Next Article in Journal
Significant Study of Fuzzy Fractional Inequalities with Generalized Operators and Applications
Previous Article in Journal
Fejér-Type Fractional Integral Inequalities Involving Mittag-Leffler Function
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Control of the Bifurcation Behaviors of Delayed Fractional-Order Neural Networks with Cooperation–Competition Topology

School of Mathematics and Physics, Qingdao University of Science and Technology, Qingdao 266061, China
Fractal Fract. 2024, 8(12), 689; https://doi.org/10.3390/fractalfract8120689
Submission received: 21 October 2024 / Revised: 12 November 2024 / Accepted: 21 November 2024 / Published: 24 November 2024

Abstract

In the real world, the competition and cooperation relationship exists in numerous systems. For instance, the competition–cooperation structure of a biological neural network is determined by the excitatory and inhibitory effects of neurons. The dynamic behaviors of a neural network model with a competition–cooperation structure are studied in this article, focusing particularly on the bifurcation and control problems. By selecting time delay as the parameter, a new sufficient condition for Hopf bifurcation is given and the impact of the fractional order on bifurcation behavior is determined for the network. Furthermore, a time-delay feedback controller is introduced to manage Hopf bifurcation behaviors, and, meanwhile, the stability domain is expanded. Our findings indicate that both fractional order and time delay play a crucial role in controlling the stability and Hopf bifurcation of the given model. Lastly, the accuracy of our theoretical results is verified through several numerical simulations, and the impact of control parameters on the bifurcation behavior of the network model is discussed in detail.

1. Introduction

Neural networks, mathematical models inspired by biological neural networks, consist of numerous interconnected neurons forming a highly nonlinear dynamic system. They are extensively utilized in diverse fields, such as pattern recognition, data mining, image processing, and natural language processing, to address complex problems [1,2,3,4,5,6,7]. The functionality of these networks largely depends on their system dynamics properties. For instance, the associative memory and optimization computing power of famous networks directly rely on the global stability of network dynamics. Consequently, the dynamic behaviors of neural networks have attracted considerable interest from scholars across various fields and have yielded a wealth of research findings. He et al. [8] introduced a novel full-dimensional nonlinear state feedback control strategy to optimize neural dynamics, then established conditions for local stability and Hopf bifurcation in binary-tree-structure and multi-delay diffusion neural network models. Lu et al. [9] constructed a new delayed neural network with a radial ring structure and bidirectional coupling, and they studied its stability and Hopf bifurcation dynamic behavior. Nie et al. [10] addressed the problem of coexistence and the dynamical behaviors of multiple equilibria for competitive neural networks.
Despite the significant interest in neural network dynamics in recent years, most research has focused only on the cooperation aspect of neuron relationships, neglecting the competition aspect present in neurons with negative connection weights. Competition is a ubiquitous phenomenon in daily life, manifesting in social networks as hostile or friendly relationships between individuals. Cooperation–competition topology networks have a wide range of specific applications, such as financial market prediction, intelligent optimization, and social network analysis [11,12,13]. Similarly, in neural networks, neurons can regulate information transmission through both promotion and inhibition, making it unrealistic to consider only cooperative scenarios. Therefore, studying neural networks with a cooperation–competition relationship [14,15,16,17] provides a more accurate reflection of real-world conditions. Exploring how to use the cooperation–competition structure of neural networks to study system dynamics is a worthwhile endeavor.
Fractional calculus, as an extension of classical integer-order differential and integral calculus, has sparked great interest, due to its potential applications in various fields, such as bioengineering, robotics, control, and medical problems. The memory and inheritance characteristics of different materials and processes in fractional-order differential equations make them more accurate than integer-order models when modeling real-world phenomena. Therefore, fractional calculus has been integrated into neural networks. Over the past few years, stability analysis, bifurcation and chaos, synchronization, and other properties of fractional-order dynamic systems have received increasing attention, leading to many significant achievements. Xu et al. [18] studied the problem of exponential bipartite synchronization of fractional-order multilayer signed networks via hybrid impulsive control. Siami [19] provided the classical secant condition for the stability analysis of cyclic interconnected commensurate fractional-order systems. Li et al. [20] investigated the stability control of fractional-order memory resistive neural networks with reaction–diffusion terms. Yin et al. [21] analyzed the discrete fractional-order LIF model and found that by adjusting the fractional order, neurons and neural network primitives can independently respond to weak signals at a certain frequency, enabling network primitives to achieve ordered cluster discharge. Huang et al. [22] investigated the bifurcation problem in a class of two delayed fractional-order neural networks with three neurons.
Among these efforts, the aim of bifurcation control is to design a controller capable of modifying bifurcation characteristics to achieve desirable dynamic behaviors. And the design and application of fractional-order control systems have always been highly concerned [23,24]. References [25,26,27,28] studied the fractional-calculus observer of the Lipschitz system and used the obtained state estimation for the final feedback control, providing insights for the control method of fractional-order systems with time delay and laying the foundation for fractional calculus and its application in neural networks. There have also been several studies on the application of bifurcation control controllers in different fields [29,30,31,32,33]. Cheng et al. [34] proposed a hybrid control strategy and considered the problem of Hopf bifurcation control for a complex network model with time delays. Lin et al. [35] studied a four-dimensional fractional-order two-gene regulatory network with multiple delays, and they introduced a fractional-order PD controller to maintain the stability of the network. Lu et al. [36] developed a fractional-order proportional–derivative controller to manage the bifurcation of a delayed fractional-order predator–prey system, effectively postponing the occurrence of Hopf bifurcation through the appropriate setting of control parameters. Wang et al. [37] designed a new hybrid controller, applying negative feedback control and time-delay feedback control, to regulate a dual-delay internet congestion control system accessed by a single resource. Based on the above research, it has proven necessary to introduce topological structures into neural networks for research, especially in the field of competition and cooperation.
Motivated by this framework, we incorporated graph theory into a multi-delay network model featuring cooperation–competition interactions, as a signed graph has both positive and negative interactions. Then, a bifurcation analysis was conducted on a delayed fractional-order neural network with a generalized cooperation–competition topology. This work offers several key contributions:
(1) The Caputo fractional derivative was employed, to expand the delayed neural network model with a generalized cooperation–competition topology into a fractional order model.
(2) The introduction of directed graphs concretizes the competitive–cooperation relationships. Unlike previous studies, we used graph connectivity to describe the relationships between neurons where the connectivity matrix plays a crucial role.
(3) Based on the connection relationship of the directed graph, the ranges of α and β were determined, to provide sufficient conditions for the stability and Hopf bifurcation of the cooperative competitive network, with time delay as the bifurcation parameter.
(4) To postpone the occurrence of Hopf bifurcation, a nonlinear delay feedback controller was designed for a delayed fractional-order neural network model with a generalized cooperation–competition topology.
The rest of the paper is organized as follows. Section 2 introduces Caputo fractional-order derivatives and some fundamental theoretical concepts. A delayed fractional-order neural network model with a generalized cooperation–competition topology is established in Section 3. In Section 4, the stability of uncontrolled systems and the existence conditions of Hopf bifurcation are investigated. The dynamic behavior of the system after implementing a time-delay feedback controller are also discussed. In Section 5, several numerical simulations are conducted to validate the previously obtained theoretical results. The effects of fractional-order and controller parameters on system dynamics are analyzed. Our conclusions are summarized in Section 6.

2. Preliminaries

Definition 1 
([38]). The Caputo fractional-order derivative operator is defined as follows:
D x ϕ 0 c f ( x ) = 1 Γ ( n ϕ ) 0 x ( x s ) n ϕ 1 f ( n ) ( s ) d s
where n 1 < ϕ < n , when k Z + . In particular, when 0 < ϕ < 1 ,
D x ϕ 0 c f ( x ) = 1 Γ ( 1 ϕ ) 0 x ( x s ) ϕ f ( s ) d s
Definition 2 
([38]). We give the Laplace transform of the Caputo fractional derivative, which is defined as
L { 0 c D x ϕ f ( x ) ; s } = s ϕ F ( s ) l = 0 k 1 s ϕ l 1 f ( l ) ( 0 )
where k 1 ϕ k when k Z + . If f ( l ) ( 0 ) = 0 , l = 1 , 2 , , n then L { 0 c D x ϕ f ( x ) ; s } = s ϕ F ( s ) .
Next, we provide some basic concepts of a signed graph: a signed graph can be conveniently described by graphs G = ( Q , V , E ) , where Q = { q 1 , q 2 , . . . q n } represents the node set, and where V Q × Q = { ( q i , q j ) | q i , q j Q } represents the edge set. The matrix E = [ E i j ] R N × N is the adjacency matrix of a signed graph G, defined as follows: if ( q i , q j ) V then E i j 0 ; if ( q i , q j ) V then E i j = 0 . Due to the cooperation–competition relationship between nodes, there is the following definition: if the relationship between node q i and q j is competitive (cooperative) then E i j < 0 ( E i j > 0 ) .
Lemma 1. 
We consider an autonomous system D q x ( t ) = B x ( t ) ; if all the eigenvalues s i ( i = 1 , 2 , n ) of B = ( b i j ) n n satisfy | a r g ( s i ) | > q π 2 and the characteristic equation d e t ( s I B ) = 0 has no purely imaginary roots for any τ > 0 then the zero solution of system is Lyapunov globally asymptotically stable.
Definition 3. 
If a node set Q of a signed graph can be divided into two subgroups Q 1 and Q 2 , and if, among them, Q 1 and Q 2 meet Q 1 Q 2 = Q and Q 1 Q 2 = , so that E i j 0 for q i , q j Q k ( k 1 , 2 ) and E i j 0 for q i Q k , q j Q ξ ( k ξ | k , ξ 1 , 2 ) then the signed graph is called a structural equilibrium graph.
Definition 4. 
The signum function is defined as
s g n ( ε ) = 1 , ε < 0 0 , ε = 0 1 , ε > 0 .

