Professional Documents
Culture Documents
a balebaidi.blaise@yahoo.com
Abstract
In this work we highlight the capabilities of particles swarm optimization (PSO) method to identify parameters
of hyperelastic models. We compared this method with Genetic Algorithm (GA) method, Least Squares (LS)
method, Pattern Search Algorithm (PSA) method, Beda-Chevalier (BC) method and the Levenberg-Marquardt
(LM) method. Four classic hyperelastic models are used to test the dierent methods through parameters identi-
cation. Then, we compare the ability of these models to reproduce experimental Treloar data in simple tension,
biaxial tension and pure shear.
Keywords: Particle Swarm Optimization; identication; hyperelastic model.
1
1 Introduction
Elastomers are used widely in industrial applications, for instance: in the aeronautics (manufacture of the tires,
the supports engine), building (gum of damping), medicine (prostheses, implants) and clothing (shoes, helmets).
Indeed, elastomers have a great capacity of stretching (about a few hundreds percent) in elastic domain. However,
Industrial use of rubber requires characterizations that need modeling behavior and a very large variety of models
was proposed to this purpose in the literature [1-4].
These models often present the diculties to predict the behavior of the rubber-like materials with the same
set of material parameters, i.e. to reproduce experimental data for dierent loading conditions (uniaxial or biaxial
extension, simple or pure shear) [1]. This problem could come by the hyperelastic models, i.e. the choice of the strain
energy functions [2] but also by the choice of the material parameters [1,5] and nally, by the identication method.
In the literature [6-10], many methods have been proposed to optimize the material parameters of hyperelastic
models: the gradient method, the relaxation method, the Newton method, the Levenberg-Marquardt (LM) method,
the least squares method (LS), the Genetic Algorithms (GA) method, the Stage by Stage method or Beda-Chevalier
method (BC), the Stage by Stage method combined with the least squares method and the Pattern Search Algorithm
(PSA) method. The main drawback of these methods is often a poor prediction of the biaxial or pure shear response
of models when parameters are identied with uniaxial data.
T. Beda and Y. Chevalier compared the Stage by Stage method to least squares (LS) method [7]. This comparison
show that, Beda-Chevalier method can reduce or limit the number and value of negative parameters in hyperelastic
models which reduce the possibility of negative values in strain energy density function, leading to models with a
higher stability and LS method can provide negative values to parameters which might not correspond to the material
parameters of the hyperelastic models. In order to controle the parameters signs and consequently their values, T.
Beda combine Stage by Stage method with the least squares method and compare this method to Beda-Chevalier
method and least squares method [8]. More recently, Yifeng Wu et al. compared PSA method to LM method [9].
The present paper proposes a comparison of the (PSO) method to ve other optimization methods on their ability
to optimize the parameters of the hyperelastic model to t experimental data of Treloar [11] in: simple tension,
biaxial tension and pure shear. In section 2, we briey summarized the formulation of each model considered. In
section 3, optimizations methods adopted to determine material parameters are described. Afterwards in section 4,
comparison methods are established and nally, in section 5 results and discussions close the paper.
N is the number of terms considered for the Ogden model, λi is the principal stretch, µp and αp are undetermined
parameters of the model such as:
µp α p > 0 (3)
This model is chosen as the hyperelastic model owing to its favorable performances under large deformations, but the
determination the parameters is too dicult. For the determination the parameters, the approximations are made
on the terms.
2
equibiaxial extension.
λa βa − βao βa sinhβao
W 8−ch
= µn √ + ln (4)
n βao sinhβa
q
where λa = I31 is the mean stretch of the material and βa = L−1 λa
√
n
indicate the inverse function Langevin.
The quantities µ, n indicate the dierents material parameters.
3 Optimization methods
The procedure of the identication consist in making coincide a theoretical solution σmodel resulting from a model
with the experimental curves of Treloar [11] in simple tension, biaxial tension and pure shear represent by the couple
of the point (λexp , σexp ), with i=1, 2, 3.....N the number of the experimental points. To quantify the gap between
the two responses we generally denes the objective function φ as:
φ = ΣN 2
i=1 (σexp (λi ) − σmodel (λi )) . (6)
Then, identication consists in nding the best parameters such as φ = 0. In reality, errors persist and we only tend
to minimize the function as much as possible.
