Neurocomputing: Hongjian Liu, Lifeng Ma, Zidong Wang, Yurong Liu, Fuad E. Alsaadi

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

JID: NEUCOM

ARTICLE IN PRESS [m5G;February 4, 2020;9:59]


Neurocomputing xxx (xxxx) xxx

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

An overview of stability analysis and state estimation for memristive


neural networksR
Hongjian Liu a,b, Lifeng Ma c,∗, Zidong Wang d, Yurong Liu e, Fuad E. Alsaadi f
a
Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment, Ministry of Education, Anhui Polytechnic University, Wuhu 241000,
China
b
School of Mathematics and Physics, Anhui Polytechnic University, Wuhu 241000, China
c
School of Automation, Nanjing University of Science and Technology, Nanjing 210094, China
d
Department of Computer Science, Brunel University London, Uxbridge, Middlesex UB8 3PH, United Kingdom
e
Department of Mathematics, Yangzhou University, Yangzhou 225002, China
f
Department of Electrical and Computer Engineering, Faculty of Engineering, King Abdulaziz University, Jeddah 21589, Saudi Arabia

a r t i c l e i n f o a b s t r a c t

Article history: This paper gives a review of recent advances on memristive neural networks with emphasis on the is-
Received 8 November 2019 sues of stability analysis and state estimation. First, the concept of memristive neural network is recalled
Revised 3 January 2020
with a brief introduction of its background. Then, certain types of frequently seen neural networks are re-
Accepted 18 January 2020
viewed comprehensively with latest progress. Some engineering-oriented phenomena that appear exten-
Available online xxx
sively in the context of networked systems are introduced and summarized, including random dynamics,
Communicated by Steven Hoi time-delays and network-induced incomplete information, etc. From different perspectives, several tech-
niques explored for designing the required state estimators of memristive neural networks are discussed
Keywords:
Neural networks in detail. Some latest progress regarding the stability analysis and state estimation problems for discrete
Memristive neural networks time memristive neural networks are presented. Finally, we provide the conclusions and point out certain
Time-delay future research directions.
Network-induced phenomena
© 2020 Elsevier B.V. All rights reserved.
State estimation,

1. Introduction Artificial neural network, also referred to as neural network,


is a mathematical or computational model that mimics the struc-
The rapid development of artificial intelligence (AI) has been ture and function of biological especially human brain neural
profoundly changing our daily life as well as the human society. networks. This idea of mimicking the function of biological brain
Nowadays, the relevant disciplines of AI technology have been directly affects the development of AI technology. It should be
promoted from different aspects including but not limited to noted that it is of great significance for the development of AI
theoretical modeling, technological innovation and software and to better realize the “intelligence-like-brain”, thereby completing
hardware upgrades, etc. To date, the AI technology has been the decision-making behavior, program behavior and reflection
triggering a chain of breakthroughs and promoting various fields behavior of the human brain. In view of this, researchers have paid
of society from networked and digital to intelligent. The artificial considerable attention to the key link of ANN to realize human
neural network (ANN) is one of the key cornerstones of the brain bionics, namely “bionic synapse”. It is widely known that
development of AI technology, which has once again attracted the synapse of human brain neurons is not only the transmission
widespread attention all over the world. channel of information, but also the basic unit of human brain
learning and storing information [86,91,125].
On the other hand, along with the appearance of deep neural
network (DNN) algorithm, the promotion of parallel computing
R
This work was supported in part by the National Natural Science Foundation of of graphics processor and the emergence of big data, it has been
China under Grants 61773209, 61773017, 61873148 and 61933007, the AHPU Youth found that the traditional CMOS transistor is difficult to meet
Top-notch Talent Support Program, the Natural Science Foundation of Universities in the current requirements of mass data computation due to its
Anhui Province under Grants gxyqZD2019053 and KJ2019A0160, the Natural Science
physical defects such as large size, high energy consumption and
Foundation of Jiangsu Province under Grant BK20190021, and the Six Talent Peaks
Project in Jiangsu Province under Grant XYDXX-033. inability of multiple storage. Moreover, in practical engineering,

Corresponding author. neural networks is always implemented by a circuit where the
E-mail addresses: [email protected] (H. Liu), [email protected] (L. Ma). neuron’s synapse is often realized by a resistor. Nevertheless, the

https://doi.org/10.1016/j.neucom.2020.01.066
0925-2312/© 2020 Elsevier B.V. All rights reserved.

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

2 H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx

Fig. 1. Framework of the survey.

usage of resistors has certain limitations. First, resistor is a volatile analysis and state estimation. The rest of the paper is orga-
device which means that the state information of the neural nized as follows. Section 2 introduces the background of the
circuit would disappear with the disappearance of voltage. Second, memristive neural networks. Section 3 provides the overview of
a neuron often corresponds to multiple synapses, and the large recent progress on the study of MNNs under certain well-studied
size of the resistors inevitably leads to a significant reduction in engineering-oriented complexities. Several frequently utilized de-
the integration density of NNs, which may exceed the acceptable sign techniques are presented in Section 4 when facing synthesis
range in some cases. These defects pose many difficulties for the issues relevant to MNNs. Section 5 gives some latest progress on
realization of “bionic synapse”, and thus, largely hinder the theory the stability analysis and state estimation of MNNs. Section 6 gives
and technology from being applied. The theory of memristor and our conclusion. The framework of the overview is shown in Fig. 1.
its physical realization have brought a new dawn to conquer the
aforementioned bottlenecks during the development of ANN. 2. Background on memristive neural networks
The memristor-based neural network, like biological brain,
has the ability to handle multiple tasks at the same time. Most ANNs have now been widely utilized in image processing [10],
importantly, the memristor-based neural network does not require pattern recognition [88], and dynamic optimization [153] , and
repeated data movement when processing large amounts of data, accordingly, the analysis and synthesis issues of ANNs have also
which is particularly suitable for machine learning systems. There- been research hotspots in the field of neural networks. During the
fore, memristor is a better choice to implement neural networks past few decades, researchers have devoted a large amount of ef-
in comparison to the traditional computer system structure. What fort in the area, resulting in a series of results published in the lit-
is more exciting is that, a neural network chip has been created erature, see, e.g. [48,60,62–64,68,69,93,96,106,116,118,132,164,169].
by Dmitri Strukov and his colleagues from California University It should be pointed out that memristive neural network (MNN)
and New York State University using only memristor [49], which is is a special type of ANNs that is based on the “memristor bionic
an important step towards creating larger-scale neural networks. synapse”, which replaces the traditional resistors with memristors
Note that nowadays, aside from academia, MNNs have also been in the circuit realization process. Nowadays, thanks to the fast
attracting tremendous research attention from the industrial development of computer science, on one hand, the digital tech-
engineering society due to the fact that the memristive neural nology is widely applied in modern industrial control systems;
networks have shown promising application prospects in many On the other hand, for the purpose of obtaining better control
areas of engineering including but not limited to, signal process- performance or simplifying the design procedure, the continuous
ing, pattern recognition, image processing, associative memory, systems or processes are usually discretized firstly before system
combinatorial optimization and control engineering. design.
In this paper, we give a review of recent advances on mem- As for mathematical model of MNNs for continuous or discrete
ristive neural networks with emphasis on the issues of stability cases, it can be described, respectively, by dynamical differential or

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx 3

with strong nonlinearity. Consequently, it is a new challenge to


investigate the state estimation issue for MNNs while taking full
consideration of the impacts on the performance from the net-
work communication capability and system nonlinear dynamics.
In summary, it is of profound significance, from both theoretical
research and engineering practice, to study the stability analysis
as well as state estimation problems of MNNs.

