Hindawi Publishing Corporation

Advances in Mathematical Physics

Volume 2015, Article ID 439526, 12 pages
http://dx.doi.org/10.1155/2015/439526

Research Article
A Numerical Method for Solving Fractional Differential
Equations by Using Neural Network

Haidong Qu1 and Xuan Liu2

1
Department of Mathematics, Hanshan Normal University, Chaozhou 521041, China
2
Department of Basic Education, Hanshan Normal University, Chaozhou 521041, China

Correspondence should be addressed to Haidong Qu; [email protected] and Xuan Liu; [email protected]

Received 5 February 2015; Revised 27 April 2015; Accepted 27 April 2015

Academic Editor: Fawang Liu

Copyright © 2015 H. Qu and X. Liu. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

We present a new method for solving the fractional differential equations of initial value problems by using neural networks
which are constructed from cosine basis functions with adjustable parameters. By training the neural networks repeatedly the
numerical solutions for the fractional differential equations were obtained. Moreover, the technique is still applicable for the coupled
differential equations of fractional order. The computer graphics and numerical solutions show that the proposed method is very
effective.

1. Introduction (PSO) algorithm along with feedforward ANN obtained

the numerical solutions for fractional differential equations.
Recently, fractional differential equations have gained con- But the convergence of the algorithm has not been proven,
siderable importance due to their frequent appearance appli- and this method is only applied to the single fractional
cations in fluid flow, rheology, dynamical processes in self- differential equations. In this paper, we construct two dif-
similar and porous structures, diffusive transport akin to dif- ferent neural networks based on cosine functions and obtain
fusion, electrical networks, probability and statistics, control the conditions of algorithm convergence.
theory of dynamical systems, viscoelasticity, electrochemistry The first neural network (NU) is applied to linear and
of corrosion, chemical physics, optics and signal processing nonlinear fractional differential equations of the form
[1–7], and so on. These applications in interdisciplinary sci-
ences motivate us to try to find out the analytic or numerical 𝛼
𝐷0+ 𝑦 (𝑥) = 𝑓 (𝑥, 𝑦 (𝑥)) , 0 < 𝑥 ≤ 1, 0 < 𝛼 ≤ 1 (1)
solutions for the fractional differential equations. But for
most ones it is difficult to find out or even have exact solu- with initial condition as follows:
tions. Thus, necessarily, the numerical techniques are applied
to the fractional differential equations. 𝑦 (0) = 𝐶, (2)
Now, many effective methods for solving fractional
differential equations have been presented, such as nonlinear 𝛼
where 𝐷0+ is the Caputo fractional derivatives of order 𝛼.
functional analysis method including monotone iterative
The second neural network (NU) is applied to the frac-
technique [8, 9], topological degree theory [10], and fixed
tional coupled differential equations of the form
point theorems [11–13]. Also, numerical solutions are
obtained by the following methods: random walk [2], matrix 𝛼
approach [14], the Adomian decomposition method and 𝐷0+ 𝑦 (𝑥) = 𝑓 (𝑥, 𝑦 (𝑥) , 𝑧 (𝑥)) ,
variational iteration method [15], HAM [16–19], homotopy 0 < 𝑥 ≤ 1, 0 < 𝛼 ≤ 1, (3)
perturbation method (HPM) [20], and so forth. Not long ago,
𝛼
in [21], Raja et al. by applying Particle Swarm Optimization 𝐷0+ 𝑧 (𝑥) = 𝑔 (𝑥, 𝑦 (𝑥) , 𝑧 (𝑥)) ,
2 Advances in Mathematical Physics

with initial conditions as follows: The generalized Mittag-Leffler function is defined by

𝑦 (0) = 𝐶1 ,
∞
(4) 𝑥𝑘
𝑧 (0) = 𝐶2 , E𝛼,𝛽 (𝑥) := ∑ , (𝑥, 𝛽 ∈ 𝐶, 𝛼 > 0) . (8)
𝑘=0
Γ (𝛼𝑘 + 𝛽)
𝛼
where 𝐷0+ is the Caputo fractional derivatives of order 𝛼.
The solutions for the above two problems are written as Definition 4. The functions Sin𝛼,𝛽 (𝑥), Cos𝛼,𝛽 (𝑥) (𝑥, 𝛽 ∈
cosine basis functions, whose parameters can be adjusted 𝐶, 𝛼 > 0) are defined by
to minimize an appropriate error function. So we need
to compute the gradient of the error with respect to the
∞
network parameters. By adjusting the parameters repeatedly, 𝑥2𝑘−1
we obtain the numerical solutions when the error values are Sin𝛼,𝛽 (𝑥) = ∑ (−1)𝑘+1 ,
𝑘=1
Γ (𝛼 (2𝑘 − 1) + 𝛽)
less than the required accuracy or the training times reach (9)
maximum. ∞
𝑥2𝑘
𝑘
Cos𝛼,𝛽 (𝑥) = ∑ (−1) .
𝑘=0
Γ (𝛼 (2𝑘) + 𝛽)
2. Definitions and Lemma
Obviously, Euler’s equations have the following forms:
Definition 1 (see [22]). The Riemann-Liouville fractional
integral of order 𝛼 ∈ 𝑅, 𝛼 > 0 of a function 𝑓(𝑥) ∈ 𝐶𝜇 , 𝜇 ≥ −1
E𝛼,𝛽 (𝑖𝑥) = Cos𝛼,𝛽 (𝑥) + 𝑖Sin𝛼,𝛽 (𝑥) ,
is defined as
(10)
𝛼 1 𝑥 𝑓 (𝑡) 𝑑𝑡 E𝛼,𝛽 (−𝑖𝑥) = Cos𝛼,𝛽 (𝑥) − 𝑖Sin𝛼,𝛽 (𝑥) .
(𝐼0+ 𝑓 (𝑡)) (𝑥) := ∫ , (𝑥 > 0) . (5)
Γ (𝛼) 0 (𝑥 − 𝑡)1−𝛼
Lemma 5. If Sin𝛼,𝛽 (𝑥) and Cos𝛼,𝛽 (𝑥) are defined as in
Definition 2 (see [22]). The Riemann-Liouville and Caputo Definition 4, then
fractional derivatives of order 𝛼 ∈ 𝑅, 𝛼 > 0 are given by

RL 𝛼 𝑑 𝑛 𝑛−𝛼 𝛼
𝐷𝑎+ (𝑥 − 𝑎)𝛽−1 Sin𝜇,𝛽 [𝜆 (𝑥 − 𝑎)𝜇 ]
𝐷0+ 𝑓 (𝑥) := ( ) 𝐼0+ 𝑓 (𝑥)
𝑑𝑥 (11)
𝑥 𝑛 = (𝑥 − 𝑎)𝛽−𝛼−1 Sin𝜇,𝛽−𝛼 [𝜆 (𝑥 − 𝑎)𝜇 ] ,
1 𝑑 𝑓 (𝑡) 𝑑𝑡
= ( ) ∫ 𝛼−𝑛+1
,
Γ (𝑛 − 𝛼) 𝑑𝑥 0 (𝑥 − 𝑡) 𝛼
𝐷𝑎+ (𝑥 − 𝑎)𝛽−1 Cos𝜇,𝛽 [𝜆 (𝑥 − 𝑎)𝜇 ]
(6) (12)
𝑑 𝑛
𝛼
𝐷0+ 𝑛−𝛼
𝑓 (𝑥) := 𝐼0+ ( ) 𝑓 (𝑥) = (𝑥 − 𝑎)𝛽−𝛼−1 Cos𝜇,𝛽−𝛼 [𝜆 (𝑥 − 𝑎)𝜇 ] .
𝑑𝑥
1 𝑥
(𝑑/𝑑𝑡)𝑛 𝑓 (𝑡) 𝑑𝑡 1
= ∫ , ̃ 𝑞) := ∫ 𝑥𝑝−1 (1−
Proof. The beta function was defined by 𝛽(𝑝,
Γ (𝑛 − 𝛼) 0 (𝑥 − 𝑡)𝛼−𝑛+1 0
𝑥)𝑞−1 𝑑𝑥, and we have the following equation:
where (𝑛 = [𝛼] + 1, 𝑥 > 0).
Γ (𝑝) Γ (𝑞)
Definition 3 (see [22]). The classical Mittag-Leffler function 𝛽̃ (𝑝, 𝑞) = . (13)
is defined by Γ (𝑝 + 𝑞)
∞
𝑥𝑘
E𝛼 (𝑥) := ∑ , (𝑥 ∈ 𝐶, 𝛼 > 0) . (7) Then according to the definition of Caputo fractional deriva-
𝑘=0
Γ (𝛼𝑘 + 1) tives, we have