To illustrate this method, Considering a Money-Rivlin model dened by (Eq. 1) and given by the experimental data
(λi , σexp (λi )), i=1,2,3.....N, with 1 ≤ λi ≤ λi+1 . The nominal stress in simple tension is given by:
1 C01
σmodel (λi ) = 2(λi − 2 )(C10 + ). (8)
λi λi
3
3.2 Levenberg-Marquart (LM) method.
On the basis of the Newton method, Marquardt proposed a method, which bears the name of Levenberg-
Marquardt [17], this method is very much used in the problems of identication by least squares method [18-19].
The principle of this method consists in adding to the problem of minimization dened by the (Eq. 6) the constrain
kδu k = Ro and Lagrange multiplier µ. The new function to be optimized becomes:
~ (δu , µ) = ΣN
U 2 2 2
i=1 kσexp (λi ) − fo − ∇f k + µ(kδk − δo ). (10)
Ro represent an arbitrary ray. As for the method of least squares, δ corrects the point obtained with the preceding
step.
To illustrate this method, let us consider Money-Rivlin model again. By considering the experimental data of Treloar,
represented by: (λi , σexp (λi )), i=1,2,3.....N, with 1 ≤ λi ≤ λi+1 .
Let us pose λ = λo + δλ and µ → µ + 1. The equation to be optimized is dened by:
σmodel (λo + δλ) = σmodel (λo ) + ∇σmodel (λ) = f~o + ∇f~~δ.(11)
ϕ = ΣN 2 2 2
i=1 kσexp (λi ) − fo − ∇f k + µ(kδk − δo ). (12)
The algorithm then consists in adjusting µ with each iteration, so that ϕ at the point running decreases between two
successive iterations. This minimum of ϕ being reached by obtaining respectively parameters C10 and C01 . We shall
therefore dene our Algorithm in Matlab [16] as follows:
• Let µ(r−1) denote the value of µ from the previous iteration. Initially let µ(0) = 10−3 , say.
• Compute ϕ(µ(r−1) ) and ϕ( µ ν ), with ν > 1.
(r−1)
• If ϕ( µ ν )> ϕ(r) , and ϕ(µ(r−1) ) > ϕ(µ(r) ), increase µ by successive multiplication by µ until for some smallest
(r−1)
k ϕ(µ(r−1) ν k ) ≤ ϕ(µ(r) ).
With LM method the choice of the initial guess is arbitrary and this method requires good judgment the experience
established. The convergence of LM method depends often the regularity of the Jacobian matrix and choice of the
line search.
yo (x) = y(x). where yk (x) is the function to be identied in the sub-domain Vk . This method evaluates the hyperelastic
material parameters by following increasing order of the variable x. To illustrate this method, consider Mooney-Rivlin
model, with nominal stress dened by Eq. 8. Let us pose Y (λi ) = 2( σmodel (λi )
1
−X 2 )
with Xi = λ1i , We obtains:
Xi i
4
3.4 Genetic Algorithm (GA) method.
Genetic Algorithms (GA) have been applied for parameter identication for a few years [6,10,21]. These
methods were proposed by Holland [22] and were widely diused by the work of Goldberg [23]. Based on natural
selection and genetics [24], GA searches for the global maximum of a function without having any knowledge of the
problem. Starting from multiple guess points (chromosomes/ individuals) forming the initial population, GA ts
each individual directly to the objective function. Then, selects parents, by an appropriate method, and reproduces
through natural operators (crossover, mutation) evolving to a new better tting population. The process stops when
the algorithm stabilizes, i.e. all individuals share the same genes. Thus, all the individuals in the population returns,
relatively, the same tting value so convergence is achieved. Herein, we consider Moneey-Rivlin model and we use
Genetic Algorithm in binary code under Matlab [16], on a computer Dell, 1.8 GHz and 4 Go RAM. After used these
dierents step we obtain the results mentioned in table 2, table 3 and table 4. For more explanation on GA, we can
also read [25].
For iteration k, we dene a trial point as any point of the form λk i = λk + sk i , where λk is the current increment. sk i
represent the trial step given by sk i = 4k Φ(λi ), with k = 1, 2, 3....N and where 4 represents the naplacian operator
. With Money-Rivlin model, we implemented PSA method in Matlab [16] by respecting the following process.
a) a compute Φk (λi )
b) Determine a step sk
c) compute pk = pk (λk ) − pk (λk + sk )
d)if pk > 0, and λk+1 = λk + sk , then evaluate parameters C10 , and C01 .
