Application of Intelligent Paradigm through Neural Networks for Numerical Solution of Multiorder Fractional Differential Equations

In this study, the intelligent computational strength of neural networks (NNs) based on the backpropagated Levenberg-Marquardt (BLM) algorithm is utilized to investigate the numerical solution of nonlinear multiorder fractional differential equations (FDEs). The reference data set for the design of the BLM-NN algorithm for different examples of FDEs are generated by using the exact solutions. To obtain the numerical solutions, multiple operations based on training, validation, and testing on the reference data set are carried out by the design scheme for various orders of FDEs. The approximate solutions by the BLM-NN algorithm are compared with analytical solutions and performance based on mean square error (MSE), error histogram (EH), regression, and curve fitting. This further validates the accuracy, robustness, and efficiency of the proposed algorithm.


Introduction
Mathematicians have regarded the theory of fractional calculus as a branch of pure mathematics for nearly three centuries. However, several researchers have recently discovered that noninteger derivatives and integrals are more useful for modelling phenomena with inherited and memory properties than integer orders [1][2][3][4]. Fractional differential equations (FDEs) are used to model various problems in science, engineering, economics, biological sciences, and applied mathematics [5][6][7][8]. FDEs are more complex than their integer order since the fractional operators are nonlocal and have weakly singular kernels [9][10][11][12][13]. e complications in integer order introduce significant computational difficulties for numerical methods to obtain solutions for such equations.
Fractional differential equations have wide application in the fields of science and engineering. Some recent applications include fractional-order financial systems [14], electrical circuits [15], nuclear magnetic resonance [16], fractional-order Bloch system [17], fractional-order Lorenz system [18], hepatitis B disease in medicine [19], pollution levels in a lake [20], and fractional-order Chua's system [21]. Due to the high usage of FDEs, several numerical and analytical methods have been proposed. Bhrawy [22][23][24] uses spectral methods based on Jacobi, Chebyshev, and Legendre polynomials over a bounded domain for an approximate solution of FDE's. Atabakzadeh [25] and Tripathi [26] use the operational matrix of Caputo fractional-order derivatives for Chebyshev polynomials and fractional integration of the generalized hat basis functions to solve systems of FDEs. Baleanu [27] in 2013 uses modified generalized Laguerre collocation methods and the Tau method based on semi-infinite interval to calculate the approximate solution for linear and nonlinear FDEs. Ahmadian [28,29] applied the Jacobi operational matrix to study a class of linear fuzzy FDEs. e spectral approximation method is used by Li [30] to compute the fractional derivative and integral and also presents the pseudo-spectral approximation technique for some classes of FDEs. Esmaeili [31] developed a numerical technique in which the properties of the Caputo derivative were used to reduce the fractional differential equation into a Volterra integral equation.
Recently, the use of spectral methods to solve various types of differential and integral equations has gained interest due to their wide applicability in both finite and infinite domains [32][33][34][35].
In recent times, stochastic numerical techniques based on artificial intelligence have been developed to solve stiff nonlinear problems arising in various fields. Such stochastic computing techniques use artificial neural networks to model approximate solutions. ese numerical solvers have wide applications in various fields including petroleum engineering [50], heat transfer [51], civil engineering [52], wire coating dynamics [53], and diabetic retinopathy classification [54]. e abovementioned techniques inspire the authors to explore and incorporate soft computing architectures as an alternative, precise, and feasible way for solving nonlinear multiorder fractional differential equations. e main purpose of this article is to obtain approximate solutions for FDEs using artificial neural networks based on the Levenberg-Marquard algorithm. Some highlighted features of the given study are illustrated as follows: (i) Novel applications of neuroheuristic techniques based on backpropagated Levenberg-Marquardt neural networks (BLM-NNs) are presented to obtain numerical solutions for different classes of nonlinear multi-order fractional differential equations.
(ii) e processes of training, validation, and testing are carried out by generating a reference solution or data set by using an analytical solution for different cases and examples of FDEs.
(iii) e performance of the proposed scheme is incorporated by fitting the approximate solutions with the reference solution. e absolute error between the targeted data and approximate solutions illustrates the worth and accuracy of the BLM-NN algorithm.
(iv) Convergence analysis based on mean square errors of the objective function, regression analysis, and histogram plots are employed to study the complexity, robustness, and correctness of the design scheme.
(v) e advantage of the proposed design is that it does not require any initial parameter settings. It has simple and smooth implementation with exhaustive applicability and stability.