3. Model Description

Considering a complex dynamic neural network composed of N neurons, the state equations of each node in a delayed fractional order neural network with cooperation–competition topology can be described by the following differential equation:
D q m x m ( t ) = a x m ( t ) + b f ( x m ( t ) ) + c f ( x m ( t τ m ) ) + d n N m | G m n | ( s g n ( G m n ) x n ( t ) x m ( t ) ) ,
where x m ( t ) ( m = 1 , 2 , N ) stand for the state of the mth neuron, a is the self-feedback coefficient, b and c is a constant and represents the connection weight, d > 0 stands for the coupling strength, f ( x ( · ) ) is the activation function, τ is the time delay at time t, N m is the neighbor set of node x m ( t ) , and G = ( G m n ) N × N denotes the symmetric adjacency matrix of the signed graph, describing the cooperation–competition interaction between neurons.
Remark 1. 
When there is a cooperative relationship between neuron m and neuron n, then G m n > 0 . When there is a competitive relationship between neuron m and neuron n, then G m n < 0 . Therefore, the term G m n ( x n ( t ) x m ( t ) ) represents the information exchange between the cooperation neighbor m and neuron n. For the convenience of analysis, we take G m n = ± 1 .
Remark 2. 
When studying competition–cooperation relationships, the introduction of directed graphs concretizes competition–cooperation patterns. Unlike previous studies, we used graph connectivity to describe the relationships between neurons, whereas previous research used connectivity coefficients to describe the relationships between neurons, which made the model complex and difficult to implement. The method proposed in this paper is simple and highly feasible, with adjacency matrix G = ( G m n ) N × N being a key factor.
The structure of competition–cooperation is found in practice commonly. In a real neural network, it often contains a large number of neurons; usually, each neuron is independent but they interact with each other, which is a manifestation of the complexity of complex systems. Lower-dimensional neural networks are the foundation for understanding the dynamics of large-scale networks. This article mainly discusses a fractional-order neural network model with generalized cooperation–competition topology and four neurons. The topology diagram is shown in Figure 1. The solid line here represents the cooperative relationship between two neurons, while the dashed line represents the competitive relationship. So, the above system becomes
D q 1 x 1 ( t ) = a x 1 ( t ) + b f ( x 1 ( t ) ) + c f ( x 1 ( t τ 1 ) ) + d ( x 2 ( t ) x 1 ( t ) ) d ( x 4 ( t ) x 1 ( t ) ) d ( x 3 ( t ) x 1 ( t ) ) D q 2 x 2 ( t ) = a x 2 ( t ) + b f ( x 2 ( t ) ) + c f ( x 2 ( t τ 2 ) ) + d ( x 1 ( t ) x 2 ( t ) ) d ( x 4 ( t ) x 2 ( t ) ) d ( x 3 ( t ) x 2 ( t ) ) D q 3 x 3 ( t ) = a x 3 ( t ) + b f ( x 3 ( t ) ) + c f ( x 3 ( t τ 3 ) ) + d ( x 4 ( t ) x 3 ( t ) ) d ( x 1 ( t ) x 3 ( t ) ) D q 4 x 4 ( t ) = a x 4 ( t ) + b f ( x 4 ( t ) ) + c f ( x 4 ( t τ 4 ) ) d ( x 2 ( t ) x 3 ( t ) ) + d ( x 3 ( t ) x 4 ( t ) ) d ( x 1 ( t ) x 4 ( t ) ) d ( x 2 ( t ) x 4 ( t ) ) .
For the convenience of analysis, we make the following assumptions:
(H1) 
f C R ,   f ( 0 ) = 0 ,   f ( 0 ) 0 ,
(H2) 
τ = τ 1 = τ 3 = τ 2 = τ 4 .
Figure 1. Topology diagram of neural networks with generalized cooperation–competition topology.
Figure 1. Topology diagram of neural networks with generalized cooperation–competition topology.
Fractalfract 08 00689 g001
Remark 3. 
As is well known, the local stability and Hopf bifurcation threshold of a system depend on the linear part of the system. It is necessary for us to linearize network (2) and explore its stability conditions and bifurcation criteria hypothesis (H1). There is usually a time delay in the information transmission between neurons, which is reflected in network (2). Assumption (H2) actually indicates that the time delay between neurons is equal.
Remark 4. 
The adjacency matrix G is crucial in a cooperative–competitive topology relationship. If the competitive and cooperative topological relationship between neurons satisfies an asymmetric adjacency matrix then the model that satisfies the condition will be a simplified version of the model studied in this paper.
Obviously, it is easy to see that system (2) has a unique equilibrium point, which is defined as E ( 0 , 0 , 0 , 0 ) , so a Taylor expansion is performed on the nonlinear function f ( x ( · ) ) at E, and its linear term is taken.
For this article, we mainly considered the study of the stability and Hopf bifurcation of system (3), and we used a delay feedback controller to delay the start of Hopf bifurcation and expand the stability domain of the system (3):
D q 1 x 1 ( t ) = a x 1 ( t ) + b x 1 ( t ) + c x 1 ( t τ ) + d ( x 2 ( t ) x 1 ( t ) ) d ( x 4 ( t ) x 1 ( t ) ) d ( x 3 ( t ) x 1 ( t ) ) , D q 2 x 2 ( t ) = a x 2 ( t ) + b x 2 ( t ) + c x 2 ( t τ ) + d ( x 1 ( t ) x 2 ( t ) ) d ( x 4 ( t ) x 2 ( t ) ) d ( x 3 ( t ) x 2 ( t ) ) , D q 3 x 3 ( t ) = a x 3 ( t ) + b x 3 ( t ) + c x 3 ( t τ ) + d ( x 4 ( t ) x 3 ( t ) ) d ( x 1 ( t ) x 3 ( t ) ) d ( x 2 ( t ) x 3 ( t ) ) , D q 4 x 4 ( t ) = a x 4 ( t ) + b x 4 ( t ) + c x 4 ( t τ ) + d ( x 3 ( t ) x 4 ( t ) ) d ( x 1 ( t ) x 4 ( t ) ) d ( x 2 ( t ) x 4 ( t ) ) ,
where b = b f ( 0 ) , c = c f ( 0 ) .

