSimpson method, and certain
subjectistosolved
system
equations
x 1 = 0.5x1 + u
Assignment #5
Example 1: A Simple Optimal Control Problem
5
J = x2 (1)
x 2 = u2 + x1 u(68)
+ x21
s section, a simple optimal control problem is solved with HermiteSimpson
method, and certai
4
are set to speed-up
the computation
time direct
and improving
accuracy.
Consider
the problem of findin
Solve
#1 conditions
with
transcription
and
collocation
(a) discrete solution and analytical time history of x1
and toProblem
the boundary
with fixed initial and final time [0, 1] to minimize the cost
0.9
x
1a
x 1 = 0.5x1 + u
0.8
x
x
(0)
=
1
1
1
0.7
(69)
J
=
x
(1)
(68
5
2
0
0.1x2 (0)
0.2 =
0.30 0.4
0.5
0.6
0.7
0.8
0.9
1
x 2 = u2 + x1 u + x21
4
(b) discrete solution and analytical time history of u
0
with the following bound box of state and control variables
ons
u
Dynamics
Obj.1 +
fcn.
x 1 = 0.5x
u
init., final time
b. c.
log10(abs(error))
log10(abs(error))
10
x1 (t) 10
2
0
0.4
0.5
0.6
0.7
0.8
0.9
1
(69
5 2 0.1 0.2 0.3(70)
2
x2 (t)
of10
x 2 = u + x1 u + x10
x2 (0)
= 0 method
(c) error
x and u located at discrete points
Use
Euler
for direct 4 01
10 u(t) 10
x error
transcription
5
x
of
state
and
control
variables
u error
o the boundary conditions
10
Provide analytic gradients and zero
15
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
10initial
x1 (t)
10
(71)
guess
x
(0)
=
1
The analytical solution of this1 problem is
(d) reintegrate dynamics with discrete control solution
(70
x2 (t) numerical
10
(72)
0
10Compare
vs analytical
x2 (0) = 0
x error
5
cosh(1
t)
solution
10
u(t) 10
(73)
x1 (t) = 10
15cosh(1)
the followingbound
of statebetween
and control
Makebox
trade-off
CPUvariables
time
0
0.1
0.2
0.3(74)
0.4
0.5
0.6
0.7
0.8
0.9
1
time
and solution accuracy
(tanh(1 t) + 0.5) cosh(1 t)
u(t)
=
10
x
(t)
10
1
Figure 11: Simulation results of Example 1, 100 points.(71
his problem is
cosh(1)
10 x2 (t) 10
(72
cosh(1 t)
half of the CPU time (17.2 s vs 8.7 s). In case 3, providing analytical Jacobin and Hes
x1 (t) =
u(t)
10of iterations and
(73
3.1 Solution by10
HermiteSimpson
method
both
the number
the computational time, which drops dramatically
cosh(1)
The structure of Jacobian is shown in Figure 3.2(a), where the rows are constraints
x1 (0) = 1
Low-Thrust Earth-Mars Transfer (1/2)
Optimal control problem:!
r =
Minimize:
Subject to:
r+u
3
r
r(t0 ) = rE (t0 ) v(t0 ) = vE (t0 )

r(tf ) = rM (tf ) v(tf ) = vM (tf )
Transcription technique:
Simple shooting
v1
v2
rM
Note: Simple shooting is

multiple shooting when N = 1
vM
vE
rE
Low-Thrust Earth-Mars Transfer (2/2)
Cubic splines for u(t) built on four points
Optimization variables:
M =4
Earths ephemerides are used to set i.c. for the integration of the
shooting method
constraints on x0 automatically satisfied
t0, tf , u1 , ..., u4
( dim(X) = 14 )
First guess: ballistic Lamberts arc
first guess
solution
Controlled Traj. in Relative Dynamics
Given the equations of the relative dynamics:
Minimize:
Subject to:
2ny 3n2 x
y + 2nx
z + n2 z
=
=
=
0
0
0
r(t0 ) = r0
r(tf ) = rf
v(t0 ) = v0
v(tf ) = vf
Note: r(t0 ), v(t0 ) = 0 , r(tf ), v(tf ) = 0
Formation depl.
r(t0 ), v(t0 ) = 0 , r(tf ), v(tf ) = 0
Formation reconf.
r(t0 ), v(t0 ) = 0 , r(tf ), v(tf ) = 0
Docking
Formation Flying Deployment
Transcription technique: Simple shooting

