DESIGN SUITABLE FEED FORWARD NEURAL NETWORK TO SOLVE TROESCH'S PROBLEM

Sci.Int.(Lahore),31(1),41-48,2019 ISSN 1013-5316;CODEN: SINTE 8 41
January-February
DESIGN SUITABLE FEED FORWARD NEURAL NETWORK TO SOLVE
TROESCH'S PROBLEM
Luma Naji Mohammed Tawfiq1
and Othman M. Salih
Department of Mathematics, College of Education for Pure Science Ibn Al-Haitham, University of Baghdad, Baghdad, Iraq.
1
Email: luma.n.m@ihcoedu.uobaghdad.edu.iq
ABSTRACT: Purpose – The aim of this research is to find accurate solution for the Troesch’s problem by using high
performance technique based on parallel processing implementation.
Design/methodology/approach – Feed forward neural network is designed to solve important type of differential equations that
arises in many applied sciences and engineering applications. The suitable designed based on choosing suitable learning rate,
transfer function, and training algorithm. The authors used back propagation with new implement of Levenberg - Marquardt
training algorithm. Also, the authors depend new idea for choosing the weights. The effectiveness of the suggested design for
the network is shown by using it for solving Troesch problem in many cases.
Findings – New idea for choosing the weights of the neural network, new implement of Levenberg - Marquardt training
algorithm which assist to speeding the convergence and the implementation of the suggested design demonstrates the
usefulness in finding exact solutions.
Practical implications – The authors have applied the suggested design of feed forward neural network to solve Troesch’s
problem which is an important equation in Magnetostatic Fields, Electrochemical, a Magnetic Field and others.
Originality/value – In this paper, main types of the differential equation have been solved by using parallel processing
technique based on feed forward neural network.
The design has been accompanied with new implementation of training algorithm. This process constructs an efficient dealing
to get the exact solutions of the non-linear problems which is easy to implement.
In addition to that, the accompanied regularization design with the used design proved its efficiency in handling many
problems especially ill-posed problems.
KEYWORDS: Ordinary Differential Equation, Troesch’s Problem, Feed Forward Neural Network, Back propagation Training Algorithms,
Levenberg - Marquardt algorithm.
1. INTRODUCTION
Artificial Neural Networks (Ann) have achieved employment
in many specialties: neurosciences, arithmetic, statistics,
physical processes, information technology, and
manufacturing. Greatest of the preceding workings in solving
differential equations by using Ann is limited to the item of
resolving the linear system of algebraic equations which
product from the discretization of the domain. The reduce of
the networks power function delivers the result to the system
of equations [1]. Mohsen and Behnam [2] used modified Ann
is called relax state to solve the Burger's equation in one
dimensional partial differential equation. Hamid et al., [3]
employed a numerical technique combines with Ann's for
solving the Lane-Emden equations. Many researchers such
[4-6] used multilayer perceptron and Radial basis function
neural networks for solving differential equations of various
types. Mall and Chakraverty [7] used Feed Forward Neural
Network (FFNN) for solving ordinary differential equations
(ODE) with regression based algorithm and discuss score
with random initial weights for various number of neuron in
hidden layers. Modjtaba et al., in [8] used Ann's has been
expanded for solving the Stokes problem. In [9] Ladislav
developed a new kind of Anns which is differential
polynomial neural networks, it’s implementation on
principles, which are employed in the human brain learning.
Yashtini and Malek [10] suggested a recurrent neural
networks technique for solving a type of monotone variation
inequalities problem. Vartziotis et al., [11] used Ann to solve
system of ordinary and partial differential equations to supply
health professionals with explanations for surgical planning
and actual-time decision action. Ahmad Jafarian et al., [12]
suggest combines the neural networks technique with the
power series method to insert an effective iterative approach
to product an approximate polynomial solution for particular
kind of fractional Volterra integro-differential equation.