Solution Methodology
In the field of artificial intelligence (AI), supervised machine learning refers to a collection of algorithms that describe a predictive model based on data set with known outcomes. e method is learned through the uses of an efficient teaching algorithm, such as artificial neural networks, which use optimization procedures to minimise the error function. e infrastructure of the proposed BLM-NN algorithm is based on two fundamental steps. In the first step, a data set of 1201 points is generated by using an analytical solution from 0 to 6 with a 0.005 step size. In the next step, the Levenberg-Marquardt framework of fitting tool "nftool" from the neural network toolbox of MATLAB R2018a is used to approximate the solutions with 75% training, 15% validation, and 15% testing. e suggested structure of the BLM-NN algorithm with 60 neurons is shown in Figure 1. A summary of the working procedure of the design scheme is presented through the flow chart in Figure 2. e performance of a design scheme is measured through the performance indicators in terms of mean square error (MSE) of fitness function, regression R 2 , error histograms, and absolute errors (AE). e mathematical formulation of the MSE, R 2 , and AE is given as follows: and AE � u j (t) − u j (t) , j � 1, 2, . . . , k.
Here, u j , u j , and u j denote the reference, approximate, and mean of the solution at jth input and k is the number of mesh points. e desired value of MSE and AE for perfect fitting is equal to zero, while the value of R 2 is one.

Numerical Experimentation
In Step I

End
In first step, a data set of 1201 points is generated by using analytical solution from 0 to 6 with 0.005 step size for each example of fractional differential equation.    Figure 2: A complete overview of the working procedure of the BLM-NN algorithm.
In order to find approximate solutions for various orders of (3), the BLM-NN algorithm is executed using "nftool" in the MATLAB package. e performance and convergence of the mean square error (MSE) of the objective function are shown in Figure 3. It can be seen that the best validated performance for υ � 1.2, 1.4, 1.6, and 1.8 are 1.0846e − 10, 8.5718e − 11, 9.7898e − 11, and 1.7456e − 10 which are attained at 1000 epoch. Table 1 demonstrates the approximate solution for each case of Example 1. Further, the fitting of approximate solutions with analytical solutions is plotted in Figure 4. e absolute errors between targeted data and obtained solutions for multiple orders of (3) are illustrated in Figures 4 and 5, respectively. e values of AE for each case lie around 10 − 5 to 10 − 6 , 10 − 5 to 10 − 7 , 10 − 5 to 10 − 7 , and 10 − 5 to 10 − 8 , respectively. Table 2 represents the measure of convergence for each testing, validation, training, gradient, mu, and complexity analysis in terms of time taken by the system to achieve the desired results. It can be seen that the values for the gradient for each case lie around 10 − 4 to 10 − 7 , while the maximum time taken by the system is 5 seconds. e training state of operators during the process of optimization for Example 1 is shown in Figure 6. e accuracy and efficiency of the proposed algorithm is shown by the results of regression as dictated in Figure 7.
Example 2. Consider the following nonlinear multiterm nonhomogenous fractional differential equation as [56] subjected to the following equation: where α 1 and α 2 are 0.75 and 1.25, respectively. e exact solution of (5) is u(x) � x 3 /3. e approximate solution obtained by the proposed algorithm for (5) are shown in Figure 8(a). In addition, Figure 8(b) shows the accuracy of the solutions in terms of residual errors. It shows the accuracy of the solutions as the errors are approaching zero.
Further, to validate the efficiency, absolute errors in solutions of BLM-NN are dictated through Table 3. It can be observed that the results of the design scheme overlap the exact solutions with minimum absolute errors as compared to the Haar wavelet collocation method [56] and the Bernoulli wavelet operational matrix method [57].
Example 3. Consider the following nonlinear multiterm fractional differential equation as follows: where 0 < α 1 ≤ 1, 1 < α 2 ≤ 2 and f(x) are defined as follows: with initial conditions as follows: for , α 1 � 0.5, and α 2 � 1.5; the exact solution of (7) is y(t) � 2 − 1/2x 2 . Figure 9 depicts the comparison of exact and numerical solutions obtained by the design algorithm for (7). e results calculated by the BLM-NN algorithm are compared with those obtained by the generalized block pulse operational matrix method [58] as shown in Table 4. e absolute errors lie around 10 − 7 to 10 − 8 . e values of the performance function in terms of mean square error are shown in Table 2.
e results in terms of computational complexity and absolute errors show the accuracy of the  proposed algorithm in calculating solutions to fractional differential equations.
Example 4. Consider the following system of fractional differential equation: with the following equation:    Computational Intelligence and Neuroscience e exact solutions of (10) and (11) are given as follows:   Computational Intelligence and Neuroscience 7

Conclusion
In this paper, we have designed an integrated soft computing technique based on supervised learning. e computational strength of neural networks is utilized by the backpropagated Levenberg-Marquardt (BLM) algorithm to find approximate solutions for nonlinear multiorder fractional differential equations. e working procedure of BLM-NN algorithms is categorized into two steps in which the reference solution is generated by using analytical solutions. Furthermore, the data set is used by the BLM algorithm for validation, testing, and training of approximate solutions. Multiple figures, in terms of approximate solutions, curve fitting of analytical solutions and output data, error histograms, and regression and convergence of performance, are plotted to validate the efficiency of the design scheme. e tabulated data and figures dictate the accuracy, efficiency, and robustness of the design paradigm.
In the future, the authors would like to extend the concept of soft computing based on neural networks to solve the mathematical models represented by partial differential equations and partial fractional differential equations.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon reasonable request.

Conflicts of Interest
e authors declare that they have no conflicts of interest. Output ~= 1*Target + -6.3e-08 All: R=1 (f )