5
data by (λi , σexp (λi )) and values of the velocity by V elocitymax = 1.5 and V elocitymin = 0.35. After we denes
dierents positions of the particles by C10 = C10min + (C10max − C10min ). ∗ rand(1, N );
C01 = C01min + (C01max − C01min ). ∗ rand(1, N ); with theses velocity
denes by V elocityC10 = V elocitymin + (V elocitymax − V elocitymin ). ∗ rand(1, N ) and
V elocityC01 = V elocitymin + (V elocitymax − V elocitymin ). ∗ rand(1, N ).
b)Compute φ dened by Eq. 6.
c)Up to date meter new positions and speeds of the particles by following code.
for j = 1 : N
i < 1500, represent iteration.
c1 = 0.9 − 0.5/(1500 − i);
c2 = 1.2;
c1 and c2 represent independent uniform random numbers.
Γ1 = c1 ∗ rand(1);
Γ2 = c2 ∗ rand(1);
V elocityC10 (j) = Γ1 ∗ V elocityC10 (j) + Γ2 ∗ (C10 (i) − C10 (j));
V elocityC01 (j) = Γ1 ∗ V elocityC01 (j) + Γ2 ∗ (C01 (i) − C01 (j));
C10 (j) = C10 (j) + VC10 (j);
C01 (j) = C01 (j) + VC01 (j);
d)Evaluation parameters by: C10 evolution(i) = C10 classe(1);
C01 evolution(i) = C01 classe(1);
For more details to see appendix of this paper.
q.PNG
TY.PNG
The PSO method updates the velocity vector for each particle then adds that velocity to the particle posi-
tion or values. Velocity updates are inuenced by both the best global solution associated with the lowest cost
ever found by a particle and the best local solution associated with the lowest cost in the present population.
If the best local solution has a less cost than the cost of the current global solution, the best local solution
replaces the best global solution (Figure 2). The particle velocity is reminiscent of local minimizers that use
derivative information, because velocity is the derivative of position. The PSO is able to tackle tough objective
functions with many local minima. The particle swarming becomes evident as the generations pass (Figure 2b).
6
4 Comparison the optimisation methods for parameter identication
The evaluation of the parameters of the models (table 2, table 3 and table 4) according to the various
methods of optimization used is done about to following criteria (table 1).
Figure 3, Figure 4, Figure 5 and Figure 6 present the histograms which describe the inuence of the computing
time and of the precision of the nal error of the function φ during the optimisation of the material parameters
of each model in simple tension (ST), biaxial tension (BT) and pure shear (PS). Indeed, By using the values
of table 1 above, we obtain the folowing gures:
The Figure 3a shows that in simple tension and biaxial tension, the PSO method evaluates more quickly the
Mooney1.PNG
parameters of the mooney model. Unlike in pure shear, the PSA method evaluates with little time the material
parameters.
In Figure 3b, the LS method gives a minimal value of the objective function in Simple tension, in biaxial tension
the PSA method provided the minimal value of the function φ, of same in pure shear GA method provided the
smallest value of the function φ.
In the optimization of the parameters of the Ogden model, the PSO method optimizes more quickly the
parameters in: simple tension, biaxial tension and pure shear (Figure 4a).
7
Ogden1.PNG
Figure 4: 3 term Ogden model: a)- CPU time, b)- Evaluation function φ
For the evaluation of the objective function, the PSO method minimizes better the function φ (Figure 4b).
In the optimization of the parameter of the 8-chains model, the PSO method evaluates more quickly the
AB1.PNG
parameters of the model in: simple Tension, biaxial tension and pure shear (Figure 5a).
In the evaluation of the objective function, the PSO method gives a minimal value of φ in simple tension, unlike
in biaxial tension it is GA method which minimizes the function φ beter and in pure shear in fact rather the
LS method reduce to minimum the function φ.
HIA1.PNG
The Figure 6a shows that in the optimization of the parameter of the HIA model, the PSO method evaluates
more quickly the parameters of the model in simple tension , biaxial tension and pure shear (Figure 6a).
But, in the evaluation of the objective function, the PSA method gives a minimal value to the function φ in
simple tension and biaxial tension, in pure shear, it is LS method which reduce the function φ in the smallest
value.
8
Table 2: Parameters Values.