Parand et al., [13] solved Lane-Emden equations by using an
Unsupervised collective neural network. Sarkhosh et al., [14]
Used the Chebyshev neural networks to get the numerical
solution of specific modules of fractional integro-differential
equation. Silvia and Robert [15] applied FFNN technique to
presented algebraic approach for illustrate a multidimensional
nonlinear function.
Nonlinear ordinary differential equations (ODEs) arise in an
extensive variety of problems in applied science and
engineering. Usually these non-linear problems are solved by
using approximate analytical methods or numerical technique
such as Homotopy perturbation method (HPM) [16],
collocation method [17], Variational iteration method (VIM)
[18], Adomian decomposition method (ADM) [19], Shooting
method [20] etc.
The Troesch’s problem is a type of ODE appear in an
inspection of the restriction of a plasma column by
corpuscular-radiation pressure [21], and furthermore in the
model of gas porous electrodes [22, 23]. The problem was 1st
modeled and resolved by Weibel [24]. Then resolved again
by analytical closed type [25] or by using shooting method
[26], and by using a Laplace remodel decomposition method
[27].
The Troesch’s problem has been widely studied and different
methods have been suggested to get the solution. Such
methods are Homotopy analysis method (HAM), HPM,
ADM, B-Spline Collocation, Shooting method and
Reproducing Kernel method (RKM) for more details see [20],
[26-31].

42 ISSN 1013-5316;CODEN: SINTE 8 Sci.Int.(Lahore),31(1),41-48,2019
January-February
In this paper, the authors design suitable FFNN with new
implementation of training algorithm to obtain the
approximate solution of the Troesch’s problem with
boundary condition, and then to illustrate the efficiency and
accuracy of suggested design some example are solved with
comparisons between the results and exact solution are
introduced
2. ARTIFICIAL NEURAL NETWORKS
A neural network is an interconnected assemblage of
simplified process elements is said to be neurons, whose
functionality is inaccurately based on the human neurons.
This neuron is arranging in the network as a layer connected
between them by the inter component connection say
weights, where its initial value can be choosing randomly.
Then adjust by suitable training algorithm.
The first layer says input layer, then hidden layers and last
layer say output layer. Each neuron in the hidden or output
layers has transfer function or activation function which is
bounded monotonically increasing, differentiable function,
that is sigmoid function. The most popular transfer function
in hidden layer is tanh. transfer function and linear (pure lin.)
transfer function in the output layer [32, 33].
There are two main connection procedures: forward
connections and feedback connections (recurrent), The
Feedback neural network (FBNN) have connection from
neurons of output layer feeding backwards to the input nodes
of former layers, or to the similar layer, while the FFNN is
linking forwards starting from input layer to hidden then
output layer [34]. Here, we suggest FFNN for solving the
problem. A general structure of Ann's is illustrated in Figure
(1).
The training of Ann's can be classified into three distinct
types. These are supervised, unsupervised, and reinforcement
learning [35]. Here, supervised training type has been
presented, based on back propagation training algorithm. The
back-propagation rule is a generality of the least mean square
rule that adjusts network weights to reduce the mean square
error among the preferred and real outputs of the network.
That is, FFNN trained to minimize the error up to an
acceptable accuracy.
Figure (1): Structure of Ann's
3. ARCHITECTURES OF SUGGESTED DESIGN
The authors suggest fully connected three layers FFNN
consist of five hidden neurons in the hidden layer with
sigmoid transfer function such tansig and linear sigmoid for
the output neuron (linsig.).