𝑥 (𝑑/𝑑𝑡)𝑛 (𝑡 − 𝑎)𝛽−1 Sin 𝜇

𝛼 1 𝜇,𝛽 [𝜆 (𝑡 − 𝑎) ]
𝐷𝑎+ (𝑥 − 𝑎)𝛽−1 Sin𝜇,𝛽 [𝜆 (𝑥 − 𝑎)𝜇 ] = ∫ 𝑑𝑡
Γ (𝑛 − 𝛼) 𝑎+ (𝑥 − 𝑡)𝛼−𝑛+1

𝑥−𝑎 (𝑑/𝑑𝜉)𝑛 𝜉𝛽−1 Sin 𝜇

1 𝜇,𝛽 [𝜆𝜉 ]
= ∫ 𝛼−𝑛+1
𝑑𝜉 (𝜉 = 𝑡 − 𝑎)
Γ (𝑛 − 𝛼) 0+ (𝑥 − 𝜉 − 𝑎)

1 (𝑑/𝑑𝑡)𝑛 𝑡𝛽−1 (𝑥 − 𝑎)𝛽−1 Sin 𝜇 𝜇

1 𝜇,𝛽 [𝜆𝑡 (𝑥 − 𝑎) ]
= ∫ (𝑥 − 𝑎) 𝑑𝑡 (𝜉 = 𝑡 (𝑥 − 𝑎))
Γ (𝑛 − 𝛼) 0+ (𝑥 − 𝑎)𝑛 (𝑥 − 𝑎)𝛼−𝑛+1 (1 − 𝑡)𝛼−𝑛+1
Advances in Mathematical Physics 3

𝑛 𝛽−1 𝑘+1 ∞
𝜇 𝜇 2𝑘−1
(𝑥 − 𝑎)𝛽−𝛼−1 1 (𝑑/𝑑𝑡) 𝑡 ∑𝑘=1 ((−1) [𝜆𝑡 (𝑥 − 𝑎) ] /Γ (𝜇 (2𝑘 − 1) + 𝛽))
= ∫ 𝑑𝑡
Γ (𝑛 − 𝛼) 0+ (1 − 𝑡)𝛼−𝑛+1

∞ 𝑘+1 𝜇 2𝑘−1
(𝑥 − 𝑎)𝛽−𝛼−1 ∑𝑘=1 (−1) [𝜆 (𝑥 − 𝑎) ] 1
(𝑑/𝑑𝑡)𝑛 𝑡𝛽−1+𝜇(2𝑘−1)
= ∫ 𝑑𝑡
Γ (𝑛 − 𝛼) Γ (𝜇 (2𝑘 − 1) + 𝛽) 0+ (1 − 𝑡)𝛼−𝑛+1
∞ 𝑘+1 𝜇 2𝑘−1 𝑛
(𝑥 − 𝑎)𝛽−𝛼−1 ∑𝑘=1 (−1) [𝜆 (𝑥 − 𝑎) ] 1
𝑡𝛽−1+𝜇(2𝑘−1)−𝑛
= ∏ [𝛽 − 𝑖 + 𝜇 (2𝑘 − 1)] ∫ 𝑑𝑡
Γ (𝑛 − 𝛼) Γ (𝜇 (2𝑘 − 1) + 𝛽) 𝑖=1 0+ (1 − 𝑡)𝛼−𝑛+1
∞ 𝑘+1 𝜇 2𝑘−1
(𝑥 − 𝑎)𝛽−𝛼−1 ∑𝑘=1 (−1) [𝜆 (𝑥 − 𝑎) ]
= 𝛽̃ (𝛽 + 𝜇 (2𝑘 − 1) − 𝑛, 𝑛 − 𝛼)
Γ (𝑛 − 𝛼) Γ (𝜇 (2𝑘 − 1) + 𝛽 − 𝑛)
∞ 𝑘+1 𝜇 2𝑘−1
(𝑥 − 𝑎)𝛽−𝛼−1 ∑𝑘=1 (−1) [𝜆 (𝑥 − 𝑎) ] Γ (𝛽 + 𝜇 (2𝑘 − 1) − 𝑛) Γ (𝑛 − 𝛼)
=
Γ (𝑛 − 𝛼) Γ (𝜇 (2𝑘 − 1) + 𝛽 − 𝑛) Γ (𝛽 + 𝜇 (2𝑘 − 1) − 𝛼)
2𝑘−1
∞
(−1)𝑘+1 [𝜆 (𝑥 − 𝑎)𝜇 ]
= (𝑥 − 𝑎)𝛽−𝛼−1 ∑ = (𝑥 − 𝑎)𝛽−𝛼−1 Sin𝜇,𝛽−𝛼 [𝜆 (𝑥 − 𝑎)𝜇 ] .
𝑘=1
Γ (𝜇 (2𝑘 − 1) + 𝛽 − 𝛼)
(14)

Then (11) holds. Similarly, we obtain (12). In particular, when where N represents the number of sample points, ‖ ⋅ ‖2 is
𝛽 = 1, 𝜇 = 1, we have Euclidean norm, and
𝛼
𝐷𝑎+ Sin1,1 [𝜆 (𝑥 − 𝑎)] 𝛼
𝑒𝑗 (𝑘) = 𝑓 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 )) − 𝐷0+ 𝑦𝑗 (𝑥𝑘 )
𝛼
= 𝐷𝑎+ sin [𝜆 (𝑥 − 𝑎)] M

−𝛼
= 𝑓 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 )) − 𝑥𝑘−𝛼 (∑𝑤𝑖,𝑗 Cos1,1−𝛼 (𝑖𝑥𝑘 )
= (𝑥 − 𝑎) Sin1,1−𝛼 [𝜆 (𝑥 − 𝑎)] , 𝑖=1 (18)
(15) M
𝛼
𝐷𝑎+ Cos1,1 [𝜆 (𝑥 − 𝑎)]
+ (𝐶 − ∑𝑤𝑖,𝑗 ) Cos1,1−𝛼 ((M + 1) 𝑥𝑘 )) ,
𝛼 𝑖=1
= 𝐷𝑎+ cos [𝜆 (𝑥 − 𝑎)]
where 𝑘 = 1, 2, . . . , N; then we can adjust the weights 𝑤𝑖,𝑗 by
= (𝑥 − 𝑎)−𝛼 Cos1,1−𝛼 [𝜆 (𝑥 − 𝑎)] .
the following equation:

𝑤𝑖,𝑗+1 = 𝑤𝑖,𝑗 + Δ𝑤𝑖,𝑗 , (19)