Fig. 2. Physical model of memristor element [121]. 2.1. Memristor and its circuit realization

Memristor, also known as memory resistor, is recognized as the


difference equations whose parameters are state-dependent. Based fourth passive electronic component discovered after resistors, ca-
on these models, the analysis and application issues of MNNs have pacitors and inductors. It was first presented in 1971 by Professor
attracted widespread attention from both academia and industry. Chua from the University of California, Berkeley [15]. The physical
A multitude of research work has been reported in the literature, realization of the memristor, however, has not made a break-
most of which, however, are relevant to the continuous MNNs through due to the limitations of the processing technology at that
and the corresponding discrete case is largely neglected despite time. It was not until 2008 that the physical entity of the first
its stronger engineering insights. This is mainly due to the lack memristor was developed by HP Labs, which consists of a double-
of appropriate paradigms capable of dealing with state-dependent layered TiO2 film [121], as shown in Fig. 2. In 2010, Professor Lv
parameters in the context of discrete memristive neural networks and his colleagues at the University of Michigan filled a memristor
(DMNNs) and the resulting unpredictable nonlinearities. In gen- made of a mixture of Ag + Si into the intersection of two metal
eral, the DMNN has not yet been adequately investigated and electrodes, and proposed the first electrical neural circuit by using
the stability analysis and estimator design problems have rarely synapse of a memristor [43], as shown in Fig. 3. Professor Lv
been discussed, which have recently become very popular in also verified in the literature [43] that the memristor is not only
communities of control science and signal processing. In addition, small in size (nanoscale), low in energy consumption, but also has
due to the fast growth of network communication technology in unique memory characteristics to the internal state of the system.
recent years, more and more systems have been implemented It is the memory characteristics that enable the memristor realize
via networks, where the computer communication network has the mechanism of synaptic plasticity, and make it the most similar
been adopted for the connection between estimators and remote electronic component to the synaptic function of the human brain,
sensors. The signal is then transmitted through the connected thereby paving the way of the research on MNNs. In the era of
communication network, which not only avoids laying a point-to- nano-scale high-density integrated circuits, the memory character-
point dedicated line, but also achieves effective remote operation, istics and learning functions of brain-like synapses of memristors
thereby enhancing the system’s flexibility and reliability. make the neural networks comprised of memristors much closer
Up to now, there have been a large number of results regarding to the human brain, and meanwhile, get rid of weight constraints
the stability of various types of neural networks including MNNs. in the circuit realization of the traditional neural network design.
On the other hand, during recent years, state estimation problem Since the first memristor invented by Hewlett-Packard Lab-
in the context of networked system has become a hotspot in oratories in the United States, the research on memristors and
the area of control theory and engineering, and has shown a memristor-based nonlinear systems has received extensive atten-
promising application prospects within various fields such as tion, especially the theoretical study of MNNs with the background
coordinated control of communication-network-based underwater of circuit systems. Recently, it has become one of the hotspots in
robots and unmanned vehicles, as well as advanced control of the field of ANN and other relevant areas, see [1,8,9,45,92] and
complex industrial systems. the references therein. Nowadays, the work regarding memristor
Notice that the current research on systems in the network en- mainly focuses on two aspects: (1) following HP Lab, researchers
vironment mainly focuses on relatively simple cases. For instance, aim to discover more economical materials; (2) based on results
most results have been obtained with certain strict assumptions provided by Professor Chua, investigation is carried out from the
that the communication ability of network is unconstrained or the aspects of establishment of system model as well as analysis and
target system is linear and time-invariant. Obviously, these results application of dynamical behavior. As for the study of system
are no longer applicable for an MNN with state-dependent param- dynamical behavior, it mainly concentrates on the following three
eters which actually can be described by a time-varying system directions: the first is the analysis of stability and attractiveness

Fig. 3. Memristor mimicking human brain neuron synapses [43].