Let us consider the following represent the 2nd
order
differential equation:
F(𝑥, y(𝑥), yʹ(𝑥), yʹʹ(x)) = 0, 𝑥 ∈ 𝐷 (1)
where 𝐷 ⊂ 𝑅, denotes the domain, and y(𝑥) is the solution to
be computed. Here, 𝐺 is the function which defines the
structure of the problem. For solving a discretized domain 𝐷
over finite set of points is considered. Thus, the problem
changed into the system of equations as follows:
F(𝑥i, y(𝑥i), yʹ(𝑥i), yʹʹ(xi)) = 0, 𝑥i ∈ 𝐷̅ (2)
Let yt(𝑥, w) denote the trail solution and then the problem can
be formulated as:
F(𝑥i, yt(𝑥i, w), (𝑥i, w), (xi, w)) = 0, 𝑥i ∈ 𝐷̅ (3)
The error function with respect to each input data is written
as:
∑ ( ( ( ) ( ) ( ))) ( )
∈ ̅
The trail solution ( ) of the problem by suggested FFNN
can be computation from the following formula:
(𝑥 ) [ ] *
( )
+ 𝑥
[(𝑥 )(𝑥 )] ̅(𝑥 ) ( )
where; ̅(x, w) is the results of the suggested FFNN.
Now, rewrite equation (5) to concerting the problem
( )
∑ *(
( )
) ( 𝑥 (𝑥 ) (
( )
))+
𝑥 (6)
while:
( ) (𝑥 ) (𝑥 ) ̅ (𝑥 )(𝑥 ) ̅
(7)
So,
̅ (𝑥 ) (𝑥 ) ̅ (𝑥 )(𝑥 ) ̅
(8)
For checking the performance of the suggested designs, the
calculating of the absolute error will be considered:
absolute error  | (𝑥 )– (𝑥)| (9)
4. BUILD TRAIN OF SUGGESTED DESIGN
The suggested FFNN initialized with randomly values of the
parameters especially, the weights then adjust by suitable
training algorithm. Here the Levenberg - Marquardt training
algorithm (LM) and its MATLAB code is (trainlm) will be
used with new implementation, the classical LM training
update is as follow:
( ) ( ) ( )
where, is the learning rate and choosing randomly in most
of networks. Here, the authors suggest rule for estimate this
value. The rule based on take value in weight space for each
iteration of the weight update equation will be controlling the
distance between wk+1 and wk.
The training process starts with randomly initialize weights
on all the connections in the FFNN. In step 2, distributing
inputs feed forwarding. Therefore, each neuron in the hidden
layer computes the summing for the inner product of its
weights with the inputs. Take suitable transfer function for

January-February
summing that value. Then its output is sent to the output layer
as an input. So, the neurons in the output layer, sums its
weighted inputs. Then its transfer function (purelin.)
computes its output which is also the output of overall
network. Hence, the error can be calculated. If the error is
less than a specific value, the training stops and the weights
are sent. Otherwise, the training proceeds to the next step.
According to the equation (10) the correction laws of the
weights between the hidden layer and the output layer are
calculated.
The sensitivity term of the output is sent back to the hidden
layer. Each hidden unit gets its input from the output layer,
and then the sensitivity is computed for updating the weights
in the hidden layer. Because there are 5 nodes in the hidden
layer, the sensitivity term is a vector. Again, the correction
laws (equation (10)) for the weights are calculated. Return to
step 2, for next iteration and so on.
There are many types of costs associated with using LM
training algorithm such derivatives, computation and storage.
For this limitation, LM can fail to converge, or it can
converge to a point that is not a minimum. To overcoming
these disadvantages, we suggest the following improvement.
The improvement based on choosing the weights to guarantee
that the Jacobian J is positive definite, to get this, we need to
satisfy the following:
Let E(x) be an n-vector of functions of ( )
Assume that the Jacobian of E is Lipschitz continuous on an
open convex set S with constant L. Then for any x, y∈ S,
‖ ( ) (𝑥) (𝑥) ( 𝑥)‖ ‖ 𝑥‖ (11)
Now, we illustrate the basic idea for new implementation for
the Levenberg - Marquardt training algorithm, where the
given data are divided into three sets of data: training, testing
and validation set. Firstly, store pattern pairs (training data)
so that when n-dimensional vector X from training set is
presented as input, the FFNN recalls m-dimensional vector Y
(output vector), but when Y is presented as input, the FFNN
recalls X.
