Copyright © 2013 IJECCE, All right reserved 1297
International Journal of Electronics Communication and Computer Engineering Volume 4, Issue 4, ISSN (Online): 2249–071X, ISSN (Print): 2278–4209
An Algorithm to Solve Separable Nonlinear Least
Square Problem
Wajeb Gharibi
Department of Computer Engineering & Networks, Jazan University, Jazan, KSA
Email: Gharibi@jazanu.edu.sa
Omar Saeed Al-Mushayt
Department of Information Systems,Jazan University, Jazan, KSA Email: oalmushayt@yahoo.com
Abstract – Separable Nonlinear Least Squares (SNLS)
problem is a special class of Nonlinear Least Squares (NLS) problems, whose objective function is a mixture of linear and nonlinear functions. SNLS has many applications in several areas, especially in the field of Operations Research and Computer Science. Problems related to the class of NLS are hard to resolve having infinite-norm metric. This paper gives a brief explanation about SNLS problem and offers a Lagrangian based algorithm for solving mixed linear-nonlinear minimization problem.
Keywords – Nonlinear Least Squares Problem,
Infinite-Norm Minimization Problem, Lagrangian Dual, Subgradient Method, Least Squares.
I. I
NTRODUCTIONSeparable Nonlinear Least Squares (SNLS) problem is a special class of Nonlinear Least Squares (NLS) problems, whose objective function is a mixture of linear and nonlinear functions. It has many applications in various areas such as: Numerical Analysis, Mechanical Systems, Neural Networks, Telecommunications, Robotics and Environmental Sciences and more [1- 10].
The existing special algorithms for these problems are derived from the variable projections scheme proposed by Golub and Pereyra [1]. However, when the linear part of variables has some bounded constraints, the methods based on variable projection strategy will be invalid. Here, we propose an unseparated scheme for NLS and an algorithm which results in solving a series of Least Squares Separable problems.
Given a set of observation
{ }
y
i , a separable nonlinear squares problem can be defined as follows:1
( , ) ( , ) (1)
n
i j j i
j
r a y a t
where
t
iare independent variables associated with the observation{ }
y
i , while thea
jand thek
-dimensionalvector
are the parameters to be determined by minimizing the LS functionalr a
( , )
. We can write the above equation in the following matrix form:2 2
2 2
( , )
( )
,
(2)
r a
y
a
where the columns of matrix
correspond to the nonlinear functions
j( , )
t
i of thek
parameters
evaluated at all the
t
i values and the vectors a and y represent the linear parameters and the observations respectively.It is easy to see that if we knew the nonlinear parameters
then the linear parameters a could be obtained by solving the Linear Least Squares problem:( ) ,
a y
which stands for the minimum-norm solution of the Linear Least Squares problem for fixed
, where( )
is the Moor-Penrose generalized inverse of( )
. By replacing this ain the original functional, we obtain:2 2
2 2
1 1
min ( ) min ( ( ) ( ) ) , (3)
2 r 2 I y
which is called Variable Projection functional [1].
The following section covers unseparated scheme for the NLS problems. Section 3 shows our proposed method. Section 4 presents numerical results for two examples and conclusion follows.
II. U
NSEPARATEDS
CHEME FOR THENLS
P
ROBLEMSConsider the following NLS problem:
2 2
1
min ( ) ( ) , (4)
2
n
x R
F x f x
where ( ) m
f x R with
( ( ))
f x
i
f x
i( )
. Many types ofiterative methods have already been designed to solve NLS problem. Most methods for NLS are based on the linear approximation of
f
in each iteration that is derived from Newton method. The main idea of Gauss-Newton method is described as follows:Suppose our current iterative point is
x
k then we obtain the next pointx
k1
x
k
d
k by solving the following Linear Least Square (LLS) problem:2 2 1
min ( ) ( ) . (5)
2 n k
k k k
d R
f x J x d
Here
J x
( )
is the Jacobian off x
( )
. We can get1
( ( ))
( )
( ( )
T( ))
( ) ( ). (6)
k k k k k k k
d
J x
f x
J x
J x
J x f x
If we compare (6) with the Newton step of (4), we will find that Gauss-Newton method uses
J x
( )
k TJ x
( )
kcontaining only the first order information of
f
, substituting the real Hessian ofF x
( )
2 2
1
( ) ( ) ( ) ( ) ( ), (7) m
T
i i i
F x J x J x f x f x
Copyright © 2013 IJECCE, All right reserved 1298
International Journal of Electronics Communication and Computer Engineering Volume 4, Issue 4, ISSN (Online): 2249–071X, ISSN (Print): 2278–4209
The efficient NLS methods, such as Levenberg-Marquardt methods and structured Quasi-Newton methods, are based on Gauss-Newton method [3]. To get the global convergence without losing good local properties, the former ones try to control the step length at each iteration by using the trust region strategy to the subproblem (5). On the other hand, the later ones reserve the first-order information
J x
( )
k TJ x
( )
k of
2F x
( )
and apply Quasi-Newton method to approximate the second term in (7) [4, 5].III. A L
AGRANGIANB
ASEDA
LGORITHMConsider the following problem; model of nonlinear functions that can depend on multiple parameters:
1 2
min
n( )
( )
(8)
n x R y R
A y x
b y
where, b y( )Rm, (generally
1 2
m
n
n
) are nonlinear and1
max
i, (
n).
i n
x
x
x
R
This type of problems is very common and has a wide range of applications in different areas [1-8].