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

4 H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx

of MNNs; the second is the analysis of periodic solution and been adopted in [5,144] to design the state estimator of continuous
homomorphic orbital behavior; and the third is its chaotic phe- MNNs with stochasticity and time-delay, respectively. In [97], the
nomenon and synchronous control. It should be emphasized that estimator has been proposed for a type of continuous MNNs with
the memristor-based system, especially the MNN with dynamical time-varying delays by resorting to the dissipation theory.
behavior, not only plays a key role in the application in real-world Again, it is worth noting that there have been so far relatively
industrial engineering, but also provides technical support for fewer results on the stability analysis and estimator design for
improving or manufacturing new memristors. As such, it is of DMNNs. For DMNN-related dynamics analysis as well as state
significance to study the stability issue for MNNs and establish a estimation, a challenge is that the nonsmooth analysis methods
unified theoretical framework. and differential inclusion theory that are established for con-
Note that because of the memory characteristics of memristors, tinuous MNNs cannot be directly extended to investigate the
memristor-based neural networks can be generally described as a discrete counterpart. Thus, novel appropriate paradigm should be
dynamical differential equation with state-dependent parameters developed to deal with the analysis as well as design issues for
[145]. Due to the unique characteristic of the memristors, the most DMNNs. However, up to now, the corresponding research has been
distinctive feature of MNNs, in comparison to conventional RNNs, far from adequate, although some limited work has appeared.
lies in the fact that their connection weights are solely reliant on For instance, in [20,71], the authors have first described MNNs
the neuron states, and such a state-dependent behavior would with state-dependent parameters by a state switching model, and
obviously have a big impact on the stability analysis and state then the state estimation issue has been handled by virtue of
estimator design. In literature, MNNs are generally described as Lyapunov theory and robust analysis.
dynamical differential equations with state-dependent parameters
[145], and therefore, the dynamic behaviors of MNNs are much 2.3. Recent progress on several types of neural networks
more complicated than conventional RNNs. Due to the unique
characteristic of the memristors, the most distinctive feature of Before presenting the review of lately advances on DMNNs, in
MNNs, in comparison to conventional RNNs, lies in the fact that this subsection, the recent research on several widely seen classes
their connection weights are solely reliant on the neuron states, of neural networks that are relevant to MNNs will be introduced,
and such a state-dependent behavior would obviously have a big including recurrent neural networks (RNNs), bidirectional associa-
impact on the stability analysis and state estimator design. In tive memory neural networks (BAMNNs) and continuous memris-
literature, MNNs are generally described as dynamical differential tive neural networks (CMNNs), all of which have been playing a
equations with state-dependent parameters [145], and therefore, pivotal role in both theoretical research and industrial engineering.
the dynamic behaviors of MNNs are much more complicated than
conventional RNNs. 2.3.1. RNNs
It is usually assumed in the theoretical analysis that the For decades, it has been generally acknowledged that RNNs
synaptic weight of MNN takes the largest absolute value, and possess many remarkable capabilities (e.g., fault tolerance, self-
the dynamical behavior is then discussed by utilizing different organizing, nonlinear function approximation and self-learning,
approaches and theories such as set-valued mapping, differen- etc.). Consequently, RNN has attracted more and more research at-
tial inclusion, non-smooth analysis, Filipov functional, M-matrix tention and found wide applications in many branches of systems
method, Lyapunov functional, ω-measure and linear matrix in- science and signal processing, such as pattern recognition, image
equality, see, e.g. [140,147,150,157,158]. For instance, in [150], a processing and combinatorial optimization [67,90,137,166], to name
detailed mathematical derivation has been proposed for RNN with but a few. It is noteworthy that the aforementioned applications
memristor delay. Since then, the research work has gradually would depend largely on the dynamic behavior of RNNs, which
started for MNN system. subsequently has simulated the extensive research concerning
dynamics analysis including but are not limited to, stability, state
2.2. Stability analysis and state estimation for MNNs estimation and synchronization. For some representative work, we
refer to [48,62,63,106,116,132,164,169] and the references therein.
So far, most of the existing results concerning MNNs are mainly The stability of RNN is one of the most studied dynamic behav-
on the analysis of dynamical behaviors such as stability, syn- iors, and many analysis paradigms have been proposed, such as in-
chronization or chaos in the continuous case. The corresponding equality analysis approach, Lyapunov functional scheme, M-matrix
investigations have been much fewer when it comes to its discrete technique, topological degree strategy and linear matrix inequality
counterpart. On the other hand, the problem of state estimation (LMI) method [32,35,112,134,146,174]. To be specific, by resorting
has always been popular in the field of complex networks [16,105]. to LMI approach, the asymptotic stability has been analyzed in
Moreover, in recent years, the corresponding state estimation [35,112] for a class of delayed neural networks. Particularly, the
issue for neural networks has stirred tremendous research inter- global stability is undoubtedly the most desirable dynamical prop-
est mainly due to the fact that it is usually impractical, if not erty for RNNs, which plays a pivotal role in practical engineering
impossible, to directly access the states of neurons. and has thus gained tremendous research interest [170].
Consequently, in these occasions, before taking the advantages Moreover, note that in an actual neural network, due to certain
from the implementation of neural networks in classification, ap- probabilistic reasons such as random fluctuations from the release
proximation and optimization, the states of the neurons should be of neurotransmitters, the synaptic transmission would probably
estimated firstly by utilizing the available network measurements, confront with stochastic noises [33]. If not appropriately dealt
see [105,133] for examples. More specifically, in [41], the authors with, such random noises could be one of the main causes for
have considered the state estimation for a class of neural networks the performance deterioration in real-world practice. Therefore,
with both stochastic and uncertain dynamics. In addition, for a stochastic RNNs have started to attract researchers’ attention,
sort of coupled stochastic networks, the data packet dropout has leading to many results published in the literature, see [71] and
been studied in [61] based on which the robust estimator has the references therein.
been designed with guaranteed performance. When it comes to In practical engineering, RNN can be implemented by a circuit.
the estimation issue for MNNs, some preliminary results have As an essential element for computation and information storage,
been obtained in [31,97,102] by utilizing the passivity theory. the neuron’s synapse is often realized by a resistor which, however,
Moreover, the framework of Philippov differential inclusion has has certain limitations. On one hand, resistor is a volatile device

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx 5

which means that the state information of the neural circuit would As such, parallel to the development/deployment of the MNNs,
disappear with the disappearance of voltage. On the other hand, much effort has been devoted to investigating the stability of
a neuron often corresponds to multiple synapses, and the large MNNs by borrowing ideas from systems and control theory. Quite
size of the resistors inevitably leads to a significant reduction in a few elegant results have been obtained from different facets, see
the integration density of RNNs, which may exceed the acceptable [89,135,141,151,152]. For example, in [141], the authors have gained
range in some cases. In view of the above-mentioned problems, a set of sufficient conditions for MNNs subject to impulse effect
it is crucial to develop a new kind of efficient component to to ensure the exponential stability. The Lagrange stability has
reduce the size of circuits while implementing the functions of been discussed in [151] for MNNs whose dynamics are affected by
both processing and information storage, which gives rise to the mixed time-delays and the addressed problem has been solved by
so-called memristors and memristive neural networks. virtue of nonsmooth analysis approach.
It should be mentioned that most of obtained results includ-
2.3.2. BAMNNs ing those aforementioned have been mainly focused on CMNNs.
Late in the 1980s, Kosko first proposed the concept of bidirec- However, in today’s digital world, most information sequences
tional associative memory neural network which was comprised are inherently discrete or should be discretized before imple-
of fully interconnected neurons on two layers [46,47]. During mentation in engineering applications. This motivates the initial
the past 20 years, researchers have devoted special efforts to investigations on the discrete-time neural networks because of
the theoretical investigations of BAMNNs with an emphasis on its powerful capability in handling sequence-based issues [32].
dynamics analysis. Moreover, BAMNNs have so far found wide Unfortunately, despite the remarkable advantages, DMNNs have
and successful applications in various industrial areas such as not been adequately investigated due mainly to the lack of suitable
combination optimization, associative memory, image processing, paradigms in a discrete-time setting to depict and deal with the
pattern recognition and so forth. It should be pointed out that state-dependent behaviors.
the dynamic behaviors (e. g., stability, chaos, synchronization, etc.)
play a paramount role in the practical applications of BAMNNs. As 3. MNNs subject to engineering-oriented complexities
a sequence, several techniques have been developed to deal with
the dynamics behavior analysis of BAMNNs [64,93,96,118] with In real-world applications especially in the networked situa-
special concerns on the analysis of stability. As for the practical tions, certain frequently-occurring engineering-related issues, such
engineering, the exponential stability is always among the most as time-delays, parameter uncertainties, random disturbances and
desirable performances of BAMNNs, which has thus captured incomplete information, have proved to be the main sources of
special interest in the last decades. Specifically, the exponential system instability as well as performance deterioration, and further
stability of neural networks has been revealed to be an indication imposed fundamentally new challenges on the study of various
of the ability to retrieve the stored pattern in many applications types of neural networks. When discussing the stability analysis
such as associative memory [163]. Some latest progress regarding and estimator design problems, these engineering-oriented phe-
the exponential stability for BAMNNs can be found in [53,154]. nomena cannot be neglected. In contrast, they must be taken
into simultaneous consideration with the neural networks dy-
2.3.3. CMNNs namics under a unified framework so as to achieve a satisfactory
It is now well known that in neural networks, synapses are the level of performance. In this section, several extensively studied
most essential units used for computing and storing information. engineering-oriented phenomena that would make the dynamics
They possess the capability to record the dynamic history, store of neural networks more complex will be introduced with certain
a series of continuous states, and most importantly, they should recently developed techniques for analysis and synthesis.
be “plastic” according to the activity of synaptic neurons. Unfor-
tunately, by a resistor in traditional recurrent and other sorts of 3.1. Stochasticity
neural networks, these functions cannot be realized, which gives
rise to the so-called MNNs. MNNs are developed by replacing It is well known that during the last decades, stochastic sys-
resistors with memristors and have the ability to conquer the tems have exerted a tremendous fascination on many researchers
aforementioned difficulties. Furthermore, as shown in many appli- from both academia and industry because of the unavoidable
cations [92], the proposed MNNs have exhibited other advantages random phenomena frequently seen in practical engineering.
over the traditional neural networks such as higher efficiency. Apparently, in the modeling of this kind of systems, the routine
Nevertheless, in spite of these merits, the memristors and the way that neglects the uncertain stochastic effects to obtain a
resulting MNNs have not received enough attention for nearly four conventional deterministic model for analysis and synthesis is no
decades ever since the concept was proposed. longer appropriate. It is now widely acknowledged that a host
The key feature of MNNs is that the connection weights cou- of complex practical systems such as engineering systems, social
pling the neurons are heavily dependent on their states. As a systems, as well as many different types of neural-network-based
consequence, MNNs show much richer dynamic behaviors than systems are more suitable to be described by stochastic differ-
those of conventional RNNs. In recent years, the dynamics anal- ential/difference equations [66,80,81]. Consequently, the stability
ysis issues for MNNs (such as stability, synchronization and so analysis and state estimation issues for dynamical systems subject
forth), especially the continuous case, have received considerable to random fluctuations or uncertain disturbances have garnered
attention and plenty of research results have been published in extensive research interest, see [3] for reference.
the literature discussing various sorts of MNNs such as memris- As for the application of a neural network, the signal transmis-
tive recurrent neural networks [135], fractional-order MNNs [11], sion among the neurons of an actual MNN always confront with
cellular MNNs [25], memristive Hopfield networks [149], chaotic random interferences or noisy fluctuations stemming from the
MNNs [165] and memristive complex-value neural networks [142], release of neurotransmitter or other probabilistic perturbations.
to name but a few. These kinds of probabilistic disturbances make the dynamics
Note that MNNs actually can be mathematically modeled by of the rustling MNNs uncertain and random, which is usually
utilizing a series of coupled differential/difference equations and referred to as stochastic MNNs. As mentioned previously, the
can be characterized by the state-dependent nonlinear systems framework established based on inclusion theory and nonsmooth
[146] whose dynamic behaviors are rather rich and complicated. analysis applicable for CMNNs cannot be directly utilized to deal