To develop the FFNN, we need to create a correlation matrix
for each pattern pair we want to store. The correlation matrix
is the matrix product of the input vector X, and the transpose
of the output vector YT
. The FFNN weight matrix is the sum
of all correlation matrices, that is,
∑ (12)
where n is the number of pattern pairs to be stored in the
training set.
Now, to guarantee the convergence of new implementation
for LM algorithm, we must satisfy the following condition
(descent condition):
( ) ( ) (13)
This is possible if the search direction is a descent direction,
that is, if E(w) = g(w), where g is the gradient of performance
function.
We note that, the necessary condition for a minimizer the
gradient of performance function at the optimal weights w∗
are equal to zero, i.e., g(w∗) = 0, which guarantee the
convergence.
5.SOLVING TROESCH’S PROBLEM BY SUGGESTED
NETWORK
Troesch’s problem arises of the confinement of a plasma
column by radiation pressure. In this section, the suggested
design of FFNN will be implemented to solve Troesch’s
problem. Consider the following nonlinear Troesch’s
problem:
yʹʹ = γ sinh(γ y);
subject to Dirichlet boundary conditions: y(0) = 0 , y(1) = 1
The analytic solution was considered in [22], [23], [36], by
using other methods as follow:
(𝑥) ( ) ,(
( )
) ( 𝑥|
( ( ))
)-
Such that:
( ) (√ )
where m is that the resolution of this equation:
( )
√
( | )
and;
( | ) ( ) ( )⁄
Such that and m are associated over the integral:
∫
√
According to the equation (5), the suggested FFNN present
solution for the problem as:
( )  ( ) ̅
The results of suggested FFNN with LM train algorithm, its
new implementation and its new implementation with
suggested rule of are given at different point in the domain
in Table (1), where and the accuracy of the results
are given in Table (2). Similarly, the results of the problem
when γ = 1, 10 can gives in Table (3) and (5) respectively and
the accuracy of that results can give in Table (4) and (6)
respectively. To assess the performance of suggested training
algorithm in comparison with others in: time of training,
number of iterations (epoch), performance of training, and
mean square error of regularization (Mse.reg.) for each value
of γ will be evaluated as in Tables (7-9).
Table 1: Results of FFNN for Troesch Problem, when ( = 0.5)
results of FFNNanalytic solution
Train (LM)New Train (LM)New Train(LM)& newya(x)x
8.030546116977177e-091.221245327087672e-13-2.220446049250313e-1600
0.0951768686437660.0951732457573780.0951751391982880.0951769020000000.1
0.1906339136858270.1906338690998790.1906338691000000.1906338691000000.2
0.2866534029977150.2866534029992350.2866534030000000.2866534030000000.3
0.3835228938457260.3835204222271020.3835225476313960.3835229288000000.4

January-February
0.4815373792656720.4815338661029420.4815373854000000.4815373854000000.5
0.5810020103667240.5810000871572980.5810023816463780.5810019749000000.6
0.6822351448612930.6822351325997310.6822351326000000.6822351326000000.7
0.7855717343611450.7855717867014340.7855717867000000.7855717867000000.8
0.8913670211473310.8913669875009290.8913692659442510.8913669875000000.9
0.9999999924699360.999999999999724111
Table 2: Accuracy of the results for Troesch’s Problem, when = 0.5
Error = | yt(x) - ya(x) |, where ya is analytic solution & yt given by
Train (LM)New Train (LM)New Train (LM) & new
8.030546116977177e-091.221245327087672e-132.220446049250313e-16
3.335623403877275e-083.656242621594141e-061.762801712310025e-06
4.458582694710778e-081.214306433183765e-132.775557561562891e-17