Problem (8) is difficult to be resolved because the nonlinearity of
A y x b y
( )
( )
and the nondifferentiality of the infinity norm [6]. It can be written as:( , )
min max ( ( )
( )) ,
i1, 2,...,
(9)
x y i
A y x b y
i
m
This can be considered equivalent to the following problem in the sense that their optimal solutions are equal:
2 ( , )
min max[( ( ) ( )) ] ,i 1, 2, ..., (10)
x y i A y x b y i m
which is equivalent to
2
( , ) {0,1} 1 1
min max [( ( ) ( )) ] (11) n i i m i i x y i
A y x b y
That could be relaxed to (12)
2 ( , ) 0
1 1
m i n m a x [ ( ( ) ( ) ) ] (1 2 )
i i m i i x y i
A y x b y
The optimal objective values of (11) and (12) are the same due to the fact that {0,1}nis the extreme points set of:
1
{ :
1;
0,
1, 2,..., }
m i i i
i
m
,Furthermore, any solvable linear programming problem always has a vertex solution.
The problem (12) has the following dual:
2
( , )
0 1
1
max min
[( ( )
( )) ]
(13)
i i m i i x y i
A y x
b y
This problem can be resolved using the subgradient method by iteratively solving its Nonlinear Least Squares subproblems [6, 7].
Algorithm
Step 1:
Choose the initial values
0 and( ,
x y
0 0)
.Step 2:
Solve the following Least Squares problem for fixed
0 using the initial solution( ,
x y
0 0)
0 2
( , ) 1
min
[( ( )
( )) ]
(14)
m
i i
x y i
A y x b y
and obtain a local optimal solution denoted by
( ,
x y
1 1)
.Step 3:
If the stop conditions satisfied, such as the variance between the current and the next obtained objective values is small enough, then stop.Otherwise, update
x
0:
x
1,y
0:
y
1,0 0 2
1 1 1
:
[( ( )
( )) ]
i i
A y x
b y
i
with 0 1 k
where
k
is the number of iterations and0
is a constant; then go to Step 2.IV. N
UMERICALR
ESULTSWe implemented the above Algorithm by MATLAB 7 using CPU Pentium IV with 2.4 GHz. We call the MATLAB function LSQNONLIN to solve Least Squares subproblems (14). The algorithm stops when the variance between the current and the next obtained objective values is less than 1e-8.
The data of the examples are produced at random with zero optimal objective values. The dimension is set m=100.
0 1 1 1
( , ,..., )
m m m
,
0
1
We ran each algorithm 10 times independently and listed the obtained average objective values with the average of the CPU time in seconds.
A. Example 1
In this example, we give fitting data for the model (Golub and Pereyra 1973 and Kaufman 1975):
1 2
1 2 3
t t
a
a e
a e
The results for this problem are given in Table 1. Table 1
Average Optimal Objective Obtained
Average Time in Seconds
0.0006 0.7
B. Example 2
The second example is given for the model (Golub and Pereyra 1973 and Kaufman 1975):
2 2 2
2 3 4 5 6 7
1 ( ) ( ) ( )
1 2 3 4
t t t
t
ae
a e
ae
a e
The results for this problem are given in Table 2.Table 2
Average Optimal Objective Obtained
Average Time in Seconds
0.0003 0.3
Copyright © 2013 IJECCE, All right reserved 1299
International Journal of Electronics Communication and Computer Engineering Volume 4, Issue 4, ISSN (Online): 2249–071X, ISSN (Print): 2278–4209
V. C
ONCLUSIONSThis paper gives a brief explanation about NLS problem supported by the given two examples.
Our proposed Lagrangian based algorithm is more efficient than general unseparated ones. Methods based on this scheme have the same convergence properties as the variable projection scheme.
A
CKNOWLEDGMENTThe first author would like to thank Professor Yong Xia for his valuable notes and comments.
R
EFERENCES[1] G. H. Golub, and V.Pereyra, “Separable nonlinear least
squares: The variable projection method and its
applications”, Inverse Problems,vol.19, pp. 1-26, 2002. [2] G. H. Golub, and V. Pereyra, “The differentiation of
pseudo-inverses and nonlinear least squares problems
whose variables separate”, SIAM Journal on Numerical
Analysis, vol.10, pp. 413-432, 1973.
[3] J.J. Moré, “The Levenberg-Marquardt algorithm:
implementation and theory”, Numerical Analysis,
Springer-Verlag, vol. 630, pp. 105-116, 1978.
[4] L. Kaufman, “A variable projected method for solving separable nonlinear least squares problems”, BIT, vol.15, pp. 49-57, 1975.
[5] X. Liu, “An efficient unseparated scheme for separable nonlinear least squares problem”, Proceedings of the
eighth national conference of operations research society of China, June 30-July2, 2006, pp. 132-137.
[6] W. Gharibi, and Y. Xia, “A dual approach for solving
nonlinear infinity-norm minimization problems with
applications in separable cases”, Numer. Math. J. Chinese
Univ. (English Ser.), Issue 3, vol. 16, pp. 265-270, 2007. [7] L. Kaufman, and V. Pereyra, “A method for nonlinear
least squares problems with separable nonlinear equality
constraints”, SIAM Journal on Numerical Analysis. vol. 15, pp. 12-20, 1979.
[8] A. Ruhe, and P. A. Wedin., “Algorithms for nonlinear least squares problems”, SIAM Reviewvol.22, pp. 318-337, 1980.
[9] E. W. Cheney, Introduction to approximation theory, McGraw-Hill Book Corporation, 1966.