3. Illustration of the Method and Application
where
3.1. The First Neural Network. To describe the method, we
consider (1) with initial condition 𝑦(0) = 𝐶. The 𝑗th trial solu- 𝜕𝐽 N
𝜕𝐽 𝜕𝑒𝑗 (𝑘)
Δ𝑤𝑖,𝑗 = −𝜇 = −𝜇 ∑
tion satisfying the initial condition is written as 𝜕𝑤𝑖,𝑗 𝜕𝑒 (𝑘) 𝜕𝑤𝑖,𝑗
𝑘=1 𝑗
M N
𝑦𝑗 (𝑥) = ∑𝑤𝑖,𝑗 cos (𝑖𝑥) = −𝜇 ∑ 𝑒𝑗 (𝑘) 𝑓𝑦 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 )) (20)
𝑖=1 𝑘=1
(16)
M −𝛼
⋅ (cos (𝑥𝑘 ) − cos ((M + 1) 𝑥𝑘 )) − (𝑥𝑘 ) 𝑒𝑗 (𝑘)
+ (𝐶 − ∑𝑤𝑖,𝑗 ) cos ((M + 1) 𝑥) ,
𝑖=1
⋅ (Cos1,1−𝛼 𝑥𝑘 − Cos1,1−𝛼 ((M + 1) 𝑥𝑘 )) .
where M represents the number of neurons and 𝑤𝑖,𝑗 are
unknown weights of the network determined in training 3.2. Convergence of the Algorithm
procedures to reduce the error function:
Theorem A. Let 𝜇 represent learning rate, let N represent the
1 󵄩󵄩 󵄩󵄩2 1 N 2
𝐽= 󵄩󵄩𝐸 󵄩󵄩 = ∑ (𝑒 (𝑘)) , number of sample points, and let M represent the number of
2 󵄩 𝑗 󵄩2 2 𝑘=1 𝑗 neurons: 𝑋 = (𝑥1 , 𝑥2 , . . . , 𝑥N ), 𝛿 < 𝑥𝑖 < 1, 1 ≤ 𝑖 ≤ 𝑁.
(17)
𝑇
Suppose |𝑓𝑦 | ≤ 𝐿 1 , |Cos1,1−𝛼 (𝑥)| ≤ 𝐿 2 on the interval (𝛿, 1)
𝐸𝑗 = (𝑒𝑗 (1) , 𝑒𝑗 (2) , . . . , 𝑒𝑗 (N)) , for 0 < 𝛿 < 1. (From Figure 10, we see that the function
4 Advances in Mathematical Physics

Cos1,1−𝛼 (𝑥) is bounded when 0 ≤ 𝛼 ≤ 1.) Then the neural Thus,

network is convergent on the interval (𝛿, 1) when
󵄩 󵄩󵄩2
1 󵄩󵄩󵄩󵄩 𝜕𝐸𝑗 𝑇 𝜕𝐸𝑗 󵄩󵄩 1 󵄩 󵄩2
0<𝜇<
N+1
. Δ𝑉𝑗 = 󵄩󵄩(𝐼 − 𝜇 ( ) ) 𝐸𝑗 󵄩󵄩󵄩 − 󵄩󵄩󵄩󵄩𝐸𝑗 󵄩󵄩󵄩󵄩2
2 (21) 2 󵄩󵄩󵄩 𝜕𝑊𝑗 𝜕𝑊𝑗 󵄩󵄩
󵄩2 2
4MN (𝐿 1 + 𝛿−𝛼 𝐿 2 )
(30)
Proof. Let 𝑊𝑗 = (𝑤1𝑗 , 𝑤2𝑗 , . . . , 𝑤M𝑗 ), and then we denote 󵄩 𝜕𝐸𝑗 𝑇 𝜕𝐸𝑗 󵄩󵄩󵄩󵄩
2
1 󵄩󵄩󵄩󵄩 󵄩 󵄩2
𝑦𝑗 (𝑋) by ≤ (󵄩󵄩(𝐼 − 𝜇 ( ) )󵄩 − 1) 󵄩󵄩󵄩󵄩𝐸𝑗 󵄩󵄩󵄩󵄩2 ,
󵄩
2 󵄩󵄩󵄩 𝜕𝑊𝑗 𝜕𝑊𝑗 󵄩󵄩󵄩󵄩𝐹
𝑦𝑗 (𝑋) = 𝑊𝑗 ⋅ 𝐺 − (𝐶 − 𝑊𝑗 ⋅ 𝐼1 ) cos ((M + 1) 𝑋) , (22)
where ‖⋅‖𝐹 is Frobenius matrix norm, defined by ‖(𝑎𝑖𝑗 )𝑛×𝑛 ‖𝐹 =
where 𝐺 = (cos(𝑋), cos(2𝑋), . . . , cos(M𝑋)) and 𝐼1 = 𝑇
(∑𝑛𝑖=1 ∑𝑛𝑗=1 𝑎𝑖𝑗2 )1/2 . Since ‖𝐸𝑗 ‖22 > 0, in order to make this neu-
M
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞ ral network converge, we have
(1, 1, . . . , 1 )𝑇 . Then according to (17), we have
󵄩󵄩 𝜕𝐸 𝑇 𝜕𝐸𝑗 󵄩󵄩󵄩󵄩
2
𝐸𝑗 󵄩󵄩
󵄩󵄩𝐼 − 𝜇 ( 𝑗 ) 󵄩󵄩 < 1, (31)
󵄩󵄩 𝜕𝑊𝑗 𝜕𝑊𝑗 󵄩󵄩󵄩󵄩𝐹
󵄩󵄩
= (𝑒𝑗 (1) , 𝑒𝑗 (2) , . . . , 𝑒𝑗 (N))

𝛼
which yields
= 𝑓 (𝑋, 𝑦𝑗 (𝑋)) − 𝐷0+ 𝑦𝑗 (𝑋) (23)
󵄩󵄩 𝜕𝐸 𝑇 𝜕𝐸𝑗 󵄩󵄩󵄩󵄩
󵄩󵄩
= 𝑓 (𝑋, 𝑦𝑗 (𝑋)) 󵄩󵄩𝐼 − 𝜇 ( 𝑗 ) 󵄩󵄩 < 1. (32)
󵄩󵄩 𝜕𝑊𝑗 𝜕𝑊𝑗 󵄩󵄩󵄩󵄩𝐹
󵄩󵄩
− 𝑋𝛼 (𝑊𝑗 𝐻 + (𝐶 − 𝑊𝑗 𝐼1 ) Cos1,1−𝛼 ((M + 1) 𝑋)) ,
Hence,
where 𝐻 = (Cos1,1−𝛼 (𝑋), Cos1,1−𝛼 (2𝑋), . . . , Cos1,1−𝛼 (M𝑋))𝑇 . 󵄩󵄩
󵄩󵄩 𝜕𝐸𝑗 𝑇 𝜕𝐸𝑗 󵄩󵄩󵄩󵄩 󵄩󵄩 𝜕𝐸 󵄩󵄩2
󵄩 󵄩
Then we have 1 > 󵄩󵄩󵄩𝐼 − 𝜇 ( ) 󵄩󵄩 ≥ 𝜇 󵄩󵄩󵄩 𝑗 󵄩󵄩󵄩 − ‖𝐼‖𝐹
󵄩󵄩 𝜕𝑊 𝜕𝑊 󵄩󵄩 󵄩󵄩 𝜕𝑊 󵄩󵄩
󵄩 𝑗 𝑗󵄩 󵄩𝐹 󵄩󵄩 𝑗 󵄩󵄩𝐹
𝜕𝐸𝑗
= 𝑓𝑦 (𝐺 − 𝐼1 cos ((M + 1) 𝑋)) (33)
𝜕𝑊𝑗 󵄩󵄩 𝜕𝐸 󵄩󵄩2
(24) 󵄩󵄩 𝑗 󵄩󵄩
= 𝜇 󵄩󵄩󵄩 󵄩󵄩 − N.
− 𝑋𝛼 (𝐻 − 𝐼1 Cos1,1−𝛼 ((M + 1) 𝑋)) . 󵄩󵄩 𝜕𝑊𝑗 󵄩󵄩󵄩
󵄩 󵄩𝐹

