A new kind of nonlinear model predictive control

algorithm enhanced by control lyapunov functions 59

A new kind of nonlinear model predictive control algorithm enhanced by
control lyapunov functions
Yuqing He and Jianda Han
x

A new kind of nonlinear model
predictive control algorithm
enhanced by control lyapunov functions

Yuqing He and Jianda Han
State Key Laboratory of Robotics, Shenyang Institute of Automation,
Chinese Academy of Sciences
P.R.China

1. Introduction
With the abilities of handling constraints and performance of optimization, model based
predictive control (MPC), especially linear MPC, has been extensively researched in theory
and applied in practice since it was firstly proposed in 1970s (Qin & Badgwell, 2003).
However, when used in systems with heavy nonlinearities, nonlinear MPC (NMPC) results
often in problems of high computational cost or closed loop instability due to their
complicated structure. This is the reason why the gaps between NMPC theory and its
applications in reality are larger and larger, and why researches on NMPC theory absorbs
numerous scholars (Chen & Shaw, 1982; Henson, 1998 ; Mayne, et al., 2000 ; Rawlings, 2000).
When the closed loop stability of NMPC is concerned, some extra strategies is necessary,
such as increasing the length of the predictive horizon, superinducing state constraints, or
introducing Control Lyapunov Functions (CLF).
That infinite predictive/control horizon (in this chapter, predictive horizon is assumed
equal to control horizon) can guarantee the closed loop stability is natural with the
assumption of feasibility because it implicates zero terminal state, which is a sufficient
stability condition in many NMPC algorithm (Chen and Shaw, 1982). In spite of the
inapplicability of infinite predictive horizon in real plants, a useful proposition originated
from it makes great senses during the development of NMPC theory, i.e., a long enough
predictive horizon can guarantee the closed loop stability for most systems (Costa & do Val,
2003; Primbs & Nevistic, 2000). Many existing NMPC algorithm is on the basis of this result,
such as Chen & Allgower (1998), Magni et al. (2001). Although long predictive horizon
scheme is convenient to be realized, the difficulty to obtain the corresponding threshold
value makes this scheme improper in many plants, especially in systems with complicated
structure. For these cases, another strategy, superinducing state constraints or terminal
constraints, is a good substitue. A typical predictive control algorithm using this strategy is
the so called dual mode predictive control(Scokaert et al., 1999 ; Wesselowske and Fierro,
2003 ; Zou et al., 2006), which is originated from the predictive control with zero terminal
state constrains and can increase its the stability region greatly. CLF is a new introduced
3
Model Predictive Control 60

concept to design nonlinear controller. It is firstly used in NMPC by Primbs et al. in 1999 to
obtain two typical predictive control algorithm with guaranteed stability.
Unfortunately, each approach above will result in huge computational burden
simultaneously since they bring either more constraints or more optimizing variables. It is
well known that the high computational burden of NMPC mainly comes from the online
optimization algorithm, and it can be alleviated by decreasing the number of optimized
variables. But this often deteriorates the closed loop stability due to the changed structure of
optimal control problem at each time step.
In a word, the most important problem during designing NMPC algorithm is that the
stability and computational burden are deteriorated by each other. Another problem,
seldom referred to but top important, is that the stability can only be guarangteed under the
condition of perfect optimization algorithm that is impossible in reality. Thus, how to design
a robustly stable and fast NMPC algorithm has been one of the most difficult problems that
many researchers are pursued.
In this chapter, we attempt to design a new stable NMPC which can partially solve the
problems referred to above. CLF, as a new introduced concept to design nonlinear controller
by directly using the idea of Lyapunov stability analysis, is used in this chapter to ensure the
stability. Firstly, a generalized pointwise min-norm (GPMN) controller (a stable controller
design method) based on the concept of CLF is designed. Secondly, a new stable NMPC
algorithm, called GPMN enhanced NMPC (GPMN-ENMPC), is given through
parameterized GPMN controller. The new algorithm has the following two advantages, 1) it
can not only ensure the closed loop stability but also decrease the computational cost
flexibly at the price of sacrificing the optimality in a certain extent; 2) a new tool of guide
function is introduced by which some extra control strategy can be considered implicitly.
Subsequently, the GPMN-ENMPC algorithm is generalized to obtain a robust NMPC
algorithm with respect to the feedback linearizable system. Finally, extensive simulations
are conducted and the results show the feasibility and validity of the proposed algorithm.

2. Concept of CLF
The nonlinear system under consideration in this chapter is in the form as:


( ) ( )
m
x f x g x u
u U R
= +
e c

(1)

where
n
x R c is state vector,
m
u R c is input vector, f(*) and g(*) are nonlinear smooth
functions with f(0) = 0. U is the control constraint.

Definition I:
For system (1), if there exists a C
1
function V(x): x c R
n
R
+

{0}, such that

1) V(0) = 0, V(x) > 0 if x 0;
2) a
1
(||x||) < V(x) < a
2
(||x||), where a
1
(*) and a
2
(*) are class K

functions;
3)
c
inf [ ( ) ( ) ( ) ( ) ] 0, {0}
m
x x
u U R
V x f x V x g x u x
e c
+ < eO ,where { : ( ) }
n
c
x R V x c O e s .
then V(x) is called a CLF of system (1). Moreover, if x can be chosen as R
n
and V(x) satisfies
the following condition,

V(x) ==> ||x||

then V(x) is called a global CLF of system (1).
If system (1) has uncertainty terms, i.e.,


( ) ( ) ( )
( )
m
x f x g x u l x
y h x
u U R
e = + +
=
e c

(2)

where e R
q
is external disturbance; l(*) and h(*) are pre-defined nonlinear smooth
functions; y is the interested output. We have the following concept of robust version CLF
called H

CLF,

Definition II,
For system (2), if there exists a C
1
function V(x): x c R
n
R
+

{0}, such that

1) V(0) = 0, V(x) > 0 if x 0;
2) a
1
(||x||) < V(x) < a
2
(||x||), where a
1
(*) and a
2
(*) are class K

functions;
3)
2 1 2
c c
1 1
inf { ( )[ ( ) ( ) ] ( ) ( ) ( ) ( ) ( )} 0
2 2
,
T T T
x x x
u R
m
V x f x g x u V x l x l x V h x h x x

e
+ + + < eO O , where c
1
>c
2
.
then V(x) is called a local H

CLF of system (2) in

1 2
c c
O O . Furthermore, V(x) is called a
global H

CLF if c
1
can be chosen + with V(x) as |x|.

Definition I and II indicate that if we can obtain a CLF or H

CLF of system (1) or (2), a

permitted control set can be found at every feasible state, and the control action inside the
set can guarantee the closed loop stability of system (1) or input output finite gain L
2

stability of system (2). Subsequently, in order to complete the controller design, what one
needs to do is just to find an approach to select a sequence of control actions from the
permitted control set, see Fig. 1.


Fig. 1. Sketch of CLF, the shadow indicates the permitted set of (x, u) ( , ) V x u

along system (1)

Input
State
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 61

concept to design nonlinear controller. It is firstly used in NMPC by Primbs et al. in 1999 to
obtain two typical predictive control algorithm with guaranteed stability.
Unfortunately, each approach above will result in huge computational burden
simultaneously since they bring either more constraints or more optimizing variables. It is
well known that the high computational burden of NMPC mainly comes from the online
optimization algorithm, and it can be alleviated by decreasing the number of optimized
variables. But this often deteriorates the closed loop stability due to the changed structure of
optimal control problem at each time step.
In a word, the most important problem during designing NMPC algorithm is that the
stability and computational burden are deteriorated by each other. Another problem,
seldom referred to but top important, is that the stability can only be guarangteed under the
condition of perfect optimization algorithm that is impossible in reality. Thus, how to design
a robustly stable and fast NMPC algorithm has been one of the most difficult problems that
many researchers are pursued.
In this chapter, we attempt to design a new stable NMPC which can partially solve the
problems referred to above. CLF, as a new introduced concept to design nonlinear controller
by directly using the idea of Lyapunov stability analysis, is used in this chapter to ensure the
stability. Firstly, a generalized pointwise min-norm (GPMN) controller (a stable controller
design method) based on the concept of CLF is designed. Secondly, a new stable NMPC
algorithm, called GPMN enhanced NMPC (GPMN-ENMPC), is given through
parameterized GPMN controller. The new algorithm has the following two advantages, 1) it
can not only ensure the closed loop stability but also decrease the computational cost
flexibly at the price of sacrificing the optimality in a certain extent; 2) a new tool of guide
function is introduced by which some extra control strategy can be considered implicitly.
Subsequently, the GPMN-ENMPC algorithm is generalized to obtain a robust NMPC
algorithm with respect to the feedback linearizable system. Finally, extensive simulations
are conducted and the results show the feasibility and validity of the proposed algorithm.

2. Concept of CLF
The nonlinear system under consideration in this chapter is in the form as:


( ) ( )
m
x f x g x u
u U R
= +
e c

(1)

where
n
x R c is state vector,
m
u R c is input vector, f(*) and g(*) are nonlinear smooth
functions with f(0) = 0. U is the control constraint.

Definition I:
For system (1), if there exists a C
1
function V(x): x c R
n
R
+

{0}, such that

1) V(0) = 0, V(x) > 0 if x 0;
2) a
1
(||x||) < V(x) < a
2
(||x||), where a
1
(*) and a
2
(*) are class K

functions;
3)
c
inf [ ( ) ( ) ( ) ( ) ] 0, {0}
m
x x
u U R
V x f x V x g x u x
e c
+ < eO ,where { : ( ) }
n
c
x R V x c O e s .
then V(x) is called a CLF of system (1). Moreover, if x can be chosen as R
n
and V(x) satisfies
the following condition,

V(x) ==> ||x||

then V(x) is called a global CLF of system (1).
If system (1) has uncertainty terms, i.e.,


( ) ( ) ( )
( )
m
x f x g x u l x
y h x
u U R
e = + +
=
e c

(2)

where e R
q
is external disturbance; l(*) and h(*) are pre-defined nonlinear smooth
functions; y is the interested output. We have the following concept of robust version CLF
called H

CLF,

Definition II,
For system (2), if there exists a C
1
function V(x): x c R
n
R
+

{0}, such that

1) V(0) = 0, V(x) > 0 if x 0;
2) a
1
(||x||) < V(x) < a
2
(||x||), where a
1
(*) and a
2
(*) are class K

functions;
3)
2 1 2
c c
1 1
inf { ( )[ ( ) ( ) ] ( ) ( ) ( ) ( ) ( )} 0
2 2
,
T T T
x x x
u R
m
V x f x g x u V x l x l x V h x h x x

e
+ + + < eO O , where c
1
>c
2
.
then V(x) is called a local H

CLF of system (2) in

1 2
c c
O O . Furthermore, V(x) is called a
global H

CLF if c
1
can be chosen + with V(x) as |x|.

Definition I and II indicate that if we can obtain a CLF or H

CLF of system (1) or (2), a

permitted control set can be found at every feasible state, and the control action inside the
set can guarantee the closed loop stability of system (1) or input output finite gain L
2

stability of system (2). Subsequently, in order to complete the controller design, what one
needs to do is just to find an approach to select a sequence of control actions from the
permitted control set, see Fig. 1.


Fig. 1. Sketch of CLF, the shadow indicates the permitted set of (x, u) ( , ) V x u

along system (1)

Input
State
Model Predictive Control 62

CLF based nonlinear controller design method is also called direct method of Lyapunov
function based controller design, and its difficulty is how to ensure the controllers
continuousness. Thus, most recently, researchers mainly pay their attentions to designing
continuous CLF based controller, and several universal formulas have been revealed.
Sontags formula (Sontag, 1989), for example, originated from the root calculation of 2
nd
-
order equation, can be written as Eq. (3) through slightly modification by Freeman (Freeman
& Kokotovic, 1996b),


2
( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
( ) ( )( )
0
0 0
T T
x x x x
x T T
x x
x
x x x x x x x
x x x x
V f V f q x V g g V
V g
u V g g V
V g
(
+ +
( =

=
(

(3)

where q(x) is a pre-designed positive definite function.
Pointwise Min-Norm (PMN) control is another well known CLF-based approach proposed
by Freeman (Freeman & Kokotovic, 1996a),


min
. . ( )[ ( ) ( ) ] ( )
u
x
u
s t V x f x g x u x
u U
o + s
e
(4)

where (x) is a pre-selected positive definite function. Controller (4) can also be explicitly
denoted as (5) if the constraint set U can be selected big enough.


[ ( ) ( ) ( )] ( ) ( )
( ) ( ) ( ) 0
( ) ( ) ( ) ( )
0 ( ) ( ) ( ) 0
T T
x x
x
T T
x x
x
V x f x x g x V x
V x f x x
u V x g x g x V x
V x f x x
o
o
o
+
+ >

+ s

(5)

(3) and (5) provide two different methods on how to design continuous and stable controller
based on CLF with respect to system (1). H

CLF with respect to system (2) is a new given

concept, and there are no methods can be used to designed robust controller based on it.
Although the closed loop stability can be guaranteed using controller (3) or controller (5),
selection of parameters q(x) or (x) is too difficult to be used in real applications. This is
mainly because these parameters heavily influence some inconsistent closed loop
performance simultaneously. Furthermore, if the known CLF is not global, the selection of
q(x) and (x) will also influence stability margin of the closed loop systems, which makes
them more difficult to be selected (Sontag, 1989; Freeman & Kokotovic, 1996a). In this
chapter, we will firstly give a new CLF based controller design strategy, which is superior
compared to the existing CLF based controller design methods referred to above.
Furthermore, the most important is that this new strategy can be used in designing robustly
stable and fast NMPC algorithm.


