Some Further Results on Fixed-Time Synchronization of Neural Networks with Stochastic Perturbations ()
1. Introduction
In the last decades, the various types of neural networks (NNs) including Hopfield NNs, cellular NNs, convolution NNs, Cohen-Grossberg NNs, BAM NNs and so on [1] [2] [3], have been introduced and broadly investigated due to their important applications in great number of fields ranging from speech recognition [4] to image encryption [5], from secure communication [6] to robotic manipulators [7], etc. As pointed out in [8], stability of NNs is prerequisite in some applications. As a result, the stability analysis of NNs has been investigated extensively by many scholars [2] [3] [8] [9] [10] [11] [12]. Duo to the security reasons or just to improve system performance in many practical applications, it is desirable that the systems solution trajectories converge to equilibrium as fast as possible [13]. Compared to the classical Lyapunov stability such as asymptotical stability or exponential stability, finite-time (FNT) stability allows the solution of an asymptotic system approaches to the equilibrium state after a some bounded time
and stays at equilibrium state any time longer than
, where
is called the settling time. Thus, FNT stability and stabilization of NNs have been investigated widely in the past two decades [10] [11] [12] [13] [14].
One of the important tasks in FNT stability is to estimate the settling time (ST)
, and it is desirable to obtain smaller upper-bound of
. However, in some cases, it is inconvenient to accurately estimate it due to its heavily dependence to the initial values of the system. So it is to better obtain FNT stability with a ST irreverent to initial conditions of the system. This issue was firstly studied by Polyakov [15] via defining a so-called fixed-time (FXT) stability which its ST is independent to systems initial conditions. Presently, FXT stability receives a hot research attention from many scholars since it has awesome applications in multi-agent systems [16], power systems [17], complex networks [18] and so on.
In addition, the synchronization of chaotic nonlinear system has received great attention in the past thirty years due to the fact that synchronization is unique in nature and plays a crucial role in many fields including biology, climatology and sociology, etc. [19]. As mentioned in [9], synchronization in neuronal systems can produce a lot of physiological mechanisms of brain functions such as attention, learning, memory formation and so on. Thus, to understand these brain functions deeply, it is an important task to study synchronization behaviors of NNs. For this reason, considerable efforts have been devoted to study the synchronization of NNs [5] [6] [20] - [25]. Especially, due to the advantage of faster convergence rate, better robustness and disturbance rejection properties, FNT and FXT synchronization of various NNs have studied recently. For example, in [9], the authors investigated the FXT synchronization of coupled discontinuous NNs by introducing new FXT stability results for dynamical systems. In [11], the authors considered FNT stabilization issue of a class of delayed memristive NNs with discontinuous right-hand side by designing two types of discontinuous controllers. In [21], the authors concerned the FXT synchronization of a class of memristor-based NNs with impulsive effects. In [26], the authors studied the FNT synchronization of a type of complex-valued NNs with distributed delays. In [27], the authors studied the FXT synchronization problem of a type of quaternion-valued NN with time delays.
However, it is worthy to note that most of the above mentioned results have only considered cases without stochastic perturbations. As depicted in [28], noises are frequently encountered in both nature and man-made systems. For instance, synaptic transmission in the real nervous systems can be seen as a noisy process which is caused by random fluctuations due to the release of neurotransmitters and other probabilistic effects. Besides, for many natural renewable energy resources such as wind or solar radiation, their availability is somewhat subject to stochastic fluctuations [19]. Therefore, recently many scholars paid their attention to study the synchronization of NNs with stochastic perturbations, and till now there are many excellent results on the complete synchronization, lag synchronization, projective synchronization and FNT synchronization of stochastic NNs with or without time-delays. But, up to now, there are very few results on the FXT synchronization of stochastic NNs.
Inspired by what mentioned above, in this paper, we considered the FXT synchronization of a type of NNs with stochastic perturbations via using some improved FXT stability results. The main contributions of this work can be stated as follows. 1) Some earlier results on FXT stability of deterministic nonlinear systems are extended to the stochastic nonlinear systems. 2) Some novel sufficient conditions guaranteeing the FXT synchronization of considered stochastic NNs are derived via introducing two types of controllers and employing some inequality techniques. Lastly, two numerical examples with simulations are provided to show the feasibility of our theoretical results.
The rest of the article is structured as follows. In Section 2, some basic assumptions together with our new FXT stability lemma are proposed. Our main results on FXT synchronization in probability are given in Section 3. Two numerical examples are provided in Section 4 and our main conclusions are given in Section 5.
Notations: The notations are quite standard. Throughout this paper,
and
denotes the set of nonnegative real numbers and the n-dimensional Euclidean space, respectively. The superscript T represents the matrix or vector transposition. The
identity matrix is denoted as
.
is the Euclidean norm in
.
denotes the set of positive integers. Moreover,
stands for the complete probability space, where
represents the filtration satisfying the usual conditions [29]. Notation
denotes for the operator of mathematical expectation corresponds to the given probability measure P.
represents the maximum eigenvalue of a real symmetric matrix.