Cubic splines for u(t) built on six points
First guess: u(t) = 0
M =6
Reference orbit:!
a = 26570 km
Mars Aero-Gravity Assist (1/4)
Gravity assist
Aero-Gravity Assist
Dynamics:
Bank angle!
=
=
=
=
Planar maneuver
= CL /CL((L/D)max)

Control parameters:!
R =
Atmospheric entry conditions
V sin
V cos cos
R cos
V cos sin
R
D
G sin
m
L cos
V 2 cos
G cos +
m
R
V 2 tan cos cos
L sin
m cos
R
Optimal control problem
Find the optimal control law, (t), the free atmospheric entry
conditions and the final time tf to!
Maximize: !
Final heliocentric velocity
Vs+
Subject
to:!
atmospheric entry conditions

must be consistent with entry
conditions in planetary
sphere of influence
( V assigned )!
Convective heating at
stagnation point
qw0 = 1.35e 8
rn
"1/2
!
"
hw
max
3.04
q
V
1
< w0
H

Transcription technique: Multiple Shooting!
100
! dim(X)
N = 11 , M = 4
Cubic splines!
RK integration
x
First
guess
using
simple
u
i 1
shooting and evolutionary

algorithms
2
i 1
u1i 1
ti
120
100
2
Soln 1
Soln 2
xi
xci
xi
u1i
2
iu
u...
i
80
Soln 2
40
100
200
300
Time [s]
400
500
600
20
0
Soln 1
100
200
uM
i
ti+1
60
xi+1
u i
1
uM
i 1
Soln 3
5
0
RK integration
ti
Soln 3
u...
i 1
h [km]
300
Time [s]
400
500
GTOC II (1/3)
Optimal control problem:!
r+u
3
r
r =
Visit four given asteroids
Maximize: J = mf /(tf
Subject to:!
t0 ) , mf = m0 e
1
Isp g0
R tf
t0
|u|d
! dep,P ) = rP (tdep,P )
- r(t
v(tdep,P ) = vP (tdep,P )
!r(tarr,P ) = rP (tarr,P )
v(tarr,P ) = vP (tarr,P )
!
- u
- tdep,Pi
umax
tarr,Pi
90 days
GTOC II (2/3)
Transcription technique: Collocation
Optimization variables:!
Four departure epochs

(Earth and three asteroids)!
Four transfer times!
State parameters deriving

from transcription
Control parameters
deriving from transcription!
dim(X)
1000
GTOC II (3/3)
Identified solution:
Collocation methods
can better describe
discontinuities
Multiple Shooting vs Collocation
Both Multiple Shooting and Collocation can be considered

robust methods, even if highly nonlinear dynamics must
be dealt with !
Advantage of Collocation w.r.t. Multiple Shooting:!
Better management of discontinuities of the control

functions
Disadvantage of Collocation w.r.t. Multiple Shooting:!
Higher number of variables
Direct Methods vs Indirect Methods
Main Advantages of Direct Methods:!
No need of deriving the equations related to the

necessary conditions for optimality!
More versatility and easier implementation in black-box

tools
Main Disadvantage of Direct Methods:!
Need of numerical techniques to effectively estimate

Hessians and Jacobians
Approximate methods!
Avoid both indirect and direct!

Suboptimal solutions
Definition of the original problem

Find
minimizing
u(t),
[ti , tf ],
J = (x(tf ), tf ) +
u = (u1 , u2 , . . . , um )
tf
L(x, u, t) dt
ti
with dynamics
and boundary conditions
x = f (x, u, t),
x(ti ) = xi
x = (x1 , x2 , . . . , xn )
(x(tf ), u(tf ), tf ) = 0
Note: control saturation, path constraints, variable

final time, etc., not considered for simplicity
Solution of the original problem

Hamiltonian of the problem
x(t),
(t), u(t),
H(x, , u, t) = L(x, u, t) +
f (x, u, t)
that satisfy the necessary conditions

x =
H
x
H
=0
u
T
under
x(ti ) = xi
(tf ) =
(1)
(x(tf ), u(tf ), tf ) = 0
t=tf
Iterative methods used to solve (1)