2.284505917771185e-127.646105970593453e-131.110223024625157e-16
3.495427386424055e-082.506572898186565e-063.811686036248041e-07
6.134327934503858e-093.519297058218740e-061.110223024625157e-16
3.546672344700852e-081.887742702044726e-064.067463783563596e-07
1.226129342501992e-082.693401057740630e-131.110223024625157e-16
5.233885469468902e-081.434186103210777e-122.220446049250313e-16
3.364733103250472e-089.289236047038685e-132.278444250602973e-06
7.530064127792002e-092.764455331316640e-130
Table 3: Results of FFNN for Troesch Problem, when ( = 1)
Results of FFNN yt(x)analytic solution
Train (LM)New Train (LM)New Train (LM) & newya(x)x
2.145591815327919e-093.723621411211298e-11-1.110223024625157e-1600
0.0817969851309970.0817969966890980.0817969966000000.0817969966000000.1
0.1645309046074170.1645308709229840.1645615612232360.1645308709000000.2
0.2491673130147700.2491668995269180.2491914140886780.2491673608000000.3
0.3367322379941660.3367317725139580.3367322092000000.3367322092000000.4
0.4283471791616600.4283471610421170.4283385692865730.4283471610000000.5
0.5252739804283990.5252740296278320.5252740296000000.5252740296000000.6
0.6289711876967390.6289705901609130.6289711434000000.6289711434000000.7
0.7411683567459790.7411683782098790.7411535630563880.7411683782000000.8
0.8639700264606520.8639731332953730.8639700206000000.8639700206000000.9
0.9999999993533091.000000000008383111
Table 4: Accuracy of the results for Troesch’s Problem, when = 1
Error = | yt(x) - ya(x) |, where ya is analytic solution &yt given by
Train(LM)New Train (LM)New Train (LM) & new
9.949124594621495e-069.681144774731365e-140
1.811043796381836e-053.637846758150350e-058.189792160412379e-17
8.716584308026689e-065.849265299548434e-146.476880911145344e-05
7.650600952155981e-073.874076623736067e-137.663632431025946e-05
1.903633290691520e-062.494911829215019e-135.448357568935505e-05
7.262222605167873e-072.736848680126060e-053.295974604355934e-17
1.774033088382254e-071.688759363077490e-049.367506770274758e-17
4.662853937686950e-085.047955415286159e-042.795214898675541e-04
1.897660294875037e-083.501574030728705e-136.938893903907228e-18
9.057067845708033e-097.757405828812125e-138.326672684688674e-17
2.423705347531779e-091.220801237877822e-122.220446049250313e-16

January-February
Table 5: Results of FFNN for Troesch Problem, when ( = 10)
results of FFNN yt(x)analytic solution
Train (LM)New Train (LM)New Train (LM)& newya(x)x
9.949124594621495e-06-9.681144774731365e-14000
5.818956203618164e-053.992153241849650e-057.629999999991810e-057.630000000000000e-050.1
1.386165843080267e-041.298999999415074e-041.946688091114535e-041.299000000000000e-040.2
3.596650600952156e-043.589000003874077e-044.355363243102595e-043.589000000000000e-040.3
9.759963667093086e-049.778999997505090e-040.0010323835756899.779000000000000e-040.4
0.0026597262222610.0026863684868010.0026590000000000.0026590000000000.5
0.0072287225966910.0073977759363080.0072289000000000.0072289000000000.6
0.0196640466285390.0201687955415290.0199435214898680.0196640000000000.7
0.0537302810233970.0537302999996500.0537303000000000.0537303000000000.8
0.1521140090570680.1521139999992240.1521140000000000.1521140000000000.9
0.9999999975762950.9999999999987791.00000000000000011
Table 6: Accuracy of the results for Troesch’s Problem, when = 10
Error = | yt(x) - ya(x) |, where ya is analytic solution &yt given by
Train(LM)New Train (LM)New Train (LM) & new
2.145591815327919e-093.723621411211298e-111.110223024625157e-16
1.146900258097716e-088.909833981718407e-115.551115123125783e-17