Noting 𝐽 = (1/2)‖𝐸𝑗 ‖22 , then we get In accordance with 0 < 𝜇 < 1, we obtain

𝜕𝐽 𝜕𝐸𝑗 𝜕𝐸𝑗
0<𝜇< 󵄩
N+1
.
Δ𝑊𝑗 = −𝜇 = −𝜇𝐸𝑗 .
𝜕𝐸𝑗 𝜕𝑊𝑗 𝜕𝑊𝑗
(25) 󵄩󵄩𝜕𝐸 /𝜕𝑊 󵄩󵄩󵄩2 (34)
󵄩󵄩 𝑗 𝑗󵄩 󵄩𝐹
Define Lyapunov function 𝑉𝑗 = (1/2)‖𝐸𝑗 ‖22 ; we have By calculating (25), we get
1 󵄩󵄩 󵄩2 1 󵄩 󵄩2 󵄩󵄩 𝜕𝐸 󵄩󵄩2
Δ𝑉𝑗 = 󵄩󵄩𝐸𝑗+1 󵄩󵄩󵄩 − 󵄩󵄩󵄩𝐸𝑗 󵄩󵄩󵄩 . (26) 󵄩󵄩 𝑗 󵄩󵄩
2 󵄩 󵄩2 2 󵄩 󵄩2 󵄩󵄩 󵄩
󵄩󵄩 𝜕𝑊 󵄩󵄩󵄩
󵄩󵄩 𝑗 󵄩󵄩𝐹
Suppose
󵄩󵄩 2 (𝐿 + 𝛿−𝛼 𝐿 ) ⋅ ⋅ ⋅ 2 (𝐿 + 𝛿−𝛼 𝐿 ) 󵄩󵄩2
𝜕𝐸𝑗 𝑇 󵄩󵄩 1 2 1 2 󵄩󵄩
𝐸𝑗+1 = 𝐸𝑗 + Δ𝐸𝑗 = 𝐸𝑗 + ( ) Δ𝑊𝑗 , 󵄩󵄩 󵄩󵄩
(27) 󵄩󵄩 . . 󵄩󵄩 (35)
𝜕𝑊𝑗 ≤ 󵄩󵄩( 󵄩 .. .. ) 󵄩󵄩
󵄩󵄩 󵄩󵄩
󵄩󵄩 󵄩󵄩
󵄩󵄩 2 (𝐿 + 𝛿−𝛼 𝐿 ) ⋅ ⋅ ⋅ 2 (𝐿 + 𝛿−𝛼 𝐿 ) 󵄩󵄩
and then in accordance with (25) that yields 󵄩 1 2 1 2 M×N 󵄩𝐹
󵄩
𝑇
𝜕𝐸𝑗 𝜕𝐸𝑗 = 4MN (𝐿 1 + 𝛿−𝛼 𝐿 2 ) .
2
𝐸𝑗+1 = (𝐼 − 𝜇 ( ) ) 𝐸𝑗 , (28)
𝜕𝑊𝑗 𝜕𝑊𝑗
Finally, we have
where
N+1
1 0<𝜇< 2
. (36)
4MN (𝐿 1 + 𝛿−𝛼 𝐿 2 )
𝐼=( d ) . (29)
1 N×N
Advances in Mathematical Physics 5

Table 1: Weights (×10−4 ) obtained along with the solution of Examples 1, 2, and 3.

Example 1 Example 2 Example 3

𝛼
1 0.7 0.5 1 0.7 0.5 1 0.7 0.5
𝑤1 4857 6310 6665 8226 9941 8880 5415 4070 5101
𝑤2 −0506 −3466 −4424 4080 0219 2914 −1589 0680 −1582
𝑤3 −4434 −2721 −1362 −3468 −0621 −3593 −5771 −5800 −3972
𝑤4 −3170 −3110 −4990 1680 1372 3715 −1309 −2908 −4142
𝑤5 1896 4204 5290 −1455 −1468 −3147 3295 3947 6096
𝑤6 5534 −0926 0222 1895 0482 1372 2815 2901 −0121
𝑤7 −6316 0182 −2151 −1350 0424 0273 −4609 −4203 −2015

Learning curve Inspection curve

1.4 1.4

1.2 1.2

1.0 1.0

0.8 0.8
y = x2

y = x2

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x x

Exact solution Exact solution

Sample points Checkpoint

Figure 1: The learning curve for Example 1. Figure 2: The inspection curve for Example 1.

3.3. Example
are done by using Lenovo T400, Intel Core 2 Duo CPU P8700,
3.3.1. Example 1. We first consider the following linear frac-
2.53 GHz, and Matlab version R2010b. The neural networks
tional differential equation:
with cosine basis functions have taken about 850 s, but the
𝛼 2 other algorithms mentioned above need to run about 2,240 s.
𝐷0+ 𝑦 (𝑥) = 𝑥2 + 𝑥2−𝛼 − 𝑦 (𝑥) , (37)
Γ (3 − 𝛼)
with condition 𝑦(0) = 0. The exact solution is 𝑦(𝑥) = 𝑥2 . 3.3.2. Example 2. We secondly consider the following linear
This equation also can be solved by the following methods: fractional differential equation:
Genetic Algorithm (GA) [21], Grünwald-Letnikov classical
numerical technique (GL) [23], and Particle Swarm Opti-
𝛼
mization (PSO) algorithm [23]. We set the parameters 𝜇 = 𝐷0+ 𝑦 (𝑥) = cos (𝑥) + 𝑥−𝛼 Cos1,1−𝛼 (𝑥) − 𝑦 (𝑥) , (38)
0.001, M = 7, and N = 10 and train the neural network
4500 times, and the weights of the network for Example
1 are given in Table 1. Figure 1 shows that sample points with condition 𝑦(0) = 1. The exact solution is 𝑦(𝑥) = cos(𝑥).
are on the exact solution curve after training is completed. We set the parameters 𝜇 = 0.001, M = 7, and N = 10 and
Then we check whether the other points also match well train the neural network 1000 times, and the weights of the
with the exact solution (see Figure 2). From Figure 3 we see network for Example 2 are given in Table 1. Figures 4, 5, and
the error values decrease rapidly. Tables 2(a) and 2(b) show 6 show that the neural network is still applicable when 𝐶 ≠ 0.
the numerical solutions and accuracy for Example 1 by the Table 3 shows the exact solution, approximate solution, and
different methods. In this paper, all numerical experiments accuracy for Example 2.
6 Advances in Mathematical Physics

Error curve Inspection curve

800 1

700 0.95
0.9
600
0.85
500
Error value

0.8

y = Cos(x)
400 0.75
0.7
300
0.65
200
0.6
100 0.55

0 0.5
0 500 1000 1500 2000 2500 3000 3500 4000 4500 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x
Training times

Figure 3: The error curve for Example 1. Exact solution

Checkpoint

Figure 5: The inspection curve for Example 2.

Learning curve
1
Error curve
0.95 1500
0.9
0.85
y = Cos (x)

0.8
1000
Error value

0.75
0.7
0.65
500
0.6
0.55
0.5
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
x 0 200 400 600 800 1000
Training times
Exact solution
Sample points Figure 6: The error curve for Example 2.
Figure 4: The learning curve for Example 2.

3.4. The Second Neural Network. To describe the method, we

consider (3) with initial conditions 𝑦(0) = 𝐶1 and 𝑧(0) = 𝐶2 .
The 𝑗th trial solutions for the problem are written as
3.3.3. Example 3. We thirdly consider the following nonlin-
ear fractional differential equation: M
𝑦𝑗 (𝑥) = ∑𝑤𝑖,𝑗 cos (𝑖𝑥)
𝑖=1
𝛼 6 Γ (3.5) 2.5−𝛼
𝐷0+ 𝑦 (𝑥) =𝑥 + 𝑥 − 𝑥𝑦2 (𝑥) , (39) M
Γ (3.5 − 𝛼) + (𝐶1 − ∑𝑤𝑖,𝑗 ) cos ((M + 1) 𝑥) ,
𝑖=1
(40)
M
5/2
with condition 𝑦(0) = 0. The exact solution is 𝑦(𝑥) = 𝑥 . We 𝑧𝑗 (𝑥) = ∑𝑞𝑖,𝑗 cos (𝑖𝑥)
set the parameters 𝜇 = 0.001, M = 7, and N = 10 and train 𝑖=1
the neural network 1000 times, and the weights of the network M
for Example 2 are given in Table 1. Table 4 shows the exact + (𝐶2 − ∑𝑞𝑖,𝑗 ) cos ((M + 1) 𝑥) ,
solution, approximate solution, and accuracy for Example 3. 𝑖=1
Advances in Mathematical Physics 7

Table 2: (a) Comparison of results for the solution of Example 1 for 𝛼 = 0.5. (b) Comparison of results for the solution of Example 1 for
𝛼 = 0.75.
(a)