4. Main Results

For this section, we selected time delay as the bifurcation parameter, and we studied the stability and Hopf bifurcation of system (3) without control and with control, respectively. The following are the main results.
  • A. Stability and Hopf bifurcation of the uncontrolled system.
In this segment, we mainly studied the dynamic behavior of (3) without control, and (3) is equivalent to (4):
D q 1 x 1 ( t ) = ( a + b 3 d ) x 1 ( t ) + c x 1 ( t τ ) + d x 2 ( t ) d x 3 ( t ) d x 4 ( t ) , D q 2 x 2 ( t ) = d x 1 ( t ) + ( a + b 3 d ) x 2 ( t ) + c x 2 ( t τ ) d x 3 ( t ) d x 4 ( t ) , D q 3 x 3 ( t ) = d x 1 ( t ) d x 2 ( t ) + ( a + b 3 d ) x 3 ( t ) + c x 3 ( t τ ) + d x 4 ( t ) , D q 4 x 4 ( t ) = d x 1 ( t ) d x 2 ( t ) + d x 3 ( t ) + ( a + b 3 d ) x 4 ( t ) + c x 4 ( t τ ) .
The characteristic matrix Δ ( s ) of system (4),
Δ ( s ) = Ξ 1 d d d d Ξ 2 d d d d Ξ 3 d d d d Ξ 4
where A = a + b 3 d and Ξ j = s q j A c e s τ ( j = 1 , 2 , 3 , 4 ) .
The characteristic equation corresponding to the above characteristic matrix is
R 1 ( s ) + R 2 ( s ) e s τ + R 3 ( s ) e 2 s τ + R 4 ( s ) e 3 s τ + R 5 ( s ) e 4 s τ = 0 ,
Multiplying e 2 s τ on both sides of Equation (6), we can obtain
R 1 ( s ) e 2 s τ + R 2 ( s ) e s τ + R 3 ( s ) + R 4 ( s ) e s τ + R 5 ( s ) e 2 s τ = 0 ,
where R 1 ( s ) , R 2 ( s ) , R 3 ( s ) , R 4 ( s ) , R 5 ( s ) is a function about s; the detailed information is presented in Appendix A, abbreviated as R 1 , R 2 , R 3 , R 4 , R 5 .
We define R i R and R i I as the real and imaginary part of R i ( i = 1 , 2 , 3 , 4 , 5 ) , respectively. And we let s = i ω = ω ( c o s π 2 + i s i n π 2 ) ( ω > 0 ) be the root of Equation (7), and (7) can be transformed into (8):
( R 1 R + i R 1 I ) ( c o s 2 ω τ + i s i n 2 ω τ ) + ( R 2 R + i R 2 I ) ( c o s ω τ + i s i n ω τ ) + ( R 3 R ( ω ) + i R 3 I ) + ( R 4 R + i R 4 I ) ( c o s ω τ i s i n ω τ ) + ( R 5 R + i R 5 I ) ( c o s 2 ω τ i s i n 2 ω τ ) = 0 .
Separating the real and imaginary parts, we obtain
( R 1 R + R 5 R ) c o s 2 ω τ + ( R 5 I R 1 I ) s i n 2 ω τ + ( R 2 R + R 4 R ) c o s ω τ + ( R 4 I R 2 I ) s i n ω τ + R 3 R = 0 , ( R 1 I + R 5 I ) c o s 2 ω τ ( R 5 R R 1 R ) s i n 2 ω τ + ( R 2 I + R 4 I ) c o s ω τ ( R 4 R R 2 R ) s i n ω τ + R 3 I = 0 .
By introducing the trigonometric function formula s i n ω τ = ± 1 c o s 2 ω τ , c o s 2 ω τ = 2 c o s 2 ω τ 1 , s i n 2 ω τ = 2 s i n ω τ c o s ω τ = 2 c o s ω τ 1 c o s 2 ω τ , the first equation of Equation (9) can be transformed into
( R 1 R + R 5 R ) ( 2 c o s 2 ω τ 1 ) + ( R 2 R + R 4 R ) c o s ω τ + R 3 R = 2 ( R 5 I R 1 I ) c o s ω τ 1 c o s 2 ω τ ( R 4 I R 2 I ) 1 c o s 2 ω τ .
Squaring both sides of Equation (10) makes it easy to obtain
α 4 c o s 4 ω τ + α 3 c o s 3 ω τ + α 2 c o s 2 ω τ + α 1 c o s ω τ + α 0 = 0 .
Let z = c o s ω τ ; then, Equation (11) becomes
α 4 z 4 + α 3 z 3 + α 2 z 2 + α 1 z + α 0 = 0 .
In the following, we need to find a condition that ensures that Equation (12) has at least one root. We define g ( z ) = α 4 z 4 + α 3 z 3 + α 2 z 2 + α 1 z + α 0 . From the above analysis, we can obtain the following lemma:
Lemma 2. 
If α 0 < 0 , i.e., ( R 1 R + R 5 R ) 2 + R 3 R 2 2 R 3 R ( R 1 R + R 5 R ) ( R 4 I R 2 I ) 2 < 0 then Equation (11) has at least one positive root.
Proof. 
Through Equation (12), it can be seen that Equation (12) is a one-dimensional quartic equation about z. If α 0 < 0 then g ( 0 ) = α 0 < 0 . Also, α 4 = 4 ( R 1 R + R 5 R ) 2 + 4 ( R 5 I R 1 I ) 2 is a constant that is always greater than or equal to 0, and lim z + g ( z ) = + . Therefore, at least one real value z 0 satisfies g ( z 0 ) = 0 . Thus, it can be concluded that Equation (11) has at least one root c o s ω τ . This complete the proof. □
We assume
c o s ω τ = h 1 ( ω ) ,
where h 1 ( ω ) is a continuous function about ω .
Similarly, by applying the trigonometric function formula c o s 2 ω τ = 1 2 s i n 2 ω τ , s i n 2 ω τ = 2 s i n ω τ 1 s i n 2 ω τ , the first equation of Equation (9) can be transformed into
( R 1 R + R 5 R ) ( 1 2 s i n 2 ω τ ) + ( R 4 I R 2 I ) sin ω τ + R 3 R = 2 ( R 5 I R 1 I ) sin ω τ 1 sin 2 ω τ ( R 2 R + R 4 R ) 1 sin 2 ω τ .
Squaring both sides of Equation (14) makes it easy to obtain
β 4 sin 4 ω τ + β 3 sin 3 ω τ + β 2 sin 2 ω τ + β 1 sin ω τ + β 0 = 0 .
Let ϱ = s i n ω τ ; then, Equation (15) becomes
β 4 ϱ 4 + β 3 ϱ 3 + β 2 ϱ 2 + β 1 ϱ + β 0 = 0 .
In the following, we need to find a condition that ensures that Equation (16) has at least one root. We denote m ( ϱ ) = β 4 ϱ 4 + β 3 ϱ 3 + β 2 ϱ 2 + β 1 ϱ + β 0 . From the above analysis, we can obtain the following lemma:
Lemma 3. 
If β 0 < 0 , i.e., ( R 1 R + R 5 R ) 2 + R 3 R 2 + 2 R 3 R ( R 1 R + R 5 R ) ( R 2 R + R 4 R ) 2 < 0 then Equation (15) has at least one root.
Proof. 
Through Equation (16), it can be seen that Equation (16) is a one-dimensional quartic equation about ϱ . If β 0 < 0 then m ( 0 ) = β 0 < 0 < 0 . Also, β 4 = 4 ( R 1 R + R 5 R ) 2 + 4 ( R 5 I R 1 I ) 2 is a constant that is always greater than or equal to 0 and lim ϱ + m ( ϱ ) = + . Therefore, at least one real value ϱ 0 satisfies m ( ϱ 0 ) = 0 . Thus, it can be concluded that Equation (15) has at least one root s i n ω τ . This complete the proof. □
Remark 5. 
According to the theorem on the existence of roots in a quadratic equation, if α 0 < 0 then g ( z ) = 0 must have a root z 0 belonging to ( 0 , + ) . Obviously, α 0 < 0 is a sufficient condition for ensuring system stability. Similarly, β 0 < 0 is quite important.
We assume
sin ω τ = h 2 ( ω ) ,
where h 2 ( ω ) is a continuous function about ω .
From s i n 2 ω τ + c o s 2 ω τ = 1 , it can be seen that
h 1 2 ( ω ) + h 2 2 ( ω ) = 1 .
In view of cos ω τ = h 1 ( ω ) , we have
τ κ = 1 ω [ arccos h 1 ( ω ) + 2 κ π ] , κ = 0 , 1 , 2 , ....
Suppose Equation (18) has at least one positive real root.
Set
τ 0 = m i n τ κ , κ = 0 , 1 , 2 , , ω 0 = ω | τ = τ 0 .
In order to derive the conditions for the occurrence of the Hopf bifurcation, the following assumptions must be given:
(H3) 
D 2 R D 1 R + D 2 I D 1 I D 1 R 2 + D 1 I 2 0 .
Lemma 4. 
Let s ( τ ) = ζ ( τ ) + i ω ( τ ) be the root of (7) satisfying ζ ( τ 0 ) = 0 , ω ( τ 0 ) = ω 0 . Then, the following transversality condition holds:
R e [ d s d τ ] | τ = τ 0 , ω = ω 0 0 .
Proof. 
Taking the derivative with respect to τ on both sides of Equation (7), we have
d s d τ = s ( 2 ( R 5 e 2 s τ R 1 e 2 s τ ) + ( R 4 e s τ R 2 e s τ ) R 3 + R 1 e 2 s τ + R 2 e s τ + R 4 e s τ τ ( 2 ( R 5 e 2 s τ R 1 e 2 s τ ) + ( R 4 e s τ R 2 e s τ ) ) ,
where R i ( i = 1 , 2 , 3 , 4 , 5 ) are the derivatives of R i .
Then,
[ d s d τ ] 1 = R 3 + R 1 e 2 s τ + R 2 e s τ + R 4 e s τ s ( 2 ( R 5 e 2 s τ R 1 e 2 s τ ) + ( R 4 e s τ R 2 e s τ ) ) τ s .
