Variational and symplectic approach of the model-free control

Loc MICHEL
GR

EI - Departement de genie electrique et genie informatique

Universite du Quebec `a Trois-Rivi`eres
C.P. 500, Trois-Rivi`eres, G9A 5H7, Canada, QC
Abstract
We propose
1
a theoretical development of the model-free control in order
to extend its robustness capabilities. The proposed method is based on
the auto-tuning of the model-free controller parameter using an optimal
approach. Some examples are discussed to illustrate our approach.
1
This work is distributed under CC license http://creativecommons.org/licenses/
by-nc-sa/3.0/
a
r
X
i
v
:
1
0
1
1
.
4
2
3
7
v
1


[
c
s
.
S
Y
]


1
8

N
o
v

2
0
1
0
1 Introduction
The model-free control methodology, originally proposed by [1], has been widely suc-
cessfully applied to many mechanical and electrical processes. The model-free control
provides good performances in disturbances rejection and an ecient robustness to
the process internal changes. To improve the eciency of the model-free control,
we propose to use the variational optimization principle in order to tune on-line the
model-free control parameter .
The proposed approach considers the model-free control law as a cost criterion
which could be minimized relating to the parameter . To minimize the cost criteria,
the variational calculus is used considering properties of the symplectic geometry. In
particular, it has been shown that the integration of a cost criteria under symplectic
geometry assumptions provides very simple integration algorithms that can preserve
physical properties of the integrated system and good convergence properties. More-
over, simplicity of this kind of integrator allows building an on-line optimization
process that can adjust the parameter according to the controlled process dynamic.
The paper is structured as follows. Section II presents an overview of the model-
free control methodology including its advantages in comparison with classical method-
ologies. Section III discusses the application of the calculus of variations to the
model-free control. Some concluding remarks may be found in Section IV.
2 Model-free control: a brief overview
2.1 General principles
2.1.1 The ultra-local model
We only assume that the plant behavior is well approximated in its operational range
by a system of ordinary dierential equations, which might be highly nonlinear and
time-varying.

The system, which is SISO, may be therefore described by the input-

output equation
E(t, y, y, . . . , y
()
, u, u, . . . , u
()
) = 0 (1)
where
u and y are the input and output variables,
E, which might be unknown, is assumed to be a suciently smooth function of
its arguments.

See [1, 2] for further details.

1
Assume that for some integer n, 0 < n ,
E
y
(n)
0. From the implicit function
theorem we may write locally
y
(n)
= E(t, y, y, . . . , y
(n1)
, y
(n+1)
, . . . , y
()
, u, u, . . . , u
()
)
By setting E = F + u we obtain the ultra-local model.
Denition 2.1 [1] If u and y are respectively the variables of input and output of a
system to be controlled, then this system admits the ultra-local model dened by:
y
(n)
= F + u (2)
where
R is a non-physical constant parameter, such that F and u are of the
same magnitude;
the numerical value of F, which contains the whole structural information, is
determined thanks to the knowledge of u, , and of the estimate of the derivative
y
(n)
.
In all the numerous known examples it was possible to set n = 1 or 2.
2.1.2 Numerical value of
Let us emphasize that one only needs to give an approximate numerical value to .
It would be meaningless to refer to a precise value of this parameter.
2.2 Intelligent PI controllers
2.2.1 Generalities
Denition 2.2 [1] If n = 1, we close the loop via the intelligent PI controller, or
i-PI controller,
u =
F

+
y

+C() (3)
where
y

is the output reference trajectory, which is determined via the rules of atness-
based control ([3, 4]);
e = y

y is the tracking error;