Numerical solution Accuracy

𝑥 𝑦(𝑥)
GL PSO GA NU GL PSO GA NU
0.1 0.01 0.0101 10−4
0.2 0.04 0.0401 0.0404 0.0396 0.0407 10−4 10−4 10−4 10−4
0.3 0.09 0.0901 0.0907 0.0917 10−4 10−4 10−3
0.4 0.16 0.1601 0.1604 0.1596 0.1621 10−4 10−4 10−4 10−3
0.5 0.25 0.2501 0.2496 0.2505 10−4 10−4 10−4
0.6 0.36 0.3602 0.3583 0.3573 0.3571 10−3 10−3 10−3 10−3
0.7 0.49 0.4902 0.4869 0.4853 10−3 10−3 10−3
0.8 0.64 0.6402 0.6362 0.6352 0.6397 10−3 10−3 10−3 10−4
0.9 0.81 0.8102 0.8069 0.8186 10−3 10−3 10−3
1 1 0.1001 0.1000 0.1004 0.1003 10−4 10−5 10−4 10−3
(b)

Numerical solution Accuracy

𝑥 𝑦(𝑥)
GL PSO NU GL PSO NU
0.1 0.01 0.0107 0.0103 0.0092 10−4 10−4 10−4
0.2 0.04 0.0413 0.0414 0.0377 10−3 10−3 10−3
0.3 0.09 0.0918 0.0928 0.0875 10−3 10−3 10−3
0.4 0.16 0.1622 0.1636 0.1592 10−3 10−3 10−4
0.5 0.25 0.2527 0.2538 0.2511 10−3 10−3 10−3
0.6 0.36 0.3631 0.3631 0.3609 10−3 10−3 10−4
0.7 0.49 0.4934 0.4918 0.4884 10−3 10−3 10−3
0.8 0.64 0.6438 0.6402 0.6373 10−3 10−4 10−3
0.9 0.81 0.8141 0.8091 0.8106 10−3 10−4 10−4
1 1 1.0044 0.9991 1.0020 10−3 10−4 10−3

Table 3: Exact solution, approximate solution, and accuracy for Table 4: Exact solution, approximate solution, and accuracy for
Example 2. Example 3.

𝛼 Numerical solution Accuracy 𝛼 Numerical solution Accuracy

𝑥 cos(𝑥) 1 0.7 0.5 1 0.7 0.5 𝑥 𝑥5/2 1 0.7 0.5 1 0.7 0.5
0.1 0.9950 0.9945 0.9967 0.9972 10−4 10−3 10−3 0.1 0.0031 0.0022 0.0055 0.0066 10−4 10−3 10−3
0.2 0.9800 0.9788 0.9852 0.9867 10−3 10−3 10−3 0.2 0.0178 0.0133 0.0234 0.0266 10−3 10−3 10−3
0.3 0.9553 0.9538 0.9620 0.9638 10−3 10−3 10−3 0.3 0.0492 0.0426 0.0566 0.0603 10−3 10−3 10−3
0.4 0.9210 0.9203 0.9249 0.9256 10−4 10−3 10−3 0.4 0.1011 0.0972 0.1075 0.1093 10−3 10−3 10−3
0.5 0.8775 0.8777 0.8758 0.8747 10−4 10−3 10−3 0.5 0.1767 0.1773 0.1783 0.1772 10−4 10−3 10−4
0.6 0.8253 0.8254 0.8196 0.8176 10−4 10−3 10−3 0.6 0.2788 0.2797 0.2733 0.2711 10−4 10−3 10−3
0.7 0.7648 0.7639 0.7607 0.7601 10−4 10−3 10−3 0.7 0.4099 0.4055 0.4010 0.4009 10−3 10−3 10−3
0.8 0.6967 0.6951 0.6990 0.7016 10−3 10−3 10−3 0.8 0.5724 0.5643 0.5712 0.5738 10−3 10−3 10−3
0.9 0.6216 0.6213 0.6286 0.6328 10−4 10−3 10−2 0.9 0.7684 0.7670 0.7832 0.7847 10−3 10−2 10−2
1 0.5403 0.5414 0.5414 0.5405 10−3 10−3 10−4 1 1 1.0064 1.0105 1.0056 10−3 10−2 10−3

where N represents the number of sample points and 𝑤𝑖,𝑗 where

and 𝑞𝑖,𝑗 are unknown weights of the network determined in
𝑦+𝑧 𝑦 𝑦 𝑦
training procedures to reduce error function: 𝐸𝑗 = (𝑒𝑗 (1) , 𝑒𝑗 (2) , . . . , 𝑒𝑗 (N) , 𝑒𝑗𝑧 (1) , 𝑒𝑗𝑧 (2) , . . . ,
𝑇
𝑒𝑗𝑧 (N)) ,
1 󵄩 𝑦+𝑧 󵄩2 1 N 𝑦 2 1N 2
𝐽 𝑦+𝑧
= 󵄩󵄩󵄩󵄩𝐸𝑗 󵄩󵄩󵄩󵄩 = ∑ (𝑒𝑗 (𝑘)) + ∑ (𝑒𝑗𝑧 (𝑘)) , (41) 𝑦 𝛼
2 2 2 𝑘=1 2 𝑘=1 𝑒𝑗 (𝑘) = 𝑓 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 )) − 𝐷0+ 𝑦𝑗 (𝑥𝑘 )
8 Advances in Mathematical Physics

= 𝑓 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 )) ⋅ (𝑔𝑧 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 ))

M M (−𝛼)
− 𝑥−𝛼 (∑𝑤𝑖,𝑗 Cos1,1−𝛼 (𝑖𝑥𝑘 ) + (𝐶1 − ∑𝑤𝑖,𝑗 ) ⋅ (cos (𝑥𝑘 ) − cos ((M + 1) 𝑥𝑘 )) − (𝑥𝑘 )
𝑖=1 𝑖=1
⋅ (Cos1,1−𝛼 𝑥𝑘 − Cos1,1−𝛼 ((M + 1) 𝑥𝑘 )))

⋅ Cos1,1−𝛼 ((M + 1) 𝑥𝑘 )) , N
𝑦
− 𝜇 ∑ 𝑒𝑗 (𝑘) 𝑓𝑧 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 )) (cos (𝑥𝑘 )
𝑘=1
𝑒𝑗𝑧 (𝑘) = 𝑔 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 )) − 𝐷0+
𝛼
𝑧𝑗 (𝑥𝑘 )
− cos ((M + 1) 𝑥𝑘 )) .
= 𝑔 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 )) (44)

M M
− 𝑥−𝛼 (∑𝑞𝑖,𝑗 Cos1,1−𝛼 (𝑖𝑥𝑘 ) + (𝐶2 − ∑𝑞𝑖,𝑗 ) 3.5. Convergence of the Algorithm
𝑖=1 𝑖=1
Theorem B. Let 𝜇 represent learning rate, let N represent the
number of sample points, and let M represent the number of
⋅ Cos1,1−𝛼 ((M + 1) 𝑥𝑘 )) . neurons: 𝑋 = (𝑥1 , 𝑥2 , . . . , 𝑥N ), 𝛿 < 𝑥𝑖 < 1, 1 ≤ 𝑖 ≤ N. Suppose
𝑦 𝑦
|𝑓𝑦 | ≤ 𝐿 1 , |𝑓𝑧 | ≤ 𝐿𝑧1 , |𝑔𝑦 | ≤ 𝐿 3 , |𝑔𝑧 | ≤ 𝐿𝑧3 , |Cos1,1−𝛼 (𝑥)| ≤ 𝐿 2
(42) on the interval (𝛿, 1) for 0 < 𝛿 < 1. Then the neural network is
convergent on the interval (𝛿, 1) when
Then we adjust the weights 𝑤𝑖,𝑗 and 𝑞𝑖,𝑗 by the following two
0
equations:
<𝜇
𝑤𝑖,𝑗+1 = 𝑤𝑖,𝑗 + Δ𝑤𝑖,𝑗 , (45)
2N + 1
(43) < 𝑦 𝑦 2 2
.
𝑞𝑖,𝑗+1 = 𝑞𝑖,𝑗 + Δ𝑞𝑖,𝑗 , 4MN ((𝐿 1 + 𝐿 3 + 𝛿−𝛼 𝐿 2 ) + (𝐿𝑧1 + 𝐿𝑧3 + 𝛿−𝛼 𝐿 2 ) )