Bringing s = i ω 0 into (23), we have
s i g n { R e [ d s d τ ] | s = i ω 0 , τ = τ 0 } = s i g n { R e [ d s d τ ] 1 | s = i ω 0 , τ = τ 0 }
R e [ d s d τ ] 1 | s = i ω 0 , τ = τ 0 = R e ( D 2 D 1 ) | s = i ω 0 , τ = τ 0 = D 2 R D 1 R + D 2 I D 1 I D 1 R 2 + D 1 I 2 ,
where
D 1 = s ( 2 ( R 5 e 2 s τ R 1 e 2 s τ ) + ( R 4 e s τ R 2 e s τ ) ) D 2 = R 3 + R 1 e 2 s τ + R 2 e s τ + R 4 e s τ The detailed proof process will be presented in Appendix B.
Therefore, under the assumptions condition ( H 3 ) , we have
R e [ d s d τ ] | s = i ω 0 , τ = τ 0 0 .
Thus, the transversality condition of system (2) near τ = τ 0 is satisfied. □
Next, we consider the dynamic behavior of the system when τ = 0 . When there is no transmission delay in system (2), Equation (7) becomes
s 4 + θ 1 s 3 + θ 2 s 2 + θ 3 s + θ 4 = 0 ,
where
θ 1 = 4 ( A + c ) , θ 2 = 6 ( A 2 d 2 + 2 A c + c 2 ) , θ 3 = 4 ( A 3 3 A d 2 + 2 d 3 + 3 A 2 c 3 c d 2 + 3 A c 2 + c 3 , θ 4 = A 4 6 A 2 d 2 3 d 4 + 4 A 3 c 12 A c d 2 + 6 A 2 c 2 6 c 2 d 2 + 4 A c 3 + c 4 + 8 A d 3 + 8 c d 3 .
In order to obtain some of the main results of this article, we made the following assumptions:
(H4) 
θ 4 > 0 ; θ 1 θ 2 > θ 3 ; θ 3 ( θ 1 θ 2 θ 3 ) > θ 1 2 θ 4 .
Lemma 5. 
Under assumption ( H 4 ) and if τ = 0 then all the characteristic roots of characteristic Equation (26) have negative real parts. Therefore, system (2) is locally asymptotically stable without a time delay.
Proof. 
It follows the assumption ( H 4 ) that all the roots s i of Equation (26) satisfy | a r g ( s i ) | > q i π 2 ( i = 1 , 2 , 3 , 4 ) . By Lemma 1, we can conclude, then, that all the characteristic roots of characteristic Equation (26) have no purely imaginary roots. Therefore, without the time delay, system (2) is locally asymptotically stable. □
Based on the above discussion, we can obtain the following main results:
Theorem 1. 
Assuming that system (2) satisfies hypothesis ( H 1 ) ( H 4 ) and satisfies conditions α 0 < 0 , β 0 < 0 , the following statement holds:
(1) 
If ( H 1 ) and ( H 4 ) hold, system (2) is locally asymptotically stable when τ = 0 .
(2) 
If ( H 1 ) , ( H 2 ) , ( H 3 ) , and α 0 < 0 , β 0 < 0 hold, system (2) is locally asymptotically stable when τ ( 0 , τ 0 ) , and it is unstable for τ > τ 0 . Then, system (2) undergoes Hopf bifurcation when τ = τ 0 .
  • B. Stability and Hopf bifurcation of the controlled system
In recent years, various controllers have been designed to address the Hopf bifurcation problem in controlling fractional-order systems. However, the exploration of the Hopf bifurcation control problem for fractional-order systems with cooperative–competitive topology is limited. To address this limitation, we introduced a time-delay feedback controller, as shown below:
U ( t ) = K 1 ( X ( t τ ) X ( t ) ) ,
where K 1 is the feedback gain coefficient.
Remark 6. 
In view of the general thought on polynomial function controllers, a nonlinear delayed feedback controller U 1 ( t ) = i = 1 3 K i ( X ( t τ ) X ( t ) ) i was designed in reference [39]. We know that the linear term K 1 ( X ( t τ ) X ( t ) ) can change the beginning of the Hopf bifurcation in fractional order models, while the nonlinear terms K 2 ( X ( t τ ) X ( t ) ) 2 and K 3 ( X ( t τ ) X ( t ) ) 3 only affect the frequency and amplitude of bifurcation oscillation, and the feedback gain coefficient K 2 has little effect on the frequency of bifurcation oscillation, while K 3 has almost no effect. Therefore, we chose to introduce a time-delay feedback controller U ( t ) = K 1 ( X ( t τ ) X ( t ) ) for bifurcation control in this article.
We add the above controller to system (2); then, system (2) will be transformed into
D q 1 x 1 ( t ) = a x 1 ( t ) + b f ( x 1 ( t ) ) + c f ( x 1 ( t τ 1 ) ) + d ( x 2 ( t ) x 1 ( t ) ) d ( x 4 ( t ) x 1 ( t ) ) d ( x 3 ( t ) x 1 ( t ) ) + K 1 ( x 1 ( t τ ) x 1 ( t ) ) , D q 2 x 2 ( t ) = a x 2 ( t ) + b f ( x 2 ( t ) ) + c f ( x 2 ( t τ 2 ) ) + d ( x 1 ( t ) x 2 ( t ) ) d ( x 4 ( t ) x 2 ( t ) ) d ( x 3 ( t ) x 2 ( t ) ) , D q 3 x 3 ( t ) = a x 3 ( t ) + b f ( x 3 ( t ) ) + c f ( x 3 ( t τ 3 ) ) + d ( x 4 ( t ) x 3 ( t ) ) d ( x 1 ( t ) x 3 ( t ) ) d ( x 2 ( t ) x 3 ( t ) ) , D q 4 x 4 ( t ) = a x 4 ( t ) + b f ( x 4 ( t ) ) + c f ( x 4 ( t τ 4 ) ) + d ( x 3 ( t ) x 4 ( t ) ) d ( x 1 ( t ) x 4 ( t ) ) d ( x 2 ( t ) x 4 ( t ) ) + K 1 ( x 4 ( t τ ) x 4 ( t ) ) .
Remark 7. 
Assuming node set V = ( x 1 , x 2 , x 3 , x 4 ) , it can be seen that node set V can be divided into two subsets V 1 and V 2 from definition 3. Therefore, we only need to control one node in each subset when adding control. For this article, we chose to add time-delay feedback controllers to the equations of node x 1 and node x 4 in system (2).
By the Taylor expansion of system (28), the linearization part of system (28) at the equilibrium point E ( 0 , 0 , 0 , 0 ) can be obtained:
D q 1 x 1 ( t ) = ( a + b 3 d K 1 ) x 1 ( t ) + ( c + K 1 ) x 1 ( t τ ) + d x 2 ( t ) d x 3 ( t ) d x 4 ( t ) , D q 2 x 2 ( t ) = d x 1 ( t ) + ( a + b 3 d ) x 2 ( t ) + c x 2 ( t τ ) d x 3 ( t ) d x 4 ( t ) , D q 3 x 3 ( t ) = d x 1 ( t ) d x 2 ( t ) + ( a + b 3 d ) x 3 ( t ) + c x 3 ( t τ ) + d x 4 ( t ) , D q 4 x 4 ( t ) = d x 1 ( t ) d x 2 ( t ) + d x 3 ( t ) + ( a + b 3 d K 1 ) x 4 ( t ) + ( c + K 1 ) x 4 ( t τ ) .
The characteristic matrix Δ 1 ( s ) of system (29),
Δ 1 ( s ) = Λ 1 d d d d Ξ 2 d d d d Ξ 3 d d d d Λ 4
where A = a + b 3 d , B = a + b 3 d K 1 , G = c + K 1 , and Λ = s q B G e s τ , ( = 1 , 4 ) .
The characteristic equation corresponding to the above characteristic matrix is
η 1 ( s ) + η 2 ( s ) e s τ + η 3 ( s ) e 2 s τ + η 4 ( s ) e 3 s τ + η 5 ( s ) e 4 s τ = 0 ,
where η 1 ( s ) , η 2 ( s ) , η 3 ( s ) , η 4 ( s ) , η ( s ) is a function about s, abbreviated as η 1 , η 2 , η 3 , η 4 , η 5 in subsequent articles.
Multiplying e 2 s τ on both sides of Equation (31), we can obtain
η 1 e 2 s τ + η 2 e s τ + η 3 + η 4 e s τ + η 5 e 2 s τ = 0 .
We define η i R and η i I as the real and imaginary part of η i ( i = 1 , 2 , 3 , 4 , 5 ) , respectively. Then, Equation (32) is equivalent to
( η 1 R + η 1 I ) e 2 s τ + ( η 2 R + η 2 I ) e s τ + ( η 3 R + η 3 I ) + ( η 4 R + η 4 I ) e s τ + ( η 5 R + η 5 I ) e 2 s τ = 0 .
We let s = i ω = ω ( c o s π 2 + i s i n π 2 ) ( ω > 0 ) be the root of Equation (31), and we substitute s = i ω = ω ( c o s π 2 + i s i n π 2 ) into Equation (31) to obtain
( η 1 R + i η 1 I ) ( c o s 2 ω τ + i s i n 2 ω τ ) + ( η 2 R + i η 2 I ) ( c o s ω τ + i s i n ω τ ) + ( η 3 R + i η 3 I ) + ( η 4 R + i η 4 I ) ( c o s ω τ i s i n ω τ ) + ( η 5 R + i η 5 I ) ( c o s 2 ω τ i s i n 2 ω τ ) = 0 .
Separating the real and imaginary parts, we obtain