Convergence depends on initial guess
Guessing i is not trivial (no physical meaning)
Difficult to treat (algebraic-differential system)
Deep knowledge of the problem required
Why approximate methods
Avoid solving problem (1)
Transform problem (1) into a simpler problem
Ease the computation of solutions
Deliver sub-optimal solutions
Examples
Direct transcription [Hargraves&Paris 1987, Enright&Conway, Betts 1998]
Generating function [Park&Scheeres, 2006]
SDRE [Pearson 1962, Wernli&Cook 1975, Mracek&Cloutier 1998]
ASRE [Cimen&Banks 2004]
...
(Time Varying) LQR

Dynamics:
x = A(t)x + B(t)u,
Objective function:
Initial condition:
1 T
1
J = x (tf )S(tf )x(tf ) +
2
2
tf
x(ti ) = xi
xT Q(t)x + uT R(t)u dt,
ti
Necessary conditions of optimality

x
= A(t)x + B(t)u,
Q(t)x
AT (t) ,
0 = R(t)u + B T (t) ,
A(t)
Q(t)
u=
B(t)R
AT (t)
(t)B T (t)
(t)B T (t)
Linear system of 2n differential equations
(2)
Solution of TVLQR by the STM

x(t) =
Exact solution of system (2):

( xi ,
xx ,
(t) =
initial state, costate)

x,
are the components

of the state transition matrix (STM)
x
STM subject to
with
xx (ti , ti )
If
xx
= In
x
n,
n,
x (ti , t)xi +
xx (ti , t)
(ti , t) =
A(t)
Q(t)
(ti , ti ) = 0n
xx (ti , t)xi
x (ti , t)
(t)B T (t)
B(t)R
AT (t)
x (ti , ti )
= 0n
(ti , t)
i,
(ti , t)
i,
(3)
(ti , t)
(ti , t)
xx
x
n,
(ti , ti ) = In
(t)
was known, it would be possible to compute x(t),

through (3), and u(t) with u = R 1 (t)B T (t)
i computed
by using (3) and the final condition (3 types)

SolveLQR.m
LQR solver available at http://www.astrodynamics.eu/Astrodynamics.eu/Software.html
Hard constrained problem (HCP)

x = A(t)x + B(t)u,
Final state given
1
J=
2
tf
x(ti ) = xi
xT Q(t)x + uT R(t)u dt,
x(tf ) = xf
ti
Statement of HCP
Write the first of (3) at t = tf ,

and solve for
; i.e.,
xf =
xx (ti , tf )xi
i (xi , xf , ti , tf )
1
x
(ti , tf )
(ti , tf ) [xf
Solving a HCP requires inverting the n
n matrix
i,
xx (ti , tf )xi ]
Soft constrained problem (SCP)

x = A(t)x + B(t)u,
Final state not given
x(ti ) = xi ,
1
1
J = xT (tf )S(tf )x(tf ) +
2
2
tf
(tf ) = S(tf )x(tf ),
xT Q(t)x + uT R(t)u dt
ti
Statement of SCP
Write (3) at t = tf ,
and solve for
i (xi , ti , tf )
x(tf )
xx (ti , tf )xi
S(tf )x(tf )
x (ti , tf )xi
(ti , tf )
i,
(ti , tf )
; i.e.,
=[
(ti , tf )
S(tf )
(ti , tf )]
Solving a SCP requires inverting the n
[S(tf )
xx (ti , tf )
n matrix [
x (ti , tf )] xi
(ti , tf )
S(tf )
(ti , tf )]
Mixed constrained problem (MCP)

Some components
x = A(t)x + B(t)u, x(ti ) = xi , xi (tf ) = xi,f ,
of final state given

(and some not)
1
1
J = xTj (tf )S(tf )xj (tf ) +
2
2
Statement of MCP (
tf
i,j (xi,j , ti , tf )=[
,j (ti , tf )
S(tf )
= S(tf )xj (tf ),
ti
xgiven,
i
Write (3) at t = tf , write i = ( i,i , i,j )T, and solve for

and for i,j using j (tf ) = S(tf )xj (tf ) (SCP); i.e.,
1
i,i (xi,i , xf,i , ti , tf )= x ,i (ti , tf ) [xf,i
j (tf )
free)
xj
i,i using
xi,f (HCP)
xx,i (ti , tf )xi,i ] ,

x ,j (ti , tf )]
[S(tf )
xx,j (ti , tf )
Solving a SCP requires inverting the

matrices x ,i and [ ,j (ti , tf ) S(tf )
x,j (ti , tf )] xi,j
,j (ti , tf )]
Idea of the method

