You are on page 1of 9
IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

A Penalty/MPQI Method for Constrained Parabolic Optimal Control Problems

M. H. FARAG 1,2 , FATHY H. RIAD 2 and W. A. HASHEM 2

1 Mathematics Department, Faculty of Science, Taif University, Hawia (888), Taif, Saudi Arabia 2 Mathematics Department, Faculty of Science, Minia University, Mina, Egypt

Abstract

This paper is concerned with the numerical implementation and testing of penalty/MPQI algorithm for solving problems in PDE-constrained optimization. It uses this method to transform a constrained optimal control problem into a sequence of unconstrained optimal control problems. It is shown that the solutions to the original constrained problem. Convergence results are proved numerically.

Keywords:

Optimal control, Exterior penalty function, parabolic Equations, MPQI technique, Distributed parameter

systems.

1.INTRODUCTION AND STATEMENT OF THE PROBLEM

This paper exposes a methodology (penalty/MPQI) allowing us to solve a constrained optimal control problem (COCP) for parabolic systems. This methodology belongs to the class of exterior penalty methods (EPMs). A penalty function approach commonly considered in finite dimensional optimization problem is employed. An augmented performance index is generally considered in penalty methods for both finite optimization problem and optimal control problem. It is constructed as the sum of the original cost function and so-called penalty functions that have some diverging asymptotic behavior when the constraints are approached by any tentative solution. This augmented performance index can then be optimized in the absence of constraints, yielding a biased estimate of the solution of the original problem. The weight of the penalty functions is gradually reduced to provide a converging sequence, hopefully diminishing the bias. The penalty function methods are computationally appealing, as they yield unconstrained problems for which a vast range of highly effective algorithms are available. In finite dimensional optimization, outstanding algorithms have resulted from the careful analysis of the choice of penalty functions and the sequence of weights. In this context, EPMs have been applied to optimal control problems by Winkler [1], Xing [2], Salim [3], Farag [4,5] and others [6-9]. In this study, we consider the following problem of minimizing the cost functional:

l T f ( v )   [ y ( x T , ;
l
T
f
( v
)
[
y
(
x T
,
;
v
)
f
(
x
)]
2 dx
[
v t
( )
 ( )]
t
2 dt
0
0
0
(1)
V
v
:
v  v t
( )
L
[0,
T
],0
 v
 v t
( )
 v
 t
[0,
T
on the set
2
min
max
subject to:
 y
 
 y 
 
 (
x )
F (x,t) , (x,t)

[0,l ]
[0,T]
  
 t
 x
 x
(2)
y(x,0)(x),
0  x  l
(3)
 y x t
(
,
)
 0,
0
 
t
T
 x
x
0
(4)
 y x t
(
,
)
(
x
)
 [ G
(
y l t
( , ))
v t
( )] ,
0
 
t
T
 x
x
l
(5)
y
(
x
,
t
)
 
,
(
x
,
t
)
[0,
l
]
[0,
T
]
0
1
(6)
F
(
x t
,
)
L
(
 
),
(
x
)
L
[0,
l
],
(
x
)
L
 [0,
l
]
where  , l,T are positive numbers,
2
2

]

are given functions. The

functions ( x) , G ( y) are continuous and their derivatives are bounded. Here (t) plays role of a guess for the

their derivatives are bounded. Here  ( t ) plays role of a guess for the
their derivatives are bounded. Here  ( t ) plays role of a guess for the

Volume 2 Issue 4 April 2014

Page 14

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm A Publisher for

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

Email:

editoriijcs@ipasj.org ISSN 2321-5992

solutions or a selection for the solutions: as (1)-(6) could have many solutions to, we choose that one which is nearest to(t) , by minimizing (1). It is proved in [10] that there is a unique weak solution to (2)-(6). The problem (1)-(6) can be understood as a vibrational formulation of the inverse problem where one wishes to find the heating (t) at a part of the boundary such that at the final time the temperature of the system reaches a given desired

goal

many applications in engineering [11-12].