3. GPMN-ENMPC
3.1 CLF based GPMN controller
Since q(x) and (x) in controller (3) and controller (5) are difficult to select, a guide function is
proposed in this subsection into the PMN controller to obtain a new CLF based nonlinear
controller with respect to system (1), in the following section, this controller will be
generated with respect to system (2). In the new controller, (x) is only used to ensure the
stability of the closed loop, while the other desired performance of the controller, for
example tracking performance, can be guaranteed by the guide function, which, as new
controller parameters, can be designed without deteriorating the stability. The following
proposition is the main result of this subsection.

Proposition I:
If V(x) is a CLF of system (1) in
c
and (x): R
n
R
m
is a continuous guide function such that
(0) = 0, then, the following controller can stabilize system (1),


( )
( ) arg min { ( ) }
( ) { | ( ) ( ) ( ) ( ) ( ), }

V
u K x
V x x
u x u x
K x y V x f x V x g x y x y U
(6)

where (x) is a positive definite function of state, and (x), called guide function, is a
continuous state function.
Proof of Proposition I:
Let V(x) be a Lyapunov function candidate for system (1), then we have

( ) ( ) ( ) ( ) ( )
x x
V x V x f x V x g x u

(7)

Substitute Eq. (6) into (7), it is not difficult to obtain the following inequality,

( ) ( ) ( ) ( ) ( ) ( )
x x
V x V x f x V x g x u x

Because (x) is a positive definite function, proposition I is proved.

Controller (6) is called Generalized Pointwise Min-Norm (GPMN) controller. The difference
between the proposed GPMN controller and the normal PMN controller of Eq. (4) can be
illustrated in Fig.2: for the normal PMN algorithm (Fig. 2a), the controller output in each
state point has the minimum permitted norm (close to the state-axis as much as possible),
while the GPMN controllers output has nearest distance from the guide function (x) (Fig.
2b). Thus, (x) in GPMN controller is actual a performance criterion which the controller is
expected to pursue, while (x) dedicates only on providing the permitted stable control
input sets.
Up to now, the design of new GPMN controller has been completed. However, in order to
use a GPMN controller in reality or in NMPC algorithm, analytical form of the solution of
Eq. (6) is necessary to be studied.
Firstly, if there are no input constraints (or the input constraint sets are big enough), the
analytical form of controller (6) can be obtained as follows, based on the projection theory,
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 63

CLF based nonlinear controller design method is also called direct method of Lyapunov
function based controller design, and its difficulty is how to ensure the controllers
continuousness. Thus, most recently, researchers mainly pay their attentions to designing
continuous CLF based controller, and several universal formulas have been revealed.
Sontags formula (Sontag, 1989), for example, originated from the root calculation of 2
nd
-
order equation, can be written as Eq. (3) through slightly modification by Freeman (Freeman
& Kokotovic, 1996b),


2
( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
( ) ( )( )
0
0 0
T T
x x x x
x T T
x x
x
x x x x x x x
x x x x
V f V f q x V g g V
V g
u V g g V
V g
(
+ +
( =

=
(

(3)

where q(x) is a pre-designed positive definite function.
Pointwise Min-Norm (PMN) control is another well known CLF-based approach proposed
by Freeman (Freeman & Kokotovic, 1996a),


min
. . ( )[ ( ) ( ) ] ( )
u
x
u
s t V x f x g x u x
u U
o + s
e
(4)

where (x) is a pre-selected positive definite function. Controller (4) can also be explicitly
denoted as (5) if the constraint set U can be selected big enough.


[ ( ) ( ) ( )] ( ) ( )
( ) ( ) ( ) 0
( ) ( ) ( ) ( )
0 ( ) ( ) ( ) 0
T T
x x
x
T T
x x
x
V x f x x g x V x
V x f x x
u V x g x g x V x
V x f x x
o
o
o
+
+ >

+ s

(5)

(3) and (5) provide two different methods on how to design continuous and stable controller
based on CLF with respect to system (1). H

CLF with respect to system (2) is a new given

concept, and there are no methods can be used to designed robust controller based on it.
Although the closed loop stability can be guaranteed using controller (3) or controller (5),
selection of parameters q(x) or (x) is too difficult to be used in real applications. This is
mainly because these parameters heavily influence some inconsistent closed loop
performance simultaneously. Furthermore, if the known CLF is not global, the selection of
q(x) and (x) will also influence stability margin of the closed loop systems, which makes
them more difficult to be selected (Sontag, 1989; Freeman & Kokotovic, 1996a). In this
chapter, we will firstly give a new CLF based controller design strategy, which is superior
compared to the existing CLF based controller design methods referred to above.
Furthermore, the most important is that this new strategy can be used in designing robustly
stable and fast NMPC algorithm.


3. GPMN-ENMPC
3.1 CLF based GPMN controller
Since q(x) and (x) in controller (3) and controller (5) are difficult to select, a guide function is
proposed in this subsection into the PMN controller to obtain a new CLF based nonlinear
controller with respect to system (1), in the following section, this controller will be
generated with respect to system (2). In the new controller, (x) is only used to ensure the
stability of the closed loop, while the other desired performance of the controller, for
example tracking performance, can be guaranteed by the guide function, which, as new
controller parameters, can be designed without deteriorating the stability. The following
proposition is the main result of this subsection.

Proposition I:
If V(x) is a CLF of system (1) in
c
and (x): R
n
R
m
is a continuous guide function such that
(0) = 0, then, the following controller can stabilize system (1),


( )
( ) arg min { ( ) }
( ) { | ( ) ( ) ( ) ( ) ( ), }

V
u K x
V x x
u x u x
K x y V x f x V x g x y x y U
(6)

where (x) is a positive definite function of state, and (x), called guide function, is a
continuous state function.
Proof of Proposition I:
Let V(x) be a Lyapunov function candidate for system (1), then we have

( ) ( ) ( ) ( ) ( )
x x
V x V x f x V x g x u

(7)

Substitute Eq. (6) into (7), it is not difficult to obtain the following inequality,

( ) ( ) ( ) ( ) ( ) ( )
x x
V x V x f x V x g x u x

Because (x) is a positive definite function, proposition I is proved.

Controller (6) is called Generalized Pointwise Min-Norm (GPMN) controller. The difference
between the proposed GPMN controller and the normal PMN controller of Eq. (4) can be
illustrated in Fig.2: for the normal PMN algorithm (Fig. 2a), the controller output in each
state point has the minimum permitted norm (close to the state-axis as much as possible),
while the GPMN controllers output has nearest distance from the guide function (x) (Fig.
2b). Thus, (x) in GPMN controller is actual a performance criterion which the controller is
expected to pursue, while (x) dedicates only on providing the permitted stable control
input sets.
Up to now, the design of new GPMN controller has been completed. However, in order to
use a GPMN controller in reality or in NMPC algorithm, analytical form of the solution of
Eq. (6) is necessary to be studied.
Firstly, if there are no input constraints (or the input constraint sets are big enough), the
analytical form of controller (6) can be obtained as follows, based on the projection theory,
Model Predictive Control 64


( )
[ ( )]
( ) , ( ) 0
( )
( ) , ( ) 0
T T
x x x
x x T T
x x x
x x
V f V g x g V
x V f V g x
u x V gg V
x V f V g x

o
o
o
+ +
+ + + >

+ + s

(8)

Secondly, if there exist input constraints, the analytical expression of controller (6) might be
very complicated or even inexistent. Thus in this subsection, only analytical form of
controller (6) with a typical super ball input constraint is researched, i.e., input constraints is
as

2 2 2
1 1
{( , , ) | }
m m
U u u u u r = + + s (9)

where (u
1
, ,u
m
) is the input vector, and r is the radius of the super ball.
In order to obtain the analytical expression of Eq. (6) with input constraint as Eq. (9), we
propose the following 4 steps (For a general control input constraint U, one can always find
a maximal inscribed super ball B of it, and then use B replacing U before continuing the
following processes):


Fig. 2a. the sketch of PMN

Fig. 2b. the sketch of GPMN
* the dashed line is the PMN controller in a) and the GPMN control in b); the solid line
denotes the guide function of (x).
Input
State
Input
State

Step1: For each state x, the following equation denotes a super plane in R
m
(uR
m
).

( ) ( ) ( ) 0
x x
V f x x V g x u (10)

Let d be the distance from zero to the super plane (10), we have,


( ) ( )
( ) ( )
x
T T
x x
V f x x
d
V g x g x V

(11)


Fig. 3a

Fig. 3b.
* Sketch of the process to build the analytic GPMN controller

Step2: From Eq. (11), the permitted stable control input set K
V
(x) in controller (6) can be
denoted as Fig. 3a, where the right figure (left figure) is the case that the super plane of (10)
intersects (does not intersect) with the super ball (9), and the region filled by the dotted line
is the permitted stable control input set. For the case denoted by the left figure of Fig. 3a, it
is easy to obtain a minimal distance from any point p to K
V
(x), and the corresponding point,
i.e., the controllers output, in K
V
(x) with minimal distance from p can also be obtained (the
( ) ( )
( )
( ) ( )
T T x
x T T
x x
V f x x
g x V
V g x g x V

( )
( ) ( )
T T
x
T T
x x
r
g x V
V g x g x V

A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 65


( )
[ ( )]
( ) , ( ) 0
( )
( ) , ( ) 0
T T
x x x
x x T T
x x x
x x
V f V g x g V
x V f V g x
u x V gg V
x V f V g x

o
o
o
+ +
+ + + >

+ + s

(8)

Secondly, if there exist input constraints, the analytical expression of controller (6) might be
very complicated or even inexistent. Thus in this subsection, only analytical form of
controller (6) with a typical super ball input constraint is researched, i.e., input constraints is
as

2 2 2
1 1
{( , , ) | }
m m
U u u u u r = + + s (9)

where (u
1
, ,u
m
) is the input vector, and r is the radius of the super ball.
In order to obtain the analytical expression of Eq. (6) with input constraint as Eq. (9), we
propose the following 4 steps (For a general control input constraint U, one can always find
a maximal inscribed super ball B of it, and then use B replacing U before continuing the
following processes):


Fig. 2a. the sketch of PMN

Fig. 2b. the sketch of GPMN
* the dashed line is the PMN controller in a) and the GPMN control in b); the solid line
denotes the guide function of (x).
Input
State
Input
State

Step1: For each state x, the following equation denotes a super plane in R
m
(uR
m
).

( ) ( ) ( ) 0
x x
V f x x V g x u (10)

Let d be the distance from zero to the super plane (10), we have,


( ) ( )
( ) ( )
x
T T
x x
V f x x
d
V g x g x V

(11)


Fig. 3a

Fig. 3b.
* Sketch of the process to build the analytic GPMN controller

Step2: From Eq. (11), the permitted stable control input set K
V
(x) in controller (6) can be
denoted as Fig. 3a, where the right figure (left figure) is the case that the super plane of (10)
intersects (does not intersect) with the super ball (9), and the region filled by the dotted line
is the permitted stable control input set. For the case denoted by the left figure of Fig. 3a, it
is easy to obtain a minimal distance from any point p to K
V
(x), and the corresponding point,
i.e., the controllers output, in K
V
(x) with minimal distance from p can also be obtained (the
( ) ( )
( )
( ) ( )
T T x
x T T
x x
V f x x
g x V
V g x g x V

( )
( ) ( )
T T
x
T T
x x
r
g x V
V g x g x V

Model Predictive Control 66

point of intersection of the super ball (9) and the beeline connecting the centre of it and p).
With respect to the case of the right figure, the maximally inscribed super ball B is used to
replace K
V
(x) (see Fig. 3b). Thus, the same processes as above can be used to obtain the
output of controller (6).
Step 3: A new permitted stable control input sets ( )
V
K x is defined,


2
( ) ( )
( ) ( )
( )
( ) ( )
{ | ( ) ( )}
( ) ( )
x
T T
x x
V
x
T T
x x
V f x x
U r
V g x g x V
K x
V f x x
u u x R x r
V g x g x V
o
o

+
>

=

+

= s

(12)
where
( ) ( )
( ) ( ) ( )
2 ( ) ( )
2 ( ) ( )
[ ( ) ( )]
( ) ( )
( )
2
T T x
x T T
T T
x x
x x
x
T T
x x
V f x x r
x g x V
V g x g x V
V g x g x V
V f x x
r
V g x g x V
R x
o

o
+
= +
+

=


It is obvious that ( ) ( )
V V
K x K x _ , thus the stability of the closed loop can be ensured from
Proposition I.
Step 4: The analytical expression of GPMN controller with super-ball input constraint can
thus be described as


( )
( ) ( ) ( ) ( )
( )
( )
[ ( ) ( )] ( )
( ) ( )
x
x x x R x
u x
R x
x x x else
x x

=

+

(13)

where (x) is the guide function of controller (6).
From the preceding procedure, it is evidently that Eq. (13) is the solution of Eq. (6) with
K
V
(x) being placed by ( )
V
K x .

3.2 GPMN-ENMPC
In order to achieve a stable NMPC with reduced computational burden, we propose to use
the GPMN to parameterize the control input sequence in NMPC. Assuming that ( , ) x 0 u is a
function of state x, where is the vector of unknown parameters, the following NMPC can
be formulated,


* *
*
( , )
arg min ( , )
( , ) ( , ( , ))
. . ( ) ( ) ( , )
( , ) , [ , ]
l
IR
t T
t
u x
J x
J x l x x d
s t x f x g x x
x U t t t T
u
0 u
u u
u 0 u t
0 u
0 u
c
+
=
=
=
= +
e e +
}

(14)

NMPC algorithm of (14) is different from the normal NMPC in the following aspect: in
normal NMPC algorithm, one tries to optimize the continuous control profile of u (Mayne et
al., 2000), while controller (14) tries to achieve good performance by optimizing the
parameter vector . Thus, the computational cost of controller (14) dependents mainly on
dimension of instead of that of control input profile in normal NMPC algorithm. The most
important problem of the latter algorithm is that its computational cost increases rapidly
with the control horizon. Based on (14), our new designed NMPC controller is introduced in
the following proposition.