𝑦+𝑧
where Proof. Let 𝑊𝑗 = (𝑤1𝑗 , 𝑤2𝑗 , . . . , 𝑤M𝑗 , 𝑞1𝑗 , 𝑞2𝑗 , . . . , 𝑞M𝑗 ), and
then we denote 𝑦𝑗 (𝑋) and 𝑧𝑗 (𝑋) by
𝑦
𝜕𝐽𝑦+𝑧 N
𝜕𝐽𝑦+𝑧 𝜕𝑒𝑗 (𝑘)
Δ𝑤𝑖,𝑗 = −𝜇 = −𝜇 ( ∑ 𝑦 𝑦+𝑧
𝜕𝑤𝑖,𝑗 𝑘=1 𝜕𝑒𝑗 (𝑘)
𝜕𝑤𝑖,𝑗 𝑦𝑗 (𝑋) = 𝑊𝑗 𝐺1,0
(46)
𝑦+𝑧 1,0
𝑧 + (𝐶1 − 𝑊𝑗 𝐼1 ) cos ((M + 1) 𝑋) ,
N
𝜕𝐽𝑦+𝑧 𝜕𝑒𝑗 (𝑘) N
𝑦
+ ∑ 𝑧 ) = −𝜇 ∑ 𝑒𝑗 (𝑘)
𝜕𝑒 (𝑘) 𝜕𝑤𝑖,𝑗
𝑘=1 𝑗 𝑘=1 𝑧𝑗 (𝑋) = 𝑊𝑗
𝑦+𝑧
𝐺0,1
(47)
𝑦+𝑧 0,1
⋅ (𝑓𝑦 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 )) + (𝐶2 − 𝑊𝑗 𝐼1 ) cos ((M + 1) 𝑋) ,

(−𝛼) respectively, where

⋅ (cos (𝑥𝑘 ) − cos ((M + 1) 𝑥𝑘 )) − (𝑥𝑘 )
𝑦+𝑧 𝑦 𝑦 𝑦
⋅ (Cos1,1−𝛼 𝑥𝑘 − Cos1,1−𝛼 ((M + 1) 𝑥𝑘 ))) 𝐸𝑗 = (𝑒𝑗 (1) , 𝑒𝑗 (2) , . . . , 𝑒𝑗 (N) , 𝑒𝑗𝑧 (1) , 𝑒𝑗𝑧 (2) , . . . ,

N 𝑒𝑗𝑧 (N)) = (𝑓 (𝑋, 𝑦𝑗 (𝑋) , 𝑧𝑗 (𝑋)) − 𝐷0+

𝛼
𝑦𝑗 (𝑋) ,
− 𝜇 ∑ 𝑒𝑗𝑧 (𝑘) 𝑔𝑦 (𝑥𝑘 , 𝑦𝑗 (𝑥𝑘 ) , 𝑧𝑗 (𝑥𝑘 ))
𝛼
𝑘=1 𝑔 (𝑋, 𝑦𝑗 (𝑋) , 𝑧𝑗 (𝑋)) − 𝐷0+ 𝑧𝑗 (𝑋)) = 𝑓 (𝑋,
⋅ (cos (𝑥𝑘 ) − cos ((M + 1) 𝑥𝑘 )) , 𝑦+𝑧
𝑦𝑗 (𝑋) , 𝑧𝑗 (𝑋)) − 𝑋𝛼 (𝑊𝑗 𝐻1,0
𝑧
𝜕𝐽𝑦+𝑧 N
𝜕𝐽𝑦+𝑧 𝜕𝑒𝑗 (𝑘) 𝑦+𝑧 1,0
Δ𝑞𝑖,𝑗 = −𝜇 = −𝜇 ( ∑ 𝑧 + (𝐶1 − 𝑊𝑗 𝐼1 ) Cos1,1−𝛼 ((M + 1) 𝑋)) + 𝑔 (𝑋,
𝜕𝑞𝑖,𝑗 𝜕𝑒 (𝑘) 𝜕𝑞𝑖,𝑗
𝑘=1 𝑗
𝑦+𝑧
𝑦 𝑦𝑗 (𝑋) , 𝑧𝑗 (𝑋)) − 𝑋𝛼 (𝑊𝑗 𝐻0,1
N
𝜕𝐽 𝑦+𝑧 𝜕𝑒𝑗 (𝑘) N
+∑ 𝑦 ) = −𝜇 ∑ 𝑒𝑗𝑧 (𝑘) 𝑦+𝑧 0,1
𝜕𝑒
𝑘=1 𝑗 (𝑘) 𝜕𝑞𝑖,𝑗 𝑘=1 + (𝐶2 − 𝑊𝑗 𝐼1 ) Cos1,1−𝛼 ((M + 1) 𝑋)) ,
Advances in Mathematical Physics 9

+ 𝑔𝑦 (𝐺1,0 − 𝐼11,0 cos ((M + 1) 𝑋))

1,0
𝐻 = (Cos1,1−𝛼 (𝑋) , Cos1,1−𝛼 (2𝑋) , . . . ,
− 𝑋𝛼 (𝐻0,1 − 𝐼10,1 Cos1,1−𝛼 ((M + 1) 𝑋)) . (49)
𝑇 𝑦+𝑧 𝑦+𝑧
M Define Lyapunov function 𝑉𝑗 = (1/2)‖𝐸𝑗 ‖22 ; then simi-
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞
Cos1,1−𝛼 (M𝑋) , 0, 0, . . . , 0) , larly to the proof of Theorem A we get
2N + 1
0<𝜇< 󵄩 .
󵄩󵄩𝜕𝐸 /𝜕𝑊𝑦+𝑧 󵄩󵄩󵄩2
𝑦+𝑧 (50)
M 󵄩󵄩 𝑗 𝑗 󵄩󵄩𝐹
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞
𝐻0,1 = (0, 0, . . . , 0, Cos1,1−𝛼 (𝑋) , Cos1,1−𝛼 (2𝑋) , . . . , 𝑦+𝑧 𝑦+𝑧
By simply calculating 𝜕𝐸𝑗 /𝜕𝑊𝑗 , we have

𝑇 󵄩󵄩 A ⋅⋅⋅ A B ⋅⋅⋅ B 󵄩󵄩2

󵄩󵄩 𝑦+𝑧 󵄩󵄩2 󵄩󵄩󵄩󵄩 󵄩󵄩
󵄩󵄩
󵄩󵄩 𝜕𝐸𝑗 󵄩󵄩 󵄩
󵄩 󵄩󵄩
Cos1,1−𝛼 (M𝑋)) , 󵄩󵄩 󵄩󵄩 ≤ 󵄩󵄩2 ( .. .. .. .. ) 󵄩󵄩
󵄩󵄩 󵄩 󵄩󵄩 . . . . 󵄩󵄩
󵄩󵄩 𝜕𝑊𝑗𝑦+𝑧 󵄩󵄩󵄩
󵄩 󵄩𝐹 󵄩󵄩󵄩󵄩 󵄩󵄩
󵄩 (51)
󵄩󵄩 A ⋅⋅⋅ A B ⋅ ⋅ ⋅ B M×2N 󵄩󵄩󵄩𝐹

𝐺1,0 = (cos (𝑋) , cos (2𝑋) , . . . , cos (M𝑋) , = 4MN (A2 + B2 ) ,

𝑦 𝑦
where A = 𝐿 1 + 𝐿 3 + 𝛿−𝛼 𝐿 2 and B = 𝐿𝑧1 + 𝐿𝑧3 + 𝛿−𝛼 𝐿 2 , and
𝑇 finally we obtain
M
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞
0, 0, . . . , 0) , 0
<𝜇
(52)
M 2N + 1
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞ < .
𝐺0,1 = (0, 0, . . . , 0, cos (𝑋) , cos (2𝑋) , . . . , 𝑦 𝑦 2
4MN ((𝐿 1 + 𝐿 3 + 𝛿−𝛼 𝐿 2 ) + (𝐿𝑧1 + 𝐿𝑧3 + 𝛿−𝛼 𝐿 2 ) )
2