Re-write the nonlinear problem as
original dynamics
factorized dynamics
x = f (x, u, t),
J = (x(tf ), tf ) +
tf
x = A(x, t)x + B(x, u, t)u,
L(x, u, t) dt
original objective function
ti
J=
1 T
1
x (tf )S(x(tf ), tf )x(tf ) +
2
2
tf
factorized objective function
xT Q(x, t)x + uT R(x, t)u dt
ti
Idea: to use state-dependent matrices
A(x, t), B(x, u, t), Q(x, t), R(x, t)
such that for given arguments x(t), u(t) they depend on time only; i.e.,
A(x(t), t), B(x(t), u(t), t), Q(x(t), t), R(x(t), t)
A(t), B(t), Q(t), R(t)
The algorithm: iterations

Iteration 0 - Find x(0) (t), u(0) (t) solving Problem 0
x = xi , u = 0
x (0) = A(xi , t)x(0) + B(xi , 0, t)u(0) ,

1
1
J = x(0)T (tf )S(xi , tf )x(0) (tf ) +
2
2
tf
x(0)T Q(xi , t)x(0) + u(0)T R(xi , t)u(0) dt
ti
x = x(i
Iteration i - Find x(i) (t), u(i) (t) satisfying Problem i

x (i) = A(x(i
1 (i)T
J= x (tf )S(x(i
2
1)
(t), t)x(i) + B(x(i
1
1)
(i)
(tf ), tf )x (tf )+
2
tf
ti
1)
x(i)TQ(x(i
(t), u(i
1)
1)
1)
, u = u(i
1)
(t), t)u(i) ,
(t), t)x(i)+u(i)TR(x(i
1)
(t), t)u(i) dt
The algorithm: convergence

Problem i = TVLQR
Each problem corresponds to a time-varying linear
quadratic regulator (TVLQR)

The method requires solving a series of TVLQR
Iterations terminate when, for given ,
||x(i)
x(i
1)
||
(i)
= max {|xj (t)

t [ti , tf ]
(i 1)
xj
(t)|, j = 1, . . . , n}
the difference between each component of

the state, evaluated for all times, changes by
less than between two successive iterations
Numerical examples
!
Low-thrust dynamics in central vector field

Low-thrust rendez-vous
Low-thrust orbital transfer
Low-thrust stationkeeping of GEO satellites
Rendez-vous: statement and factorization

Dynamics
Rotating frame
x 1
x 2
x 3
x 4
= x3 ,
= x4 ,
= 2x4
= 2x3
with
r=
x1 radial displacement
x2 transversal displacement
(1 + x1 )(1/r3 1) + u1 ,
x2 (1/r3 1) + u2 ,
(x1 + 1)2 + x22
State
Control
x = (x1 , x2 , x3 , x4 )
u = (u1 , u2 )
Normalized units
length unit = orbit radius
time unit = 1/
Initial condition
xi = (0.2, 0.2, 0.1, 0.1)

[Park&Scheeres, 2006]
Factorization (dynamics)
x 1
x 2
x 3
x 4
0
0
0
0
0
f (x1 , x2 )(1 + 1/x1 )
0
f (x1 , x2 )
1
0
0
2
0
1
2
0
A(x)
with
f (x1 , x2 ) =
1/[(x1 + 1)2 + x22 ]3/2 + 1
x1
x2
x3
x4
0
0
1
0
0
0
0
1
B
u1
u2
Rendez-vous: HCP and SCP definition

HCP
1
J=
2
tf
SCP
1 T
1
J = x (tf )Sx(tf ) +
2
2
u u dt
T
ti
xf = (0, 0, 0, 0), ti = 0, tf = 1
tf
uT u dt
ti
S = diag(25, 15, 10, 10), ti = 0, tf = 1

x(tf ) free
Factorization (objective function)

with
1
1
J = xT (tf )S(tf )x(tf ) +
2
2
tf
ti
Termination tolerance
(valid for all examples show)
= 10
Q = 04
4,
R = I2
(S not defined in HCP)
+ 1)2 + x22]3/2 + 1. Thus, the problem is put into the pseudo-LQR form
and B as in (41), and by setting in Q = 044 and R = I22 .
0.3
ound replicate those already know in literature,9, 10 so indicating the ef0.05

