Modification Of Levenberg-Marquardt Algorithm For Solve Two Dimension Partial Differential Equation

In this paper we presented a new way based on neural network has been developed for solutione of two dimension partial differential equations . A modified neural network use to over passing the Disadvantages of LM algorithm, in the beginning we suggest signaler value decompositions of Jacobin matrix (J) and inverse of Jacobin matrix( J-1), if ( ) a matrix rectangular or singular Secondly, we suggest new calculation of μk , that isk= E (w)2 look the nonlinear execution equations E(w) = 0 has not empty solution W* and we refer ‖ ‖ to the second norm in all cases ,whereE(w): is continuously differentiable and E(x) is Lipeschitz continuous, that is= E(w 2)E(w 1) L w 2w 1 ,where L is Lipeschitz constant.


Introduction
We have been many methods developed so far to solve differential equations.a few of them generate a result in the type of an range that contain the value of the result at a chosen grouping of points.The result in analytic type and make over the novel problem generally in a system of linear equations.the majority of the preceding work in solving differential equations using Nearul networks is constrained to the case.The result of a linear structure of equations is mapped onto the manner of a Hopfield Neural network.The decreasing of the network's force function provide the solution to the structure of equations.
Here object we outlook the problem from a special angle.We at in attendance a universal manner for solve partial differential equations (PDEs), that depends on function approximate capability of feedforword Nearul networks and outcome in the configuration of a solutione write in a differentiable, near exact form.This figure employ a feedforword nearul network as the basic approximate element, whose parameters' (weights and biases) are use to decreasing an suitable error function.in the direction of teach the network we use optimization technique, which in twirl want the calculation of the grade of the error with value to the network parameters.In the projected move toward the mold function is articulated as the sum of two requisites: the first part satisfy the initial / boundary conditions and contain no modifiable parameters'.The second part involve a feedforward nearul network to be train so as to suit the differential equation by Lee, H. & Kang, I.,1990 , Wang, L. & Mendel J.M.,1990 and Yentis, R. and Zaghoul, M.E.,1996.

Levnberg-Marquardt algorithm (LM)
The algorithm of Levnberg-Marquardt (LM) is an purified procedure that locate the least of a more than variable function that is uttered as the sum of squares of non-linear real-valued functions by K. Levenberg ,1994 and D.W. Marquardt ,1963.It has happen to a typical procedure for non-linear leastsquares problems, extensively adopt in a wide range of discipline.
For LM algorithm, the performance index to be optimized is defined by H.D. Mittelmann ,2004 "Where , consistes of all weights of the network, is the desired value of the output and the pattern, is the actual value of the output and the pattern, N is the number of the weights, P is the number of pattern, and K is the number of the network output".
"Equation (1) can be written its "Where , - Where E is the cumulative error vector ( for all pattern) from equeation (2) the jacobian matrix is define as= "And the weights are calculated using the following equation "Where I is identite unit matrix, is the learing parameter and is jacobian of m output error with respect to n weights of the Neural network.The parameter is automatically adjusted at each iteration in order to secure convergence, the LM algorithm requires computation of the jacobian matrix at each iteration step and the inversion of square matrix, the dimension of which is N*N"by B. M. wilamowski and S. Iplikei ,2001.

Modification Of High Performance Training Algorithms:
In this part we will improve the training algorithm LM by overcoming some of the disadvantages that have emerged in the LM standard training algorithm, in the beginning we suggest singalur value decompositions of Jacobin matrix (J) and inverse of Jacobin matrix( J -1 ), if ( ) a matrix rectangular or singular and secondly we suggest new calculation of parameter μk that isk=E(w) 2 and we refer ‖ ‖ to the second norm in all cases ,whereE(w): is and E(x) is Lipeschitz continuous, that is E(w 2)- E(w 1)Lw2-w1 , where L is Lipeschitz constant.

Illustration of the Method
Consider the following 2 nd order of two-dimensional of partial defferential equeation for the function ( by Abdul-Majid Wazwaz , 2009.The network treale use a net of ten equidistant points in x [0,1] , y [0,1] .Figure (1) expand the analytic and nearul solutions activation functions.The neural results with f activation functions with modify Levnberg-Marquardt algorithmic rule (trainmlm) introduced in

Figure 1 :
Figure 1:exact and approximate solution of example2 using : modify trainlm and stander trainlm algorithms Table (1) , In table (2) Accurace of solutions, execution of the trail with period and time insert in table (3) and table (4) gave the parameter of initial values.

Table 7 : execution of the trail with period and time
()