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

6 H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx

with their discrete counterparts. If not taken into account, these countered in engineering practice, such as missing measurements,
random effects would become the main sources of degradation of channel fading, and signal quantization.
system performance. As such, much effort has been dedicated to
exploring the stochastic MNNs and to date, the stability analysis 3.3.1. Missing measurements
has aroused some preliminary research attention. A few scattered The measurements missing may be caused by many factors
pioneering results have appeared, see, e.g. [55,70,102] and the such as network congestion, instantaneous sensor failures, and
references therein. limited channel bandwidth in real-world practical engineering. In
this case, the measurement data of the system will be lost or par-
tially lost. In recent years, researchers have been working on the
3.2. Time-delays
state estimation problems subject to measurement missing, see,
e.g. [21,36,39,87,99,103,117,138,143,148], among which the literature
Time-delay is now widely recognized as an inherent charac-
[117] have introduced a stochastic variable that obeys Bernoulli
teristic of neural networks due to physical hardware limitations,
distribution to describe the measurements missing. In [115], such
which might bring about certain complicated dynamic behaviors
a phenomenon has been expressed by a Markovian jump process.
(e. g., oscillation, divergence, etc.) and sometimes even leads to
A more general approach that is capable of characterizing both
instability [94,95,108,109]. Based on the occurrence mode, time-
totally and partially missing measurements has been developed in
delays can be generally classified into two types, namely, discrete
[143] where a random variable is utilized that obeys an arbitrary
time-delay and distributed time-delay [13,14,65,79]. Note that real-
distribution with known probability. This approach is able to
world engineering systems are often confronted with simultaneous
cope with situations in the context of multiple sensors where the
presence of different types of time-delays [40,62,78,82]. Therefore,
missing probability of different sensors is different, and a diagonal
ever since the introduction in [84], researchers and engineers have
matrix can be constructed whose diagonal entries indicate the
poured attention to the delayed neural networks and developed
missing probabilities of the corresponding sensors [143]. The con-
quite a few schemes for stability analysis issues, such as integral
secutive data packet dropout case has been discussed in [99] by
inequality approach [48], matrix inequality method [120,173] and
resorting to the Bernoulli distributed random parameters. Based on
descriptor model transformation technique [27].
these established models, much work has then been done to the
Along with the research on stability, the state estimation has
study of state estimation problem subject to different performance
gradually become a hotspot in both theoretical research and
specifications. For instance, in [122], a class of optimal filters has
practical engineering for delayed neural networks, and some work
been designed for the system by means of the Riccati equation
have been done for traditional RNNs [61,133] and MNNs [71]. For
approach when multiple consecutive measurements missing oc-
CMNNs, time-delays such as constant delays, time-varying delays
curs. With the help of the linear matrix inequality technique, the
and distributed delays have been adopted to characterize the net-
H2 filtering problem has been solved in [100] for systems with
work latency during the data transmissions due to certain physical
multiple consecutive packet losses, whereas the corresponding H∞
constraints. The impacts imposed on the dynamical behaviors by
filtering problem has been examined in [101].
the time-delays have been extensively studied [30,151], among
which the characteristics of global exponential synchronization
3.3.2. Channel fading
for a class of MNNs subject to multiple time-delays has been
In the context of digital data transmission via networks, the
examined in [30] by means of LMI approach.
phenomenon of channel fading is inevitably occurring. This is
Recently, various kinds of time-delays have been taken into
mainly because the communication channel has certain physical
account when dealing with estimation issues for different neural
constraints such as limited bandwidth. Moreover, such fading
networks. The time-varying delay has been considered in [42] for
measurements always happen in a probabilistic way due to the
static neural networks, while in [44], the addressed discrete
random variation in the network environment [37]. For recurrent
neural networks have been assumed to be affected by the mixed
neural networks, it is often the case that the measurement out-
time-delays. In [4], the random delays have been examined for
put suffers signal fading when being transmitted via unreliable
discrete stochastic neural networks, where the problem is solved
communication network channels for further processing [17,18,22].
by using a delay-distribution-dependent method. Employing Lya-
Although there exist some initial publications concerning the state
punov functional in combination with LMI approach, time-varying
estimation for RNNs by using fading measurements (e.g. [35,112]),
delays have been discussed in [133] for the estimator design. It
the relevant results on DMNNs have been much fewer due
should be emphasized that, two time-delays usually occurring in
probably to the level of mathematical difficulties/complexities.
neural networks (namely, leakage and probabilistic delays) start to
stir particular interest yet still remain unexplored, see [12,162] and
3.3.3. Signal quantization
the references therein. Moreover, an army of work has been done
Due to the limited capacity of transmission channel, systems
regarding RNNs with probabilistic time-varying delays, see, for
communicating via networks always confront with signal quan-
instance, [4,112,119,162]. Nevertheless, the corresponding study
tization of the transmitted data [38,83,110,167,172]. It should be
of DMNNs has been relatively scattered despite certain limited
emphasized that in real-world applications especially in the con-
results such as [71], and there still exists a huge gap towards
text of networked systems, the signal quantization has proved to
the investigation on the estimator design for different classes of
be one of the main sources of system instability as well as per-
DMNNs, let alone the case where various engineering-oriented
formance deterioration, and further imposed fundamentally new
complexities like multiple time-delays are also involved.
challenges on the study of networked systems. When discussing
the analysis, design and application problems for neural networks,
3.3. Network-induced incomplete information such network-induced quantization effect cannot be neglected. As
a result, the relevant research has recently been garnering research
In the network environment, the stability analysis and state interest, see, e.g. [98,111] for some recent results, among which
estimation often suffer from the phenomenon of network-induced two main techniques have been utilized to deal with the quantiza-
incomplete information, which is would lead to performance tion effects. One is to convert the quantization effect into parame-
deterioration. In this subsection, we shall discuss several types of ter uncertainties whereas the other is to transfer the quantization
network-induced incomplete information that are extensively en- function to nonlinearities constrained within known sector bound.