2. Problem Formulation and Preliminaries
Consider a class of n-dimensional stochastic NNs depicted by the following equation
(1)
where
represents the state vector of the network at time t,
stands for the activation function of neurons;
with
represents to the self-feedback connection weight matrix;
represents the connection weight matrix between neurons;
denotes neuron input vector;
is an n-dimensional Brown motion defined on a complete probability space
with a natural filtration
generated by
. For more explanations about stochastic process, please see the works [19] [22] [29].
The corresponding response system of the drive system (1) is given by
(2)
where
denotes the state vector of the response system,
is the feedback controller to be introduced. Other parameters
and
are the same as defined in system (1).
In the paper, we assume that the following assumptions are satisfied for the system (1).
Assumption 1 The neuron activation functions
in system (1) satisfy the Lipschitz condition. That is, for each i there exists a positive constant
such that
where
.
Assumption 2 For
, there exists a matrix
with appropriate dimensions such that the following inequality holds true
Now let
be the synchronization error between drive-response systems (1) and (2), then the error dynamical system can be derived as follows:
(3)
where
and
.
Furthermore, to obtain our main results, we give some related properties of stochastic perturbation, which can be found in [29].
Denote by
the set of all nonnegative functions
on
which its first order derivative exist for t and second order derivative exist for z. For each
, we define an operator
from
to R given as
where
,
,
.
Now consider the following general stochastic nonlinear system:
(4)
where
is state vector of system,
is an n-dimensional Brown motion defined on a complete probability space
.
and
are nonlinear vector-valued continuous functions and they satisfy the condition
.
For convenience, we denote by
the solution of stochastic nonlinear system (4) satisfy the initial value
. Also, in order to get our main results in this part, we state here some needed definitions and lemmas as follows.
Definition 1 (FXT stable in probability [30] ). The zero solution of stochastic system (4) is called to be FNT stable in probability, if the following conditions hold true.
1) FNT attractiveness in probability. That is, for any initial conditions
, the equation
is satisfied, where
is ST function defined as
;
2) Stability in probability: For every pair of scalers
and
, there exists a positive constant
such that
.
Definition 2 (FXT stable in probability [30] ). The zero solution
of system (4) is said to be globally FXT stable in probability, if the following statements are satisfied for all the initial states
.
1) The zero solution
is globally stochastically FNT stable in probability.
2) Mathematical expectation of ST function
is independent on the initial state
of (4) and its upper bound is bounded by a positive constant
. That is,
for all
.
Now, we introduce the following lemmas about FXT stability.
Lemma 1 [30]. Assume that
is a positive definite Lyapunov function,
is a continuous function and it belongs to set
, where
denotes set of the bounded functions which is defined be
where M is positive constant. If the following inequality is satisfied for all
(5)
then the zero solution
of system (4) is globally stochastically FXT stable in probability and the its ST function
can be estimated as
.
Lemma 2 [20]. For system (4), if there exists a positive definite function
and positive numbers
satisfying
,
such that
(6)
then the zero solution of system (4) is globally stochastically FXT stable in probability, and its ST
can be estimated as
Lemma 3 [31]. Suppose
is a positive definite function, and it satisfies the following conditions.
1)
;
2) For any solution
of system (4), following inequality hold true
(7)
for some
and
. Then the zero solution of system (4) can achieve FXT stability, and its corresponding ST
can be estimated by
(8)
Lemma 4 [23]. Suppose
is a C-regular function such that
(9)
where
, and
, then the following results are true.
1) If
, the zero solution of system (4) is FXT stable and its settling-time
is estimated by
where
.
2) If
, the zero solution of system (4) is FXT stable and its ST
is estimated by
where
stands for the incomplete beta function ratio for
and
, which is defined by
here
is the beta function given by
Lemma 5 [23]. For system (4), assume
is a C-regular function. If there exist constants
,
,
,
and
satisfying
such that
the zero solution of system (4) is FXT stable and its ST can be estimated by
Lemma 6 [24]. For system (4), assume that
is a C-regular function. If there exist constants
,
and
satisfying
such that
Then the zero solution of system (4) is FXT stable and its corresponding ST can be estimated as
, where
here
.
Lemma 7 [21]. If
, then
Lemma 8 [22]. Let v and z be any two column vectors in
, then the following matrix inequality is satisfied for any positive definite matrix
.
3. Main Results
In this section, based on the FXT stability results introduced in above section, we will derive some sufficient criteria for the FXT synchronization between the drive-response systems (1) and (2). To this, first we design the controller
in response system (2) as follows:
(10)
where
is a positive constant matrix.
and
are the tunable constants, and p and q are the real numbers such that
.