Proposition II:
Assuming V(x) is a known CLF of system (1),
c
is the stability region of V(x), then
controller (14) with the following GPMN controller ( , ) x 0 u ,


( )
( , ) ( , ) arg min { ( , ) }
V
u K x
x u x u x 0 u u u
e
= = (15)

(u(x,) is the GPMN control and (x,) the guide function in Eq. (6)), is stable in
c
.
Furthermore, if V(x) is a global CLF, controller of (14) combined with (15) is stable over R
n
.
(14), combined with (15), is called GPMN-Enhanced NMPC (GPMN-ENMPC).
Proof of Proposition II:
At any time instant t, by assuming that
*
is the optimal parameters at t, control input at t
can be represented as u(x,
*
). From Proposition I, we can conclude that the control inputs
u(x,
*
) can guarantee a negative definite ( ) V x

. Due to the randomicity of t, GPMN-ENMPC

actually makes the ( ) V x

negative in any time instant, which means that the closed loop
stability of controller (14) and (15) is guaranteed.

3.3 Selection of (x,)
Theoretically, (x,) in (15) can be selected in any forms since it does not influence the closed
loop stability, which is guaranteed by GPMN. However, it is natural that (x,) will
influence other closed loop performances of the GPMN-ENMPC except the stability.
Since optimality is the main concern in designing NMPC algorithm, the Bellmans
Optimization Principle (BOP, Lewis & Syrmos, 1995) is used to design (x,) in this sub-
section.
In BOP, with the following quadratic cost function,
( , ) ( )
t T
T T
t
J x x Px u Qu d u t
A
+
= +
}
(16)
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 67

point of intersection of the super ball (9) and the beeline connecting the centre of it and p).
With respect to the case of the right figure, the maximally inscribed super ball B is used to
replace K
V
(x) (see Fig. 3b). Thus, the same processes as above can be used to obtain the
output of controller (6).
Step 3: A new permitted stable control input sets ( )
V
K x is defined,


2
( ) ( )
( ) ( )
( )
( ) ( )
{ | ( ) ( )}
( ) ( )
x
T T
x x
V
x
T T
x x
V f x x
U r
V g x g x V
K x
V f x x
u u x R x r
V g x g x V
o
o

+
>

=

+

= s

(12)
where
( ) ( )
( ) ( ) ( )
2 ( ) ( )
2 ( ) ( )
[ ( ) ( )]
( ) ( )
( )
2
T T x
x T T
T T
x x
x x
x
T T
x x
V f x x r
x g x V
V g x g x V
V g x g x V
V f x x
r
V g x g x V
R x
o

o
+
= +
+

=


It is obvious that ( ) ( )
V V
K x K x _ , thus the stability of the closed loop can be ensured from
Proposition I.
Step 4: The analytical expression of GPMN controller with super-ball input constraint can
thus be described as


( )
( ) ( ) ( ) ( )
( )
( )
[ ( ) ( )] ( )
( ) ( )
x
x x x R x
u x
R x
x x x else
x x

=

+

(13)

where (x) is the guide function of controller (6).
From the preceding procedure, it is evidently that Eq. (13) is the solution of Eq. (6) with
K
V
(x) being placed by ( )
V
K x .

3.2 GPMN-ENMPC
In order to achieve a stable NMPC with reduced computational burden, we propose to use
the GPMN to parameterize the control input sequence in NMPC. Assuming that ( , ) x 0 u is a
function of state x, where is the vector of unknown parameters, the following NMPC can
be formulated,


* *
*
( , )
arg min ( , )
( , ) ( , ( , ))
. . ( ) ( ) ( , )
( , ) , [ , ]
l
IR
t T
t
u x
J x
J x l x x d
s t x f x g x x
x U t t t T
u
0 u
u u
u 0 u t
0 u
0 u
c
+
=
=
=
= +
e e +
}

(14)

NMPC algorithm of (14) is different from the normal NMPC in the following aspect: in
normal NMPC algorithm, one tries to optimize the continuous control profile of u (Mayne et
al., 2000), while controller (14) tries to achieve good performance by optimizing the
parameter vector . Thus, the computational cost of controller (14) dependents mainly on
dimension of instead of that of control input profile in normal NMPC algorithm. The most
important problem of the latter algorithm is that its computational cost increases rapidly
with the control horizon. Based on (14), our new designed NMPC controller is introduced in
the following proposition.

Proposition II:
Assuming V(x) is a known CLF of system (1),
c
is the stability region of V(x), then
controller (14) with the following GPMN controller ( , ) x 0 u ,


( )
( , ) ( , ) arg min { ( , ) }
V
u K x
x u x u x 0 u u u
e
= = (15)

(u(x,) is the GPMN control and (x,) the guide function in Eq. (6)), is stable in
c
.
Furthermore, if V(x) is a global CLF, controller of (14) combined with (15) is stable over R
n
.
(14), combined with (15), is called GPMN-Enhanced NMPC (GPMN-ENMPC).
Proof of Proposition II:
At any time instant t, by assuming that
*
is the optimal parameters at t, control input at t
can be represented as u(x,
*
). From Proposition I, we can conclude that the control inputs
u(x,
*
) can guarantee a negative definite ( ) V x

. Due to the randomicity of t, GPMN-ENMPC

actually makes the ( ) V x

negative in any time instant, which means that the closed loop
stability of controller (14) and (15) is guaranteed.

3.3 Selection of (x,)
Theoretically, (x,) in (15) can be selected in any forms since it does not influence the closed
loop stability, which is guaranteed by GPMN. However, it is natural that (x,) will
influence other closed loop performances of the GPMN-ENMPC except the stability.
Since optimality is the main concern in designing NMPC algorithm, the Bellmans
Optimization Principle (BOP, Lewis & Syrmos, 1995) is used to design (x,) in this sub-
section.
In BOP, with the following quadratic cost function,
( , ) ( )
t T
T T
t
J x x Px u Qu d u t
A
+
= +
}
(16)
Model Predictive Control 68

and J
*
(x
0
,) denoting the optimal value function of J(x
0
,) in state x
0
, the following controller
of system (1) is optimal,

*
* 1
1
( ) ( )
2
T T
J
u Q g x
x

c
=
c
(17)

Unfortunately, in most applications, it is impossible to obtain J
*
(x
*
,).
Based on the Stone-Weierstrass theorem (Brinkhuis & Tikhomirov, 2005), any continuous
function defined in a bounded set can be uniformly approximated by a polynomial function,


*
1
1
1
* 1
1 ; , , 1
, , 0
( , , ) ( , , ) ( , , )
n
n
n
J n
k n k v v n
v v
v v k
v v
B x x J p x x
k k
>
+ + s
=

(18)
where

1 1
1
; , , 1 1 1
1
1 1 2 1
( , , ) (1 ) ,
, ,
!
,
, , ! ! !( )!
n n
n
v k v v v
k v v n n n
n
n n n
k
p x x x x x x
v v
k
k
v v v v v k v v

| |
=
|
\ .
| |
=
|

\ .

(19)
and

*
1
*
, , 1 1
( 1, , )
lim ( , , ) ( , , )
n
i
J
k k n n
k
i k
B x x J x x

=
=

(20)

Thus, take the coefficients of the Bernstein polynomial as the parameters , and select
optimally using the NMPC algorithm, a quasi-optimal function closed to J
*
(x
*
,) can be
obtained. That means we can complete the design of GPMN-ENMPC algorithm by taking


1 1
1
1
, , ; , , 1
, , 0
( , ) ( , , )
n n
n
n
v v k v v n
v v
v v k
x p x x u
>
+ + s
=

(21)

where
1
,...,
n
v v
,
1
,...,
n
v v 0 and
1
...
n
v v + + k are the parameters to be optimized, k is the
order of the Bernstein polynomial, and


1 2
, ,...,
1
[ ]
k
n
k k k
n
u

= (22)

It should be noted that the order of the Bernstein polynomial determines the consequent
optimization cost, i.e., the higher the order is, the higher the computational cost is. About
the GPMN-ENMPC, we have the following remarks:
Remark-1: Selection of (x,) as Eq. (21) provides a feasible way to complete the GPMN-
ENMPC of (20) and (21). By this way, the computation cost is controllable, namely, one can
select the order of k to meet the CPU capability of a specific real system. This makes the
GPMN-ENMPC feasible to be implemented.


Remark-2: The selection of k does not influence the closed loop stability, which has already
guaranteed by the GPMN scheme. But there still exist trade-offs between computation cost
and the optimal performance which is determined by (x,).

Remark-3: Compared to nominal NMPC algorithm, the huge computational burden
problem of GPMN-ENMPC algorithm is improved due to the following two reasons: 1) the
dimension of optimizing variables is one of key elements which increase the computational
burden of NMPC, while that of GPMN-ENMPC algorithm is independent of the predictive
horizon; 2) online considerations of control input constraints are not necessary in GPMN-
ENMPC algorithm since it can be dealt with offline during designing GPMN controller.

3.4 The Feasibility of GPMN-ENMPC
Another important problem, normally called the feasibility problem of NMPC, is that
general NMPC algorithm may not guarantee that a control set always exists to meet all of
the input and state constraints, while the proposed GPMN-ENMPC can guarantee such a
control sequence always exists. This is because for any , from the proposition-I, one can
always obtain a stable GPMN controller, i.e., u(x,) of (6) meeting all input and state
constraints. Therefore, by Eq. (14) and (15), there will always exist a feasible control
u = ( , ) x , and the task left is just to find an optimal parameter set of to minimize the cost
function of J(x,) in Eq. (14).

4. H

GPMN-ENMPC
In section 3, GPMN-ENMPC algorithm is introduced with respect to system (1). In this
section, it will be generalized to deal with the disturbed system as Eq. (2). Firstly, an H

controller with partially known disturbances is given, and then it is used to design
H

GPMN controller, which followed by the designing process of H

GPMN-ENMPC.

4.1 H

Control With Partially Known Disturbances

Suppose the following two assumptions are satisfied with respect to system (2),

Assumption I:
System (2) is static feedback linearizable, i.e., there exists a state feedback controller u = k(x)
such that (2) can be transformed into a linear system without considering .

Assumption II:
The disturbances of system (2) are partially obtainable, i.e., the variables can be used to
construct controller.

Assumption II is reasonable because the uncertainty information can often be measured or
estimated in reality (He & Han, 2007; Chen, 2004). Moreover, the tracking problem of
general nonlinear system, where is composed of known desired trajectory, can also be
modeled as Eq. (2). However, the higher order derivative of the disturbances with respective
to time is often difficult to be obtained due to the heavy additive noise. Thus, the
disturbances are often partially obtainable.
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 69

and J
*
(x
0
,) denoting the optimal value function of J(x
0
,) in state x
0
, the following controller
of system (1) is optimal,

*
* 1
1
( ) ( )
2
T T
J
u Q g x
x

c
=
c
(17)

Unfortunately, in most applications, it is impossible to obtain J
*
(x
*
,).
Based on the Stone-Weierstrass theorem (Brinkhuis & Tikhomirov, 2005), any continuous
function defined in a bounded set can be uniformly approximated by a polynomial function,


*
1
1
1
* 1
1 ; , , 1
, , 0
( , , ) ( , , ) ( , , )
n
n
n
J n
k n k v v n
v v
v v k
v v
B x x J p x x
k k
>
+ + s
=

(18)
where

1 1
1
; , , 1 1 1
1
1 1 2 1
( , , ) (1 ) ,
, ,
!
,
, , ! ! !( )!
n n
n
v k v v v
k v v n n n
n
n n n
k
p x x x x x x
v v
k
k
v v v v v k v v

| |
=
|
\ .
| |
=
|

\ .

(19)
and

*
1
*
, , 1 1
( 1, , )
lim ( , , ) ( , , )
n
i
J
k k n n
k
i k
B x x J x x

=
=

(20)

Thus, take the coefficients of the Bernstein polynomial as the parameters , and select
optimally using the NMPC algorithm, a quasi-optimal function closed to J
*
(x
*
,) can be
obtained. That means we can complete the design of GPMN-ENMPC algorithm by taking


1 1
1
1
, , ; , , 1
, , 0
( , ) ( , , )
n n
n
n
v v k v v n
v v
v v k
x p x x u
>
+ + s
=

(21)

where
1
,...,
n
v v
,
1
,...,
n
v v 0 and
1
...
n
v v + + k are the parameters to be optimized, k is the
order of the Bernstein polynomial, and


1 2
, ,...,
1
[ ]
k
n
k k k
n
u

= (22)

It should be noted that the order of the Bernstein polynomial determines the consequent
optimization cost, i.e., the higher the order is, the higher the computational cost is. About
the GPMN-ENMPC, we have the following remarks:
Remark-1: Selection of (x,) as Eq. (21) provides a feasible way to complete the GPMN-
ENMPC of (20) and (21). By this way, the computation cost is controllable, namely, one can
select the order of k to meet the CPU capability of a specific real system. This makes the
GPMN-ENMPC feasible to be implemented.


Remark-2: The selection of k does not influence the closed loop stability, which has already
guaranteed by the GPMN scheme. But there still exist trade-offs between computation cost
and the optimal performance which is determined by (x,).

Remark-3: Compared to nominal NMPC algorithm, the huge computational burden
problem of GPMN-ENMPC algorithm is improved due to the following two reasons: 1) the
dimension of optimizing variables is one of key elements which increase the computational
burden of NMPC, while that of GPMN-ENMPC algorithm is independent of the predictive
horizon; 2) online considerations of control input constraints are not necessary in GPMN-
ENMPC algorithm since it can be dealt with offline during designing GPMN controller.