C() is of the form K
P
+ K
I

. K
P
, K
I
are the usual tuning gains.
Equation (3) is called model-free control law or model-free law.
The i-PI controller 3 is compensating the poorly known term F. Controlling the
system therefore boils down to the control of a precise and elementary pure integrator.
The tuning of the gains K
P
and K
I
becomes therefore quite straightforward.
2
2.2.2 Classic controllers
See [5] for a comparison with classic PI controllers.
2.2.3 Applications
See [6, 7, 8, 9, 10, 11, 12, 13] for already existing applications in various domains.
2.3 Numerical dierentiation of noisy signals
Numerical dierentiation, which is a classic eld of investigation in engineering and
in applied mathematics, is a key ingredient for implementing the feedback loop 3.
Our solution has already played an important role in model-based nonlinear control
and in signal processing (see [14] for further details and related references).
The estimate of the 1
st
order derivative of a noisy signal y reads (see, e.g., [15])

y =
3!
T
3

T
0
(T 2t)y(t)dt where [0,T] is a quite short time window.

This window
is sliding in order to get this estimate at each time instant.
Denoising of y leads to the estimate y =
2!
T
2

T
0
(2T 3t)y(t)dt
The above results are the basis of our estimation techniques. Important theoretical
developments, which are of utmost importance for the computer implementation, may
be found in [16]. We establish the following hypothesis:
Hypothesis 2.1 [2] The estimate of the derivative is realized with an high bandwidth
and is not biased; in particular compared to the noise.
2.4 A rst academic example: a stable monovariable linear
system
Introduce as in [1, 2] the stable transfer function
(s + 2)
2
(s + 1)
3
(4)
2.4.1 A classic PID controller
We apply the well known method due to Broda (see, e.g., [17]) by approximating
System 4 via the following delay system
Ke
s
(Ts + 1)

It implies in other words that we obtain real-time techniques.

3
K = 4, T = 2.018, = 0.2424 are obtained thanks to graphical techniques. The
gain of the PID controller are then deduced [17]: K
P
=
100(0.4+T)
120K
= 1.8181, K
I
=
1
1.33K
= 0.7754, K
D
=
0.35T
K
= 0.1766.
2.4.2 i-PI.
We are employing y = F + u and the i-PI controller
u = [F]
e
+ y

+C()
where
[F]
e
= [ y]
e
u,
y

is a reference trajectory,
= y

y,
C() is an usual PI controller.
2.4.3 Numerical simulations
Figure 1(a) shows that the i-PI controller behaves only slightly better than the classic
PID controller (Fig. 1(b)). When taking into account on the other hand the age-
ing process and some fault accommodation there is a dramatic change of situation:
Figure 1(c) indicates a clear cut superiority of our i-PI controller if the ageing pro-
cess corresponds to a shift of the pole from 1 to 1.5, and if the previous graphical
identication is not repeated (Fig. 1(d)).
2.4.4 Some consequences
It might be useless to introduce delay systems of the type
T(s)e
Ls
, T R(s), L 0 (5)
for tuning classic PID controllers, as often done today in spite of the quite
involved identication procedure.
This example demonstrates also that the usual mathematical criteria for robust
control become to a large irrelevant.
As also shown by this example some fault accommodation may also be achieved
without having recourse to a general theory of diagnosis.
4
(a) i-PI control (b) PID control
(c) i-PI control (d) PID control
Figure 1: Stable linear monovariable system (Output (); reference (- -); denoised
output (. .)).
2.5 From the analytical mechanics to the model-free control
The concept of symplecticity, described very largely by [18] gives a rigorous math-
ematical framework to the variational problems including the concept of symmetry.
Lagrangian mechanics besides is completely formalized thanks to variational calcula-
tion within the framework of the symplectic spaces properties. It introduces, for the
mechanical systems, the concept of Hamiltonian and the generalized coordinates p and
q (resp. proportional to the position and the acceleration), which using the Euler-
Lagrange equation and the Hamilton-Jacobi equation, allows to retrieve the basic
principle of dynamics. Thus, in the case of a conservative mechanical system inde-
pendent of time, the Hamiltonian energy is preserved and induces the conservation of
certain properties in symplectic spaces [19]. Although the symplecticity is a relatively
abstract concept, much of the works (let us quote e.g. [19, 20, 21, 22, 23, 24, 25])
established the possibility of using symplectic integration algorithms under certain
conditions i.e able to preserve the Lagrangian structure of the mechanical systems,
and to preserve the Hamiltonian in particular in the case of the systems that are
independent of time.
5
3 Introduction to the variational model-free con-
trol
3.1 Possibility of controlling
The fundamental goal of a variational optimization of the model-free control law (3)
is to increase the robustness/the tracking of the reference by an online adjustment of
the parameter . Starting from the denition of a cost criterion, we justify the use
of variational methods by the fact that they allow to calculate the parameter as a
solution of the optimization problem according to the time.
Considered application Consider an application in power electronics [13] where
we have to control the output voltage of a power converter according to an output
reference. Consider the dc/dc buck converter (Fig. 2) where u is the duty-cycle and
E = 20V L = 1 mH, C = 10 F. The load is a simple varying resistor R.
Figure 2: Classical buck converter.
u is the input of the converter to control. The controlled output voltage is noted y and
the output reference voltage is noted y