,
, and
(11)
Then, under controller (10), the error system (3) can be rewritten as follows:
(12)
Now let
, then based on the FXT controller (12), the following results can be derived.
Theorem 1. Suppose that the Assumptions 1 and 2 are satisfied, if the control gain matrix
satisfy the following matrix inequality
(13)
where
is an arbitrary
positive matrix. Then the drive-response networks (1) and (2) can be FXT synchronized in probability via controller (10), and its corresponding ST can be estimated by
where
and
are respectively given in Lemma 3 and Lemma 4 with the parameters
,
,
,
and
.
Proof. First, we construct the following Lyapunov function
(14)
Then, by calculating the
along the trajectories of error system (3), we get
(15)
By Lemma 8 and Assumption 1, we obtain the following inequality:
(16)
By Assumption 2, we have
(17)
Also, it is not difficult to check that
(18)
and
(19)
Let
, and using the well-known Schur complement equivalence [32] to
, which is defined in (15), we obtain
or
. Thus, by substituting (16), (17), (18) and (19) to (15), we have
(20)
By Lemma 7, we can obtain that
(21)
and
(22)
In view of (22), (21) and (22), we can have
(23)
where
and
.
Therefore, we can conclude from the Lemmas 3 and 4 that the origin of error system (3) is FXT stable in probability and its ST can be estimated by
where
and
are given in Lemmas 3 and 4 respectively, and their parameters are chosen as
,
,
,
,
. The proof is achieved.
When
in the controller (10), we have a following Corollary from Theorem 1 and Lemma 6.
Corollary 1. Suppose that
in the controller (10), if Assumption 1, Assumption 2 and matrix inequality (15) are satisfied, then the drive-response networks (1) and (2) can be FXT synchronized in probability under controller (10), and its ST is estimated by
, where
is given in
Lemma 6 with the parameters
,
,
,
and
.
Proof. Similar to proof of Theorem 1, we know that the inequality (24) is satisfied under the conditions of Corollary 1. Thus from Lemma 6, we can obtain that the conclusions of Corollary 1 hold true. The proof is completed.
In the following, we will realize the fixed time synchronization between the systems (1) and (2) via designing a simplified controller given as follows
(24)
where parameters
and q are the same as defined in controller (10).
Theorem 2. Suppose that
is an arbitrary
positive matrix and the Assumptions 1 and 2 hold true, then the drive-response systems (1) and (2) will realize FXT synchronization in probability via controller (24). Moreover, its corresponding ST can be estimated as
, where
here
, and the
are given in Lemmas 3 and 4, and their parameters are chosen as
,
,
,
and
.
Proof. We again chose the Lyapunov function as
. Then, under controller (24), we have
(25)
By Equations (16)-(19), we have
(26)
Introducing (21) and (22) to (26) yields
(27)
According to Lemma 4, the drive-response networks (1) and (2) will achieve FXT synchronization in probability. In addition, its ST
can estimate through following analysis.
1) If
, then from Lemma 4, we can get that
.
2) If
, then from Lemma 4 again, we can have that
.
3) If
, then from Lemmas 3 and 4, we can obtain that
.
where the parameters of
and
are chosen as
,
,
,
and
.
The proof of Theorem 2 is completed.
When
in the controller (24), we have a following result from Theorem 2 and Lemma 4.
Corollary 2. Suppose that
in controller (10) and the Assumptions 1 and 2 are satisfied, then the drive-response systems (1) and (2) will achieve FXT synchronization in probability via controller (24). Moreover, its corresponding ST can be estimated as
, where
here
,
and
are respectively given in Lemmas 4, 5 and 6, and their parameters are chosen as
,
,
,
,
and
.
Proof. From the proof of Theorem 2, we know that the inequality (28) is satisfied. Thus according to the Lemma 4, the drive-response networks (1) and (2) can be FXT synchronized in probability. In addition, its ST
can estimate through following analysis.
1) If
, then from Lemma 5, we can get that
.
2) If
, then from Lemma 4 again, we can have that
.
3) If
, then from Lemma 6, we can obtain that
.
Where the parameters of
and
are chosen as
,
,
,
,
,
and
. The proof is completed.
Remark 1. As known to all, when study the synchronization issue of nonlinear systems, it is an important task to design a controller, and it is desirable to design the controller as simple as possible in order to save the control cost. In early published works [21] [23] [31] [33], however, the authors realized FXT synchronization of NNs via employing a type of hybrid controller
which composed of one linear term
and two nonlinear terms
. However, as we done in Theorem 2 and Corollary 2, when the linear term
of
removed, the other two terms
still insures the FXT synchronization of considered networks. Thus the results of Theorem 2 and Corollary 2 are simpler and have a better applicability.