G(y)y corresponding to Stefan-Boltzmann radiation condition seems to have

f

0

(

)

x . The particular case

4

2.MODIFIED FUNCTIONAL AND ITS GRADIENT FORMULA

The constrained optimal control problem (1)-(6) is converted to an unconstrained control problem by adding a penalty

function [13] to the cost functional (1), yielding the modified function

, k

(

v

where

1

( y)

,

r

k

)

 

(

v

max

0

)

f

(

v

)

P

k

(

v

),

y ( x, t) ;0

2

,

2

( y)

P

(

v

)

r

k

l

T



0

0

1

(

y

)

 

2

(

y

)

dx dt

and r 0,k 1,2,

k

are positive numbers and

max

y (x, t)

1

lim r

k 

k

  

.

,k

;0

2

(

v

,

r

k

)

as follows

The gradient of the cost functional with respect to the control variable can be efficiently calculated by means of the adjoint system. Let us now begin with the strong version of the Lagrangian, where we adjoin all constraints by separate multipliers (x,t),(0,t), (l,t), (x,0), , and perform partial integrations with respect to space and time.

L

(

y

,

,

x

,

t

)

l

T



0

0

l

0

y

y

t

(

x

,0)

l

0

[

y

x

(

(

x

)

(

x

,

T

;

v

)

x

)

y

x

  F

(

x

,0)

dx

f

0

(

x

(

x t

,

)

)]

2

(

dx

x t

,

)

T

0

r

k

1

T

0

y

(0,

t

)

x

(0,

[

(

t

v

(

t

)

y

)

)



dt

T

0

{

[

G

(

y

(

l

,

t

))

v

(

t

)]

(

x

)

y

(

l

,

t

)

x

}

(

l

,

t

)

dx

2

(

(

y

)

t

)]

2

dt

dx dt

(8)

The vibrational of the functional L ( y (x,t), (x,t), x,t) is as follows

L

l

0

(

l

y

T



0

[

0

,

,

x

,

t

)

2

2

l

0

[

y

(

T

0

[

x T

,

v

(

t

)

;

v

)

f

0

( )]

t

(

x

)]

v

(

t

)

y

(

x T

,

;

dt

v

T

0

)

[

dx

v

( )]

t

2

l

0

 

(

l



t

x

)

(

 ( , x

l

t

)

x

)

 

x

  r

G

(

y

(

l

,

t

)

k

y

  ( )

1

y

y

( ,

l

t

)



2

(

y

y

)     

y

(

x T

,

;

v

)]

2

dx

T

0

(

x T

,

)

y

(

x T

,

)

dx

l

 

0

(

x

,0)

dt

y

(

l

,

t

)

dt

 

y

(

x

,

t

)

dx dt

y

(

x

,0)

dx

(9)

 y ( x , t ) dx dt  y ( x ,0) dx (9)
 y ( x , t ) dx dt  y ( x ,0) dx (9)

Volume 2 Issue 4 April 2014

Page 15

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

In (9), differentiation with respect to the state

time lead, by means of usual vibrational arguments, to the following adjoint system

y (x,t)

and additional partial integration with respect to space and



t

 

x

(

x

)

 

x

  r

k

  y

1

(

)

y



2

(

y

)

y

,

(

(x,T)2 [ y(x,T;v)f (x)], 0 x l

0

 (0, ) t  0  x  x t ( , ) 
 (0, )
t
 0
 x
 x t
(
, )
 G
( y
( , ))
l t
( x )
 
 l t
(
,
)
 x
 y
x 
l

 

 

,

0  

t

T

x t

,

)



(10)

(11)

(12)

By helping the obtaining adjoint system (10)-(12), we are going to give the following theorem for computing the

gradient of the modified functional

Theorem 1 Let the above conditions of the optimal control problem (2)-(5) and (7) hold. Then the functional

,k

(

v

,

r

k

)

.

,k

(

v r

,

k

)

is

Freshet differentiable



, k

(

v r

,

k

)

v

L

2

[0,

]

T

. Moreover, Freshet derivative at v(t ) of the functional



,

k

(

v

,

r

k

)

v

can be defined by the solution

(

x t

,

)

1,0

W

2

(

)

of the adjoint problem (10)-(12) as follows:



,

k

(

v r

,

k

)

v



v

( )

v



H

(

( , ),

y x t

( , ),

x t

( ))

v t

v

where H ( y ( x, t) , ( x, t ) , v(t ) )

H

(

( , ),

y x t

x t

( , ),

( ))

v t

 

is defined by

( )

v t

( )

t

2

l t

( ,

)

( )

v t

(13)

(14)

Proof: The proof is similar to that of Theorem 5 [14].