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx 7

4. Design techniques would probably lead to network congestions as well as data colli-
sions in the situations of simultaneous multiple accesses. Accord-
The complexity of neural networks dynamics makes it ex- ingly, the protocols for communication are adopted to regulate the
tremely challenging to establish systematic framework for estima- data transmission sequence of different nodes [6,155,168]. Some
tor design with desired performances and required specifications. of the extensively used communication protocols in industrial
Apparently, it is impractical, if not impossible, to develop one par- applications include the Round-Robin (RR) protocol [28,127], the
ticular procedure capable of tackling all kinds of neural networks; Weighted Try-Once-Discard (WTOD) protocol [23,128] and stochas-
therefore, a host of tools and mechanisms have been proposed tic communication (SC) protocol [113,123], to name just a few.
to deal with stability analysis and estimator design for specific To date, although some pioneering results have been reported
types of neural networks. In the sequel, we will investigate several regarding the state estimation for MNNs, see e.g. [70,71], little ef-
techniques that have been widely applied in estimator design of fort has been devoted to the constrained communication between
neural networks to meet different design demands. the estimator and the network output. In fact, due to the large size
of the MNN and the complexity of the tasks to be accomplished,
4.1. Event-triggering mechanisms the amount of the measurement data of the network output tends
to be huge, and this has placed a great deal of burden on the
In most of the existing literature, it is always assumed that transmission networks of limited capacity [76]. Due primarily
the state estimator is time-triggered rather than event-triggered. to the limited bandwidth of the network, data collision may
However, in the network environment, it is more attractive to de- probably happen. For the purpose of avoiding this phenomenon
sign the state estimator with an event-triggering strategy for the and coordinating the network traffic, an effective approach widely
purpose of saving network resources, thereby reducing the heavy employed in the practical engineering is to employ communication
communication burden [85]. In addition, the event-triggering protocols like those aforementioned ones, whose mechanisms will
scheme can largely reduce the activating frequency of electronic be introduced as follows.
components in the digital control circuit, and help extend the
working life of the circuit. It has proved that the event-triggered 4.2.1. RR protocol
scheduling scheme is more suitable for state estimation of MNNs As a periodic scheduling scheme, the so-called RR protocol has
in the network environment. In view of this, particular inter- been widely used in communication, load balancing and server
est has been poured to state estimation with event-triggering resource allocation [76]. The RR protocol is also named as the
mechanism, and quite a few results have been published, see, time-division multiple access (TDMA) protocol or Token Ring
e.g. [26,50,51,58,69,104,114,124,139,161,171]. For example, the protocol. Under the regulation of the RR protocol, the allowable
event-triggering approach has been used in [139] for estimator transmission steps of all the nodes are determined by applying a
design of delayed RNN. fixed circular order. Up to now, many researchers have introduced
It should be pointed out that the most significant feature of the RR protocol to govern the data transmission when dealing
the event-triggering mechanism is that the input of the desired the network-based communication problems and designing the
estimator (i.e., the measurement output) will not be updated remote estimator with an equal allocation scheme of resource,
until a specific event occurs, and the transmission signal between thereby avoiding data congestion and saving communication
consecutive updating interval remains unchanged. However, the energy [57,107,129,156,177]. Recently, the RR protocol has been ex-
desired update instant is unknown and should be determined tensively utilized in the estimation problems for neural networks
according to the pre-specified triggering condition which usually such as the ANNs in [19,76], Markovian jump neural networks in
depends on time, system state and sampled data [26]. The de- [54] and MNN in [159].
sign of triggering condition plays a pivotal role in the estimator
synthesis and the subsequent analysis of the resulting impacts on 4.2.2. WTOD protocol
stability and other performances. The so-called Weighted Try-Once-Discard protocol is catego-
It is noteworthy that most of the event-triggering mechanisms rized to the sort of quadratic protocols. By applying the WTOD
employed in the existing literature are static, that is, the threshold protocol, the allowable signal transmission steps to certain nodes
parameters in the triggering condition are fixed [70,126]. Recently, are assigned by some predetermined quadratic selection principles,
in order to further achieve the target of energy conservation, which differs from the “periodic assignment” behavior of the RR
the dynamic event-triggering mechanisms have been applied protocol. WTOD protocol has been widely applied to govern the
to many systems [29,59] by adjusting the threshold parameters data propagation over the channels between the neural networks
dynamically at each checking instant. Intuitively, using a dynamic and its remote state estimator. Note that the WTOD protocol
event-triggering mechanism would decrease the total number of is a classical dynamic regulating scheme in associated with the
released events than using a static one. From this point of view, we importance of different missions, and obviously, it is more suitable
can conclude that the employment of the static event-triggering for tasks such as resource allocation than those static scheduling
mechanisms possibly results in unnecessary data transmissions schemes, see e.g. [23,127,176]. As such, it is necessary to study
and exhibits a conservative solution. It is worth mentioning that how the WTOD protocol can be employed to schedule the pos-
the static type of event-triggered mechanism has not been applied sibly huge data communication between the MNN and its state
extensively in state estimation for MNNs, let alone the case of estimator, which still remains open and challenging.
dynamic triggering schemes.
4.2.3. SC protocol
4.2. Network communication protocols The so-called SC protocol has been first investigated in
[123] for continuous-time systems and [24] for discrete-time ones.
In most existing results relevant to the state estimation issues As a matter of fact, SC protocol is classified to the CSMA/CA
of systems in the network environment, an assumption has usually protocols, which should be implemented according to the rule of
been made that all the network nodes could simultaneously be “listen before talk” or “sense before transmit”. It is worth noting
allowed to transmit signals via communication channel. It is, that the CSMA/CA protocol may be considered and applied as a
however, basically impractical for network-based systems since modified version of the CSMA protocol. In practical application,
real-world networks will inevitably face limited bandwidth which such a protocol is able to prevent the network congestion and data