. The equivalent transfer function H = y/u

of this buck converter to control is of the second order and is stable.
3.2 Discrete model-free control law and associated properties
In practice, when the implementation of a i-PI controller is carried out in a digital
way like a micro-controller or FPGA for example, it is necessary to discretize the i-PI
controller of the def. (2.2).
u
k
=
1

F
k1
y
(n)

+C()|
k
=
1

y
(n)

k1
u
k1
. .. .
F
k1
y
(n)

+C()|
k
(6)
u
k
= u
k1

1

y
(n)

k1
y
(n)

+ C()|
k
(7)
6
Denition 3.1 For any discrete moment t
k
, k N, one denes the discrete con-
troller i-PI.
u
k
= u
k1

1

y
(n)

k1
y
(n)

+ C()|
k
(8)
where
y

is the output reference trajectory;

= y

y is the tracking error;

C is a usual corrector PI where K
P
, K
I
are the usual tuning gains.
The discrete intelligent controller is also called discrete model-free control law or
discrete model-free law.
Theorem 3.1 [26] If the function f(x, t) in the equation x = f(x, t) is continuous
in t and if it exists a positive constant l such that:
||f(x
2
, t) f(x
1
, t)|| l||x
2
x
1
|| (9)
for all x
1
and x
2
in the neighborhood B
r
of the origin, and for all t [t
0
, t
0
+T], T
R
+
, then the dierential equation has a single solution x(t) for suciently small initial
states and a suciently short time period T. The condition (9), is called Lipschitz
condition and the constant l is the Lipschitz constant. If (9) is satised, one says that
f is locally Lipschitz.
The example presented Fig. 4 shows, in the case of a variable reference with a
load disturbance, which occurs at (t = 0,003 s), and a simple corrector gain K
p
, the
existing dierence between a model-free control law with constant and dependent
on the time.
Hypothesis 3.1 The output reference voltage is K
y
-Lipschitz and varies slowly ac-
cording to the time.
The hyp. 3.1 implies that u and thus y and are Lipschitz if the model-free
control law is stable. In the same manner, it is appropriate to limit the variations on
the possible drifts of the model H.
Hypothesis 3.2 The drifts of the H parameters are Lipschitz.
In previous model-free control applications (e.g. [13]), the static value of does
not inuence signicantly the stability of the model-free control. However, when
varies, it proves that it is possible to act on the performances of the model-free control
by adjusting on-line the parameter. The objective is thus to be able to dene a
variation law of such that the dynamic performances of the model-free control are
improved.
7
3.2.1 Model-free control -K

degenerated
In order to assume correctly the variations of within the discrete model-free law
(8) and with an aim of satisfying the integration conditions of the Euler-Lagrange
equation, which will allow the -optimization of the model-free law, it is advisable to
dene the dynamic of . Consequently, we consider the addition of the derivative
in (8)

.
Denition 3.2 When the system works in closed-loop, one denes the u
k
-degenerate
model-free law for n = 1 such that:
u
k
= u
k1

dy
dt

k1

dy
c
dt

+ C()|
k
+

d
dt

(10)
where = 1/

, K

and N that dene the dynamics of variation of .