Models Parameters GA PSO BC PSA LM LS
Mooney C10 0.0923 0.1504 0.075 0.103 0.1473 0.162
C01 0.1451 0.014746 0.148 0.1351 0.000090 5.90×10−3
3 terms µ1 0.63 0.00127 0.6174 0.37282 0.45262 -0.011
Ogden µ2 1.2×10−3 0.597 0.001176 1.0209×10−5 2.842×10−5 0.62995
µ3 -0.01 -0.011 -0.00078 −8.2747×10−3 −8.2746×10−3 0.00127
α1 1.3 5 1.3 1.8464 1.71617 -2
α2 5 1.3 5 7.29012 6.80992 1.3
α3 -2 -2 -5.4 -2.09521 -1.98197 5
Boyce et µ 0.3074 0.2874 0.294 0.294 0.314 0.328
Aruda n 25.4 25.4 26.4 27.4 28.4 29.5
HIA µ 0.3474 0.3874 0.3074 0.38 0.382 0.3874
n 27 27.5 25.5 28.5 29 28.5
K 0.001266 0.001 0.001266 0.01266 0.01266 0.001266
η 0.1 0.15 0.09 0.1 0.1 0.1
ζ 0.127 0.0177 0.0187 0.127 0.127 0.127
ξ 0.419 0.419 0.419 0.419 0.419 0.418
9
Table 4: Parameters values Pure Shear.
Models Parameters GA PSO BC PSA LM LS
Mooney C10 0.172 0.180 0.181 0.170 0.169 0.171
C01 0.002251 0.00030746 0.0002102 0.002351 5.30×10−4 0.000090
Ogden µ1 0.63 0.00147 0.6174 0.00122 0.42262 0.00152
3 terms µ2 1.3×10−3 0.599 0.001176 0.627 1.842×10−5 0.607
µ3 -0.01 -0.011 -0.00078 −0.021 −8.2746×10−3 -0.011
α1 1.3 5 1.4 5 1.3 5
α2 5 1.3 5 1.3 5 1.3
α3 -2 -2 -2.9 -2 -2 -2
Boyce et µ 0.2924 0.30 0.31 0.29 0.28 0.31
Aruda n 25.5 27 28.5 24 22.5 29.5
HIA µ 0.3974 0.3874 0.4074 0.38 0.382 0.3874
n 34.5 27.5 42.5 40.5 32.5 34.5
K 0.001266 0.001 0.001266 0.01266 0.01266 0.001266
η 0.1 0.15 0.12 0.1 0.1 0.1
ζ 0.127 0.0177 0.137 0.127 0.127 0.127
ξ 0.419 0.419 0.419 0.419 0.419 0.418
Mooney.PNG
Figure 7: Mooney model: a)- Simple tension, b)- Biaxial tension and c)- Pure shear
10
8b) shows that all the six methods optimize the parameters of the 3 terms Ogden model by covering the whole
of the deformation. In pure shear (Figure 8c), all the six methods of optimizations used, give to the 3 terms
Ogden model a good correlation with the experimental data of Treloar.
Ogden.PNG
Figure 8: 3 terms Ogden model: a)- Simple tension, b)- Biaxial tension and c)- Pure shear
BOYCE.PNG
Figure 9: 8-chain model: a)- Simple tension, b)- Biaxial tension and c)- Pure shear
11
HIA.PNG
Figure 10: HIA model: a)- Simple tension, b)- Biaxial tension and c)- Pure shear
5.3 Discussions
Comparing the dierent tting methods according the criteria mentioned in table 1, it is clear that the PSO
method as the GA, BC, PSA, LM and LS methods is also able to optimize hyperelastic material parameters
under multi loading conditions. As far as convergence consumption is concerned, we obtains fast convergence
towards the best solution with this method (Figure 2b). However, the reduced amount of overall time, the robust
characterization and value of the nal error justify the use of PSO method. sometimes, the PSO method can
articially generate a problem of convergence at high strain values, because the tting of the model concerned
can shows a local maximum which is not in the data behavior (Figure 7b and Figure 8a), this is a limitation
produced by the PSO method.
With GA method the results are dubious: in spite of a signicant number of generation and it is impossible
to be assured that the solution obtained is the best, but we are sure to have approached the optimal solution,
without the certainty to have reached it.
BC method is very tiresome for a high number of parameter. But, this method remains ecient in the
determination the initial guess points. The initial guess points allows to nd perfectly the parameter space for
the best approximation, i.e, Interval in which one can have the best parameters.
PSA method requires a machine which have great capacity, with a rather high time of calculation (table 2).
LM method remains very tiresome for a high number of parameter and depend also the derivative of the
function. By implementing this method under Matlab [16] one notes that it is just a combination the gradient
method and Gauss-Newton method.