3.4 The Feasibility of GPMN-ENMPC
Another important problem, normally called the feasibility problem of NMPC, is that
general NMPC algorithm may not guarantee that a control set always exists to meet all of
the input and state constraints, while the proposed GPMN-ENMPC can guarantee such a
control sequence always exists. This is because for any , from the proposition-I, one can
always obtain a stable GPMN controller, i.e., u(x,) of (6) meeting all input and state
constraints. Therefore, by Eq. (14) and (15), there will always exist a feasible control
u = ( , ) x , and the task left is just to find an optimal parameter set of to minimize the cost
function of J(x,) in Eq. (14).

4. H

GPMN-ENMPC
In section 3, GPMN-ENMPC algorithm is introduced with respect to system (1). In this
section, it will be generalized to deal with the disturbed system as Eq. (2). Firstly, an H

controller with partially known disturbances is given, and then it is used to design
H

GPMN controller, which followed by the designing process of H

GPMN-ENMPC.

4.1 H

Control With Partially Known Disturbances

Suppose the following two assumptions are satisfied with respect to system (2),

Assumption I:
System (2) is static feedback linearizable, i.e., there exists a state feedback controller u = k(x)
such that (2) can be transformed into a linear system without considering .

Assumption II:
The disturbances of system (2) are partially obtainable, i.e., the variables can be used to
construct controller.

Assumption II is reasonable because the uncertainty information can often be measured or
estimated in reality (He & Han, 2007; Chen, 2004). Moreover, the tracking problem of
general nonlinear system, where is composed of known desired trajectory, can also be
modeled as Eq. (2). However, the higher order derivative of the disturbances with respective
to time is often difficult to be obtained due to the heavy additive noise. Thus, the
disturbances are often partially obtainable.
Model Predictive Control 70

Based on assumption I, system (2) can be changed into the following equations through
some coordination transformation,


1 2 1
1 1
1
( )
( ) ( ) ( )
n n
z z F z
z f z g z u F z
y z
= + A
= + + A
=

(23)
where z = [z
1
,z
2
,,z
n
]
T
is the new state variable.
An H

robust controller for system (23) can be designed based on the following Theorem,

Theorem I:
Consider system (23), if there exists a control u = u
1
(z) and a radially unbounded function
V(x) to satisfy the following inequality,


1
1 1 1 1 1 2 2
1
2
1 2 1
2
[ ( ) ( ) ( )] ( ) ( ) ( )
( ) ( ) ( ) 0
i n
n
T
T T T
z i z n
i
T T T T
n
V z V f z g z u z V F z F z F z
F z F z F z V z

+
=
( + + +

( + s

z
z
(24)

Then, controller


1 ( 1) ( ) 1
1 1 1 1 1 2 1
1
1 1 1 1
2 2 2 2
2
( 1) ( 1)
( )[ ( ) ( ) ( ) ( , , , ) ] [ ( )]
( ) ( , ) ( ) ( , )
( ) ( , , ) ( ) ( , , ) 2
{
( ) ( , , , ) ( ) ( , , , )
n
n
i n i
z
i
n n
n n n n
u g z f z g z u f z F z V g z
F z F z F z F z
F z F z F z F z
V
F z F z F z F z

=

= + + +
(
(

(
(
(

(

z
}
T
T
V
(
(
(
(
(
(

z
(25)

can make the system (23) finite gain L
2
stable from + to y, and the gain is less than or
equal to . is a new defined signal to further attenuate the disturbances.
Proof of Theorem I:
Define new variables,


1 1
2 2 1
1
( )
1
( )
( )
n
n i
n n i
i
z z
z z F z
z z F z

=
=
=
=

(26)

Then, system (23) can be written as


1 2 1
2 3 2
( 1) ( ) ( 1)
1 1
1
1
( , )( )
( , , )( )
( ) ( ) ( , , , ) ( , , , )( )
n
i n i n
n i n
i
z z F z
z z F z
z f z g z u F z F z
y z




=
= + A +
= + A +
= + + A +
=

(27)

where

1
( )
1
( 1)
( )
( , , , ) ( )
i
n i
i j i
j
j
j j
z z F z
F z F z

= +

Let
( ) ( ) | V z V z
=
=
z z
(28)

where | |
T
1 2 n
z = z z z . Computing the HJI equation (Khalil, 2002) of system (27)
with respect to ( ) V z , we have,


1
( 1) ( )
1 1 1 2
1 1
( 1)
1 2 2
( 1) 2
1 2 1
[ ( ) ( ) ( , , , ) ]
2
( , ) ( , , ) ( , , , )
( , ) ( , , ) ( , , , )
i n
n n
i n i
z i z
i i
T
T T n
n
T T n T
n
V z V f z g z u F z
V F z F z F z
F z F z F z V z

+
= =

+ + +
(

( +





z
z
(29)

Thus, combine controller (25) and Eq. (29), we have,


1
( 1) ( )
1 1 1 2
1 1
( 1)
1 2
2
( 1) 2
1 2 1
1
1 1 1 2
1
(29) [ ( ) ( ) ( , , , ) ]
2
( , ) ( , , ) ( , , , )
( , ) ( , , ) ( , , , )
2
{ [ ( ) ( ) ]

+
= =

+
=
= + + +
(

( +

= + + +

i n
i n
n n
i n i
z i z
i i
T
T T T n
n
T T T n T
n
n
z i z
i
V z V f z g z u F z
V F z F z F z
F z F z F z V z
V z V f z g z u
z
z
1 2
2
1 2 1
( ) ( ) ( )
( ) ( ) ( ) }|
0
=
(

( +

s

T
T T T
n
T T T T
n z z
V F z F z F z
F z F z F z V z
z
z
(30)
Based on theorem 5.5 in reference (Khalil, 2002), controller (25) can make system (23) finite
gain L
2
stable from + to y, and the L
2
gain is less than or equal to .

A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 71

Based on assumption I, system (2) can be changed into the following equations through
some coordination transformation,


1 2 1
1 1
1
( )
( ) ( ) ( )
n n
z z F z
z f z g z u F z
y z
= + A
= + + A
=

(23)
where z = [z
1
,z
2
,,z
n
]
T
is the new state variable.
An H

robust controller for system (23) can be designed based on the following Theorem,

Theorem I:
Consider system (23), if there exists a control u = u
1
(z) and a radially unbounded function
V(x) to satisfy the following inequality,


1
1 1 1 1 1 2 2
1
2
1 2 1
2
[ ( ) ( ) ( )] ( ) ( ) ( )
( ) ( ) ( ) 0
i n
n
T
T T T
z i z n
i
T T T T
n
V z V f z g z u z V F z F z F z
F z F z F z V z

+
=
( + + +

( + s

z
z
(24)

Then, controller


1 ( 1) ( ) 1
1 1 1 1 1 2 1
1
1 1 1 1
2 2 2 2
2
( 1) ( 1)
( )[ ( ) ( ) ( ) ( , , , ) ] [ ( )]
( ) ( , ) ( ) ( , )
( ) ( , , ) ( ) ( , , ) 2
{
( ) ( , , , ) ( ) ( , , , )
n
n
i n i
z
i
n n
n n n n
u g z f z g z u f z F z V g z
F z F z F z F z
F z F z F z F z
V
F z F z F z F z

=

= + + +
(
(

(
(
(

(

z
}
T
T
V
(
(
(
(
(
(

z
(25)

can make the system (23) finite gain L
2
stable from + to y, and the gain is less than or
equal to . is a new defined signal to further attenuate the disturbances.
Proof of Theorem I:
Define new variables,


1 1
2 2 1
1
( )
1
( )
( )
n
n i
n n i
i
z z
z z F z
z z F z

=
=
=
=

(26)

Then, system (23) can be written as


1 2 1
2 3 2
( 1) ( ) ( 1)
1 1
1
1
( , )( )
( , , )( )
( ) ( ) ( , , , ) ( , , , )( )
n
i n i n
n i n
i
z z F z
z z F z
z f z g z u F z F z
y z




=
= + A +
= + A +
= + + A +
=

(27)

where

1
( )
1
( 1)
( )
( , , , ) ( )
i
n i
i j i
j
j
j j
z z F z
F z F z

= +

Let
( ) ( ) | V z V z
=
=
z z
(28)

where | |
T
1 2 n
z = z z z . Computing the HJI equation (Khalil, 2002) of system (27)
with respect to ( ) V z , we have,


1
( 1) ( )
1 1 1 2
1 1
( 1)
1 2 2
( 1) 2
1 2 1
[ ( ) ( ) ( , , , ) ]
2
( , ) ( , , ) ( , , , )
( , ) ( , , ) ( , , , )
i n
n n
i n i
z i z
i i
T
T T n
n
T T n T
n
V z V f z g z u F z
V F z F z F z
F z F z F z V z

+
= =

+ + +
(

( +





z
z
(29)

Thus, combine controller (25) and Eq. (29), we have,


1
( 1) ( )
1 1 1 2
1 1
( 1)
1 2
2
( 1) 2
1 2 1
1
1 1 1 2
1
(29) [ ( ) ( ) ( , , , ) ]
2
( , ) ( , , ) ( , , , )
( , ) ( , , ) ( , , , )
2
{ [ ( ) ( ) ]

+
= =

+
=
= + + +
(

( +

= + + +

i n
i n
n n
i n i
z i z
i i
T
T T T n
n
T T T n T
n
n
z i z
i
V z V f z g z u F z
V F z F z F z
F z F z F z V z
V z V f z g z u
z
z
1 2
2
1 2 1
( ) ( ) ( )
( ) ( ) ( ) }|
0
=
(

( +

s

T
T T T
n
T T T T
n z z
V F z F z F z
F z F z F z V z
z
z
(30)
Based on theorem 5.5 in reference (Khalil, 2002), controller (25) can make system (23) finite
gain L
2
stable from + to y, and the L
2
gain is less than or equal to .

Model Predictive Control 72

Furthermore, can be used to further attenuate the disturbances which are partially
obtainable from assumption II by the following equation,


( )
( ) ( )
( )
B s
s s
A s
= A (31)

where s is the Laplace operator. Thus, the new external disturbances + can be denoted as,


( ) ( )
( ) ( ) ( )
( )
A s B s
s s s
A s

+
A + = A (32)

From Eq. (32), proper A(s) and B(s) is effective for attenuating the influence of external
disturbances on the closed loop system. Thus, we have designed an H

controller (25) and

(31) with partially known uncertainty information.

4.2 H

GPMN Controller Based on Control Lyapunov Functions

In this sub-section, by using the concept of H

CLF, H

GPMN controller is designed as

following proposition,

Proposition III:
If V(x) is a local H

CLF of system (23), and (x): R

n
R
m
is a continuous guide function such
that (0) = 0, then, the following controller, called H

GPMN, can make system (23) finite

gain L
2
stable from A to output y,


( )
( ) arg min{ ( ) }
H
V
H
u K x
u x u x

e
= (33)
where

2
1 1
( ) { ( ) : [ ( ) ( ) ] ( ) ( ) ( ) ( ) ( )}
2 2
H T T T
V x x x
K x u U x V f x g x u V l x l x V h x h x x o

= e + + + s (34)

Proof of Proposition III can be easily done based on the definition of finite gain L
2
stability
and H

CLF. The analytical form of controller (33) can also be obtained as steps in section 3.
Here only the analytical form of controller without input constraints is given,

2
1 1
[ ( ) ]
2 2
0
( )
0
T T T T T
x x x
H
T T
x x
V f ll V g h h g V
u x
V gg V
o

+ + + +

>
=

(35)
where
; ( ); ( ); ( );
( ); ( ); ( ); ( )
x x
x x
V f V g f f x g g x x
x V V x h h x l l x
o
o o
= + + = = =
= = = =


It is not difficult to show that H

GPMN satisfies inequality (24) of Theorem I, thus, it can be

used as u
1
(z) in controller (25) to bring the advantages of H

GPMN controller to the robust

controller in section 4.1.

4.3 H

GPMN-ENMPC
As far as the external disturbances are concerned, nominal model based NMPC, where the
prediction is made through a nominal certain system model, is an often used strategy in
reality. And the formulation of it is very similar to non-robust NMPC, so dose the GPMN-
ENMPC.


Fig. 4. Structure of new designed RNRHC controller

However, for disturbed nonlinear system like Eq. (23), GPMN-ENMPC algorithm can
hardly be used in real applications due to weak robustness. Thus, in this subsection, we will
combine it to the robust controller from sub-section 4.1 and sub-section 4.2 to overcome the
drawbacks originated from both GPMN-ENMPC algorithm and the robust controller (25)
and (35). The structure of the new parameterized H

GPMN-ENMPC algorithm based on

controller (25) and (35) is as Fig. 4.
Eq. (36) is the new designed H

GPMN-ENMPC algorithm. Compared to Eq. (14), it is easy

to find out that the control input in the H

GPMN-ENMPC algorithm has a pre-defined

structure given in section 4.1 and 4.2.
Uncertain No-
nlinear System
Feedback lineari-
zation z=T(x)
Robust control-
er with partially
obtainable distu-
rbances (25)
RGPMN
HGPMN controller
(35)
z
GPMN-ENMPC
*
u1(z)
*
( , )
( )
H
x
u z

( , )
( )
H
x
u z

x
x
Feedback lineari-
zation z=T(x)
z
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 73

Furthermore, can be used to further attenuate the disturbances which are partially
obtainable from assumption II by the following equation,


( )
( ) ( )
( )
B s
s s
A s
= A (31)

where s is the Laplace operator. Thus, the new external disturbances + can be denoted as,


( ) ( )
( ) ( ) ( )
( )
A s B s
s s s
A s

+
A + = A (32)

From Eq. (32), proper A(s) and B(s) is effective for attenuating the influence of external
disturbances on the closed loop system. Thus, we have designed an H

controller (25) and

(31) with partially known uncertainty information.