3.NUMERICAL PROCEDURE

3.1 PQI Technique

The partial quadratic interpolation method [15] or shortly (PQI) technique may be considered as a second-order

method, the principal idea of this method is to approximate the objective function f (u) about certain point

n ,from which one may determine approximations to the gradient and

Hessian of this function as [ ( )]

from the Hessian matrix

technique. The following flowchart of the Partial Quadratic Interpolation Technique is as follows in Figure 1

and

, using a particular application of the cholesky

second degree polynomial in the space

A

p

u

(

u

r

by

)]

r

A

(

r

)]

b

n

u

[

A

p

(

u

and [

r )]

in

n

u

r

respectively. One then extract a positive definite matrix [

p

 

n

the subspace

x

r

u

r

.

3.2 MPQI Technique

For Modified PQI

[16] , in PQI technique we set

1,

1

2

,

1

4

,

and we take the first value of

which satisfies the

condition

f u

(

r

1

)

f

(

r

u

) ,

r

u 

n

.

However, the value of

taken by this way may not be the optimal value of

 n . However, the value of  taken by this way may not be the
 n . However, the value of  taken by this way may not be the

Volume 2 Issue 4 April 2014

Page 16

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

. Since there is a great possibility that the optimal value

*

 

1

1

lies between these values, i.e. between

1,

2

,

4

,

*

to get the optimal step size

Let us approximate following:

f

( )

t

we suggest the following modification:

r

 u

r

)

f u

(

by a polynomial of second degree

f

(

)

p

2

(

)

1

h

h

2

L

2

f

f

0

1

f

2

(15)

where h is the interval of the interpolation,

L

2

1

2

2

3

1

0

4

2

0

1

1

f

2

i

from (21) and (22) we have

)

1

d

h

p

 

h

(

)

2

2

1

1

2

3

1

(

0

2

1

0

1

2

1

2

3

f

0

d

4

f

1

f

2

h

2 f
3

0

4 (

f

0

2

f

1

f

2

)

p (

then,

*

f

f

0

1

f

2

2 f

1

1

2

f u

(

r

f

2

)

i

r

u

) ,

i

0,

1

2

(16)

, 1

*

2

(

1

h

2

2

f

(18)

0

(17)

4

f

and

1

p

2

(

)

over the interval [0,1] as

L

2

is the Lagrange matrix where

f

2

)

and taking

h

1

2

we obtain

f 2 ) and taking h  1 2 we obtain Figure 1: Partial Quadratic Interpolation

Figure 1: Partial Quadratic Interpolation Method

h  1 2 we obtain Figure 1: Partial Quadratic Interpolation Method Volume 2 Issue 4
h  1 2 we obtain Figure 1: Partial Quadratic Interpolation Method Volume 2 Issue 4

Volume 2 Issue 4 April 2014

Page 17

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

3.3 Iterative Algorithm (PMPQI) and its Convergence for COCP The outlined of the algorithm (PMPQI) for solving the constrained optimal control problem (COCP) are as follows:

1) Given

2)

k

0 ,

*

0 ,

At each iteration k do

r k

0 ,

0

and

v

(

k

)

V

.

Solve state system (2)-(4), then find

y

(.,

v

Minimize

v

(

End do.

(k)

)

to find optimal control

k

)

v

.

(

*

k

1)

using PMPQI technique.

3)

If

4) Set

v

(

k

(

v

1)

(k

1)

v

(

)

k



)

,

r

k

(

v

1

(k)

)

*

r

k

,

k

, then Stop, else, go to Step 4.

 

k

1

and go to Step 2.

The following flowchart of the penalty function method combined with the Modified Partial Quadratic Interpolation Technique (Figure 2) is as follows

:

Quadratic Interpolation Technique (Figure 2) is as follows : Figure 2: Penalty Modified Partial Quadratic Interpolation

Figure 2: Penalty Modified Partial Quadratic Interpolation Method

The following theorem represents the main contribution of the convergence theory for the control sequence, generated by the above numerical algorithm for the unconstrained optimal control problem.

Theorem 2:

Let

above numerical algorithm for any increasing sequence of values r

{

v

k

}

be a sequence of minimizers to the unconstrained optimal control problem problem which generated by the

converge to the optimum solution

.

. Then

{

v

k

}

k

v * of the constrained optimal control problem as

lim r

k 

k

  

Proof: The proof is similar to that of Theorem 4 [17].