( η 1 R + η 5 R ) c o s 2 ω τ + ( η 5 I η 1 I ) s i n 2 ω τ + ( η 2 R + η 4 R ) c o s ω τ + ( η 4 I η 2 I ) s i n ω τ + η 3 R = 0 , ( η 1 I + η 5 I ) c o s 2 ω τ ( η 5 R η 1 R ) s i n 2 ω τ + ( η 2 I + η 4 I ) c o s ω τ ( η 4 R η 2 R ) s i n ω τ + η 3 I = 0 .
By introducing the trigonometric function formula s i n ω τ = ± 1 c o s 2 ω τ , c o s 2 ω τ = 2 c o s 2 ω τ 1 , s i n 2 ω τ = 2 s i n ω τ c o s ω τ = 2 c o s ω τ 1 c o s 2 ω τ , the first equation of Equation (35) can be transformed into
( η 1 R + η 5 R ) ( 2 c o s 2 ω τ 1 ) + ( η 2 R + η 4 R ) c o s ω τ + η 3 R = ( η 5 I η 1 I ) 2 c o s ω τ 1 c o s 2 ω τ ( η 4 I η 2 I ) 1 c o s 2 ω τ .
Squaring both sides of Equation (36) makes it easy to obtain
γ 4 c o s 4 ω τ + γ 3 c o s 3 ω τ + γ 2 c o s 2 ω τ + γ 1 c o s ω τ + γ 0 = 0 .
Let χ = c o s ω τ ; then, Equation (37) becomes
γ 4 χ 4 + γ 3 χ 3 + γ 2 χ 2 + γ 1 χ + γ 0 = 0 .
In the following, we need to find a condition that ensures that Equation (38) has at least one root. We denote n ( χ ) = γ 4 χ 4 + γ 3 χ 3 + γ 2 χ 2 + γ 1 χ + γ 0 , noting that lim χ + n ( χ ) = + .
From the above analysis, we can obtain the following lemma:
Lemma 6. 
If γ 0 < 0 , i.e., ( η 1 R + η 5 R ) 2 + η 3 R 2 2 η 3 R ( η 1 R + η 5 R ) ( η 4 I η 2 I ) 2 < 0 then Equation (38) has at least one root.
Proof. 
Through Equation (38), it can be seen that Equation (38) is a one-dimensional quartic equation about χ . If γ 0 < 0 then n ( 0 ) = γ 0 < 0 . Due to γ 4 = 4 ( η 1 R + η 5 R ) 2 + 4 ( η 5 I η 1 I ) 2 being a constant that is always greater than or equal 0 and lim χ + n ( χ ) = + , at least one real value χ 0 satisfies n ( χ 0 ) = 0 . Thus, it can be concluded that Equation (37) has at least one root c o s ω τ . This complete the proof. □
We assume
c o s ω τ = H 1 ( ω ) ,
where H 1 ( ω ) is a continuous function about ω .
Similarly, by applying the trigonometric function formula c o s ω τ = ± 1 s i n 2 ω τ , c o s 2 ω τ = 1 2 s i n 2 ω τ , s i n 2 ω τ = 2 s i n ω τ c o s ω τ = 2 s i n ω τ 1 s i n 2 ω τ , we can obtain
( η 1 R + η 5 R ) ( 1 2 s i n 2 ω τ ) + ( η 4 I η 2 I ) sin ω τ + η 3 R = 2 ( η 5 I η 1 I ) sin ω τ 1 sin 2 ω τ ( η 2 R + η 4 R ) 1 sin 2 ω τ .
Squaring both sides of Equation (40) makes it easy to obtain
ξ 4 sin 4 ω τ + ξ 3 sin 3 ω τ + ξ 2 sin 2 ω τ + ξ 1 sin ω τ + ξ 0 = 0 .
Let υ = s i n ω τ ; then, Equation (41) becomes
ξ 4 υ 4 + ξ 3 υ 3 + ξ 2 υ 2 + ξ 1 υ + ξ 0 = 0 .
In the following, we need to find a condition that ensures that Equation (42) has at least one root. We denote ψ ( υ ) = ξ 4 υ 4 + ξ 3 υ 3 + ξ 2 υ 2 + ξ 1 υ + ξ 0 , noting that lim υ + ψ ( υ ) = + .
From the above analysis, we can obtain the following lemma:
Lemma 7. 
If ξ 0 < 0 , i.e., ( ( η 1 R + η 5 R ) 2 + η 3 R 2 + 2 η 3 R ( η 1 R + η 5 R ) ( η 2 R + η 4 R ) 2 then the Equation (42) has at least one root.
Proof. 
Through Equation (42), it can be seen that Equation (42) is a one-dimensional quartic equation about υ . If ξ 0 < 0 then ψ ( 0 ) = ξ 0 < 0 . Also, due to ξ 4 = 4 ( η 1 R ( ω ) + η 5 R ( ω ) ) 2 + 4 ( η 5 I ( ω ) η 1 I ( ω ) ) 2 being a constant that is always greater than or equal 0 and lim υ + ψ ( υ ) = + , at least one real value υ 0 satisfies ψ ( υ 0 ) = 0 . Thus, it can be concluded that Equation (41) has at least one root s i n ω τ . This completes the proof. □
We assume
sin ω τ = H 2 ( ω ) ,
where H 2 ( ω ) is a continuous function about ω .
From s i n 2 ω τ + c o s 2 ω τ = 1 , it can be seen that
H 1 2 ( ω ) + H 2 2 ( ω ) = 1 .
In view of cos ω τ = H 1 ( ω ) , we have
τ ϑ = 1 ω [ arccos H 1 ( ω ) + 2 ϑ π ] , ϑ = 0 , 1 , 2 , ....
We suppose that Equation (44) has at least one positive real root.
We set
τ * = m i n ( τ ϑ ) , ϑ = 0 , 1 , 2 , , ω * = ω | τ = τ * .
In order to derive the conditions for the occurrence of Hopf bifurcation, the following assumptions must be given:
(H5) 
E 2 R E 1 R + E 2 I E 1 I E 1 R 2 + E 1 I 2 0 .
Lemma 8. 
Let s ( τ ) = ς ( τ ) + i ω ( τ ) be the root of (32) satisfying ς ( τ * ) = 0 , ω ( τ * ) = ω * . Then, the following transversality condition holds:
R e [ d s d τ ] | τ = τ * , ω = ω * 0 .
Proof. 
Taking the derivative with respect to τ on both sides of Equation (32), we have
d s d τ = s ( 2 ( η 5 e 2 s τ η 1 e 2 s τ ) + ( η 4 e s τ η 2 e s τ ) η 3 + η 1 e 2 s τ + η 2 e s τ + η 4 e s τ τ ( 2 ( η 5 e 2 s τ η 1 e 2 s τ ) + ( η 4 e s τ η 2 e s τ ) ) ,
where η i ( i = 1 , 2 , 3 , 4 , 5 ) are the derivatives of η i .
Because of s i g n { R e [ d s d τ ] | s = i ω * , τ = τ * } = s i g n { R e [ d s d τ ] 1 | s = i ω * , τ = τ * }
Then
[ d s d τ ] 1 = η 3 + η 1 e 2 s τ + η 2 e s τ + η 4 e s τ s ( 2 ( η 5 e 2 s τ η 1 e 2 s τ ) + ( η 4 e s τ η 2 e s τ ) ) τ s .
Bringing s = i ω into (49), we have
R e [ d s d τ ] 1 | s = i ω * , τ = τ * = R e ( E 2 E 1 ) | s = i ω * , τ = τ * = E 2 R E 1 R + E 2 I E 1 I E 1 R 2 + E 1 I 2 ,
where
E 1 = s ( 2 ( η 5 e 2 s τ η 1 e 2 s τ ) + ( η 4 e s τ η 2 e s τ ) ) , E 2 = η 3 + η 1 e 2 s τ + η 2 e s τ + η 4 e s τ .
The detailed proof process will be presented in Appendix C.
Therefore, under the assumptions of condition ( H 5 ) , we have
R e [ d s d τ ] | s = i ω * , τ = τ * 0 .
Thus, the transversality condition of system (28) near τ = τ * is satisfied. □
Next, we consider the dynamic behavior of the system when τ = 0 . When there is no transmission delay in system (28), Equation (32) becomes
s 4 + Π 1 s 3 + Π 2 s 2 + Π 3 s + Π 4 = 0 ,
where the detailed process of Π 1 , Π 2 , Π 3 , Π 4 will be presented in Appendix D.
In order to obtain some of the main results of this article, we made the following assumptions:
(H6) 
Π 4 > 0 , Π 1 Π 3 1 Π 2 > 0 , Π 1 Π 3 0 1 Π 2 Π 4 0 Π 1 Π 3 > 0
Lemma 9. 
Under assumption ( H 6 ) and if τ = 0 , then all the characteristic roots of characteristic Equation (52) have negative real parts. Therefore, without time delay, system (28) is locally asymptotically stable.
Proof. 
It follows the assumption ( H 6 ) that all the roots s i of Equation (52) satisfy | a r g ( s i ) | > q i π 2 ( i = 1 , 2 , 3 , 4 ) . By Lemma 1, we can conclude, then, that all the characteristic roots of characteristic Equation (52) have no purely imaginary roots. Therefore, without time delay, system (28) is locally asymptotically stable. □
Based on the above discussion, we can obtain the following main results:
Theorem 2. 
Assuming that system (28) satisfies hypothesis ( H 1 ) , ( H 2 ) , ( H 5 ) , ( H 6 ) and satisfies conditions γ 0 < 0 , ξ 0 < 0 , the following statement holds:
(1) 
If ( H 1 ) and ( H 6 ) hold, system (28) is locally asymptotically stable when τ = 0 .
(2) 
If ( H 1 ) , ( H 2 ) , ( H 5 ) , and γ 0 < 0 , ξ 0 < 0 hold, system (28) is locally asymptotically stable when τ ( 0 , τ * ) , and it is unstable for τ > τ * . Then, system (28) undergoes Hopf bifurcation when τ = τ * .