0.4
0
0.05
0.1
0.15
0.2
0.25
0.4
0.2
0
ed method.
xx1
xx 3
1
0.9586
0.5660
Iter CPU Time (s)

5
6
0.447
0.492
0.968
0.966
2
error
log
10
HCP
SCP
0.964
0.962
CPU Time referred to an Intel

Core 2 Duo 2 GHz with 4 GB
RAM running Mac OS X 10.6
2.5
2
0.4
0.96
0.958
1
uu1
JJ
0.2
Table 1. Rendez-vous solutions details

Problem
results (HCP)
uu 2
x4
xx 2
x4
e have been successfully solved with the ASRE method described above.Rendez-vous:
s of the
0.5
0.3 HCP and SCP, whose solutions are shown
0.3 in Figures 1 and 2,
J is0.25
the objective function associated
solution, Iter is x
xi to the optimal
0.2
0
i
arried0.2out to converge to the optimal solution, and0.1the CPU Time x
(comf
0.5
an Intel
Core
2
Duo
2
GHz
with
4
GB
RAM
running
Mac
OS
X
10.6.
0.15
0
1
9 . (The machine specifications and the termination
in (14)
0.1 is set to 10
0.1
1.5
other0.05twoxproblems
presented next, too).
f
0.2
4
6
8
3
iter
iter
10
1
3
iter
iter
SolveNL.m
+ 1)2 + x22]3/2 + 1. Thus, the problem is put into the pseudo-LQR form
and B as in (41), and by setting in Q = 044 and R = I22 .
results (SCP)
uu 2
x4
xx 2
x4
e have been successfully solved with the ASRE method described above.Rendez-vous:
s of0.25the HCP and SCP, whose solutions are shown
0.5
0.1 in Figures 1 and 2,
xi
J is the objective function associated
to the optimal
solution, Iter is xi
0.05
0
x(tf )
0.2 out to converge to the optimal solution, and the CPU Time (comarried
0
0.5
an Intel Core 2 Duo 2 GHz with 4 GB RAM running
Mac OS X 10.6.
0.05
0.15
in (14) is set to 109 . (The machine specifications
0.1 and the termination
1
other0.1two problems presented next, too).
0.15
x(tf )
1.5
0.2
ound replicate those already know in literature,9, 10 so indicating the ef2
0.05
0.25
1.5
0.05
0.1
0.15
0.2
0.25
0.3
0.2
0.1
0
0.1
0.2
ed method.
xx 1
xx 3
1
Table 1. Rendez-vous solutions details

Iter CPU Time (s)
0.569
0.5685
0.9586
0.5660
5
6
0.447
0.492
log
error 10
HCP
SCP
0.5695
0.568
JJ
Problem
0.5675
CPU Time referred to an Intel

Core 2 Duo 2 GHz with 4 GB
RAM running Mac OS X 10.6
0.5
0.5
4
6
8
0.567
10
0.5665
0.566
1
u1
u1
iter
iter
12
1
iter
iter
SolveNL.m
Orbital transfer: statement