4.2 H

GPMN Controller Based on Control Lyapunov Functions

In this sub-section, by using the concept of H

CLF, H

GPMN controller is designed as

following proposition,

Proposition III:
If V(x) is a local H

CLF of system (23), and (x): R

n
R
m
is a continuous guide function such
that (0) = 0, then, the following controller, called H

GPMN, can make system (23) finite

gain L
2
stable from A to output y,


( )
( ) arg min{ ( ) }
H
V
H
u K x
u x u x

e
= (33)
where

2
1 1
( ) { ( ) : [ ( ) ( ) ] ( ) ( ) ( ) ( ) ( )}
2 2
H T T T
V x x x
K x u U x V f x g x u V l x l x V h x h x x o

= e + + + s (34)

Proof of Proposition III can be easily done based on the definition of finite gain L
2
stability
and H

CLF. The analytical form of controller (33) can also be obtained as steps in section 3.
Here only the analytical form of controller without input constraints is given,

2
1 1
[ ( ) ]
2 2
0
( )
0
T T T T T
x x x
H
T T
x x
V f ll V g h h g V
u x
V gg V
o

+ + + +

>
=

(35)
where
; ( ); ( ); ( );
( ); ( ); ( ); ( )
x x
x x
V f V g f f x g g x x
x V V x h h x l l x
o
o o
= + + = = =
= = = =


It is not difficult to show that H

GPMN satisfies inequality (24) of Theorem I, thus, it can be

used as u
1
(z) in controller (25) to bring the advantages of H

GPMN controller to the robust

controller in section 4.1.

4.3 H

GPMN-ENMPC
As far as the external disturbances are concerned, nominal model based NMPC, where the
prediction is made through a nominal certain system model, is an often used strategy in
reality. And the formulation of it is very similar to non-robust NMPC, so dose the GPMN-
ENMPC.


Fig. 4. Structure of new designed RNRHC controller

However, for disturbed nonlinear system like Eq. (23), GPMN-ENMPC algorithm can
hardly be used in real applications due to weak robustness. Thus, in this subsection, we will
combine it to the robust controller from sub-section 4.1 and sub-section 4.2 to overcome the
drawbacks originated from both GPMN-ENMPC algorithm and the robust controller (25)
and (35). The structure of the new parameterized H

GPMN-ENMPC algorithm based on

controller (25) and (35) is as Fig. 4.
Eq. (36) is the new designed H

GPMN-ENMPC algorithm. Compared to Eq. (14), it is easy

to find out that the control input in the H

GPMN-ENMPC algorithm has a pre-defined

structure given in section 4.1 and 4.2.
Uncertain No-
nlinear System
Feedback lineari-
zation z=T(x)
Robust control-
er with partially
obtainable distu-
rbances (25)
RGPMN
HGPMN controller
(35)
z
GPMN-ENMPC
*
u1(z)
*
( , )
( )
H
x
u z

( , )
( )
H
x
u z

x
x
Feedback lineari-
zation z=T(x)
z
Model Predictive Control 74


* *
*
( , )
arg min ( , )
( , ) ( ( ), ( ))
. . ( ) ( )
( ) ( , )
H
u U
t T
t
H
u u x
J x u
J x u l x u d
s t x f x g x u
u t u x

(36)

5. Practical Considering
Both GPMN-ENMPC algorithm and H

GPMN-ENMPC algorithm can be divided into two

processes, including the implementation process and the optimization process as Fig.5.


Fig. 5. The process of (H

)GPMN-ENMPC

The implementation process and the optimization process in Fig. 5 are independent. In
implementation process, the (H

)GPMN scheme is used to ensure the closed loop (L

2
)
stability, and in the optimization process, the optimization algorithm is responsible to
improving the optimality of the controller. And the interaction of the two processes is
realized through the optimized parameter * (from optimization process to implementation
process) and the measured states (form implementation process to optimization process).

5.1 Time Interval Between Two Neighboring Optimizing Process
Sample time in controller implemented using computer is often very short, especially in
mechatronic system. This is very challenging to implement complicated algorithm, such as
GPMN-ENMPC in this chapter. Fortunately, the optimization process of the new designed
controller will end up with a group of parameters which are used to form a stable
(H

)GPMN controller, and the optimization process itself does not influence the closed loop
stability at all. Thus, theoretically, any group of optimized parameters can be used for
several sample intervals without destroying the closed loop stability.
Computing control input
based on (H

)GPMN scheme
Computing the optimal
parameter
*
by solving an
optimal control problem
Optimized parameter
*

Implementation process
Current state x
t


Optimization process

Fig.6 denotes the scheduling of (H

)GPMN-ENMPC algorithm. In Fig.6, t is the current time

instant; T is the prediction horizon; T
S
is the sample time of the (H

)GPMN controller; and T

I

is the duration of every optimal parameter
*
(t), i.e., the same parameter
*
is used to
implement the (H

)GPMN controller from time t to time t+T

I
.


Fig. 6. Scheduling of ERNRHC

5.2 Numerical Integrator
How to predict the futures behavior is very important during the implementation of any
kind of MPC algorithms. In most applications, the NMPC algorithm is realized by
computers. Thus, for the continuous systems, it will be difficult and time consuming if some
accurate but complicated numerical integration methods are used, such as Newton-Cotes
integration and Gaussian quadratures, etc. In this chapter, we will discretize the continuous
system (1) as follows (take system (1) as an example),

( ) ( ( )) ( ( )) ( )
O O O O O O O
x kT T f x kT T g x kT u kT T (37)

where T
o
is the discrete sample time. Thus, the numerical integrator can be approached by
the operation of cumulative addition.

5.3 Index Function
Replace x(kT
o
) with x(k), the index function can be designed as follows,


0
0
* *
0
( ( ), ) ( ( ), )
k N
T
c l l c
i k
J x k J x i

(38)

where k
0
denotes the current time instant; N is the predictive horizon with N=Int(T/T
o
) (here
Int(*) is the operator to obtain a integer nearest to *);
c
is the parameter vector to be
optimized at current time instant; and
l
*
is the last optimization result; Q, Z, R are constant
matrix with Q>0, Z>0, and R0.
The new designed item
l
T*
Z
l
*
is used to reduce the difference between two neighboring
optimized parameter vector, and improve the smoothness of the optimized control inputs u.

T
T
I

T
S


t
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 75


* *
*
( , )
arg min ( , )
( , ) ( ( ), ( ))
. . ( ) ( )
( ) ( , )
H
u U
t T
t
H
u u x
J x u
J x u l x u d
s t x f x g x u
u t u x

(36)

5. Practical Considering
Both GPMN-ENMPC algorithm and H

GPMN-ENMPC algorithm can be divided into two

processes, including the implementation process and the optimization process as Fig.5.


Fig. 5. The process of (H

)GPMN-ENMPC

The implementation process and the optimization process in Fig. 5 are independent. In
implementation process, the (H

)GPMN scheme is used to ensure the closed loop (L

2
)
stability, and in the optimization process, the optimization algorithm is responsible to
improving the optimality of the controller. And the interaction of the two processes is
realized through the optimized parameter * (from optimization process to implementation
process) and the measured states (form implementation process to optimization process).

5.1 Time Interval Between Two Neighboring Optimizing Process
Sample time in controller implemented using computer is often very short, especially in
mechatronic system. This is very challenging to implement complicated algorithm, such as
GPMN-ENMPC in this chapter. Fortunately, the optimization process of the new designed
controller will end up with a group of parameters which are used to form a stable
(H

)GPMN controller, and the optimization process itself does not influence the closed loop
stability at all. Thus, theoretically, any group of optimized parameters can be used for
several sample intervals without destroying the closed loop stability.
Computing control input
based on (H

)GPMN scheme
Computing the optimal
parameter
*
by solving an
optimal control problem
Optimized parameter
*

Implementation process
Current state x
t


Optimization process

Fig.6 denotes the scheduling of (H

)GPMN-ENMPC algorithm. In Fig.6, t is the current time

instant; T is the prediction horizon; T
S
is the sample time of the (H

)GPMN controller; and T

I

is the duration of every optimal parameter
*
(t), i.e., the same parameter
*
is used to
implement the (H

)GPMN controller from time t to time t+T

I
.


Fig. 6. Scheduling of ERNRHC

5.2 Numerical Integrator
How to predict the futures behavior is very important during the implementation of any
kind of MPC algorithms. In most applications, the NMPC algorithm is realized by
computers. Thus, for the continuous systems, it will be difficult and time consuming if some
accurate but complicated numerical integration methods are used, such as Newton-Cotes
integration and Gaussian quadratures, etc. In this chapter, we will discretize the continuous
system (1) as follows (take system (1) as an example),

( ) ( ( )) ( ( )) ( )
O O O O O O O
x kT T f x kT T g x kT u kT T (37)

where T
o
is the discrete sample time. Thus, the numerical integrator can be approached by
the operation of cumulative addition.

5.3 Index Function
Replace x(kT
o
) with x(k), the index function can be designed as follows,


0
0
* *
0
( ( ), ) ( ( ), )
k N
T
c l l c
i k
J x k J x i

(38)

where k
0
denotes the current time instant; N is the predictive horizon with N=Int(T/T
o
) (here
Int(*) is the operator to obtain a integer nearest to *);
c
is the parameter vector to be
optimized at current time instant; and
l
*
is the last optimization result; Q, Z, R are constant
matrix with Q>0, Z>0, and R0.
The new designed item
l
T*
Z
l
*
is used to reduce the difference between two neighboring
optimized parameter vector, and improve the smoothness of the optimized control inputs u.

T
T
I

T
S


t
Model Predictive Control 76

6. Numerical Examples
6.1 Example1 (GPMN-ENMPC without control input constrains)
Consider the following pendulum equation (Costa & do Va, 2003),

1 2
2
1 2 1 1
2
2 2
1 1
19.6sin 0.2 sin 2 0.2cos
4/ 3 0.2cos 4/ 3 0.2cos
x x
x x x x
x u
x x
=

= +

(39)

A local CLF of system (39) can be given as,


| |
1
1 2
2
151.57 42.36
( )
42.36 12.96
T
x
V x x Px x x
x
( (
= =
( (

(40)
Select

2 2
1 2
( ) 0.1( ) x x x o = + (41)

The normal PMN control can be designed according to (5) as,

2
1
1 1 2
2
1 2 1
1 2 2 1 2
2
1
2 2
2 1 2
( )(4/ 3 0.2cos )
( ) 0
0.4cos (42.36 12.96 )
0 ( ) 0
19.6sin 0.2 sin 2
( ) 2[(151.57 42.36 ) (42.36 12.96 ) ]
4/ 3 0.2cos
(10.54 1.27 )
x x
x
u x x x
x
x x x
x x x x x x
x
x x x

<

= +

= + + + +

+ +
(42)

Given initial state x
0
= [x
1
,x
2
]
T
= [-1,2]
T
, and desired state x
d
= [0,0]
T
, time response of the
closed loop for PMN controller is shown in Fig. 7 in solid line. It can be seen that the closed
loop with PMN controller (42) has a very low convergence rate for state x
1
. This is mainly
because the only regulable parameter to change the closed loop performance is (x), which is
difficult to be properly selected due to its great influence on the stability region.
To design GPMN-ENMPC, two different guide functions are selected based on Eq. (21),


0,0 1 2 1,0 1 0,1 2
( , ) (1 ) x x x x x u u u u = + + (43)

2 2 2
0,0 1 2 0,1 2 1,0 1 1 2 1,1 1 2 0,2 2 2,0 1
( , ) (1 ) 2( )(1 ) 2 x x x x x x x x x x x u u u u u u u = + + + + + (44)

CLF V(x) and (x) are given in Eq. (40) and Eq. (41), and others conditions in GPMN-
ENMPC are designed as follows,


2
0
20 0
( 0.01 )
0 1
T
T
J x x u dt
(
= +
(

}
(45)


2
2
2
1 2 1
2
1
1
2
1
20 0
( , ) 0.01 ; ( ) ;
19.6sin 0.2 sin 2
0 1
4/ 3 0.2cos
0
( ) ; 0.1 0.2cos
4/ 3 0.2cos
T
x
l x u x x u f x
x x x
x
g x z I x
x
(
(
(
= + =

(
(

(

(
(
= =
(
(

(46)

Integral time interval T
o
in Eq. (37) is 0.1s. Genetic algorithm (GA) in MATLAB toolbox is
used to solve the online optimization problem. Time response of GPMN-ENMPC algorithm
with different predictive horizon T and approaching order are presented in Fig. 7, where the
dotted line denotes the case of T = 0.6s with guide function (43), and the dashed line is the
case of T = 1.5s with guide function (44). From Fig. 7, it can be seen that the convergence
performance of the proposed NMPC algorithm is better than PMN controller, and both the
prediction horizon and the guide function will result in the change of the closed loop
performance.
The improvement of the optimality is the main advantage of MPC compared with others
controller. In view of this, we propose to estimate the optimality by the following index
function,

2
0
20 0
( 0.01 )
0 1
T
J lim x x u dt
I
I
(
= +
(

}
(47)

0 1 2 3 4 5
-1
-0.5
0
x
1
0 1 2 3 4 5
0
1
2
time (s)
x
2

PMN
ENMPC (1,0.6)
ENMPC (2,1.5)

Fig. 7. Time response of different controller, where the (a,b) indicates that the order of
( , ) x u is a, and the predictive horizon b

The comparison results are summarized in Table 1, from which the following conclusions
can be obtained, 1) GPMN-ENMPC has better optimizing performance than PMN controller
in terms of optimization. 2) In most cases, GPMN-ENMPC with higher order (x,) will
usually result in a smaller cost than that with lower order (x,). This is mainly because
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 77

6. Numerical Examples
6.1 Example1 (GPMN-ENMPC without control input constrains)
Consider the following pendulum equation (Costa & do Va, 2003),