The parameter veries the def. 2.1.
The model-free law (10) seems to be a redenition of the original concept (2.2)
taking into account the variations of the parameter .
3.2.2 Model-free local control law u
k
-implicite
We establish a non-recursive expression and implicitly dependant on F. According
to the def. 2.2, and under the hyp. 2.1, we have in the discrete time:
u
k
=
1

F
k1

dy
c
dt

d
dt

+ C()|
k
(11)
that can be rewritten:

u
k
= F
k1
+
dy
c
dt

k
+

d
dt

C()|
k
(12)

u
k
+ F
k1
=
dy
c
dt

k
+

d
dt

C()|
k
(13)
According to the th. 3.1, we call, u
k
-implicite model-free discrete law, the following
expression:
dy
dt
....
F
k1
+

u
k1

k1

dy
c
dt

k
=

C()

k
+

d
dt

(14)

One may consider, in future work, higher derivative order of

8
(a) = 14, K
p
= 2
(b) (t) = 14 + 10
6
t, K
p
= 2
Figure 3: Comparison between constant and variable.
9
d
dt

k1
=

C()

k
+

d
dt

(15)
Let us dene
d
dt
=
dy
dt

k1

dyc
dt

k
like the derivative of the tracking error using the
th. 3.1 via the hyp. 3.2.
When the corrector C is dened by a single constant gain C K
p
, then it exists
a linear relation between and the output and its derivative. It comes then:
Denition 3.3 The model-free Lagrangian veries:
1
K
p

d
dt

k1
+ K

d
dt

= |
k
(16)
with =
1

+
K
p
R

+
K

N
In this expression, the parameters K
p
and K

are constants. Since the variable to

nd is , and
d
dt
are considered as constants over a small time.
3.3 Model-free Lagrangian
3.3.1 Criterion L

Based on the -K

-degenerated expression, we dene the variational model-free cri-

terion L

such that:
L

(y y

)d =

1
K
p

d
dt

k1
+ K

d
dt

d (17)
The expression -K

-degenerated provides a direct equivalence relation between

the traditional expression of the reference tracking y y

and the model-free control.

The correction gain K
p
is here a scaling factor.
3.3.2 Elements of variational calculation
The fundamental theorem of calculus of variation denes the existence of a solution
to the problem of integrals of functions minimization via the resolution of the Euler-
Lagrange equation.
Theorem 3.2 The Euler-Lagrange equation [27]: Let J : C
2
[x
0
, x
1
] R, be a func-
tional of the form:
J(y) =

x
1
x
0
f(x, y, y

)dx (18)
10
where f has continuous partial derivatives of second order with respect to x, y and y

,
and x
0
< x
1
. Let:
S = {y C
2
[x
0
, x
1
] : y(x
0
) = y
0
et y(x
1
) = y
1
} (19)
where y
0
and y
1
are given real numbers. If y S is an extremal for J, then
d
dx

f
y

f
y
= 0 (20)
for all x [x
0
, x
1
].
The Euler-Lagrange equation, is thus a dierential equation, generally nonlinear,
that must be satised for any extremal function y(x). In other words, if y(x) is
solution of the Euler-Lagrange equation, then the functional J, dened as the integral
on [x
0
, x
1
] of a f, functional dependant on x, y and y

, is minimal.
3.3.3 Numerical resolution of the criterion L

Conjecture 3.1 Under certain conditions to dene, the Euler-Lagrange equation,

applied to the variational model-free Lagrangian (17), establishes the possibility of
controlling the tracking error y y

permanently by the on-line adjustment of .