3.370741694097568e-082.298425338942423e-113.069032323585463e-05
4.778522963433396e-084.612730821473843e-072.405328867841061e-05
2.879416560741532e-084.366860422155838e-071.110223024625157e-16
1.816165967616357e-084.211658799491147e-118.591713426819858e-06
4.917160156825418e-082.783218100432805e-110
4.429673927663913e-085.532390874307680e-070
2.145402089315240e-089.878764473114643e-121.481514361179048e-05
5.860652185774029e-093.112695372564645e-060
6.466911450786483e-108.382627925129782e-120
Table 7: The details of training of FFNN, when = 0.5
Mse.reg.TimesepochsPerformanceTraining rule
7.4593e-009000002331133.36e-33New Train (LM) & new
2.3352e-0080:00:1916713.49e-25New Train(LM)
4.1541e-0110:00:1925005.08e-10Train(LM)
Table 8: The details of training of FFNN, when = 1
2.9127e-0120:00:3429465.26e-25New Train (LM)
7.9862e-0160:01:23103419.76e-15Train(LM)
Table 9: The details of training of FFNN, when = 10
8.5078e-0130:00:4338394.22e-26New Train (LM
7.8462e-0160:01:32113319.59e-15Train(LM)

January-February
There are several previous studies have shown the importance
to solve this equation such: The Adomian decomposition
method (ADM) [19], the variational iteration method (VIM)
[18], the modified homotopy perturbation method (MHPM)
[30], a combination of the ADM with the reproducing kernel
method (RKM) [28], and others. Table (10), represents a
comparison between those solutions of Troesch’s Problem
when γ = 10.
It is worth mentioning that the main advantages of the
suggested design of FFNN in comparison to the other
methods are the fact that there is no need to construct a
homotopy and solve the corresponding equations in HPM. No
need for the large evaluating the Lagrange multiplier like the
VIM. Without any restricted assumption for nonlinear terms
like finding Adomian polynomials to deal with the nonlinear
terms in ADM. Moreover, the FFNN provided good accuracy
even for a few number of neurons and layers. Moreover, the
suggested FFNN provided good accuracy even for γ = 0.5
and 1. Also, Figures (2-4) illustrate the efficiency of
implemented suggested network wherein demonstrate its
results in validation, training, and testing where γ = 0.5, 1 and
10 respectively.
Table 10: Comparison between the different methods for solving Troesch’s problem, γ = 10
x ADM-RKM VIM ADM MHPM
0.1 0.0000576 0.1186109866 667081.1874 17.61750
0.2 0.0001902 0.4461962517 1333955.1189 33.69333
0.3 0.0005676 3.8003366781 1999860.1189 46.78583
0.4 0.0016654 79.89147273 2661970.7366 55.65333
0.5 0.0048331 1880.3539472 3310585.4201 59.35417
0.6 0.0137488 41642.365193 3914127.8659 57.34667
0.7 0.0374013 878764.64189 4374578.5342 49.58917
0.8 0.0936540 18064027.967 4406724.4178 36.64000
0.9 0.2189270 366613074.02 3290268.6374 19.75750
Figure 2: Accuracy of suggested network in validation,
training, and testing where γ = 0.5
Figure 3: Accuracy of suggested network in validation,
training, and testing where γ = 1

January-February
Figure 4: Accuracy of suggested network in validation, training,
and testing where γ = 10
6. CONCLUSION
In this work, a feed forward neural network (FFNN) with new
implementation of training algorithm is successfully applied
to get an accurate solution for Troesch’s Problem. This
technique appears to be very promising to get the solution. its
characterized by applying without any restrictive assumptions
for nonlinear terms, discretization or perturbation techniques.
The numerical results showed that the FFNN accurate,
reliable and better than other methods such HPM, ADM and
VIM.