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

8 H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx

collision by means of employing the acknowledgments signal from Lyapunov functional approach, sufficient conditions have been
the communication network to mark each transmission. Up to gained, guaranteeing the exponential synchronization of the ad-
now, SC protocol has started to attract interest in both theoretical dressed DMNNs despite the existence of time-delays. It should
research and engineering application, see, e.g. [2,130,175] for some be pointed out that the conservatism of the obtained results
latest publications. has been reduced due to the adoption of delay partitioning
method as well as the free-weighting matrix technique.
4.3. Set-membership technique • An H∞ state estimator has been designed in [136] for discrete-
time stochastic memristive BAMNNs with mixed time-delays.
In reality, the MNN can be realized by very large scale integra- Both discrete and distributed time-delays have been taken
tion (VLSI) circuits, and the connection weights are implemented into account. A set of new switching functions have been
by the memristors [146]. Unfortunately, it is often the case that designed to account for the state-dependent characteristics of
the VLSI circuits are subject to both the device noises (e.g., ther- the memristive connection weights in the discrete-time setting,
mal noise, shot noise, flicker noise) and the external noises (e.g., which facilitates the dynamics analysis of the addressed MNNs.
substrate noise, power/ground bounce, crosstalk). In particular, • For the DMNNs subject to both time-delays and fading mea-
these external noises, if not appropriately handled, might largely surements, a delay-distribution-dependent scheme has been
deteriorate the performance and the reliability of integrated cir- developed in [74] to deal with the H∞ state estimation prob-
cuits [34]. From the practical point of view, in many cases, these lems. A state estimation algorithm capable of achieving both
external noises are not stochastic but deterministic, unknown yet stochastic stability and pre-specified noise attenuation level
bounded within certain sets. As such, the set-membership state has been proposed by applying switching function technique
estimation algorithm is introduced to confine the state estimate in combination with LMI approach.
in certain desired regions centered at the true value of state of • For the purpose of energy saving, an event-triggering mech-
targeted neuron. As one of the most popular state estimation anism has been implemented in [70] to handle the H∞ state
approaches, the set-membership state estimation issue has been estimation issue for delayed MNNs subject to missing mea-
well investigated, and a rich body of literature has been available surements. An event-triggered H∞ estimator has been designed
(see e.g. [7,56,77,176]). Nevertheless, the corresponding results for with both exponential mean square stability and H∞ crite-
neural networks have been much limited, see e.g. [159], and the rion, by means of Lyapunov–Krasovskii functional method
study for MNNs of set-membership state estimation has been far and stochastic analysis technique, and the desired estimator
from adequate at the moment. gains have been characterized as the solutions to certain
easily-solved matrix inequalities.
4.4. Non-fragile algorithm • In [75], two different types of time-delays have been taken
into consideration for the stability analysis of DMNNs. Both
In the course of estimator execution, due to various reasons probabilistic delays and leakage delays are taken into account.
(e.g., digital rounding errors, finite computer word length, and By considering the state-dependent characteristics of the net-
inaccuracy of analog to digital conversion), the actual execution work parameters, the sufficient conditions have been obtained
parameters of the estimator often slightly differs from the ex- for the globally exponentially mean-square stability of the
pected ones. It is generally acknowledged that small even tiny addressed DMNNs via virtue of Lyapunov theory.
drifts/variations of the estimator parameters would result in great • A framework in the H∞ sense has been established in [71] to
damage to the overall system dynamics [52,131,160]. Accordingly, solve state estimation for DMNNs subject to probabilistic de-
over the past several years, the so-called non-fragile estimation lays. The time-delays under consideration are characterized by
problem has become a research focus that aims to construct an a Bernoulli-distributed random sequence. A robust estimation
effective estimator which is insensitive to the admissible estimator algorithm, in terms of LMI, has been developed with guarantee
execution errors. For example, in [52], a non-fragile finite-time of both mean square stability and required noise attenuation
l2 − l∞ state estimator has been designed for a type of Markov level.
jump neural networks with unreliable communication links. Nev-
ertheless, in contrast to the rich literature regarding traditional 6. Conclusions and future work
RNNs, the available results on the non-fragile state estimation
problems for MNNs has been very few, not to mention MNNs with In this paper, the stability analysis and state estimation prob-
various engineering-oriented complexities. lems for MNNs have been reviewed with some recent advances for
discrete time-delay cases. Latest results on stability analysis and
5. Latest progress estimator design issues for different types of MNNs subject to var-
ious engineering-oriented complexities especially network-induced
Very recently, the stability analysis and state estimation for phenomena have been surveyed. It can be concluded that, so far,
DMNNs have been receiving ever-increasing research interest most existing literature regarding the stability analysis problem
and some elegant results have been reported. In this section, we for MNNs have focused on CMNNs, while the corresponding re-
summarize some of the newest work with respect to this topic, sults for discrete-time case have been much fewer. Unfortunately,
most of which are the authors’ recent work. as is well known, in nowadays digital world, almost all signals
are digitalized before being transmitted for further processing.
• In [72], the exponentially global stability has been discussed for Therefore, discrete signals are more practical and appealing than
the DMNNs with time-varying delay. For the neural networks continuous-time ones in simulating and applications of MNNs.
under consideration, the activation functions are assumed to be Despite its clear engineering significance, the relevant research on
slope-bounded. By employing the theories of set-valued map- DMNNs has been far from adequate, in spite of certain limited
ping, difference inclusions as well as homeomorphic mapping, initial work reported in the literature.
the existence of the equilibrium point of the discrete-time Furthermore, based on the literature review, some related areas
MRNNs has been examined in terms of a set of LMIs. for potential future research work are mentioned as follows.
• For the DMNNs with time-varying delay, in [73], the syn- In practical engineering, there are still some more complicated
chronization problem has been solved. By virtue of piecewise yet important kinds of network-induced phenomena (e. g., data