4.NUMERICAL RESULTS

In this section we illustrate PMPQI technique for the optimal control problem (7),(2)-(4) with unconstrained (UOCP) and constrained (COCP) in the following example. Let be a number in (0, ] . Then

y

(

x t

,

y

t

)

e

2

t

(

 

x

cos (

x )

y

x  

x

)

is the solution of the problem

, ( x,t)

[ 0,1]

[0,T ]

(19)

  x ) is the solution of the problem , ( x , t )
  x ) is the solution of the problem , ( x , t )

Volume 2 Issue 4 April 2014

Page 18

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014 y (x,0) cos( x), 0  x  l
Volume 2, Issue 4, April 2014
y (x,0) cos( x),
0  x  l
 y
(
x t
,
)
 y x t
(
,
)
 2
0,
 (
x )

e
 x
 x
x  0
x  1
y (x,t)
 0
1
2
2
4
 t
4
 t
Let
v t
( )
 e
cos
(
)
e
sin(
)
set
G( y)   y
4 .To test our method we set

(20)

t sin (  ) , 0   t T  y ( x
t sin (
)
,
0
 
t
T
 y
(
x t
,
)
 x
. We see that
2

f
(
x
)
y
(
x T
,
)
e
0

x 1

T

(21)

( )

v t

(22)

4

y



cos(

( )

v t

(1, )

t

)

. Thus, in this example we

is set to be 1

and the space and time gird sizes are taken to be 0.05 and 0.05, respectively. To analyze the effects of the initial guess on the reconstruction of the control function, the algorithm was run with three initial guesses functions in figures 3, 4

for UOCP and COCP respectively. We tested our algorithm for , then

x

. In this example, T

e

4

2

t

,

f

0

(

x

)

e

2

T

cos(

x

)

.

  4 2 t , f 0 ( x )  e 2  

Figure 3: Values of control function v(t) with 3 initial guesses (UOCP)

function v ( t ) with 3 initial guesses (UOCP) Figure 4: Values of control function

Figure 4: Values of control function v(t) with 3 initial guesses (COCP)

Figure 4: Values of control function v ( t ) with 3 initial guesses (COCP) Volume
Figure 4: Values of control function v ( t ) with 3 initial guesses (COCP) Volume

Volume 2 Issue 4 April 2014

Page 19

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

The numerical results are given in figures 5-9. In figure 5 shows the estimated control function compared with the exact one for UOCP and COCP. But in figures 6 and 7 show the values of the functional and gradient functional

,k

(

v

,

r

k

)

various iterations numbers very close to one another and close to zero.

numbers very close to one another and close to zero. Figure 5: Values of control function

Figure 5: Values of control function v(t) for UOCP, COCP and exact control

function v ( t ) for UOCP, COCP and exact control Figure 6: The values of

Figure 6: The values of function

,k

(

v

,

r

k

)

various iteration numbers

function   , k ( v , r k ) various iteration numbers Figure 7:

Figure 7: The values of



,

k

(

v

,

r

k

)

v

various iteration numbers

Figure 7: The values of   , k ( v , r k ) 
Figure 7: The values of   , k ( v , r k ) 

Volume 2 Issue 4 April 2014

Page 20

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

But in Figures 8 and 9 show the comparison between the exact state results and the present numerical results for

y

j

i

in

x i h

and t j

when

h

1

20

,

1

20

20 at

N

M

8

for UOCP and COCP. We note that

the values of

y

i j for COCP problem converge to exact solution more accurate than UCOP problem.

converge to exact solution more accurate than UCOP problem. Figure 8: The values of state y

Figure 8: The values of state

y

j

i

at

i

j 8 for UOCP

values of state y j i at i  j  8 for UOCP Figure 9:

Figure 9: The values of state

y

j

i

at

i

j 8 for COCP

References

[1] M. Winkler, The Quadratic Penalty Method for Solving Control-Constrained Optimal Control Problems,83 rd GAMM Annual Meeting,Darmstadt,2012 [2] A-Q. Xing and C-H Wang, Applications of the exterior penalty method in constrained optimal control problems, Optimal Control Applications and Methods, 10(4), 1989, 333-345. [3] M. S. Salim, Numerical Studies of Optimal Control Problems and Its Applications, Ph. D. Thesis, Assiut University, (1990). [4] M. H. Farag, Computing optimal control with a quasilinear parabolic partial differential equation}, Surveys in Mathematics and its Applications, 23, 2009, 34-44. [5] M. H. Farag, A. B. Shamardan and H. H. Saleh, On an optimal control problem for a class of distributed parameter systems, ARPN Journal of Science and Technology, Vol 3, No.12 Dec 2013. [6] M. H. Farag, T. A. Talaat and E. M. Kamal, Existence and uniqueness solution of a class of Quasilinear Parabolic boundary Control Problems, Cubo Math. J.,(15)2,2013, 111--119. [7] G. M. Bahaa and M. M. Tharwa, Optimal boundary control for infinite variables parabolic systems with time lags given in integral form, Iran. J. of Sci. and Tech., (3), 2012, 277-291. [8] T. M. El-Gindy, Numerical Studies in Optimal Control Theory, Ph. D. Thesis, University of Wales, (1977).

Studies in Optimal Control Theory, Ph. D. Thesis, University of Wales, (1977). Volume 2 Issue 4
Studies in Optimal Control Theory, Ph. D. Thesis, University of Wales, (1977). Volume 2 Issue 4

Volume 2 Issue 4 April 2014

Page 21

IPASJ International Journal of Computer Science (IIJCS) Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm Email:

IPASJ International Journal of Computer Science (IIJCS)

Web Site: http://www.ipasj.org/IIJCS/IIJCS.htm

Email: editoriijcs@ipasj.org ISSN 2321-5992

A Publisher for Research Motivation

Volume 2, Issue 4, April 2014

[9] M. H. Farag and S. H. Farag, On the derivation of gradient formulae by solving the conjugated boundary value problem for an optimal control problem, J. Egypt. Math. Soc., 4, 1996, 23-30. [10] E. J. P. G. Schmidt, Boundary control for the heat equation with nonlinear boundary codition,J. Diff.

Eq.,78,1989,89-121.

[11] D. N. Hao, Methods for inverse heat conduction problems, Peter Lang, Frankfurt am Main, 1989. [12] E. Sachs, A parabolic control problem with a boundary condition of the Stefan-Boltzmann type, ZAMM, 58, 1978,

443-449.

[13] S. S. Rao, Optimization: Theory and Applications, Wiley Eastern Limited, (1984). [14] M. H. Farag, Necessary Optimality Conditions for Constrained Optimal Control Problems Governed by Parabolic Equations, Journal of Vibration and Control, 200,V. 39, 949-963, [15] T. M. El-Gindy, M. S. Salim and A. R. Ibrahim, A modified partial quadratic interpolation method for unconstrained optimization, J. CONCRETE AND APPLICABLE MATHEMATICS, 11(1), 2013, 136-146. [16] A. A. Radwan, Acceleration of partial quadratic interpolation technique, M. Sc. Thesis, University of Assuit,

(1980).

[17] F. Vasiliev, Methods for solving external problems, Nauka, Moscow, 1981.A. Bonnaccorsi, “On the Relationship between Firm Size and Export Intensity,” Journal of International Business Studies, XXIII (4), pp. 605-635, 1992

AUTHORS

M. H. FARAG – Professor of Mathematics (Numerical Analysis and Optimal control PDEs), Mathematics

M.

H. FARAG – Professor of Mathematics (Numerical Analysis and Optimal control PDEs), Mathematics

and Statistics Department, Faculty of Science, Taif University, Hawia (P.O. 888), Kingdom of Saudi Arabia (KSA). OR Mathematics Department, Faculty of Science, Minia University, Minia, Egypt.

Fathy H. RIAD – Lecturer of Mathematical Statistics, Mathematics Department, Faculty of Science, Minia University,

Fathy H. RIAD – Lecturer of Mathematical Statistics, Mathematics Department, Faculty of Science, Minia University, Minia, Egypt.

W. A. HASHEM - M.S. degree Student in Mathematics. He received the B.S. degree in

W.

A. HASHEM - M.S. degree Student in Mathematics. He received the B.S. degree in Mathematics

Department, Faculty of Science, Minia University, Minia, Egypt. His research interests include Numerical

Analysis, Partial differential equations and Optimal control PDEs.

Numerical Analysis, Partial differential equations and Optimal control PDEs. Volume 2 Issue 4 April 2014 Page
Numerical Analysis, Partial differential equations and Optimal control PDEs. Volume 2 Issue 4 April 2014 Page

Volume 2 Issue 4 April 2014

Page 22