The Euler-Lagrange equation thus allows to determine such that the criterion
L

is miminum. The continuous Euler-Lagrange reads:

d
dt

=
L

(21)
The integration of the variational model-free Lagrangian (16) provides the evo-
lution law of according to the time such that the integral of the error y y

is
minimal.
Let us dene the discrete Euler-Lagrange equation according to [28] [29] where
D
1
and D
2
indicate the discrete derivation of the Lagrangian:
D
2
L
d
(
k1
,
k
, h) + D
1
L
d
(
k
,
k+1
, h) = 0 (22)
We have:

K
p

k
d
dt
+ K

k1

k
h

+

K
p

k
d
dt
+ K

k+1

k
h

= 0 (23)
then:

k
d
dt
+ K

k+1
2
k
+
k1
h

= 0 (24)
11
The discrete symplectic integrator derived from (24) is written (expression of
k+1
according to
K
and
k1
):
h

d
dt

k
+ K

(
k+1
2
k
+
k1
) = 0 (25)
then:

d
dt
2

k
+
k1
=
k+1
(26)
Notice that the dependence with respect to the corrector K
p
does not exist any-
more. Equation (10), called control law forms with (16), the symplectic model-free
control or i-PIS corrector (by extension of i-PI).
Denition 3.4 When the system works in closed-loop, we dene the intelligent con-
troller that is symplectic type or i-PIS controller by:

u
k
= u
k1

1

dy
dt

k1

dy
c
dt

+ K
p
K

d
dt

d
dt
2

k

k1
=
k+1

0
(resp.
0
) is initial condition
(27)
We preserved the natural notations of (10) (continuous into ) and (16) (discrete
form in

). Let us note that now seems a state variable of the model-free control.
Remarks
instability of the strong load transients When there are disturbances (vari-
ations of model or load), the output may have to change quickly, this implies
strong variations of the output derivative. However, under its conditions, al-
though a Lipschitz condition is imposed on , the assumptions requested to
apply the Euler-Lagrange equation may not be satised. Simulations of the
3.5 show in particular that, if R is varying very slowly, then at each level of R,
oscillates then is stabilized; more R varies fast, more stabilization is slow.
justication of the nature of (26) In order to interpret the form of the
equation of the control, which is of the damped harmonic oscillator type,
let us look at the output buck converter signal in case of the i-PI (e.g. Fig.
5). Those appear naturally oscillating and the role of the derivative in (3) is
to attenuate these oscillations in order to reach the permanent state. It thus

The sign of the derivative is not yet dened; this point belongs to the future studies in the
symplectic model-free control.
12
appears rather intuitive that to counterbalance these oscillations, it would be
enough to vary the coecient in the shape of an oscillator, which species
well the law of control.
independence with respect to the initial condition on the nature of
the Euler-Lagrange equation implies a certain independence with respect to
the initial conditions

. This independence is in particular supposed to be
true according to the simulations of the 3.5. We can also conjecture that
independence with respect to the initial means that is auto-adjusted even
if it undergoes strong variations.
critical integration step h The integration step h seems a critical element
and must be selected suciently small to allow the preservation of the condition
of integration of the th. 3.1.
3.4 Symplecticity of the numerical integration methods
Let us consider the damped mass system moving according to the x-coordinate for
which we have the following Lagrangian:
L

x,
dx
dt

=
1
2
m

dx
dt

2
+
1
2
kx
2
(28)
m and k are respectively the mass and the constant of the spring. Figures 4(a)
and 4(b) present respectively the traditional response of the system according to the
time and the comparison of the energy of (28) calculated via the dierent integration
methods. This example

shows that the symplectic integration of Euler, correspond-

ing to the equation of control (26) is at the same time easy to implement and
ecient regarding the preservation of the energy.
3.5 Application of the symplectic model-free control
Let us consider the buck converter for a slope output reference. Figure 5 presents the
comparison between the i-PI and the i-PIS without impact of load; in all the cases,
we took = 1 and K

= 2. Let us note that the static error tends to being slightly

decreased in the case of the i-PIS, thanks to the harmonic compensation of the
coecient. Figure 6 presents the case of the i-PIS corrector for two initial values of
for a progressive impact of load

of about 10 . Let us note that the variations of

The eect of the initial conditions on the derivative of is unknown.

From the work of Mr. Benjamin Stephens, Ph.D. student in the Robotics Institute at Carnegie
Mellon University.