LS method optimize easily the parameters of the models less complex (Mooney-Rivlin model, Gent and
Thomas model), but for the complex models, this method requires a high number of iteration (Ogden model,
HIA model).
We can also see the limits of the dierent models to reproduce complectly the dierent experimental curves.
For example Mooney-Rivlin model is only the most ecient for reproduce the moderate strain; HIA model and
Ogden model are appropriated for large strain.
6 Conclusion
In this paper, material parameters for the hyperelastic model and its optimizations methods have been studied.
The material parameters have been optimized from simple tension, biaxial tension and pure shear with Treloar
data. Then, we focused on the elaboration PSO method for the optimization the rheological parameters and
we also showed the inuence of the optimizations methods on the rheological parameters on the hyperelastic
models. For this purpose, the following conclusions have been made:
• The PSO is most adapted for approximation under constraint and to provide appropriate estimated param-
eters for higher values of variable.
• The PSO algorithm used allows optimized with a very good approximation and in reasonable computing
time the hyperelastic material parameters.
• According to the tting results of the experiment data from Treloar, optimization of the parameters of the
hyperelastic model depends the choice of the model and the nonlinearities problem.
12
Appendix A
%Optimization of the Mooney model by PSO method in Matlab
%% Optimisation by PSO method.
all clear all clc
N=100; %Population number
Ngenerations =40;
C10min = 0.0921;
C10max = 0.0924;
C01min = 0.1449;
C01max = 0.1451;
V itmax = 1.5;
V itmin = −1.5;
Errorparticuleinit =1000;
Errorglobaleinit =1000;
C10 = C10min + (C10max − C10min ). ∗ rand(1, N );
V itesseC10 = V elocitymin + (V elocitymax − V elocitymin ). ∗ rand(1, N );
C01 = C01min + (C01max − C01min ). ∗ rand(1, N );
V elocityC01 = V elocitymin + (V elocitymax − V elocitymin ). ∗ rand(1, N );
M attote rror = zeros(1, N );
M aterror = zeros(1, length(x));
for j=1:N
for h=1:length(x)
Ycal = 2 ∗ (x(h) − 1/(x(h)2 )) ∗ (C10 (j) + C01 (j)/(x(h)));
error1 = (Ycal − Yref (h))2 ;
M aterror (h) = error1;
end
Error1 = max(abs(M aterror ));
M attoterror (j) = Error1 ;
end
end
for i = 1 : Ngenerations
13
% Calculation of the news good position and velocity
for j = 1 : N
if M attoterror (j) < Errorparticuleinit
Errorparticuleinit = M attoterror (j)
Cp10good = C10 (j);
Cp01good = C01 (j);
end
Cglo10good = C10classe (1);
Cglo01good = C01classe (1);
end
Cp01good = C01 (j);
Cp10good = C10 (j);
V elocityC10 (j) = gama1 ∗ V elocityC10 (j) + gama2 ∗ (M atCp10 good (i) − C10 (j)) + gama2 p ∗ (M atCglo10 (i) − C10 (j)) ;
V elocityC01 (j) = gama1 ∗ V elocityC01 (j) + gama2 ∗ (M atCp01 good (i) − C01 (j)) + gama2 p ∗ (M atCglo01 (i) − C01 (j)) ;
C10 (j) = C10 (j) + V elocityC10 (j) ;
C01 (j) = C01 (j) + V elocityC01 (j) ;
end
% Evaluation parameters
for q=1:N
for h=1:length(x)
Ycal = 2 ∗ (x(h) − 1/(x(h)2 )) ∗ (C10 (q) + C01 (q)/(x(h))) ;
error1 = (Ycal − Yref (h))2 ;
M aterror (h) = error1 ;
end
Error1 = max(M aterror ) ;
M attoterror (j) = Error1 ;
end
end
C10evolution (i) = C10classe (1) ;
C01evolution (i) = C01classe (1) ;
14
generation(i)=i;
end
Errorf inal = Errorclasse (1);
References
[1] G. Marckmann, E. Verron, Comparison of hyperelastic models for rubber-like materials, Rubber Chem.
Technol. 79, (2006), pp. 835-858.
[2] T. Beda, An approach for hyperelastic model-building and parameters estimation a review of constitutive
models, European Polymer Journal, vol. 50, (2014), pp. 97-108.
[3] L. Gornet and G. Marckmann, A new isotropic hyperelastic strain energy function in terms of invariants
and its derivation into a pseudo-elastic model for Mullins eect: Application to nite element analysis,
Taylor and Francis Group, London, (2012), pp. 265-271.