REFERENCES
[1] Hoda, S. A. and Nagla, H. A., “Neural network methods
for mixed boundary value problems,” International
Journal of Nonlinear Science, 11(3): 312-316(2011)
[2] Hayati, M. and Karami, B., “Feedforward neural
network for solving partial differential equations,”
Journal of Applied Sciences, 7(19): 2812-2817(2007)
[3] Jalab, H. A., Ibrahim, R. W., Murad, S. A., Melhum, A.
I. and Hadid, S. B., “Numerical solution of Lane-Emden
equation using neural network,” In AIP Conference
Proceedings, 1482(1): 414-418 AIP (2012, September)
[4] Kumar, M. and Yadav, N., “Multilayer perceptrons and
radial basis function neural network methods for the
solution of differential equations: a survey,” Computers
& Mathematics with Applications, 62(10): 3796-
3811(2011)
[5] Tawfiq, L. N. M., “On Training of Artificial Neural
Networks,” Al-Fatih journal, 1(23): 130-139(2005)
[6] Tawfiq, L. N. M., “Design and training artificial neural
networks for solving differential equations,” [Ph. D.
thesis]. University of Baghdad, College of Education-
Ibn-Al-Haitham, Baghdad, Iraq(2004)
[7] Mall, S. and Chakraverty, S., “Comparison of artificial
neural network architecture in solving ordinary
differential equations,” Advances in Artificial Neural
Systems, 2013: 12(2013)
[8] Baymani, M., Kerayechian, A. and Effati, S., “Artificial
neural networks approach for solving stokes
problem,” Applied Mathematics, 1(04): 288(2010)
[9] Zjavka, L., “Construction and adjustment of differential
polynomial neural network,” Journal of Engineering
and Computer Innovations, 2(3): 40-50(2011)
[10] Yashtini, M. and Malek, A., “Solving complement-arity
and variational inequalities problems using neural
networks,” Applied Mathematics and Comput-
ation, 190(1): 216-230(2007)
[11] Vartziotis, F., Fotiadis, D. I., Likas, A. and Lagaris, I.
E., “A portable decision making tool for health
professionals based on neural networks,” Health
Informatics Journal, 9(4): 273-282(2003)
[12] Jafarian, A., Rostami, F., Golmankhaneh, A. K. and
Baleanu, D., “Using ANNs approach for solving
fractional order Volterra integro-differential
equations,” International Journal of Computational
Intelligence Systems, 10(1): 470-480(2017)
[13] Parand, K., Roozbahani, Z. and Bayat Babolghani, F.,
“Solving nonlinear Lane-Emden type equations with
unsupervised combined artificial neural
networks,” International Journal of Industrial
Mathematics, 5(4): 355-366(2013)
[14] Chaharborj, S. S., Chaharborj, S. S. and Mahmoudi, Y.,
“Study of fractional order integro-differential equations
by using Chebyshev Neural Network,” Journal of
Mathematics and Statistics, 13: 1-13(2017)
[15] Ferrari, S. and Stengel, R. F., “Smooth function
approximation using neural networks,” IEEE
Transactions on Neural Networks, 16(1): 24-38(2005)
[16] Mirmoradia, S. H., Hosseinpoura, I., Ghanbarpour, S.
and Barari, A., “Application of an approximate
analytical method to nonlinear Troesch’s
problem,” Applied Mathematical Sciences, 3(32): 1579-
1585(2009)
[17] Salih, H., Tawfiq, L. N. M., Yahya, Z. R. and Zin, S.