By respect of the Lipschitz condition on .

13
(a) Speed and position of the mass according to the time.
(b) Comparison of energy resulting from the various integration methods.
Figure 4: Symplectic integration.
14
(a) i-PIS control for
0
= 3
(b) i-PI control for
0
= 3
Figure 5: Comparison between i-PI and i-PIS without any load impact (R does not
change).
15
(a)
0
= 3
(b)
0
= 10
Figure 6: Comparison between two initial values of : the load resistor increases
gradually from R = 10 at t = 0, 003 s such that R = 18 at t = 0, 01 s.
16
allow a priori only an adjustment of the i-PIS sensitivity since the initial condition
on appears as a median value around which is adjusted. It is thus advisable to
be able to choose properly the initial value of .
Compared to the i-PIS corrector, one will note the importance of the Lipschitz
condition in the stabilization of this kind of transient.
4 Towards a formalization of the model-free con-
trol?
We implemented a procedure of auto-adjustment based on the calculus of variation
and from the symplectic geometry considerations. The presented results are encour-
aging but there remain much of challenges before arriving to a completely optimal
i-PIS form. Future work should be directed towards formal proofs of stability and
dynamic performances.
Acknowledgement
The author is sincerely gratefully to Prof. Michel Fliess, Dr. Cedric Join and Prof.
Pierre Sicard for their kind attention and technical support about the model-free
control methodology.
References
[1] M. Fliess and C. Join, Commande sans mod`ele et commande ` a mod`ele re-
streint, e-STA, vol. 5 (n

4), pp. 1-23, 2008 (available at http://hal.inria.

fr/inria-00288107/en/).
[2] M. Fliess and C. Join, Model-free control and intelligent PID controllers: To-
wards a possible trivialization of nonlinear control ?, 15
th
IFAC Symp. System
Identif., Saint-Malo, 2009 (available at http://hal.inria.fr/inria-00372325/
en/).
[3] M. Fliess, J. Levine, P. Martin and P. Rouchon, Flatness and defect of non-
linear systems: introductory theory and examples, Int. J. Control, vol. 61, pp.
1327-1361, 1995.
[4] H. Sira-Ramrez and S. Agrawal, Dierentially Flat Systems, Marcel Dekker,
2004.
17
[5] B. dAndrea-Novel, M. Fliess, C. Join, H. Mounier and B. Steux, A mathematical
explanation via intelligent PID controllers of the strange ubiquity of PIDs, 18
th
Medit. Conf. Control Automation, Marrakech, 2010 (available at http://hal.
inria.fr/inria-00480293/en/).
[6] B. dAndrea-Novel, C. Boussard, M. Fliess, O. El Hamzaoui, H. Mounier and B.
Steux, Commande sans mod`ele de vitesse longitudinale dun vehicule electrique,
6
e
Conf. Internat. Francoph. Automatique, Nancy, 2010 (available at http://hal.
inria.fr/inria-00463865/en/).
[7] S. Choi, B. dAndrea-Novel, M. Fliess and H. Mounier, Model-free
control of automotive engine and brake for stop-and-go scenario, 10
th
IEEE Conf. Europ. Control Conf., Budapest, 2009 (available at http:
//hal.inria.fr/index.php?halsid=1trv221lj8gke47u27bk03n2t6&view_
this_doc=inria-00395393&version=1).
[8] P.-A. Gedouin, C. Join, E. Delaleau, J.-M. Bourgeot, S. Arbab-Chirani and
S. Calloch, Model-free control of shape memory alloys antagonistic actuators,
17
th
IFAC World Congress, Seoul, 2008 (available at http://hal.inria.fr/
inria-00261891/en/).
[9] P.-A. Gedouin, C. Join, E. Delaleau, J.-M. Bourgeot, S. Arbab-Chirani and
S. Calloch, A new control strategy for shape memory alloys actuators, 8
th
Europ. Symp. Martensitic Transformations, Prague, 2009 (available at http:
//hal.inria.fr/inria-00424933/en/).
[10] C. Join, J. Masse and M. Fliess,