1 2
2
1 2 1 1
2
2 2
1 1
19.6sin 0.2 sin 2 0.2cos
4/ 3 0.2cos 4/ 3 0.2cos
x x
x x x x
x u
x x
=

= +

(39)

A local CLF of system (39) can be given as,


| |
1
1 2
2
151.57 42.36
( )
42.36 12.96
T
x
V x x Px x x
x
( (
= =
( (

(40)
Select

2 2
1 2
( ) 0.1( ) x x x o = + (41)

The normal PMN control can be designed according to (5) as,

2
1
1 1 2
2
1 2 1
1 2 2 1 2
2
1
2 2
2 1 2
( )(4/ 3 0.2cos )
( ) 0
0.4cos (42.36 12.96 )
0 ( ) 0
19.6sin 0.2 sin 2
( ) 2[(151.57 42.36 ) (42.36 12.96 ) ]
4/ 3 0.2cos
(10.54 1.27 )
x x
x
u x x x
x
x x x
x x x x x x
x
x x x

<

= +

= + + + +

+ +
(42)

Given initial state x
0
= [x
1
,x
2
]
T
= [-1,2]
T
, and desired state x
d
= [0,0]
T
, time response of the
closed loop for PMN controller is shown in Fig. 7 in solid line. It can be seen that the closed
loop with PMN controller (42) has a very low convergence rate for state x
1
. This is mainly
because the only regulable parameter to change the closed loop performance is (x), which is
difficult to be properly selected due to its great influence on the stability region.
To design GPMN-ENMPC, two different guide functions are selected based on Eq. (21),


0,0 1 2 1,0 1 0,1 2
( , ) (1 ) x x x x x u u u u = + + (43)

2 2 2
0,0 1 2 0,1 2 1,0 1 1 2 1,1 1 2 0,2 2 2,0 1
( , ) (1 ) 2( )(1 ) 2 x x x x x x x x x x x u u u u u u u = + + + + + (44)

CLF V(x) and (x) are given in Eq. (40) and Eq. (41), and others conditions in GPMN-
ENMPC are designed as follows,


2
0
20 0
( 0.01 )
0 1
T
T
J x x u dt
(
= +
(

}
(45)


2
2
2
1 2 1
2
1
1
2
1
20 0
( , ) 0.01 ; ( ) ;
19.6sin 0.2 sin 2
0 1
4/ 3 0.2cos
0
( ) ; 0.1 0.2cos
4/ 3 0.2cos
T
x
l x u x x u f x
x x x
x
g x z I x
x
(
(
(
= + =

(
(

(

(
(
= =
(
(

(46)

Integral time interval T
o
in Eq. (37) is 0.1s. Genetic algorithm (GA) in MATLAB toolbox is
used to solve the online optimization problem. Time response of GPMN-ENMPC algorithm
with different predictive horizon T and approaching order are presented in Fig. 7, where the
dotted line denotes the case of T = 0.6s with guide function (43), and the dashed line is the
case of T = 1.5s with guide function (44). From Fig. 7, it can be seen that the convergence
performance of the proposed NMPC algorithm is better than PMN controller, and both the
prediction horizon and the guide function will result in the change of the closed loop
performance.
The improvement of the optimality is the main advantage of MPC compared with others
controller. In view of this, we propose to estimate the optimality by the following index
function,

2
0
20 0
( 0.01 )
0 1
T
J lim x x u dt
I
I
(
= +
(

}
(47)

0 1 2 3 4 5
-1
-0.5
0
x
1
0 1 2 3 4 5
0
1
2
time (s)
x
2

PMN
ENMPC (1,0.6)
ENMPC (2,1.5)

Fig. 7. Time response of different controller, where the (a,b) indicates that the order of
( , ) x u is a, and the predictive horizon b

The comparison results are summarized in Table 1, from which the following conclusions
can be obtained, 1) GPMN-ENMPC has better optimizing performance than PMN controller
in terms of optimization. 2) In most cases, GPMN-ENMPC with higher order (x,) will
usually result in a smaller cost than that with lower order (x,). This is mainly because
Model Predictive Control 78

higher order (x,) indicates larger inherent optimizing parameter space. 3) A longer
prediction horizon will usually be followed by a better optimal performance.

J
ENMPC PMN
x
0
= (-1,2) x
0
= (0.5,1)
x
0
= (-
1,2)
x
0
=
(0.5,1)
k = 1 k = 2
K =
1
k =
2
----
T=0.6 29.39 28.87 6.54 6.26 + +
T=0.8 23.97

23.83

5.02

4.96

+ +
T=1.0 24.08 24.07 4.96 4.90 + +
T=1.5 26.31 24.79 5.11 5.28 + +
Table 1. the cost value of different controller
* k is the order of Bernstein polynomial used to approach the optimal value function; T is the
predictive horizon; x
0
is the initial state

Another advantage of the GPMN-ENMPC algorithm is the flexibility of the trade offs
between the optimality and the computational time. The computational time is influenced
by the dimension of optimizing parameters and the parameters of the optimizing algorithm,
such as the maximum number of iterations and the size of the population (the smaller these
values are selected, the less the computational cost is). However, it will be natural that the
optimality maybe deteriorated to some extent with the decreasing of the computational
burden. In preceding paragraphs, we have researched the optimality of GPMN-ENMPC
algorithm with different optimizing parameters, and now the optimality comparisons
among the closed loop systems with different GA parameters will be done. And the results
are listed in Table 2, from which the certain of the optimality loss with the changing of the
optimizing algorithms parameters can be observed. This can be used as the criterion to
determine the trade-off between the closed loop performance and the computational
efficiency of the algorithm.

OP
G=100
PS=50
G=50
PS=50
G=50
PS=30
G=50
PS=20
G=50
PS=10
cost

26.2 28.1

30.8 43.5 45.7
Table 2. The relation between the computational cost and the optimality
*x
0
= (-1,2), T=1.5, k = 1, OP means Optimization Prameters, G means Generations, PS means
Population Size

Finally, in order to verify that the new designed algorithm is improved in the computational
burden, simulations comparing the performance of the new designed algorithm and
algorithm in (Primbs, 1999) are conducted with the same optimizing algorithm. Time
interval of two neighboured optimization (T
I
in Table 3) in Primbs algorithm is important
since control input is assumed to be constant at every time slice. Generally, large time
interval will result in poor stability.
While our new GPMN-ENMPC results in a group of controller parameter, and the closed loop
stability is independent of T
I
. Thus different T
I
is considered in these simulations of Primbs

algorithm and Table 3 lists the results. From Table 3, the following items can be concluded: 1)
with same GA parameters, Primbs algorithm is more time-consuming and poorer in optimality
than GPMN-ENMPC. This is easy to be obtained through comparing results of Ex-2 and Ex-5; 2)
in order to obtain similar optimality, GPMN-ENMPC takes much less time than Primbs
algorithm. This can be obtained by comparing results of Ex-1/Ex-4 and Ex-6, as well as Ex-3 and
Ex-5. The reasons for these phenomena have been introduced in Remark 3.


Algorithm in (Primbs, 1999) GPMN-ENMPC
Ex-1 Ex-2 Ex-3 Ex-4 Ex-5 Ex-6
TI 0.1 0.05 0.1
OP
G=100
PS=50
G=50
PS=50
G=100
PS=50
G=50
PS=50
G=50
PS=50
G=50
PS=30
Average Time
Consumption
2.2075 1.8027 2.9910 2.2463 1.3961 0.8557
Cost

31.2896 35.7534 27.7303 31.8055 28.1 31.1043
Table 3. Performance comparison of GPMN-ENMPC and Primbs algorithm
*x
0
= (-1,2), TI means time interval of two neighbored optimization; OP means Optimization
Prameters; G means Generations, PS means Population Size. Other parameters of GPMN-
ENMPC are T=1.5, k = 1

6.2 Example 2 (GPMN-ENMPC with control input constraint)
In order to show the performance of the GPMN-ENMPC in handling input constraints, we
give another simulation using the dynamics of a mobile robot with orthogonal wheel
assemblies (Song, 2007). The dynamics can be denoted as Eq. (48),

( ) ( ) x f x g x u = + (48)
where
2
4 6 2
4
2 6 4
6
6
5 5 5 5 5
5 5 5 5 5
2.3684 0.5921
( )
2.3684 0.5921
0.2602
0 0 0
0.8772( 3sin cos ) 0.8772*2cos 0.8772( 3sin cos )
0 0 0
( )
0.8772( 3 cos sin ) 0.8772*2sin 0.8772( 3 cos sin )
0 0 0
(
(

(
(
=
(

(
(
(

(


=

x
x x x
x
f x
x x x
x
x
x x x x x
g x
x x x x x
-1.4113 -1.4113 -1.4113
(
(
(
(
(
(
(
(
(


A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 79

higher order (x,) indicates larger inherent optimizing parameter space. 3) A longer
prediction horizon will usually be followed by a better optimal performance.

J
ENMPC PMN
x
0
= (-1,2) x
0
= (0.5,1)
x
0
= (-
1,2)
x
0
=
(0.5,1)
k = 1 k = 2
K =
1
k =
2
----
T=0.6 29.39 28.87 6.54 6.26 + +
T=0.8 23.97

23.83

5.02

4.96

+ +
T=1.0 24.08 24.07 4.96 4.90 + +
T=1.5 26.31 24.79 5.11 5.28 + +
Table 1. the cost value of different controller
* k is the order of Bernstein polynomial used to approach the optimal value function; T is the
predictive horizon; x
0
is the initial state

Another advantage of the GPMN-ENMPC algorithm is the flexibility of the trade offs
between the optimality and the computational time. The computational time is influenced
by the dimension of optimizing parameters and the parameters of the optimizing algorithm,
such as the maximum number of iterations and the size of the population (the smaller these
values are selected, the less the computational cost is). However, it will be natural that the
optimality maybe deteriorated to some extent with the decreasing of the computational
burden. In preceding paragraphs, we have researched the optimality of GPMN-ENMPC
algorithm with different optimizing parameters, and now the optimality comparisons
among the closed loop systems with different GA parameters will be done. And the results
are listed in Table 2, from which the certain of the optimality loss with the changing of the
optimizing algorithms parameters can be observed. This can be used as the criterion to
determine the trade-off between the closed loop performance and the computational
efficiency of the algorithm.

OP
G=100
PS=50
G=50
PS=50
G=50
PS=30
G=50
PS=20
G=50
PS=10
cost

26.2 28.1

30.8 43.5 45.7
Table 2. The relation between the computational cost and the optimality
*x
0
= (-1,2), T=1.5, k = 1, OP means Optimization Prameters, G means Generations, PS means
Population Size

Finally, in order to verify that the new designed algorithm is improved in the computational
burden, simulations comparing the performance of the new designed algorithm and
algorithm in (Primbs, 1999) are conducted with the same optimizing algorithm. Time
interval of two neighboured optimization (T
I
in Table 3) in Primbs algorithm is important
since control input is assumed to be constant at every time slice. Generally, large time
interval will result in poor stability.
While our new GPMN-ENMPC results in a group of controller parameter, and the closed loop
stability is independent of T
I
. Thus different T
I
is considered in these simulations of Primbs

algorithm and Table 3 lists the results. From Table 3, the following items can be concluded: 1)
with same GA parameters, Primbs algorithm is more time-consuming and poorer in optimality
than GPMN-ENMPC. This is easy to be obtained through comparing results of Ex-2 and Ex-5; 2)
in order to obtain similar optimality, GPMN-ENMPC takes much less time than Primbs
algorithm. This can be obtained by comparing results of Ex-1/Ex-4 and Ex-6, as well as Ex-3 and
Ex-5. The reasons for these phenomena have been introduced in Remark 3.


Algorithm in (Primbs, 1999) GPMN-ENMPC
Ex-1 Ex-2 Ex-3 Ex-4 Ex-5 Ex-6
TI 0.1 0.05 0.1
OP
G=100
PS=50
G=50
PS=50
G=100
PS=50
G=50
PS=50
G=50
PS=50
G=50
PS=30
Average Time
Consumption
2.2075 1.8027 2.9910 2.2463 1.3961 0.8557
Cost

31.2896 35.7534 27.7303 31.8055 28.1 31.1043
Table 3. Performance comparison of GPMN-ENMPC and Primbs algorithm
*x
0
= (-1,2), TI means time interval of two neighbored optimization; OP means Optimization
Prameters; G means Generations, PS means Population Size. Other parameters of GPMN-
ENMPC are T=1.5, k = 1

6.2 Example 2 (GPMN-ENMPC with control input constraint)
In order to show the performance of the GPMN-ENMPC in handling input constraints, we
give another simulation using the dynamics of a mobile robot with orthogonal wheel
assemblies (Song, 2007). The dynamics can be denoted as Eq. (48),

( ) ( ) x f x g x u = + (48)
where
2
4 6 2
4
2 6 4
6
6
5 5 5 5 5
5 5 5 5 5
2.3684 0.5921
( )
2.3684 0.5921
0.2602
0 0 0
0.8772( 3sin cos ) 0.8772*2cos 0.8772( 3sin cos )
0 0 0
( )
0.8772( 3 cos sin ) 0.8772*2sin 0.8772( 3 cos sin )
0 0 0
(
(

(
(
=
(

(
(
(

(


=

x
x x x
x
f x
x x x
x
x
x x x x x
g x
x x x x x
-1.4113 -1.4113 -1.4113
(
(
(
(
(
(
(
(
(


Model Predictive Control 80

1 2 3 4 5 6
; ; ; ; ;
w w w w w w
x x x x x y x y x x = = = = = = ; x
w
, y
w
,
w
are respective the x-y positions
and yaw angle; u
1
, u
2
, u
3
are motor torques.
Suppose that control input is limited in the following closed set,

U = {( u
1
, u
2
, u
3
)|( u
1
2
+ u
2
2
+ u
3
2
)
1/2
20} (49)

System (48) is feedback linearizable, and by which we can obtain a CLF of system (48) as
follows,
( )
T
V x x Px = (50)
where
1.125 0.125 0 0 0 0
0.125 0.156 0 0 0 0
0 0 1.125 0.125 0 0
0 0 0.125 0.156 0 0
0 0 0 0 1.125 0.125
0 0 0 0 0.125 1.156
P
(
(
(
(
=
(
(
(
(
(



The cost function J(x) and (x) are designed as,


0
0
2 2 2 2 2 2 2 2 2
1 3 5 2 4 6 1 2 3
2 2 2 2 2 2
1 2 3 4 5 6
( ) (3 3 3 5 5 5 ) ( 1) ( 1);
( ) 0.1( ); =0.1
t T
T
t
J x x x x x x x u u u dt k Z k
x x x x x x x Z I
u u
o
+
= + + + + + + + + +
= + + + + +
}
(51)

System (48) has 6 states and 3 inputs, which will introduce large computational burden if
using the GPMN-ENMPC method. Fortunately, one of the advantages of GPMN-ENMPC is
that the optimization does not destroy the closed loop stability. Thus, in order to reduce the
computation burden, we reduce the frequency of the optimization in this simulation, i.e.,
one optimization process is conducted every 0.1s while the controller of (13) is calculated
every 0.002s, i.e., T
I
= 0.1s, T
s
= 0.002s.