5. Numerical Simulation

In this section, we will provide two numerical simulation examples to verify the main results obtained in the previous section and the effectiveness of the added delay feedback controller. In this paper, the activation function was chosen as f ( x ) = t a n h ( x ) and its selection met the assumption ( H 1 ) . The parameter values of the system were set to fixed values, as shown in Table 1:
Example 1. 
In this example, we mainly considered the following uncontrollable fractional-order systems with cooperation–competition relationships:
D 0.92 x 1 ( t ) = x 1 ( t ) + 0.3 f ( x 1 ( t ) ) 1.5 f ( x 1 ( t τ 1 ) ) + 0.5 ( x 2 ( t ) x 1 ( t ) ) 0.5 ( x 4 ( t ) x 1 ( t ) ) 0.5 ( x 3 ( t ) x 1 ( t ) ) , D 0.85 x 2 ( t ) = x 2 ( t ) + 0.3 f ( x 2 ( t ) ) 1.5 f ( x 2 ( t τ 2 ) ) + 0.5 ( x 1 ( t ) x 2 ( t ) ) 0.5 ( x 4 ( t ) x 2 ( t ) ) 0.5 ( x 3 ( t ) x 2 ( t ) ) , D 0.77 x 3 ( t ) = x 3 ( t ) + 0.3 f ( x 3 ( t ) ) 1.5 f ( x 3 ( t τ 3 ) ) + 0.5 ( x 4 ( t ) x 3 ( t ) ) 0.5 ( x 1 ( t ) x 3 ( t ) ) 0.5 ( x 2 ( t ) x 3 ( t ) ) , D 0.75 x 4 ( t ) = x 4 ( t ) + 0.3 f ( x 4 ( t ) ) 1.5 f ( x 4 ( t τ 4 ) ) + 0.5 ( x 3 ( t ) x 4 ( t ) ) 0.5 ( x 1 ( t ) x 4 ( t ) ) 0.5 ( x 2 ( t ) x 4 ( t ) ) .
Obviously, we can see that the above system had a unique equilibrium point E ( 0 , 0 , 0 , 0 ) . We calculated τ 0 = 1.8885 and ω 0 = 0.8069 . Figure 2 shows the waveform plots and phase diagrams of the neurons in system (53). From Figure 2, we find that the system approached a stable equilibrium point E ( 0 , 0 , 0 , 0 ) when τ = 1.6 < τ 0 = 1.8885 , and when τ = 1.9 > τ 0 = 1.8885 the system lost stability and generated bifurcation periodic solutions near the equilibrium point, as shown in Figure 3 and Figure 4. Based on the previous assumptions ( H 1 ) ( H 4 ) , the conclusion of Theorem 1 can be confirmed.
Next, we use the control variable method to discuss the effect of orders on the Hopf bifurcation critical value of the uncontrollable system (54). When q i ( i = 1 , 2 , 3 , 4 ) changed, its corresponding bifurcation point τ 0 and frequency critical value ω 0 could be easily determined, as shown in Table 2 and Table 3. At the same time, as shown in Figure 5, we plotted the curve of the bifurcation value changing with order. From the figure, it is evident that the bifurcation value gradually decreased and then increased as the order increased; the bifurcation value was smallest when q i = 0.5 .
Example 2. 
In this example, we mainly considered the following controllable fractional-order system with cooperation–competition relationships:
D 0.92 x 1 ( t ) = x 1 ( t ) + 0.3 f ( x 1 ( t ) ) 1.5 f ( x 1 ( t τ 1 ) ) + 0.5 ( x 2 ( t ) x 1 ( t ) ) 0.5 ( x 4 ( t ) x 1 ( t ) ) 0.5 ( x 3 ( t ) x 1 ( t ) ) + K 1 ( x 1 ( t τ ) x 1 ( t ) ) , D 0.85 x 2 ( t ) = x 2 ( t ) + 0.3 f ( x 2 ( t ) ) 1.5 f ( x 2 ( t τ 2 ) ) + 0.5 ( x 1 ( t ) x 2 ( t ) ) 0.5 ( x 4 ( t ) x 2 ( t ) ) 0.5 ( x 3 ( t ) x 2 ( t ) ) , D 0.77 x 3 ( t ) = x 3 ( t ) + 0.3 f ( x 3 ( t ) ) 1.5 f ( x 3 ( t τ 3 ) ) + 0.5 ( x 4 ( t ) x 3 ( t ) ) 0.5 ( x 1 ( t ) x 3 ( t ) ) 0.5 ( x 2 ( t ) x 3 ( t ) ) , D 0.75 x 4 ( t ) = x 4 ( t ) + 0.3 f ( x 4 ( t ) ) 1.5 f ( x 4 ( t τ 4 ) ) + 0.5 ( x 3 ( t ) x 4 ( t ) ) 0.5 ( x 1 ( t ) x 4 ( t ) ) 0.5 ( x 2 ( t ) x 4 ( t ) ) + K 1 ( x 4 ( t τ ) x 4 ( t ) ) .
It is obvious that system (54) had a unique equilibrium, E * ( 0 , 0 , 0 , 0 ) . In order to better demonstrate the effectiveness of the bifurcation control of the delay feedback controller proposed in this article, we adopted all the parameters and orders set in Example 1. To compare with Figure 3 and Figure 4, we still took τ = 1.9 and the feedback gain coefficient k 1 = 0.4 . We calculated that the bifurcation point τ 0 = 3.1056 and the critical frequency ω 0 = 0.4773 . Figure 5 and Figure 6 show the waveform plots and phase diagrams of the neurons in system (54). From Figure 5, we can see that the system approached a stable equilibrium point E * ( 0 , 0 , 0 , 0 ) when τ = 1.9 < τ * = 3.1056 and when the τ = 3.2 > τ * = 3.1056 the system lost stability and generated a bifurcation periodic solution near the equilibrium point, as shown in Figure 7 and Figure 8. Based on the previous assumptions ( H 1 ) ( H 2 ) , ( H 5 ) ( H 6 ) , the conclusion of Theorem 2 can be confirmed.
Remark 8. 
From Figure 2, it can be seen that the uncontrolled system (53) was unstable and that periodic solutions would appear when τ = 1.9 . Figure 5 shows that at τ = 1.9 , the controlled system (54) was locally asymptotically stable at the equilibrium point. Comparing Figure 2 with Figure 5, the results show the effectiveness of our designed delay feedback controller.
In addition, we investigated the effect of feedback gain parameter k 1 on the dynamic behavior of the system. Figure 9 and Figure 10 show that with the increase of k 1 in system (54) the bifurcation control of the controller became more pronounced.
Remark 9. 
From Figure 10, it can be seen that as the feedback gain coefficient k 1 increases, the Hopf bifurcation value τ * of the system increases. The yellow region is the stable region of system (54), which will undergo Hopf bifurcation on the boundary curve. When the values of k 1 and τ exceed the region, system (54) will bifurcate out of its periodic solution from the equilibrium point. This indicates that as the feedback gain coefficient k 1 increases, the start of the Hopf bifurcation can be delayed, which broadens the stability threshold of the control system, and the bifurcation control of the controller becomes more pronounced.
In order to further analyze the impact of changes in the fractional order of the system on bifurcation, we provide a sensitivity analysis of the system to order. As shown in Figure 11, q 1 has stronger sensitivity to the bifurcation critical value.

6. Conclusions

For this article, we investigated the bifurcation control problem of fractional-order neural networks with general cooperation–competition topology. Our research mainly focused on the following aspects. Firstly, self-feedback delay was selected as the bifurcation parameter to analyze the bifurcation dynamic behavior of the uncontrolled system, and a new sufficient condition was established to ensure the stability of the network model and the existence of Hopf bifurcation. Secondly, the influence of fractional order on the bifurcation critical value of uncontrolled systems was discussed. Then, a time-delay feedback controller was introduced to study the stability of its controlled system and the existence of Hopf bifurcation. Finally, the influence of feedback gain coefficient on the controlled system was discussed. The results indicate that when q i ( i = 1 , 2 , 3 , 4 ) ( 0 , 0.5 ) , reducing the order will delay the start of Hopf bifurcation and broaden its stability threshold. When q i ( i = 1 , 2 , 3 , 4 ) ( 0.5 , 1 ) , reducing the order will delay the start of Hopf bifurcation and broaden its stability threshold. In addition, for uncontrolled systems, increasing the feedback gain coefficient k 1 results in more pronounced bifurcation control of the controller.
In future research, we will introduce more types of controllers to broaden the stability thresholds of network models.

Funding

This research received no external funding.

Data Availability Statement

Data are contained within the article.

Conflicts of Interest

The author declares no conflicts of interest.