Remark 2. Similar to the most of the published works on FXT stabilization and synchronization, in Theorem 1 and Corollary 1, we have achieved to FXT synchronization by designing a types of controller
which composed of one linear term
and two nonlinear terms
. However, it is worth to note that, all of the three tunable parameters
and
take effect in the estimation of the upper-bound of ST, and this can be seen the main advantages of Theorem 1 and Corollary 1 compared with the early published results [23] [31] [34] [35] [36].
Remark 3. As mentioned above, it is better to realize the FXT synchronization via using simple controller
. However, it is not difficult to see that compared the results of Theorem 2 and Corollary 2, the results of Theorem 1 and Corollary 1 gives smaller estimation when the control gains
in controller
, and the bigger
results in a smaller ST estimation. Therefore, the linear term
should be added in accordance with the convergence time T to be short and the control cost not to be high, considering the designer requirements.
4. Numerical Examples and Simulations
In this section, the following two numerical examples are provided to illustrate the effectiveness of the established theoretical results in above sections.
Example 1. For n = 3, consider the FXT synchronization between drive-response systems (1) and (2) with the following system parameters:
,
,
and
Set the initial values of system (1) in Example 1 as
,
and
, then the numerical simulation of system (1) with above parameters are illustrated in Figure 1, which shows that it has a chaotic attractor.
It is not difficult check that the Assumptions 1 and 2 are satisfied with
and
. By using the LMI Toolbox in Matlab, the following solution are obtained for matrix inequality (15)
Thus inequality (15) is also satisfied and
. Now choosing two different set of parameters as follows: (i)
,
,
and
; (ii)
,
,
and
. Then, from Theorem 1 and Corollary 1, the derive system (1) is FXT stochastic synchronized to response system (2) under the controller (10). The time evolution of synchronization errors between systems (1) and (2) for above two different set of parameters are shown in Figure 2 and Figure 3, respectively, where the initial conditions of response systems (2) are randomly chosen in
. For case (i), from Theorem 1, by simple calculations we get
,
. Thus
. For case (ii), since
, we can get from Corollary 1 that
, while Lemma 2 and Lemma 3 give the ST estimations
and
, respectively.
Example 2. For n = 3, consider the FXT synchronization between drive-response systems (1) and (2) under the controller (24) with the following system parameters:
,
,
and
Figure 1. The transient behavior of system (1) in Example 1.
Figure 2. Evaluation of synchronization errors for case (i).
Figure 3. Evaluation of synchronization errors for case (ii).
Now set the initial values of system (1) in Example 2 as
,
and
, then the numerical simulation of system (1) with above parameters are illustrated in Figure 4, which shows that it also has a chaotic attractor.
It is not difficult check that
,
. Letting
, then by simple calculation we can get that
. Now choosing two different set of parameters such that (i)
,
,
and
; (ii)
,
,
and
. Then, all the conditions of Theorem 2 and Corollary 2 are satisfied for case (i) and case (ii). Therefore, from Theorem 2 and Corollary 2, the derive system (1) is FXT stochastic synchronized to response system (2) under the controller (24). The time evolution of synchronization errors between systems (1) and (2) for above two different set of parameters are shown in Figure 5 and Figure 6 respectively, where the initial conditions of response systems (2) are randomly chosen in
. For case (i), since
, we can calculate from Theorem 2 that
. Thus
. For case (ii), since
and
, we can get from Corollary 2 that
, while Lemma 1 and gives the ST estimation
for case (i) and
for case (ii), respectively.
Remark 4. From the above two examples, we can see that the settling time estimations obtained through Theorems 1, 2 and Corollaries 1, 2 are more accurate compared to the early published results [37] [38] [39] [40] [41]. From this point, results obtained in this paper are more general and have better applicability.
Figure 4. The transient behavior of system (1) in Example 2.
Figure 5. Evaluation of synchronization errors for case (i).
Figure 6. Evaluation of synchronization errors for case (ii).
5. Conclusions
In this paper, first, some recently developed new results on the FXT stability of deterministic dynamical systems are extended to stochastic dynamical systems. First, some earlier results on FXT stability of deterministic nonlinear systems are extended to the stochastic nonlinear systems. Then, based on these results, some simple sufficient conditions insuring the FXT synchronization of considered networks are derived by introducing two types of FXT controllers and utilizing some inequality techniques. Finally, our theoretical results are illustrated via giving two numerical examples with their Matlab simulations.
Recently, the FXT stability and synchronization of impulsive neural networks have been studied. However, there are very few works on the FXT synchronization issue of the stochastic neural networks with impulsive effects; this issue may be somewhat challenging since we have to deal with the effects of caused by impulsive term and stochastic perturbations at the same time, and it will be one of our future studying directions.
Funding
This work was supported by the National Innovation Training Program for College Students (Grant no. 202010755075).
Acknowledgements
Not applicable.