Dynamics
x 1
x 2
x 3
x 4
Rotating frame (polar

coordinates)
x1 radial distance
= x3 ,
= x4 ,
= x1 x24 1/x21 + u1 ,
= 2x3 x4 /x1 + u2 /x1 .
State
Control
x = (x1 , x2 , x3 , x4 )
u = (u1 , u2 )
x2 angular phase
Normalized units
Objective function
length unit = initial orbit radius
time unit = 1/
(initial orbit)
Initial condition
1
J=
2
tf
uT u dt
ti
ti = 0
tf =
xi = (1, 0, 0, 1)
Final conditions
Problem A
xf = (1.52, , 0,
1/1.52)
Problem B
xf = (1.52, 1.5 , 0,
1/1.52)
The two HPC have been solved with the ASRE method. The solution
Table 2 whose columns have the same meaning as in Table 1. It can be seen
transfer:
results
an increasedOrbital
computational burden
is required to solve
this problem, esp
x2,f = 1.5. The two solutions reported in Table 2 are shown in Figures 3
0
0
0
0
1
0
0
1
0x1 x4
2x4 /x1 0
90
, B(x) =
90
30
180
150
210
Problem A
270
u1
120
1
30
60
30
210
240
300
270
330
210
control
Contol
-0.5
300
240
270
-1
0
1
u2
0.5
330
180
u
0.5
90 2
150
300
47.865
60
180
330
240
Contol
control
B x2,f = 1.5 4.8665 123
120
Iter CPU Time (s)
,A Qx2,f==04 40.5298
, R=
2.
22 I2 6.262
60
150
Problem
120
Results
0
0
0
0
1
0
0 1/x1
Problem B
A(x) =
0
0
1/x31
0
Table 2. EarthMars transfers details
Contol
Factorization
(a) Transfer trajectory.
2
Time
(b) Control pr
0.5
2
1
0
2
Time
time
3
0
Figure 3. Optimal EarthMars transfer (x2,f = )

1
Time
time
Stationkeeping: statement
x 1
x 2
x 3
x 4
x 5
x 6
= x4 ,
State
Dynamics
= x5 ,
x = (x1 , x2 , x3 , x4 , x5 , x6 )
= x6 ,
Control
1
2
2
+ x1 x6 + x1 (x5 + 1) cos x3 + a1 (x1 , x2 , x3 ) + u1 ,
=
2
u = (u1 , u2 , u3 )
x1
x4
a2 (x1 , x2 , x3 )
u2
= 2x6 (x5 + 1) tan x3 2 (x5 + 1) +
+
,
x1
x1 cos x3
x1 cos x3
a3 (x1 , x2 , x3 ) u3
x4
+ ,
= 2 x6 (x5 + 1)2 sin x3 cos x3 +
x1
x1
x1
Initial condition
Rotating frame (spherical coordinates)
xi = (1, 0.05
180/ , 0.05
x1 radial distance
x2 longitude deviation
x3 latitude
180/ , 0, 0, 0) ti = 0
Final condition
xf,1 = 1
xf,j free j = 2, . . . , 6
tf =
Objective function
Q = diag(0, 1, 1, 1, 1, 1),
R = diag(1, 1, 1),
S = 100 diag(1, 1, 1, 1, 1)
Normalized units
length unit = GEO radius
time unit = 1/ (initial orbit)
Reference longitude = 60 E
Perturbations
a1 , a2 , a3
Stationkeeping: factorization
Factorization
033
A(x) =
a41
0
a61
0
0
0
033
I33
0 0
0 a54
0 a64
a45
a55
a65
a46
a56
a66
B(x) =
1
0
0
0
r2
1
cos2
0
0
1
r2
with
a41
a45
a46
a54
a55
= x13 + 1 x26 + ( 2 x25 + 2 3 x5 + 1) cos2 x3 , a56

1
2
= [(1
a61
2 )x1 x5 + 2(1
3 )] cos x3 ,
= (1
a64
1 )x1 x6 ,
1
= x21 2(1
a65
2 ) x1 x5 ,
a66
= 2 1 x5 tan x3 2 2 xx41 ,
1,
2,
3,
1,
Can vary in [0, 1]
2,
1free parameters
= [2 + 2(1
1 )x5 ] tan x3 ,
= 2x1 1 sin 2x3 ,
1
= 2(1
1 ) x1 x6 ,
= [ 12 x5 1] sin 2x3 ,
= 2 1 xx41
[Topputo&Bernelli-Zazzera 2011]
Stationkeeping: results
RESULTS
Manoeuver
Free drift
TC
TFD
TSK
x 10
uu 2
Trajectory during 1 control cycle:

TFD of 2.5 days
TC of 0.5 day
uu 3
xx 3
0.5
0.5
Free drift (2.5 days)
Maneuver (0.5 day)

free drift
manoeuver
start control
end control
1 month simulation
uu 1
1.5
Control cycle
1.5
10
xx 2
2
5
x 10
x 10
2
1
0
0
x 10
1
0
1
0
x 10
5
0
5
0
10
15
20
25
10
15
20
25
10
15
20
25
time
Time
1 year stationkeeping simulated in [Topputo&Bernelli-Zazzera 2011]
Final Remarks
All the previous numerical techniques for optimization are

eventually based on the use of the Newton method:
Direct methods
Solution of the nonlinear system of

equations related to the necessary
conditions for optimality
Indirect methods
Solution of the boundary value

problem on the DAE system
Approx methods
Solution of TVLQR, no need of first

guess solution, but suboptimal
They suffer of the same disadvantages:!
Local convergence, i.e., they tend to converge to solutions

close to the supplied first guesses!
Need of good first guesses for the solution
Note: global optimization is another matter!
Selected references
Direct Methods:!
Enright, P.J., and Conway, B.A., Discrete Approximation to Optimal Trajectories Using Direct
Transcription and Nonlinear Programming, J. of Guid., Contr., and Dyn., 15, 9941001, 1992!
Betts, J.T., Survey of Numerical Methods for Trajectory Optimization, J. of Guid., Contr., and Dyn., 21,
193207, 1998!
Betts, J. T., Practical Methods for Optimal Control Using Nonlinear Programming, Society for Industrial
and Applied Mathematics, Philadelphia, PA, 2001!
Conway, B.A., Space Trajectory Optimization, Cambridge University Press, 2010
Indirect Methods:!
L. S. Pontryagin, V. G. Boltyanskii, R. V. Gamkrelidze, and E. F. Mishchenko, The Mathematical

Theory of Optimal Processes, John Wiley & Sons, New York, NY, USA, 1962!
Bryson, A.E., Ho, Y.C., Applied Optimal Control, Hemisphere Publishing Co., Washigton, 1975
ASRE Method:!
Cimen, T. and Banks, S.P., Global optimal feedback control for general nonlinear systems with
nonquadratic performance criteria, Systems and Control Letters, vol. 53, no. 5, pp. 327346, 2004.

OTS Talk Topputo PDF

Încărcat de

Informații document

Descriere originală:

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

OTS Talk Topputo PDF

Încărcat de

Drepturi de autor:

Formate disponibile

An overview of optimal control

methods with applications

Stardust - Opening Training School!

Founded by 5 partners in January 2008

Used to minimize the

Considerable savings (10-20%) compared to standard methods

Optimal Control Problem (1/2)

Consider the following dynamical system:!

is the state vector and

is the control vector

Determine the m control functions such that the following

where the initial and final state vectors,

Optimal Control Problem (2/2)

In addition to the previous statements suppose that the

Boundary conditions at final time tf : !

Path constraints on the control variables:!

Two classical solution methods:!

Indirect methods: based on reducing the optimal control

Direct methods: based on reducing the optimal control

Example: Low-Thrust Earth-Mars Transfer

Given the dynamics of the controlled 2 body problem:

Subject to equality constraints:

r(t0 ) = rE (t0 ) v(t0 ) = vE (t0 )

and the inequality

Indirect Methods (1/6)

Reconsider the optimal control problem:

Given the dynamical system!

Constraints are added to the performance index J by

p -dimensional vector of constants

for the final constraints!

two n- and a q -dimensional vectors of functions and

Indirect Methods (2/6)

Augmented performance index:

The dynamics is included in the augmented performance

Moreover, pertaining the costate variables for the path

Indirect Methods (3/6)

Integrating by parts the term

The problem is then reduced to identify a stationary point of

Indirect Methods (4/6)

Indirect Methods (5/6)

The problem consists on finding the functions

The previous equations read:

Indirect Methods (6/6)

The previous differential-algebraic system must be coupled to

and to the p + q + 1 additional constraints

The optimal control problem is reduced to a boundary value

timal control problem

Low-Thrust Transfer to Halo Orbit (1/4)

Low-Thrust Transfer to Halo Orbit (2/4)

In canonical form, the dynamics reads:

Performance index: minimize the quadratic functional

, fixed final time

Low-Thrust Transfer to Halo Orbit (3/4)

Processing the last algebraic equation leads to:

which can be inserted in the differential equations

All constraints simply reduce to:

The original problem is reduced to a simple Two Point

Low-Thrust Transfer to Halo Orbit (4/4)

Solution of the TPBVP:!

Transcribe the dynamics! (x,

Evaluate the control parameters

Couple the transcribed dynamics with the constraints on