Appendix A

R 1 ( s ) = s q 1 + q 2 + q 3 + q 4 A ( s q 1 + q 2 + q 3 + s q 1 + q 2 + q 4 + s q 2 + q 3 + q 4 + s q 1 + q 3 + q 4 ) + ( A 2 d 2 ) ( s q 1 + q 2 + s q 1 + q 3 + s q 1 + q 4 + s q 2 + q 3 + s q 2 + q 4 + s q 4 + q 3 ) + ( A 3 + 3 A d 2 2 d 2 ) ( s q 1 + s q 4 + s q 2 + s q 3 ) + A 4 6 A 2 d 2 + 8 A d 3 3 d 4 , R 2 ( s ) = c ( s q 1 + q 2 + q 3 + s q 1 + q 2 + q 4 + s q 2 + q 3 + q 4 + s q 1 + q 3 + q 4 ) + 2 A c ( s q 1 + q 2 + s q 1 + q 3 + s q 1 + q 4 + s q 2 + q 3 + s q 2 + q 4 + s q 4 + q 3 ) + ( 3 A 2 c + 3 c d 2 ) ( s q 1 + s q 4 + s q 2 + s q 3 ) + 4 A 3 c 12 A c d 2 + 8 c d 3 , R 3 ( s ) = c 2 ( s q 1 + q 2 + s q 1 + q 3 + s q 1 + q 4 + s q 2 + q 3 + s q 2 + q 4 + s q 4 + q 3 ) 3 A c 2 ( s q 1 + s q 4 + s q 2 + s q 3 ) + 6 A 2 c 2 6 c 2 d 2 , R 4 ( s ) = c 3 ( s q 1 + s q 4 + s q 2 + s q 3 ) + 4 A c 3 , R 5 ( s ) = c 4 .

Appendix B

R e [ d s d τ ] 1 | s = i ω 0 , τ = τ 0 = R e ( D 2 ( s ) D 1 ( s ) ) | s = i ω 0 , τ = τ 0 = R e ( D 2 ( s ) D 1 ( s ) ) | s = i ω 0 , τ = τ 0 = R e ( D 2 R + i D 2 I D 1 R + i D 1 I ) | s = i ω 0 , τ = τ 0 = R e ( ( D 2 R + i D 2 I ) ( D 1 R i D 1 I ) ( D 1 R + i D 1 I ) ( D 1 R i D 1 I ) ) | s = i ω 0 , τ = τ 0 = D 2 R D 1 R + D 2 I D 1 I D 1 R 2 + D 1 I 2 .

Appendix C

R e [ d s d τ ] 1 | s = i ω * , τ = τ * = R e ( η 3 + η 1 e 2 s τ + η 2 e s τ + η 4 e s τ s ( 2 ( η 5 e 2 s τ η 1 e 2 s τ ) + ( η 4 e s τ η 2 e s τ ) ) ) | s = i ω * , τ = τ * = R e ( E 2 E 1 ) | s = i ω * , τ = τ * = R e ( E 2 R + i E 2 I E 1 R + i E 1 I ) | s = i ω * , τ = τ * = R e ( ( E 2 R + i E 2 I ) ( E 1 R i E 1 I ) ( E 1 R + i E 1 I ) ( E 1 R i E 1 I ) ) | s = i ω * , τ = τ * = E 2 R E 1 R + E 2 I E 1 I E 1 R 2 + E 1 I 2 ,
E i R and E i I as the real and imaginary part of E i ( i = 1 , 2 ) , respectively.

Appendix D

Π 1 = 2 ( A + c + B + G ) , Π 2 = A 2 + 4 A B + B 2 + 2 A c + 4 B c + c 2 6 d 2 + 4 A G + 2 B G + 4 c G + G 2 , Π 3 = 2 A 2 B 2 A B 2 4 A B c 2 B 2 2 B c 2 + 6 A d 2 + 6 B d 2 + 6 c d 2 8 d 3 2 A 2 G 4 A B G 4 A c G 4 B c G 2 c 2 G + 6 d 2 G 2 A G 2 2 c G 2 , Π 4 = A 2 B 2 + 2 A B 2 c + B 2 c 2 A 2 d 2 4 A B d 2 B 2 d 2 2 A c d 2 4 B c d 2 c 2 d 2 + 4 ( A + B + G + c ) d 3 + 2 B G ( A 2 + c 2 d 2 ) 3 d 4 + 4 A G ( B c d 2 ) 4 c d 2 G + A 2 G 2 + 2 A G 2 c + ( c 2 d 2 ) G 2 .