Etude preliminaire dune commande sans

mod`ele pour papillon de moteur, J. europ. syst. automat., vol. 42, pp. 337-354,
2008 (available at http://hal.inria.fr/inria-00187327/en/).
[11] C. Join, G. Robert and M. Fliess, Vers une commande sans mod`ele pour
amenagements hydroelectriques en cascade, 6
e
Conf. Internat. Francoph. Au-
tomat., Nancy, 2010 (available at http://hal.inria.fr/inria-00460912/en/).
[12] J. Villagra and C. Balaguer, Robust motion control for humanoid robot exible
joints, 18
th
Medit. Conf. Control Automation, Marrakech, 2010.
[13] L. Michel, C. Join, M. Fliess, P. Sicard and A. Cheriti, Model-free control
of dc/dc converter, in 2010 IEEE 12th Workshop on Control and Modeling
for Power Electronics, pp.1-8, June 2010. (available at http://hal.inria.fr/
inria-00495776/).
[14] M. Fliess, C. Join and H. Sira-Ramrez, Non-linear estimation is easy, Int. J.
Model. Identif. Control, vol. 4, pp. 12-27, 2008 (available at http://hal.inria.
fr/inria-00158855/en/).
18
[15] F.A. Garca Collado, B. dAndrea-Novel, M. Fliess and H. Mounier, Analyse
frequentielle des derivateurs algebriques, XXII
e
Coll. GRETSI, Dijon, 2009 (avail-
able at http://hal.inria.fr/inria-00394972/en/).
[16] M. Mboup, C. Join and M. Fliess, Numerical dierentiation with annihilators
in noisy environment, Numer. Algor., vol. 50, pp. 439-467, 2009.
[17] D. Dindeleux, Technique de la regulation industrielle, Eyrolles, 1981.
[18] J. Marsden and T.S. Ratiu, Introduction to Mechanics and Symmetry: A Basic
Exposition of Classical Mechanical Systems. Springer, 1 January 1999.
[19] R.J. ONeale, Preservation of phase space symplectic structure in symplectic
integration, Ph.D. dissertation thesis, Massey University, New Zeland, 2009.
[20] M. Leok, Foundations of Computational Geometric Mechanics, Ph.D. disser-
tation thesis, California Institute of Technology, CA, 2004.
[21] A.M. Bloch, N.E. Leonard, J.E. Marsden, Controlled Lagrangians and the stabi-
lization of mechanical systems. I. The rst matching theorem, IEEE Transactions
on Automatic Control, vol.45, no.12, pp.2253-2270, Dec. 2000.
[22] A.M.Bloch, Eui Chang Dong ; N.E. Leonard and J.E. Marsden , Controlled
Lagrangians and the stabilization of mechanical systems. II. Potential shaping,
IEEE Transactions on Automatic Control, vol.46, no.10, pp.1556-1571, Oct 2001.
[23] J. E. Marsden, S. Pekarsky, S. Shkoller and M. West, Variational methods,
multisymplectic geometry and continuum mechanics, Journal of Geometry and
Physics, Volume 38, Issues 3-4, pp. 253-284, June 2001.
[24] A. N. Hirani, Linearization methods for variational integrators and Euler-
Lagrange equations, Master thesis, California Institute of Technology, 2000.
[25] J.E. Marsden and M. West, Discrete mechanics and variational integrators,
Acta Numerica 10, pp.357-514, 2001.
[26] J.-J E. Slotine and W. Li, Applied nonlinear control, Prentice-Hall, 1991.
[27] B. Van Brunt, The calculus of variations, Ed. Springer, 2006.
[28] G. Vilmart,

Etude dintegrateurs geometriques pour des equations

dierentielles, Ph.D. dissertation thesis,

Ecole Normale Superieure de Cachan,
France, 2008.
[29] M. West, Variational integrators, Ph.D. dissertation thesis, California Institute
of Technology, CA, 2004.
19