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx 9

disorder, sensor saturations, etc.) that have not been studied in [17] D. Ding, Z. Wang, B. Shen, H. Dong, Envelope-constrained H∞ filtering with
a unified framework of neural networks. Therefore, the stability fading measurements and randomly occurring nonlinearities: the finite hori-
zon case, Automatica 55 (2015) 37–45.
analysis and state estimation problems for MNNs taking into [18] D. Ding, Z. Wang, B. Shen, H. Dong, H∞ state estimation with fading measure-
account more engineering-oriented complexities still remain open ments, randomly varying nonlinearities and probabilistic distributed delays,
and challenging. Moreover, it is necessary to consider certain other Int. J. Robust Nonlinear Control 25 (2015) 2180–2195.
[19] D. Ding, Z. Wang, Q. Han, G. Wei, Neural-network-based output-feedback con-
performance indices than H∞ specification in relevant studies. trol under round-robin scheduling protocols, IEEE Trans. Cybern. 49 (6) (2019)
These performances include, but are not limited to, accuracy, 2372–2384.
robustness, reliability that can be simultaneously considered in a [20] S. Ding, Z. Wang, J. Wang, H. Zhang, H∞ state estimation for memristive neu-
ral networks with time-varying delays: The discrete-time case, Neural Netw.
unified framework. Also, the state estimation problems could be
84 (2016) 47–56.
considered for DMNNs subject to many other types of delays such [21] H. Dong, Z. Wang, H. Gao, Robust H∞ filtering for a class of nonlinear net-
as neutral delay. worked systems with multiple stochastic communication delays and packet
dropouts, IEEE Trans. Signal Process. 58 (4) (2010) 1957–1966.
Another future research direction is to further study the prac-
[22] H. Dong, Z. Wang, S.X. Ding, H. Gao, Event-based filter design for a class
tical application of the DMNNs in many branches of science and of nonlinear time-varying systems with fading channels and multiplicative
industrial engineering due to the remarkable merits and wide im- noises, IEEE Trans. Signal Process. 63 (13) (2015) 3387–3395.
plementation in image processing/recognition, machine learning, [23] M.C.F. Donkers, W.P.M.H. Heemels, N. van de Wouw, L. Hetel, Stability analysis
of networked control systems using a switched linear systems approach, IEEE
reinforcement learning and so forth. Trans. Autom. Control 56 (9) (2011) 2101–2115.
[24] M.C.F. Donkers, W.P.M.H. Heemels, D. Bernardini, A. Bemporad, V. Shneer, Sta-
Declaration of Competing Interest bility analysis of stochastic networked control systems, Automatica 48 (4)
(2012) 917–925.
[25] S. Duan, X. Hu, Z. Dong, L. Wang, P. Mazumder, Memristor-based cellular non-
The authors declare that they do not have any financial or linear/neural network: design, analysis, and applications, IEEE Trans. Neural
nonfinancial conflict of interests. Netw. Learn. Syst. Syst. 26 (6) (2015) 1202–1213.
[26] Y. Fan, G. Feng, Y. Wang, C. Song, Distributed event-triggered control of multi-
-agent systems with combinational measurements, Automatica 49 (2) (2013)
CRediT authorship contribution statement
671–675.
[27] E. Fridman, New Lyapunov-Krasovskii functionals for stability of linear re-
Hongjian Liu: Conceptualization, Writing - original draft, Writ- tarded and neutral type systems, Syst. Control Lett. 43 (4) (2001) 309–319.
ing - review & editing. Lifeng Ma: Writing - original draft. Zidong [28] M. Gao, S. Yang, L. Sheng, D. Zhou, Fault diagnosis for time-varying systems
with multiplicative noises over sensor networks subject to round-robin pro-
Wang: Methodology, Supervision. Yurong Liu: Writing - review & tocol, Neurocomputing 346 (2019) 65–72.
editing. Fuad E. Alsaadi: Writing - review & editing. [29] A. Girard, Dynamic triggering mechanisms for event-triggered control, IEEE
Trans. Autom. Control 60 (7) (2015) 1992–1997.
References [30] Z. Guo, S. Yang, J. Wang, Global exponential synchronization of multiple
memristive neural networks with time delay via nonlinear coupling, IEEE
[1] A. Afifi, A. Ayatollahi, F. Raissi, Efficient hybrid CMOS-nano circuit design Trans. Neural Netw. Learn. Syst. Syst. 26 (6) (2015) 1300–1311.
for spiking neurons and memristive synapses with STDP, in: Proceedings of [31] J. Guo, Z. Meng, Z. Xiang, Passivity analysis of stochastic memristor-based
the IEICE Transactions on Fundamentals of Electronics, Communications and complex-valued recurrent neural networks with mixed time-varying delays,
Computer Sciences, E93A, 2010, pp. 1670–1677. Neural Process. Lett. 47 (3) (2018) 1097–1113.
[2] F.E. Alsaadi, Y. Luo, Y. Liu, Z. Wang, State estimation for delayed neural net- [32] Q.-L. Han, Y. Liu, F. Yang, Optimal communication network-based H∞ quan-
works with stochastic communication protocol: the finite-time case, Neuro- tized control with packet dropouts for a class of discrete-time neural net-
computing 281 (2018) 86–95. works with distributed time delay, IEEE Trans. Neural Netw. Learn. Syst. Syst.
[3] K. Åström, Introduction to Stochastic Control Theory, Acadamic Press, New 27 (2) (2016) 426–434.
York and London, 1970. [33] S. Haykin, Neural Networks, Englewood Cliffs, USA: Prentice-Hall, 1994.
[4] H. Bao, J. Cao, Delay-distribution-dependent state estimation for discrete-time [34] P. Heydari, M. Pedram, Capacitive coupling noise in high-speed VLSI cir-
stochastic neural networks with random delay, Neural Netw. 24 (1) (2011) cuits, IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst. 24 (3) (2005) 478–
19–28. 488.
[5] H. Bao, J. Cao, J. Kurths, A. Alsaedi, H∞ state estimation of stochastic memris- [35] N. Hou, H. Dong, Z. Wang, W. Ren, F.E. Alsaadi, H∞ state estimation for
tor-based neural networks with time-varying delays, Neural Netw. 99 (2018) discrete-time neural networks with distributed delays and randomly oc-
79–91. curring uncertainties through fading channels, Neural Netw. 89 (2017) 61–
[6] N.W. Bauer, M.C.F. Donkers, N. van de Wouw, W.P.M.H. Heemels, Decentral- 73.
ized observer-based control via networked communication, Automatica 49 (7) [36] F.O. Hounkpevi, E.E. Yaz, Robust minimum variance linear state estimators
(2013) 2074–2086. for multiple sensors with different failure rates, Automatica 43 (7) (2007)
[7] A. Benavolia, D. Piga, A probabilistic interpretation of set-membership filter- 1274–1280.
ing: application to polynomial systems through polytopic bounding, Automat- [37] J. Hu, Z. Wang, D. Chen, F.E. Alsaadi, Estimation, filtering and fusion for
ica 70 (2016) 158–172. networked systems with network-induced phenomena: new progress and
[8] K. Cantley, A. Subramaniam, H. Stiegler, R. Chapman, E. Vogel, Hebbian learn- prospects, Inf. Fusion 31 (2016) 65–75.
ing in spiking neural networks with nanocrystalline silicon TFTs and memris- [38] J. Hu, Z. Wang, G.-P. Liu, H. Zhang, Variance-constrained recursive state es-
tive synapses, IEEE Trans. Nanotechnol. 10 (5) (2011) 1066–1073. timation for time-varying complex networks with quantized measurements
[9] K. Cantley, A. Subramaniam, H. Stiegler, R. Chapman, E. Vogel, Neural learn- and uncertain inner coupling, IEEE Trans. Neural Netw. Learn. Syst. Syst.
ing circuits utilizing nano-crystalline silicon transistors and memristors, IEEE (2019), doi:10.1109/TNNLS.2019.2927554.
Trans. Neural Netw. Learn. Syst. Syst. 23 (4) (2012) 565–573. [39] J. Hu, Z. Wang, S. Liu, H. Gao, A variance-constrained approach to recursive
[10] H. Chen, Y. Hung, C. Chen, T. Li, C. Chen, Image-processing algorithms realized state estimation for time-varying complex networks with missing measure-
by discrete-time cellular neural networks and their circuit implementations, ments, Automatica 64 (2016) 155–162.
Chaos Solitons Fractals 29 (5) (2006) 1100–1108. [40] J. Hu, H. Zhang, X. Yu, H. Liu, D. Chen, Design of sliding-mode-based control
[11] J. Chen, Z. Zeng, P. Jiang, Global Mittag-Leffler stability and synchronization for nonlinear systems with mixed-delays and packet losses under uncertain
of memristor-based fractional-order neural networks, Neural Netw. 51 (2014) missing probability, IEEE Trans. Syst. Man Cybern. Syst. (2019), doi:10.1109/
1–8. TSMC.2019.2919513.
[12] Y. Chen, Z. Fu, Y. Liu, F.E. Alsaadi, Further results on passivity analysis of [41] H. Huang, G. Feng, J. Cao, Robust state estimation for uncertain neural
delayed neural networks with leakage delay, Neurocomputing 224 (2017) networks with time-varying delay, IEEE Trans. Neural Netw. 19 (8) (2008)
135–141. 1329–1339.
[13] Y. Chen, S. Fei, Y. Li, Robust stabilization for uncertain saturated time-delay [42] H. Huang, G. Feng, J. Cao, State estimation for static neural networks with
systems: a distributed-delay-dependent polytopic approach, IEEE Trans. Au- time-varying delay, Neural Netw. 23 (2010) 1202–1207.
tom. Control 62 (7) (2017) 3455–3460. [43] S. Jo, T. Chang, I. Ebong, B. Bhadviya, P. Mazumder, W. Lu, Nanoscale mem-
[14] Y. Chen, Z. Wang, S. Fei, Q.L. Han, Regional stabilization for discrete time-de- ristor device as synapse in meuromorphic systems, Nano Lett. 10 (4) (2010)
lay systems with actuator saturations via a delay-dependent polytopic ap- 1297–1301.
proach, IEEE Trans. Autom. Control 64 (3) (2019) 1257–1264. [44] X. Kan, H. Shu, Z. Li, Robust state estimation for discrete-time neural net-
[15] L. Chua, Memristor-the missing circuit element, IEEE Trans. Circuit Theory 18 works with mixed time-delays,linear fractional uncertainties and successive
(5) (1971) 507–519. packet dropouts, Neurocomputing 135 (2014) 130–138.
[16] D. Ding, Z. Wang, H. Dong, H. Shu, Distributed H∞ state estimation [45] H. Kim, M. Sah, C. Yang, T. Roska, L. Chua, Neural synaptic weighting with a
with stochastic parameters and nonlinearities through sensor networks: the pulse-based memristor circuit, IEEE Trans. Circuits Syst. I: Regul. Pap. 59 (1)
finite-horizon case, Automatica 48 (8) (2012) 1575–1585. (2012) 148–158.