References

  1. Bartocci, E.; Gol, E.; Haghighi, I.; Belta, C. A formal methods approach to pattern recognition and synthesis in reaction diffusion networks. IEEE Trans. Control Netw. Syst. 2018, 5, 308–320. [Google Scholar] [CrossRef]
  2. Kulkarni, S.; Rajendran, B. Spiking neural networks for handwritten digit recognition-supervised learning and network optimization. Neural Netw. 2018, 103, 118–127. [Google Scholar] [CrossRef] [PubMed]
  3. Jabeen, D.; Iqrar, T.; Khan, M. Multidimensional signal processing using quaternion complex Hadamard transform in sequency domain. Electron. Lett. 2018, 54, 1435–1436. [Google Scholar] [CrossRef]
  4. Jia, C.; Wang, S.; Zhang, X. Content-aware convolutional neural network for in-loop filtering in high efficiency video coding. IEEE Trans. Image Process. 2019, 28, 3343–3356. [Google Scholar] [CrossRef]
  5. Carpenter, G. Neural network models for pattern recognition and associative memory. Neural Netw. 1989, 2, 243–257. [Google Scholar] [CrossRef]
  6. Song, H.; Dai, J.; Sheng, G. GIS partial discharge pattern recognition via deep convolutional neural network under complex data source. IEEE Trans. Dielectr. Electr. Insul. 2018, 25, 678–685. [Google Scholar] [CrossRef]
  7. Liu, J.; Gong, M.; He, H. Deep associative neural network for associative memory based on unsupervised representation learning. Neural Netw. 2019, 113, 41–53. [Google Scholar] [CrossRef]
  8. He, J.; Xiao, M.; Zhao, J.; Wang, Z.; Yao, Y.; Cao, J. Tree-structured neural networks: Spatiotemporal dynamics and optimal control. Neural Netw. 2023, 164, 395–407. [Google Scholar] [CrossRef]
  9. Lu, Y.; Xiao, M.; He, J.; Wang, Z. Stability and bifurcation exploration of delayed neural networks with radial-ring configuration and bidirectional coupling. IEEE Trans. Neural Netw. Learn. Syst. 2023. [Google Scholar] [CrossRef]
  10. Nie, X.; Zheng, W. Dynamical behaviors of multiple equilibria in competitive neural networks with discontinuous nonmonotonic piecewise linear activation functions. IEEE Trans. Cybern. 2016, 46, 679–693. [Google Scholar] [CrossRef]
  11. Zheng, S.; Luo, M. Competition or cooperation? ports’ strategies and welfare analysis facing shipping alliances. Transp. Res. Part E Logist. Transp. Rev. 2021, 153, 102429. [Google Scholar] [CrossRef]
  12. Zhang, X.; Zhang, Z.; Wade, M.J. Dynamical analysis of a competition and cooperation system with multiple delays. Bound. Value Probl. 2018, 2018, 111. [Google Scholar] [CrossRef]
  13. Shen, H.; Wang, X.; Duan, P.; Cao, J.; Wang, J. H bipartite synchronization control of markov jump cooperation–competition networks with reaction–diffusions. IEEE Trans. Cybern. 2023, 53, 6626–6635. [Google Scholar] [CrossRef] [PubMed]
  14. Chandra, R. Competition and collaboration in cooperative coevolution of elman recurrent neural networks for time-series prediction. IEEE Trans. Neural Netw. Learn. Syst. 2015, 26, 3123–3136. [Google Scholar] [CrossRef] [PubMed]
  15. Shen, H.; Song, Y.; Wang, J.; Park, J. H infinity state estimation for PDT-switched coupled neural networks under round-robin protocol: A cooperation-competition-based mechanism. IEEE Trans. Netw. Sci. Eng. 2023, 10, 911–921. [Google Scholar] [CrossRef]
  16. Wu, Y.; Hu, B.; Guan, Z. Consensus problems over cooperation-competition random switching networks with noisy channels. IEEE Trans. Neural Netw. Learn. Syst. 2019, 30, 35–43. [Google Scholar] [CrossRef]
  17. Lee, M.; Ahn, H.; Kwon, S.; Kim, S. Cooperative and competitive contextual effects on social cognitive and empathic neural responses. Front. Hum. Neurosci. 2018, 12, 218. [Google Scholar] [CrossRef]
  18. Xu, Y.; Lin, T.; Liu, X.; Li, W. Exponential bipartite synchronization of fractional-order multilayer signed networks via hybrid impulsive control. IEEE Trans. Cybern. 2023, 53, 3926–3938. [Google Scholar] [CrossRef]
  19. Siami, M. Stability and robustness analysis of commensurate fractional-order networks. IEEE Trans. Control Netw. Syst. 2021, 8, 1261–1269. [Google Scholar] [CrossRef]
  20. Li, R.; Cao, J.; Li, N. Stabilization of reaction-diffusion fractional-order memristive neural networks. Neural Netw. 2023, 165, 290–297. [Google Scholar] [CrossRef]
  21. Yin, H.; Tang, J.; Liu, B.; Yao, S.; Kang, Q. Chaotic resonance in discrete fractional-order LIF neural network motifs. Int. J. Bio-Inspired Comput. 2023, 21, 175–188. [Google Scholar] [CrossRef]
  22. Huang, C.; Gao, J.; Mo, S.; Cao, J. Hopf bifurcation in a fractional-order neural network with self-connection delay. Nonlinear Dyn. 2023, 111, 14335–14350. [Google Scholar] [CrossRef]
  23. Ma, X.; Tan, C.; Chen, Z.; Wong, W. Adaptive control for a class of stochastic nonlinear time-delay systems with unknown control coefficients. J. Frankl. Inst.-Eng. Appl. Math. 2024, 361, 107110. [Google Scholar] [CrossRef]
  24. Chen, Y.; Tao, G.; Yao, Y. A composite adaptive robust control for nonlinear systems with model uncertainties. Int. J. Robust Nonlinear Control 2023, 33, 5800–5818. [Google Scholar] [CrossRef]
  25. Jmal, A.; Naifar, O.; Makhlouf, A.B.; Derbel, N.; Hammami, M.A. Adaptive estimation of component faults and actuator faults for nonlinear one-sided lipschitz systems. Int. J. Robust Nonlinear Control 2020, 30, 1021–1034. [Google Scholar] [CrossRef]
  26. Naifar, O.; Jmal, A.; Makhlouf, A.B. Non-fragile H observer for lipschitz conformable fractional-order systems. Asian J. Control 2021, 24, 2202–2212. [Google Scholar] [CrossRef]
  27. Naifar, O.; Makhlouf, A.B.; Mohamed, A.K.; Hammami, M.A. Finite-time stability of linear fractional-order time-delay systems. Int. J. Robust Nonlinear Control 2019, 29, 180–187. [Google Scholar] [CrossRef]
  28. Jmal, A.; Naifar, O.; Makhlouf, A.B.; Derbel, N.; Hammami, M.A. On observer design for nonlinear caputo fractional-order systems. Asian J. Control 2018, 20, 1533–1540. [Google Scholar] [CrossRef]
  29. Zazo, M.; Rademacher, J. Bifurcation control for a ship maneuvering model with nonsmooth nonlinearities. Siam J. Control Optim. 2023, 61, 225–251. [Google Scholar] [CrossRef]
  30. Zhao, X.; Zhang, Z.; Zhang, C.; Wang, X.; Guo, Z. Bifurcation behavior analysis and stability region discrimination for series-parallel architecture electric energy router. Int. J. Bifurc. Chaos 2023, 33, 2350091. [Google Scholar] [CrossRef]
  31. Li, P.; Lu, Y.; Xu, C.; Ren, J. Bifurcation phenomenon and control technique in fractional BAM neural network models concerning delays. Fractal Fract. 2023, 7, 7. [Google Scholar] [CrossRef]
  32. Piccirillo, V. LP optimal feedback control of homoclinic bifurcation in a forced Duffing oscillator. Nonlinear Dyn. 2023, 111, 13017–13037. [Google Scholar] [CrossRef]
  33. Gray, R.; Franci, A.; Srivastava, V.; Leonard, N. Multiagent decision-making dynamics inspired by honeybees. IEEE Trans. Control Netw. Syst. 2018, 5, 793–806. [Google Scholar] [CrossRef]
  34. Cheng, Z.; Cao, J. Hybrid control of Hopf bifurcation in complex networks with delays. Neurocomputing 2014, 131, 164–170. [Google Scholar] [CrossRef]
  35. Lin, Y.; Ma, Y.; Dai, Y. Stability and Hopf bifurcation control for fractional-order two-gene regulatory network with multiple delays. IEEE Access 2023, 11, 58389–58405. [Google Scholar] [CrossRef]
  36. Lu, L.; Huang, C.; Song, X. Bifurcation control of a fractional-order PD control strategy for a delayed fractional-order prey-predator system. Eur. Phys. J. Plus 2023, 138, 77. [Google Scholar] [CrossRef]
  37. Wang, L.; Qin, W.; Zhao, Y. Dynamic properties of dual-delay network congestion control system based on hybrid control. Neural Process. Lett. 2023, 55, 9295–9314. [Google Scholar] [CrossRef]
  38. Podlubny, I. Fractional Differential Equations; Mathematics in Science and Engineering; Lightning Source Inc.: La Vergne, TN, USA, 1999. [Google Scholar]
  39. Guo, W.; Yang, J. Hopf bifurcation control of hydro-turbine governing system with sloping ceiling tailrace tunnel using nonlinear state feedback. Chaos Solitons Fractals 2017, 104, 426–434. [Google Scholar] [CrossRef]
Figure 2. The waveform plots of the neurons in system (53) for τ = 1.6 < τ 0 = 1.8885 .
Figure 2. The waveform plots of the neurons in system (53) for τ = 1.6 < τ 0 = 1.8885 .
Fractalfract 08 00689 g002
Figure 3. The waveform plots of the neurons in system (53) for τ = 1.9 > τ 0 = 1.8885 .
Figure 3. The waveform plots of the neurons in system (53) for τ = 1.9 > τ 0 = 1.8885 .
Fractalfract 08 00689 g003
Figure 4. The phase diagrams of the neurons in system (53) for τ = 1.9 > τ 0 = 1.8885 .
Figure 4. The phase diagrams of the neurons in system (53) for τ = 1.9 > τ 0 = 1.8885 .
Fractalfract 08 00689 g004
Figure 5. The waveform plots of the neurons in system (54) for τ = 1.9 < τ * = 3.1056 .
Figure 5. The waveform plots of the neurons in system (54) for τ = 1.9 < τ * = 3.1056 .
Fractalfract 08 00689 g005
Figure 6. The phase diagrams of the neurons in system (54) for τ = 1.9 < τ * = 3.1056 .
Figure 6. The phase diagrams of the neurons in system (54) for τ = 1.9 < τ * = 3.1056 .
Fractalfract 08 00689 g006
Figure 7. The waveform plots of the neurons in system (54) for τ = 3.2 > τ * = 3.1056 .
Figure 7. The waveform plots of the neurons in system (54) for τ = 3.2 > τ * = 3.1056 .
Fractalfract 08 00689 g007
Figure 8. The phase diagrams of the neurons in system (54) for τ = 3.2 > τ * = 3.1056 .
Figure 8. The phase diagrams of the neurons in system (54) for τ = 3.2 > τ * = 3.1056 .
Fractalfract 08 00689 g008
Figure 9. (ad) The waveform plots of different in k 1 on the dynamics of system (54) with q 1 = 0.92 , q 2 = 0.85 , q 3 = 0.77 , q 4 = 0.75 , and τ = 1.9 .
Figure 9. (ad) The waveform plots of different in k 1 on the dynamics of system (54) with q 1 = 0.92 , q 2 = 0.85 , q 3 = 0.77 , q 4 = 0.75 , and τ = 1.9 .
Fractalfract 08 00689 g009
Figure 10. Curve of bifurcation value changing with k 1 in system (54).
Figure 10. Curve of bifurcation value changing with k 1 in system (54).
Fractalfract 08 00689 g010
Figure 11. Sensitivity analysis of system (54).
Figure 11. Sensitivity analysis of system (54).
Fractalfract 08 00689 g011
Table 1. Parameter values of the system.
Table 1. Parameter values of the system.
Parameter q 1 q 2 q 3 q 4
Values0.920.850.770.75
Parameterabcd
Values10.3−1.50.5
Table 2. The bifurcation value changing with order q 1 , q 2 .
Table 2. The bifurcation value changing with order q 1 , q 2 .
Different OrdersDifferent q 1 Different q 2
ω 0 τ 0 ω 0 τ 0
0.10.80121.91720.81951.8735
0.20.77791.97340.79631.9269
0.30.76152.01440.77981.9663
0.40.75192.03840.77001.9894
0.50.74912.04400.76691.9954
0.60.75272.03220.77031.9844
0.70.76272.00330.78011.9572
0.80.77891.95940.79631.9150
0.90.80151.90180.81911.8590
10.83121.83120.84931.7902
Table 3. The bifurcation value changing with order  q 3 , q 4 .
Table 3. The bifurcation value changing with order  q 3 , q 4 .
Different OrdersDifferent q 3 Different q 4
ω 0 τ 0 ω 0 τ 0
0.10.83481.83830.83771.8316
0.20.81211.88830.81521.8808
0.30.79591.92530.79911.9172
0.40.78621.94710.78951.9386
0.50.78311.95270.78641.9441
0.60.78641.94210.78981.9334
0.70.79621.91580.79961.9072
0.80.81241.87490.81581.8666
0.90.83541.82040.83891.8123
10.86601.75300.86951.7452
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Cheng, Z. Control of the Bifurcation Behaviors of Delayed Fractional-Order Neural Networks with Cooperation–Competition Topology. Fractal Fract. 2024, 8, 689. https://doi.org/10.3390/fractalfract8120689

AMA Style

Cheng Z. Control of the Bifurcation Behaviors of Delayed Fractional-Order Neural Networks with Cooperation–Competition Topology. Fractal and Fractional. 2024; 8(12):689. https://doi.org/10.3390/fractalfract8120689

Chicago/Turabian Style

Cheng, Zunshui. 2024. "Control of the Bifurcation Behaviors of Delayed Fractional-Order Neural Networks with Cooperation–Competition Topology" Fractal and Fractional 8, no. 12: 689. https://doi.org/10.3390/fractalfract8120689

APA Style

Cheng, Z. (2024). Control of the Bifurcation Behaviors of Delayed Fractional-Order Neural Networks with Cooperation–Competition Topology. Fractal and Fractional, 8(12), 689. https://doi.org/10.3390/fractalfract8120689

Article Metrics

Back to TopTop