[4] A. Nguessong Nkenfack, T. Beda, Z. Q. Feng, F. Peyraut HIA: A Hybrid Integral Approach to model
incompressible isotropic hyperelastic materials-Part 1: Theory, International Journal of Non-Linear Me-
chanics 84, (2016), pp. 1-11.
[5] Jos� A. L�pez-Campos, Abraham Segade, Enrique Casarejos, Jos� R. Fern�ndez, Gustavo R.
D�as, Hyperelastic characterization oriented to nite element applications using genetic algorithms,
Advances in Engineering Software, 133, (2019), pp. 52-59.
[6] G. Marckmann, Contribution � l'�tude des �lastom�res et des membranes sou�es, Th�se de
Doctorat PhD, Ecole Centrale de Nantes, 2004.
[7] T. Beda , Optimizing the Ogden Strain Energy Expression of Rubber Materials, Journal of Engineering
Materials and Technology, 127, (2005), pp. 351-353.
[8] T. Beda, Combining Approach in Stages with Least Squares for ts of data in hyperelasticity, Comptes
Rendus Mecanique, (2006), pp. 334-628.
[9] Y. Wu, H. Wang and A. Li, Parameter identication Methods for Hyperelastic and Hyper-Viscoelastic
Models, Appl. Sci., (2016), pp. 3-13.
[10] Bal� Baidi Blaise, Gambo Betchewe and Tibi Beda, Optimization of the model of Ogden
energy by the genetic algorithm method, Appl. Rheol. 29 (1), (2019), pp. 21-29.
[11] L. R. G. Treloar, The physics of rubber elasticity, Oxford, UK: Clarendon Press, (2005).
[12] Mooney, M. J. Appl. Phys. 11, (1940), pp. 582-592
[13] R. W. Ogden, Proc. R. Soc. Lond. A326, 565 (1972).
[14] E. Arruda, M. Boyce, A three-dimensional constitutive model for the large stretch behavior of rubber
elastic materials, J. Mech. Phys. Solids 41, (1993), pp. 389-412.
[15] Morrison, Methods for nonlinear least squares problems and convergence proofs, Tracking Programs and
Orbit Determination, Proc. Jet Propulsion Laboratory Seminar, (1960), pp. 1-9.
[16] Matlab, The Language of Technical Computing. Copyright (c) 1984-1999 by The MathWorks, Inc. Version
5.3.0.10183 (R11), 2013.
[17] Donali. W., Marquardt, an algorithm for least-squares estimation of nonlinear parameters, j. soc. Indust.
Appl. Math. Vol. 11. No. 2, June, (1963), pp. 431-441.
[18] Carsten Grammes, t.c Complot source code. http://www.gnuplot.info
[19] M. I. A. Lourakis, A brief description of Levenberg-Marquardt algorithm implemented by levmar, Tenichal
Repport Institute of Computer Science, Foundation for Ressearch and Technology - Hellas, (2005).
[20] T. Beda, Paul Tchoua, Guy Edgar Ntamack, Examination of Parameters Evaluation Methods in Compu-
tational Mechanics, International Journal of Science and Technology, Volume 2 No.9, September (2012),
pp. 634-641.
[21] N. Harb, N. Labed, M. Domaszewski, F. Peyraut, A new parameter identication method of soft biological
tissue combining genetic algorithm with analytical optimization, Comput. Methods Appl. Mech. Engrg.
200, (2011), pp. 208-215.
[22] J. H. Holland, Adaptation in natural and articial systems. Ann Arbor: University of Michigan Press,
(1975).
15
[23] David Goldberg E., Genetic algorithms in search, optimization and machine learning. Reading: Addison-
Wesley (1989).
[24] John R. KOZA, Genetic programming on the programming of computers by means of natural selection,
A. Bradford Book, the MIT press cambridge, massachussetts, london, England, (1998).
[25]Bal� Baidi Blaise, Gambo Betchewe and Tibi Beda, Optimization of the model of Ogden
energy by the genetic algorithm method, Book, Lambert Academic Publishing, ISBN. 978-
620-0-23092-8, (2019), pp. 1-53.
[26] Davidson, Wc., variable metric method for minimization, SIAM journal on optimization, 1, (1991), pp.
1-17.
[27] John R. KOZA, Genetic programming on the programming of computers by means of natural selection,
A. Bradford Book, the MIT press cambridge, massachussetts, london, England, 1998.
16