M., “Solving Modified Regularized Long Wave
Equation Using Collocation Method,” Journal of
Physics: Conference Series, 1003(1): 012062, IOP
Publishing (2018)
[18] Momani, S., Abuasad, S. and Odibat, Z., “Variational
iteration method for solving nonlinear boundary value
problems,” Applied Mathematics and
Computation, 183(2): 1351-1358(2006)
[19] Deeba, E., Khuri, S. A. and Xie, S., “An algorithm for
solving boundary value problems,” Journal of
Computational Physics, 159(2): 125-138(2000)
[20] Chang, S.H., “Numerical solution of Troesch’s problem
by simple shooting method,” Applied Mathematics and
Computation, 216(11): 3303-3306(2010)
[21] Weibel, E.S., “On the confinement of a plasma by
magnetostatic fields,” The Physics of Fluids, 2(1): 52-
56(1959)
[22] Gidaspow, D. and Baker, B. S., “A model for discharge
of storage batteries,” Journal of the Electrochemical
Society, 120(8): 1005-1010(1973)
[23] Markin, V. S., Chernenko, A. A., Chizmadehev, Y. A.
and Chirkov, Y. G., “Aspects of the theory of gas
porous electrodes,” Fuel Cells: Their Electrochemi-cal
Kinetics, 22-33(1966)
[24]Tawfiq, L. N. M. and Jabber, A.K., “Steady State Radial
Flow in Anisotropic and Homogenous in Confined

January-February
Aquifers,” Journal of Physics: Conference Series,
1003(012056):1-12, IOP Publishing (2018).
[25] Roberts, S. M. and Shipman, J. S., “On the closed form
solution of Troesch's problem,” Journal of Computational
Physics, 21(3): 291-304(1976)
[26] Troesch, B. A., “A simple approach to a sensitive two-
point boundary value problem,” Journal of
Computational Physics, 21(3): 279-290(1976)
[27] Khuri, S. A., “A numerical algorithm for solving
Troesch's problem,” International Journal of Computer
Mathematics, 80(4): 493-498(2003)
[28] Geng, F. and Cui, M., “A novel method for nonlinear
two-point boundary value problems: combination of
ADM and RKM,” Applied Mathematics and
Computation, 217(9): 4676-4681(2011)
[29] Feng, X., Mei, L. and He, G., “An efficient algorithm
for solving Troesch’s problem,” Applied Mathematics
and Computation, 189(1): 500-507(2007)
[30] Mohyud-Din, S. T., “Solution of Troesch’s problem
using He’s polynomials,” Rev. Un. Mat. Argentina, 52:
143-148(2011)
[31] Khuri, S. A. and Sayfy, A., “Troesch’s problem: A B-
spline collocation approach,” Mathematical and
Computer Modelling, 54(9-10): 1907-1918(2011)
[32] Oonsivilai, A. and Saichoomdee, S., “Distance
transmission line protection based on radial basis
function neural network,” World Academy of Science,
Engineering and Technology, 60(2009)
[33] Tawfiq, L. N. and Oraibi, Y. A., “Fast Training
Algorithms for Feed Forward Neural Networks,” Ibn Al-
Haitham Journal for Pure and Applied Science, 26(1):
275-280(2017)
[34] Tawfiq, L. N. and Hussein, A. A., “Neural Network and
Mathematical Modeling of a Spherical Body of
Gas,” International Journal of Modern Computer
Science and Engineering, 2(1): 1-11(2013)
[35] Abramowitz, M. and Stegun, I. A., “Handbook of
mathematical functions: with formulas, graphs, and
mathematical tables, 55: 886, New York: Dover
publications (1972)
[36] Tawfiq, L. N. M. and Hassan, M. A., “Estimate the
Effect of Rainwaters in Contaminated Soil by Using
Simulink Technique,” In Journal of Physics:
Conference Series, 1003(012057):1-7, (2018)

DESIGN SUITABLE FEED FORWARD NEURAL NETWORK TO SOLVE TROESCH'S PROBLEM

More Related Content

What's hot (18)

Similar to DESIGN SUITABLE FEED FORWARD NEURAL NETWORK TO SOLVE TROESCH'S PROBLEM (20)

Recently uploaded (20)

DESIGN SUITABLE FEED FORWARD NEURAL NETWORK TO SOLVE TROESCH'S PROBLEM