𝑇
This completes the proof.

cos (M𝑋)) , 3.6. Example

3.6.1. Example 4. We first consider the following linear cou-
𝑇 pled fractional differential equations:
M M
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞ ⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞
𝐼11,0 = (1, 1, . . . , 1, 0, 0, . . . , 0) , 𝛼 2
𝐷0+ 𝑦 (𝑥) = 𝑥2 + 𝑥3 + 𝑥2−𝛼 − 𝑦 (𝑥) − 𝑧 (𝑥) ,
Γ (3 − 𝛼)
𝑇 0 < 𝑥 ≤ 1, 0 < 𝛼 ≤ 1, (53)
M M
⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞ ⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞⏞
𝐼10,1 = (0, 0, . . . , 0, 1, 1, . . . , 1) . 𝛼 6
𝐷0+ 𝑧 (𝑥) = 𝑥2 + 𝑥3 + 𝑥3−𝛼 − 𝑦 (𝑥) − 𝑧 (𝑥) ,
Γ (4 − 𝛼)
(48) with initial condition as follows:
𝑦 (0) = 0,
Then we have (54)
𝑧 (0) = 0.

𝜕𝐸𝑗
𝑦+𝑧 The exact solution is 𝑦(𝑥) = 𝑥2 and 𝑧(𝑥) = 𝑥3 . We set the
= 𝑓𝑦 (𝐺 1,0
− 𝐼11,0 cos ((M + 1) 𝑋)) parameters 𝛼 = 0.9, 𝜇 = 0.001, M = 7, and N = 10 and train
𝑦+𝑧
𝜕𝑊𝑗 the neural network 2000 times, and the weights of the
network for Example 4 are given in Table 5. Figures 7 and
+ 𝑓𝑧 (𝐺0,1 − 𝐼10,1 cos ((M + 1) 𝑋)) 8 show that the sample points and checkpoints are in well
agreement with the exact solutions for the problem. Figure 9
shows that the error of the numerical solutions decreases
− 𝑋𝛼 (𝐻1,0 − 𝐼11,0 Cos1,1−𝛼 ((M + 1) 𝑋)) rapidly within the first 50 training times. Table 6 shows
the exact solution, approximate solution, and accuracy for
+ 𝑔𝑧 (𝐺0,1 − 𝐼10,1 cos ((M + 1) 𝑋)) Example 4.
10 Advances in Mathematical Physics

Table 5: Weights obtained along with the solution of Examples 4 Inspection curve
and 5. 1.4

𝛼 = 0.9 Example 4 Example 5 1.2

𝑤1 /𝑞1 0.5307/0.4494 0.8215/0.7682
𝑤2 /𝑞2 −0.0592/0.0013 0.3545/ − 0.7084 1.0

y = x2 and z = x3
𝑤3 /𝑞3 −0.6414/ − 0.7451 −0.1512/ − 0.2292
𝑤4 /𝑞4 0.0052/ − 0.0392 −0.1342/ − 0.1116 0.8
𝑤5 /𝑞5 0.0559/0.3705 0.1657/0.3582
0.6
𝑤6 /𝑞6 0.3998/0.2282 −0.0815/0.0781
𝑤7 /𝑞7 −0.4141/ − 0.4014 0.0533/ − 0.2259
0.4

0.2
Learning curve
1.4 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1.2 x

Exact solution y(x) Exact solution z(x)

1.0 Checkpoint for y(x) Checkpoint for z(x)
y = x2 and z = x3

0.8 Figure 8: The inspection curve for Example 4.

0.6
Error curve
50
0.4
45
0.2 40
35
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Error value

30
x
25
Exact solution y(x) Exact solution z(x)
20
Sample point for y(x) Sample point for z(x)
15
Figure 7: The learning curve for Example 4.
10
5
3.6.2. Example 5. We second consider the following nonlin- 0
ear fractional coupled differential equations: 0 500 1000 1500 2000
Training times
𝛼 6 2
𝐷0+ 𝑦 (𝑥) = 𝑥6 + cos (𝑥) + 𝑥3−𝛼 − (𝑦 (𝑥)) Error of y(x)
Γ (4 − 𝛼)
Error of z(x)
− 𝑧 (𝑥) , 0 < 𝑥 ≤ 1, 0 < 𝛼 ≤ 1,
(55) Figure 9: The error curve for Example 4.
𝛼 3 −𝛼
𝐷0+ 𝑧 (𝑥) = 𝑥 + cos (𝑥) + 𝑥 Cos1,1−𝛼 (𝑥) − 𝑦 (𝑥)
− 𝑧 (𝑥) ,
4. Conclusion
with initial conditions as follows: In this paper, by using the neural network, we obtained the
𝑦 (0) = 0, numerical solutions for single fractional differential equa-
(56) tions and the systems of coupled differential equations of
𝑧 (0) = 1, fractional order. The computer graphics demonstrates that
numerical results are in well agreement with the exact
The exact solution is 𝑦(𝑥) = 𝑥3 and 𝑧(𝑥) = cos(𝑥). We set solutions. In (1), suppose that 𝑓(𝑥, 𝑦(𝑥)) = 𝐴(𝑥) + 𝐵(𝑥)𝑦+
the parameters 𝛼 = 0.9, 𝜇 = 0.001, M = 7, and N = 10. 𝐶(𝑥)𝑦2 ; then the problem transformed into Fractional Riccati
Numerical solutions in Table 7 show that this network can Equations (Example 3 in this paper). In (3), suppose that
also be applied to the nonlinear fractional coupled differential 𝑓(𝑥, 𝑦(𝑥)) = 𝑦(𝑥)(𝑟 − 𝑎𝑦(𝑥) − 𝑏𝑧(𝑥)) and 𝑔(𝑥, 𝑧(𝑥)) =
equations but we need more time to train the network. 𝑧(𝑥)(−𝑑 + 𝑐𝑦(𝑥)); then the problem transformed into
Advances in Mathematical Physics 11

1 Table 7: Exact solution, approximate solution, and accuracy for

Example 5.
0.8
0.6 𝛼 = 0.9 Numerical solution Accuracy
𝑥 𝑦(𝑥) 𝑧(𝑥) 𝑦(𝑥) 𝑧(𝑥) 𝑦(𝑥) 𝑧(𝑥)
0.4
0.1 0.9950 0.001 0.9926 0.0042 10−3 10−3
0.2
0.2 0.9800 0.008 0.9834 0.0148 10−3 10−3
0
y

−3
0.3 0.9553 0.027 0.9625 0.0351 10 10−3
−0.2 0.4 0.9210 0.064 0.9267 0.0689 10−3 10−3
−0.4 0.5 0.8775 0.125 0.8773 0.1220 10−4 10−3
−3
−0.6 0.6 0.8253 0.216 0.8198 0.2040 10 10−3
−0.8 0.7 0.7648 0.343 0.7604 0.3276 10−3 10−2

−1 0.8 0.6967 0.512 0.7004 0.5049 10−3 10−3

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −2
0.9 0.6216 0.729 0.6336 0.7375 10 10−3
x
1 0.5403 1 0.5475 1.0056 10−3 10−3
The graphics of Cos1,0(x) The graphics of Cos1,0.8(x)
The graphics of Cos1,0.1(x) The graphics of Cos1,1 (x)
The graphics of Cos1,0.5(x) the National Natural Science Foundations of China under
Grant no. 11247310, the Foundations for Distinguished Young
Figure 10: The graphics of the function Cos𝛼,1−𝛼 (𝑥). Talents in Higher Education of Guangdong under Grant no.
2012LYM0096, and the Fund of Hanshan Normal University
Table 6: Exact solution, approximate solution, and accuracy for under Grant nos. LY201302 and LF201403.
Example 4.