2 4 6 8 10
-5
0
5
10
15
x
1
2 4 6 8 10
-5
0
5
x
2
2 4 6 8 10
-15
-10
-5
0
5
x
3
2 4 6 8 10
-5
0
5
x
4
2 4 6 8 10
-0.5
0
0.5
1
1.5
time(s)
x
5
2 4 6 8 10
-1
0
1
time(s)
x
6
2 4 6 8 10 12 14 16 18 20
-16
-14
-12
-10
-8
-6
-4
-2
0
2
time(s)
u
3

a) states response b) control input u
1


2 4 6 8 10 12 14 16 18 20
-1
0
1
2
3
4
time(s)
u
2

2 4 6 8 10 12 14 16 18 20
-2
0
2
4
6
8
10
12
14
time(s)
u
1

c) control input u
2
d) control input u
3
Fig. 8. GPMN-ENMPC controller simulation results on the mobile robot with input
constraints

Initial States
(x1; x2; x3; x4; x5; x6)
Feedback
linearization
controller
GPMN-NMPC
(10; 5; 10; 5; 1; 0) 2661.7 1377.0
(10; 5; 10; 5; -1; 0) 3619.5 1345.5
(-10; -5; 10; 5; 1; 0) 2784.9 1388.5
(-10; -5; 10; 5; -1; 0) 8429.2 1412.0
(-10; -5; -10; -5; 1; 0) 394970.0 1349.9
(-10; -5; -10; -5; -1; 0) 4181.6 1370.9
(10; 5; -10; -5; 1; 0) 3322 1406
(10; 5; -10; -5; -1; 0) 1574500000 1452.1
(-5; -2; -10; -5; 1; 0) 1411.2 856.1
(-10; -5; -5; -2; 1; 0) 1547.5 850.9
Table 4. The comparison of the optimality

Simulation results are shown in Fig.8 with the initial state (10; 5; -10; -5; 1; 0), From Fig.8, it is
clear that GPMN-ENMPC controller has the ability to handling input constraints.
In order to evaluate the optimal performance of the GPMN-ENMPC, we proposed the
following cost function according to Eq. (51),


2 2 2 2 2 2 2 2 2
1 3 5 2 4 6 1 2 3
0
cos t lim (3 3 3 5 5 5 ) x x x x x x u u u dt

(52)

Table 4 lists the costs by feedback linearization controller and GPMN-ENMPC for several
different initial states, from which it can be seen that the cost of GPMN-ENMPC is less than
the half of the cost of feedback linearization controller when the initial is (10; 5; -10; -5; 1; 0).
And in most cases listed in Table 4, the cost of GPMN-ENMPC is about one second of that of
feedback linearization controller. Actually, in some special cases, such as the initial of (10; 5;
-10; -5; -1; 0), the cost ratio of feedback linearization controller to GPMN-ENMPC is more
than 1000000.

A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 81

1 2 3 4 5 6
; ; ; ; ;
w w w w w w
x x x x x y x y x x = = = = = = ; x
w
, y
w
,
w
are respective the x-y positions
and yaw angle; u
1
, u
2
, u
3
are motor torques.
Suppose that control input is limited in the following closed set,

U = {( u
1
, u
2
, u
3
)|( u
1
2
+ u
2
2
+ u
3
2
)
1/2
20} (49)

System (48) is feedback linearizable, and by which we can obtain a CLF of system (48) as
follows,
( )
T
V x x Px = (50)
where
1.125 0.125 0 0 0 0
0.125 0.156 0 0 0 0
0 0 1.125 0.125 0 0
0 0 0.125 0.156 0 0
0 0 0 0 1.125 0.125
0 0 0 0 0.125 1.156
P
(
(
(
(
=
(
(
(
(
(



The cost function J(x) and (x) are designed as,


0
0
2 2 2 2 2 2 2 2 2
1 3 5 2 4 6 1 2 3
2 2 2 2 2 2
1 2 3 4 5 6
( ) (3 3 3 5 5 5 ) ( 1) ( 1);
( ) 0.1( ); =0.1
t T
T
t
J x x x x x x x u u u dt k Z k
x x x x x x x Z I
u u
o
+
= + + + + + + + + +
= + + + + +
}
(51)

System (48) has 6 states and 3 inputs, which will introduce large computational burden if
using the GPMN-ENMPC method. Fortunately, one of the advantages of GPMN-ENMPC is
that the optimization does not destroy the closed loop stability. Thus, in order to reduce the
computation burden, we reduce the frequency of the optimization in this simulation, i.e.,
one optimization process is conducted every 0.1s while the controller of (13) is calculated
every 0.002s, i.e., T
I
= 0.1s, T
s
= 0.002s.

2 4 6 8 10
-5
0
5
10
15
x
1
2 4 6 8 10
-5
0
5
x
2
2 4 6 8 10
-15
-10
-5
0
5
x
3
2 4 6 8 10
-5
0
5
x
4
2 4 6 8 10
-0.5
0
0.5
1
1.5
time(s)
x
5
2 4 6 8 10
-1
0
1
time(s)
x
6
2 4 6 8 10 12 14 16 18 20
-16
-14
-12
-10
-8
-6
-4
-2
0
2
time(s)
u
3

a) states response b) control input u
1


2 4 6 8 10 12 14 16 18 20
-1
0
1
2
3
4
time(s)
u
2

2 4 6 8 10 12 14 16 18 20
-2
0
2
4
6
8
10
12
14
time(s)
u
1

c) control input u
2
d) control input u
3
Fig. 8. GPMN-ENMPC controller simulation results on the mobile robot with input
constraints

Initial States
(x1; x2; x3; x4; x5; x6)
Feedback
linearization
controller
GPMN-NMPC
(10; 5; 10; 5; 1; 0) 2661.7 1377.0
(10; 5; 10; 5; -1; 0) 3619.5 1345.5
(-10; -5; 10; 5; 1; 0) 2784.9 1388.5
(-10; -5; 10; 5; -1; 0) 8429.2 1412.0
(-10; -5; -10; -5; 1; 0) 394970.0 1349.9
(-10; -5; -10; -5; -1; 0) 4181.6 1370.9
(10; 5; -10; -5; 1; 0) 3322 1406
(10; 5; -10; -5; -1; 0) 1574500000 1452.1
(-5; -2; -10; -5; 1; 0) 1411.2 856.1
(-10; -5; -5; -2; 1; 0) 1547.5 850.9
Table 4. The comparison of the optimality

Simulation results are shown in Fig.8 with the initial state (10; 5; -10; -5; 1; 0), From Fig.8, it is
clear that GPMN-ENMPC controller has the ability to handling input constraints.
In order to evaluate the optimal performance of the GPMN-ENMPC, we proposed the
following cost function according to Eq. (51),


2 2 2 2 2 2 2 2 2
1 3 5 2 4 6 1 2 3
0
cos t lim (3 3 3 5 5 5 ) x x x x x x u u u dt

(52)

Table 4 lists the costs by feedback linearization controller and GPMN-ENMPC for several
different initial states, from which it can be seen that the cost of GPMN-ENMPC is less than
the half of the cost of feedback linearization controller when the initial is (10; 5; -10; -5; 1; 0).
And in most cases listed in Table 4, the cost of GPMN-ENMPC is about one second of that of
feedback linearization controller. Actually, in some special cases, such as the initial of (10; 5;
-10; -5; -1; 0), the cost ratio of feedback linearization controller to GPMN-ENMPC is more
than 1000000.

Model Predictive Control 82

6.3 Example 3 (H

GPMN-ENMPC)
In this section, a simulation will be given to verify the feasibility of the proposed H

GPMN-
ENMPC algorithm with respect to the following planar dynamic model of helicopter,


1
2
2 2
3
4
9.8cos sin
9.8sin
0.05 sin cos tan (0.5+0.05cos )
0.07cos tan sin tan
0.05 sin cos 0.07sin cos
y
L M
M
| u
|
| u | | |u u |
| u | u
u |u | | | |
+ A

= + A

= + +

+ + + A

= + + A

x =
(53)

where
1
,
2
,
3
,
4
are all the external disturbances, and are selected as following values,

1 2
3
4
3; 3
10sin(0.5 )
10sin(0.5 )
t
t
A = A =

A =

A =

Firstly, design an H

CLF of system (53) by using the feedback linearization method,

T
V X PX = (54)

where,

[ , , , , , , , ]
14.48 11.45 3.99 0.74 0 0 0 0
11.45 9.77 3.44 0.66 0 0 0 0
3.99 3.44 1.28 0.24 0 0 0 0
0.74 0.66 0.24 0.05 0 0 0 0
0 0 0 0 14.48 11.45 3.99 0.74
0 0 0 0 11.45 9.77 3.44 0.66
0 0 0 0 3.99 3.44 1.28 0.24
0 0 0 0 0.74 0.66 0.24 0.05
T
X x x x x y y y y
P
=
=

(
(
(
(
(
(
(
(
(
(
(
(



Thus, the robust predictive controller can be designed as Eq. (25), (35) and (36) with the
following parameters,

* * T T
1
1 2 3 4 5 6 7 8 9
10 11 12 13 14 15 16 17 18
( )
[ ( ) ( ) ( ) ( )]
( ) =
50000 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 50000 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0
T
N
T
l l o o o o o
i
x X X
J I x iT Px iT u iT Qu iT T
x x y y
x,
x x y y
P
o
u u
| | | | | | | | | | u | u
|
| | | | | | | | | | u | u
=
=
= + +
( + + + + + + + +
(
+ + + + + + + +

=

0 0 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1
1 0
; 0.1 ; 0.02 ; 1 ; 20;
0 1
o s I
Q T s T s T s N Z = I
(
(
(
(
(
(
(
(
(
(
(
(

(
= = = = =
(



Time response of the H

GPMN-ENMPC is as solid line of Fig.9 and Fig.10. Furthermore, the

comparisons between the performance of the closed loop controlled by the proposed
H

GPMN-ENMPC and some other controller design method are done. The dashed line in
Fig.9 and Fig.10 is the time response of the feedback linearization controller. From Fig.9 and
Fig.10, the disturbance attenuation performance of the H

GPMN-ENMPC is apparently
better than that of feedback linearization controller, because the penalty gain of position
signals, being much larger than other terms, can be used to further improve the ability.

0 4 8 12 16 20 24
-0.05
0
0.05
x
0 4 8 12 16 20 24
-0.05
0
0.05
x
d
o
t
0 4 8 12 16 20 24
-0.05
0
0.05
y
0 4 8 12 16 20 24
-0.05
0
0.05
y
d
o
t
0 4 8 12 16 20 24
-0.32
-0.31
-0.3
|
0 4 8 12 16 20 24
-0.1
0
0.1
|
d
o
t
0 4 8 12 16 20 24
0.32
0.33
0.34
time(s)
u
0 4 8 12 16 20 24
-0.1
0
0.1
time(s)
u
d
o
t

Fig. 9. Time response of states
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 83

6.3 Example 3 (H

GPMN-ENMPC)
In this section, a simulation will be given to verify the feasibility of the proposed H

GPMN-
ENMPC algorithm with respect to the following planar dynamic model of helicopter,


1
2
2 2
3
4
9.8cos sin
9.8sin
0.05 sin cos tan (0.5+0.05cos )
0.07cos tan sin tan
0.05 sin cos 0.07sin cos
y
L M
M
| u
|
| u | | |u u |
| u | u
u |u | | | |
+ A

= + A

= + +

+ + + A

= + + A

x =
(53)

where
1
,
2
,
3
,
4
are all the external disturbances, and are selected as following values,

1 2
3
4
3; 3
10sin(0.5 )
10sin(0.5 )
t
t
A = A =

A =

A =

Firstly, design an H

CLF of system (53) by using the feedback linearization method,

T
V X PX = (54)

where,

[ , , , , , , , ]
14.48 11.45 3.99 0.74 0 0 0 0
11.45 9.77 3.44 0.66 0 0 0 0
3.99 3.44 1.28 0.24 0 0 0 0
0.74 0.66 0.24 0.05 0 0 0 0
0 0 0 0 14.48 11.45 3.99 0.74
0 0 0 0 11.45 9.77 3.44 0.66
0 0 0 0 3.99 3.44 1.28 0.24
0 0 0 0 0.74 0.66 0.24 0.05
T
X x x x x y y y y
P
=
=

(
(
(
(
(
(
(
(
(
(
(
(



Thus, the robust predictive controller can be designed as Eq. (25), (35) and (36) with the
following parameters,

* * T T
1
1 2 3 4 5 6 7 8 9
10 11 12 13 14 15 16 17 18
( )
[ ( ) ( ) ( ) ( )]
( ) =
50000 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 50000 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0
T
N
T
l l o o o o o
i
x X X
J I x iT Px iT u iT Qu iT T
x x y y
x,
x x y y
P
o
u u
| | | | | | | | | | u | u
|
| | | | | | | | | | u | u
=
=
= + +
( + + + + + + + +
(
+ + + + + + + +

=

0 0 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 1
1 0
; 0.1 ; 0.02 ; 1 ; 20;
0 1
o s I
Q T s T s T s N Z = I
(
(
(
(
(
(
(
(
(
(
(
(

(
= = = = =
(



Time response of the H

GPMN-ENMPC is as solid line of Fig.9 and Fig.10. Furthermore, the

comparisons between the performance of the closed loop controlled by the proposed
H

GPMN-ENMPC and some other controller design method are done. The dashed line in
Fig.9 and Fig.10 is the time response of the feedback linearization controller. From Fig.9 and
Fig.10, the disturbance attenuation performance of the H

GPMN-ENMPC is apparently
better than that of feedback linearization controller, because the penalty gain of position
signals, being much larger than other terms, can be used to further improve the ability.