𝛼 = 0.9 Numerical solution Accuracy

References
𝑥 𝑦(𝑥) 𝑧(𝑥) 𝑦(𝑥) 𝑧(𝑥) 𝑦(𝑥) 𝑧(𝑥) [1] A. I. Saichev and G. M. Zaslavsky, “Fractional kinetic equations:
0.1 0.01 0.001 0.0101 0.0020 10−4 10−3 solutions and applications,” Chaos, vol. 7, no. 4, pp. 753–764,
1997.
0.2 0.04 0.008 0.0408 0.0104 10−4 10−3
[2] R. Metzler and J. Klafter, “The random walk’s guide to anoma-
0.3 0.09 0.027 0.0926 0.0301 10−3 10−3 lous diffusion: a fractional dynamics approach,” Physics Reports,
0.4 0.16 0.064 0.1641 0.0664 10−3 10−3 vol. 339, no. 1, pp. 1–77, 2000.
−3
0.5 0.25 0.125 0.2529 0.1233 10 10−3 [3] R. Metzler and J. Klafter, “Boundary value problems for frac-
tional diffusion equations,” Physica A: Statistical Mechanics and
0.6 0.36 0.216 0.3584 0.2071 10−3 10−3 its Applications, vol. 278, no. 1-2, pp. 107–125, 2000.
−3
0.7 0.49 0.343 0.4847 0.3290 10 10−2 [4] S. B. Yuste, L. Acedo, and K. Lindenberg, “Reaction front in an
0.8 0.64 0.512 0.6389 0.5033 10−3 10−3 𝐴 + 𝐵𝐶 ⃗ reaction-subdiffusion process,” Physical Review E, vol.
69, no. 3, Article ID 036126, 2004.
0.9 0.81 0.729 0.8215 0.7361 10−2 10−3
−2
[5] K. Diethelm and A. D. Freed, “On the solution of nonlinear
1 1 1 1.0126 1.0076 10 10−3 fractional order differential equations used in the modeling of
viscoplasticity,” in Scientifice Computing in Chemical Engineer-
ing II—Computational Fluid Dynamics, Reaction Engineering
fractional-order Lotka-Volterra predator-prey system. We and Molecular Properties, F. Keil, W. Mackens, H. Voss, and
will consider this problem in another paper. The neural net- J. Werther, Eds., pp. 217–224, Springer, Heidelberg, Germany,
work is a powerful method and is effective for the above two 1999.
problems, which should be also able to solve fractional partial [6] L. Gaul, P. Klein, and S. Kemple, “Damping description involv-
differential equations. ing fractional operators,” Mechanical Systems and Signal Pro-
cessing, vol. 5, no. 2, pp. 81–88, 1991.
[7] Z. Odibat and S. Momani, “Numerical methods for nonlinear
Conflict of Interests partial differential equations of fractional order,” Applied Math-
ematical Modelling, vol. 32, no. 1, pp. 28–39, 2008.
The authors declare that there is no conflict of interests
regarding the publication of this paper. [8] Z. Liu and J. Liang, “A class of boundary value problems for first-
order impulsive integro-differential equations with deviating
arguments,” Journal of Computational and Applied Mathematics,
Acknowledgments vol. 237, no. 1, pp. 477–486, 2013.
[9] Z. H. Liu, J. H. Sun, and I. Szántó, “Monotone iterative tech-
The authors would like to thank the referees for their many nique for Riemann-Liouville fractional integro-differential
constructive comments and suggestions to improve the paper. equations with advanced arguments,” Results in Mathematics,
This work was partly supported by the Special Funds of vol. 63, no. 3-4, pp. 1277–1287, 2013.
12 Advances in Mathematical Physics

[10] Z. H. Liu, N. Van Loi, and V. Obukhovskii, “Existence and

global bifurcation of periodic solutions to a class of differential
variational inequalities,” International Journal of Bifurcation and
Chaos, vol. 23, no. 7, Article ID 1350125, 2013.
[11] Z. Liu, L. Lu, and I. Szántó, “Existence of solutions for fractional
impulsive differential equations with p-Laplacian operator,”
Acta Mathematica Hungarica, vol. 141, no. 3, pp. 203–219, 2013.
[12] X. Liu and Z. Liu, “Existence results for a class of second order
evolution inclusions and its corresponding first order evolution
inclusions,” Israel Journal of Mathematics, vol. 194, no. 2, pp.
723–743, 2013.
[13] H. D. Qu and X. Liu, “Existence of nonnegative solutions for
a fractional m-point boundary value problem at resonance,”
Boundary Value Problems, vol. 2013, article 127, 2013.
[14] I. Podlubny, A. Chechkin, T. Skovranek, Y. Chen, and B. M.
Vinagre Jara, “Matrix approach to discrete fractional calculus.
II. Partial fractional differential equations,” Journal of Compu-
tational Physics, vol. 228, no. 8, pp. 3137–3153, 2009.
[15] Z. Odibat, S. Momani, and H. Xu, “A reliable algorithm of
homotopy analysis method for solving nonlinear fractional
differential equations,” Applied Mathematical Modelling, vol. 34,
no. 3, pp. 593–600, 2010.
[16] S. Das and P. K. Gupta, “Homotopy analysis method for solving
fractional hyperbolic partial differential equations,” Interna-
tional Journal of Computer Mathematics, vol. 88, no. 3, pp. 578–
588, 2011.
[17] A. Elsaid, “Homotopy analysis method for solving a class of
fractional partial differential equations,” Communications in
Nonlinear Science and Numerical Simulation, vol. 16, no. 9, pp.
3655–3664, 2011.
[18] L. Song and H. Zhang, “Solving the fractional BBM-Burgers
equation using the homotopy analysis method,” Chaos, Solitons
& Fractals, vol. 40, no. 4, pp. 1616–1622, 2009.
[19] H. Jafari, A. Golbabai, S. Seifi, and K. Sayevand, “Homotopy
analysis method for solving multi-term linear and nonlinear
diffusion-wave equations of fractional order,” Computers &
Mathematics with Applications, vol. 59, no. 3, pp. 1337–1344,
2010.
[20] S. Momani and Z. Odibat, “Homotopy perturbation method
for nonlinear partial differential equations of fractional order,”
Physics Letters. A, vol. 365, no. 5-6, pp. 345–350, 2007.
[21] M. A. Z. Raja, J. A. Khan, and I. M. Qureshi, “Evolutionary com-
putation technique for solving Riccati differential equation of
arbitrary order,” World Academy of Science, Engineering and
Technology, vol. 58, pp. 531–536, 2009.
[22] A. A. Kilbsa, H. M. Srivastava, and J. J. Trujillo, Theory and
Applications of Fractional Differential Equations, Elsevier, Ams-
terdam, The Netherlands, 2006.
[23] M. A. Z. Raja, J. A. Khan, and I. M. Qureshi, “A new stochastic
approach for solution of Riccati differential equation of frac-
tional order,” Annals of Mathematics and Artificial Intelligence,
vol. 60, no. 3-4, pp. 229–250, 2010.
Advances in Advances in Journal of Journal of
Operations Research
Hindawi Publishing Corporation
Decision Sciences
Hindawi Publishing Corporation
Applied Mathematics
Hindawi Publishing Corporation
Algebra
Hindawi Publishing Corporation
Probability and Statistics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

The Scientific International Journal of

World Journal
Hindawi Publishing Corporation
Differential Equations
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

http://www.hindawi.com

International Journal of Advances in

Combinatorics
Hindawi Publishing Corporation
Mathematical Physics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of Journal of Mathematical Problems Abstract and Discrete Dynamics in

Complex Analysis
Hindawi Publishing Corporation
Mathematics
Hindawi Publishing Corporation
in Engineering
Hindawi Publishing Corporation
Applied Analysis
Hindawi Publishing Corporation
Nature and Society
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

International
Journal of Journal of
Mathematics and
Mathematical
Discrete Mathematics
Sciences

Journal of International Journal of Journal of

Hindawi Publishing Corporation Hindawi Publishing Corporation Volume 2014

Function Spaces
Hindawi Publishing Corporation
Stochastic Analysis
Hindawi Publishing Corporation
Optimization
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014