0 4 8 12 16 20 24
-0.05
0
0.05
x
0 4 8 12 16 20 24
-0.05
0
0.05
x
d
o
t
0 4 8 12 16 20 24
-0.05
0
0.05
y
0 4 8 12 16 20 24
-0.05
0
0.05
y
d
o
t
0 4 8 12 16 20 24
-0.32
-0.31
-0.3
|
0 4 8 12 16 20 24
-0.1
0
0.1
|
d
o
t
0 4 8 12 16 20 24
0.32
0.33
0.34
time(s)
u
0 4 8 12 16 20 24
-0.1
0
0.1
time(s)
u
d
o
t

Fig. 9. Time response of states
Model Predictive Control 84

0 4 8 12 16 20 24
-20
-10
0
10
20
L
0 4 8 12 16 20 24
-20
-10
0
10
20
time(s)
M

Fig. 10. Control inputs

Simultaneously, the following index is used to compare the optimality of the two different
controllers,


0
T T
0
lim [ ( ) ( ) ( ) ( )] t J x t Px t u t Qu t d

(55)

The optimality performance of H

GPMN-ENMPC, computed from Eq. (55), is about 3280,

and the feedback linearization controller is about 5741, i.e., the H

GPMN-ENMPC has better

optimality than the feedback linearization controller.

7. Conclusion
In this paper, nonlinear model predictive control (NMPC) is researched and a new NMPC
algorithm is proposed. The new designed NMPC algorithm, called GPMN-enhancement
NMPC (GPMN-ENMPC), has the following three advantages: 1) closed loop stability can be
always guaranteed; 2) performance other than optimality and stability can be considered in
the new algorithm through selecting proper guide function; 3) computational cost of the
new NMPC algorithm is regulable according to the performance requirement and available
CPU capabilities. Also, the new GPMN-ENMPC is generalized to a robust version with
respect to input-output feedback linearizable nonlinear system with partially known
uncertainties. Finally, extensive simulations have been conducted, and the results have
shown the feasibility and validity of the new designed method.

8. References
Brinkhuis, J. & Tikhomirov, V., Optimization : insights and applications, Princeton University
Press, ISBN : 978-0-691-10287-0, Oxfordshire, United Kingdom
Chen, C. & Shaw, L., On receding horizon feedback control, Automatica, Vol. 18, No. 3, May,
1982, 349-352, ISSN : 0005-1098

Chen, H. & Allgower, F., A quasi-infinite horizon nonlinear model predictive control
scheme with guaranteed stability. Automatica, Vol. 34, No. 10, Oct, 1998, 1205-1217,
ISSN : 0005-1098
Chen, W., Disturbance observer based control for nonlinear systems, IEEE/ASME
Transactions on mechatronics, Vol. 9, No. 4, 2004, 706-710, ISSN : 1083-4435
Costa, E. & do Val, J., Stability of receding horizon control of nonlinear systems, Proceedings
of The 42nd IEEE Conference on Decision and Control, pp. 2077-2801, ISSN : 0191-2216,
HI, USA, Dec, 2003, IEEE, Maui
Freeman, R. & Kokotovic, P, Inverse optimality in robust stabilization, SIAM Journal on
Control and Optimization, Vol. 34, No. 4, Aug, 1996a, 1365-1391, ISSN : 0363-0129
Freeman, R. & Kokotovic, P.(1996b), Robust nonlinear control design: state-space and Lyapunov
techniques, Birkhauser, ISBN: 978-0-8176-4758-2, Boston
Henson, M., Nonlinear model predictive control: current status and future directions.
Computers & Chemical Engineering, Vol. 23, No. 2, Dec, 1998, 187-202, ISSN : 0098-
1354
He, Y., and Han, J., Acceleration feedback enhanced H

disturbance attenuation control,

Proceedings of The 33
rd
Annual Conference of the IEEE Industrial Electronics Society
(IECON), pp. 839-844, ISSN : 1553-572X, Taiwan, Nov, 2007, IEEE, Taipei
Khalil, H. (2002), Nonlinear systems, 3rd edition, Printice Hall, ISBN: 0-13-067389-7, NJ, USA
Lewis, F. & Syrmos, V. (1995), Optimal control, John Wiley & Sons, ISBN : 0-471-03378-2,
Bangalore, India
Magni, L., De Nicolao, G., Magnani, L. & Scattolini, R., A stabilizing model-based predictive
control algorithm for nonlinear systems. Automatica, Vol. 37, No. 9, Sep, 2001, 1351-
1362, ISSN : 0098-1354
Mayne, D., Rawlings, J., Rao, C. & Scokaert, P., Constrained model predictive control:
stability and optimality. Automatica, Vol. 36, No. 6, Jun, 2000, 789-814, ISSN : 0005-
1098
Pothin, R., Disturbance decoupling for a class of nonlinear MIMO systems by static
measurement feedback, Systems & Control Letters, Vol. 43, No. 2, Jun, 2001, 111-116,
ISSN : 0167-6911
Primbs, J. A., Nevistic, V. & Doyle, J. C., Nonlinear optimal control: a control
Lyapunov function and receding horizon perspective, Asian Journal of
Control, Vol. 1, No. 1, Jan, 1999, 14-24, ISSN: 1561-8625
Primbs, J. & Nevistic, V., Feasibility and stability of constrained finite receding horizon
control, Automatica, Vol. 36, No. 7, Jul, 2000, 965-971, ISSN : 0005-1098
Qin, S., & Badgwell, T., A survey of industrial model predictive control technology. Control
Engineering Practice, Vol. 11, No. 7, Jul, 2003, 733-764, ISSN :0967-0661
Rawlings, J., Tutorial overview of model predictive control, IEEE Control System Magazine,
Vol. 20, No. 3, Jun, 2000, 38-52, ISSN : 0272-1708
Scokaert, P., Mayne, D. & Rawlings, J., Suboptimal model predictive control (feasibility
implies stability), IEEE Transactions on Automatica Control, Vol. 44, No. 3, Mar, 1999,
648-654, ISSN : 0018-9286
Song, Q., Jiang, Z. & Han, J., Noise covariance identification-based adaptive UKF with
application to mobile robot system, Proceedings of IEEE International Conference on
Robotics and Automation (ICRA 2007), pp. 4164-4169, ISSN: 1050-4729, Italy, May,
2007, Roma
A new kind of nonlinear model predictive control
algorithm enhanced by control lyapunov functions 85

0 4 8 12 16 20 24
-20
-10
0
10
20
L
0 4 8 12 16 20 24
-20
-10
0
10
20
time(s)
M

Fig. 10. Control inputs

Simultaneously, the following index is used to compare the optimality of the two different
controllers,


0
T T
0
lim [ ( ) ( ) ( ) ( )] t J x t Px t u t Qu t d

(55)

The optimality performance of H

GPMN-ENMPC, computed from Eq. (55), is about 3280,

and the feedback linearization controller is about 5741, i.e., the H

GPMN-ENMPC has better

optimality than the feedback linearization controller.

7. Conclusion
In this paper, nonlinear model predictive control (NMPC) is researched and a new NMPC
algorithm is proposed. The new designed NMPC algorithm, called GPMN-enhancement
NMPC (GPMN-ENMPC), has the following three advantages: 1) closed loop stability can be
always guaranteed; 2) performance other than optimality and stability can be considered in
the new algorithm through selecting proper guide function; 3) computational cost of the
new NMPC algorithm is regulable according to the performance requirement and available
CPU capabilities. Also, the new GPMN-ENMPC is generalized to a robust version with
respect to input-output feedback linearizable nonlinear system with partially known
uncertainties. Finally, extensive simulations have been conducted, and the results have
shown the feasibility and validity of the new designed method.

8. References
Brinkhuis, J. & Tikhomirov, V., Optimization : insights and applications, Princeton University
Press, ISBN : 978-0-691-10287-0, Oxfordshire, United Kingdom
Chen, C. & Shaw, L., On receding horizon feedback control, Automatica, Vol. 18, No. 3, May,
1982, 349-352, ISSN : 0005-1098

Chen, H. & Allgower, F., A quasi-infinite horizon nonlinear model predictive control
scheme with guaranteed stability. Automatica, Vol. 34, No. 10, Oct, 1998, 1205-1217,
ISSN : 0005-1098
Chen, W., Disturbance observer based control for nonlinear systems, IEEE/ASME
Transactions on mechatronics, Vol. 9, No. 4, 2004, 706-710, ISSN : 1083-4435
Costa, E. & do Val, J., Stability of receding horizon control of nonlinear systems, Proceedings
of The 42nd IEEE Conference on Decision and Control, pp. 2077-2801, ISSN : 0191-2216,
HI, USA, Dec, 2003, IEEE, Maui
Freeman, R. & Kokotovic, P, Inverse optimality in robust stabilization, SIAM Journal on
Control and Optimization, Vol. 34, No. 4, Aug, 1996a, 1365-1391, ISSN : 0363-0129
Freeman, R. & Kokotovic, P.(1996b), Robust nonlinear control design: state-space and Lyapunov
techniques, Birkhauser, ISBN: 978-0-8176-4758-2, Boston
Henson, M., Nonlinear model predictive control: current status and future directions.
Computers & Chemical Engineering, Vol. 23, No. 2, Dec, 1998, 187-202, ISSN : 0098-
1354
He, Y., and Han, J., Acceleration feedback enhanced H

disturbance attenuation control,

Proceedings of The 33
rd
Annual Conference of the IEEE Industrial Electronics Society
(IECON), pp. 839-844, ISSN : 1553-572X, Taiwan, Nov, 2007, IEEE, Taipei
Khalil, H. (2002), Nonlinear systems, 3rd edition, Printice Hall, ISBN: 0-13-067389-7, NJ, USA
Lewis, F. & Syrmos, V. (1995), Optimal control, John Wiley & Sons, ISBN : 0-471-03378-2,
Bangalore, India
Magni, L., De Nicolao, G., Magnani, L. & Scattolini, R., A stabilizing model-based predictive
control algorithm for nonlinear systems. Automatica, Vol. 37, No. 9, Sep, 2001, 1351-
1362, ISSN : 0098-1354
Mayne, D., Rawlings, J., Rao, C. & Scokaert, P., Constrained model predictive control:
stability and optimality. Automatica, Vol. 36, No. 6, Jun, 2000, 789-814, ISSN : 0005-
1098
Pothin, R., Disturbance decoupling for a class of nonlinear MIMO systems by static
measurement feedback, Systems & Control Letters, Vol. 43, No. 2, Jun, 2001, 111-116,
ISSN : 0167-6911
Primbs, J. A., Nevistic, V. & Doyle, J. C., Nonlinear optimal control: a control
Lyapunov function and receding horizon perspective, Asian Journal of
Control, Vol. 1, No. 1, Jan, 1999, 14-24, ISSN: 1561-8625
Primbs, J. & Nevistic, V., Feasibility and stability of constrained finite receding horizon
control, Automatica, Vol. 36, No. 7, Jul, 2000, 965-971, ISSN : 0005-1098
Qin, S., & Badgwell, T., A survey of industrial model predictive control technology. Control
Engineering Practice, Vol. 11, No. 7, Jul, 2003, 733-764, ISSN :0967-0661
Rawlings, J., Tutorial overview of model predictive control, IEEE Control System Magazine,
Vol. 20, No. 3, Jun, 2000, 38-52, ISSN : 0272-1708
Scokaert, P., Mayne, D. & Rawlings, J., Suboptimal model predictive control (feasibility
implies stability), IEEE Transactions on Automatica Control, Vol. 44, No. 3, Mar, 1999,
648-654, ISSN : 0018-9286
Song, Q., Jiang, Z. & Han, J., Noise covariance identification-based adaptive UKF with
application to mobile robot system, Proceedings of IEEE International Conference on
Robotics and Automation (ICRA 2007), pp. 4164-4169, ISSN: 1050-4729, Italy, May,
2007, Roma
Model Predictive Control 86

Sontag, E., A universal construction of Artsteins theorem on nonlinear stabilization,
Systems & Control Letters, Vol. 13, No. 2, Aug, 1989, 117-123, ISSN : 0167-6911
Zou, T., Li, S. & Ding, B., A dual-mode nonlinear model predictive control with the enlarged
terminal constraint sets, Acta Automatica Sinica, Vol. 32, No. 1, Jan, 2006, 21-27,
ISSN : 0254-4156
Wesselowske, K. & Fierro, R., A dual-mode model predictive controller for robot
formations, Proceedings of the 42nd IEEE Conference on Decision and Control, pp.
3615-3620, ISSN : 0191-2216, HI, USA, Dec, 2003, IEEE, Maui