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

10 H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx

[46] B. Kosko, Adaptive bidirectional associative memories, Appl. Opt. 26 (23) tic memristive neural networks with both leakage and probabilistic delays,
(1987). 4947–4860. Neural Netw. 102 (2018) 1–9.
[47] B. Kosko, Bidirectional associative memories, IEEE Trans. Syst. Man Cybern. 18 [76] Y. Luo, Z. Wang, G. Wei, F.E. Alsaadi, T. Hayat, State estimation for a class of
(1) (1988) 49–60. artificial neural networks with stochastically corrupted measurements under
[48] O.M. Kwon, M.J. Park, S.M. Lee, J.H. Park, E.J. Cha, Stability for neural networks round-robin protocol, Neural Netw. 77 (2016) 70–79.
with time-varying delays via some new approaches, IEEE Trans. Neural Netw. [77] L. Ma, Z. Wang, H.-K. Lam, N. Kyriakoulis, Distributed event-based set-mem-
Learn. Syst. Syst. 24 (2) (2013) 181–193. bership filtering for a class of nonlinear systems with sensor saturations over
[49] R. Legenstein, Nanoscale connections for brain-like circuits, Nature 521 (7550) sensor networks, IEEE Trans. Cybern. 47 (11) (2017) 3772–3783.
(2015) 37–38. [78] L. Ma, Z. Wang, Y. Liu, F.E. Alsaadi, A note on guaranteed cost control for
[50] B. Li, Z. Wang, L. Ma, An event-triggered pinning control approach to synchro- nonlinear stochastic systems with input saturation and mixed time-delays,
nization of discrete-time stochastic complex dynamical networks, IEEE Trans. Int. J. Robust Nonlinear Control 27 (18) (2017) 4443–4456.
Neural Netw. Learn. Syst. Syst. 29 (12) (2018) 5812–5822. [79] L. Ma, Z. Wang, Y. Liu, F.E. Alsaadi, Distributed filtering for nonlinear
[51] B. Li, Z. Wang, L. Ma, H. Liu, Observer-based event-triggered control for non- time-delay systems over sensor networks subject to multiplicative link noises
linear systems with mixed delays and disturbances: the input-to-state stabil- and switching topology£, Int. J. Robust Nonlinear Control 29 (10) (2019)
ity, IEEE Trans. Cybern. 49 (7) (2019) 2806–2819. 2941–2959.
[52] F. Li, H. Shen, M. Chen, Q. Kong, Non-fragile finite-time l2 − l∞ state estima- [80] L. Ma, Z. Wang, Q.-L. Han, Y. Liu, Consensus control of stochastic multi-agent
tion for discrete-time Markov jump neural networks with unreliable commu- systems: a survey, Sci. China Inf. Sci. 60 (12) (2017). Art. No. 120201. doi:
nication links, Appl. Math. Comput. 271 (2015) 467–481. 10.1007/s11432- 017- 9169- 4.
[53] H. Li, H. Jiang, C. Hu, Existence and global exponential stability of periodic [81] L. Ma, Z. Wang, H.K. Lam, Mean-square H∞ consensus control for a class of
solution of memristor-based BAM neural networks with time-varying delays, nonlinear time-varying stochastic multiagent systems: the finite-horizon case,
Neural Netw. 75 (2016) 97–109. IEEE Trans. Syst. Man Cybern. Syst. 47 (7) (2017) 1050–1060.
[54] J. Li, H. Dong, Z. Wang, W. Zhang, Protocol-based state estimation for delayed [82] L. Ma, Z. Wang, Q.-L. Han, Y. Liu, Dissipative control for nonlinear Markovian
Markovian jumping neural networks, Neural Netw. 08 (2018) 355–364. jump systems with actuator failures and mixed time-delays, Automatica 98
[55] R. Li, J. Cao, A. Alsaedi, T. Hayat, Non-fragile state observation for delayed (2018) 358–362.
memristive neural networks: continuous-time case and discrete-time case, [83] L. Ma, Z. Wang, Q.-L. Han, H.K. Lam, Envelope-constrained H∞ filtering for
Neurocomputing 245 (2017) 102–113. nonlinear systems with quantization effects: the finite horizon case, Auto-
[56] Q. Li, B. Shen, Z. Wang, F.E. Alsaadi, A sampled-data approach to distributed matica 93 (2018) 527–534.
H∞ resilient state estimation for a class of nonlinear time-delay systems over [84] C.M. Marcus, R.M. Westervelt, Stability of analog neural networks with delay,
sensor networks, J. Frankl. Inst. 354 (2017) 7137–7157. Phys. Rev. A 39 (1989) 347–359.
[57] J. Li, H. Dong, Z. Wang, W. Zhang, Protocol-based state estimation for delayed [85] X. Meng, T. Chen, Event based agreement protocols for multi-agent networks,
Markovian jumping neural networks, Neural Netw. 108 (2018) 355–364. Automatica 49 (7) (2013) 2125–2132.
[58] Q. Li, B. Shen, Z. Wang, F.E. Alsaadi, An event-triggered approach to dis- [86] T. Mitchell, Machine Learning, The MeGraw-Hill Companies, New York, USA,
tributed H∞ state estimation for state-saturated systems with randomly oc- 1997.
curring mixed delays, J. Frankl. Inst. 355 (6) (2018) 3104–3121. [87] M. Moayedi, Y.K. Foo, Y.C. Soh, Adaptive Kalman filtering in networked sys-
[59] Q. Li, B. Shen, Z. Wang, T. Huang, J. Luo, Synchronization control for a class of tems with random sensor delays, multiple packet dropouts and missing mea-
discrete time-delay complex dynamical networks: a dynamic event-triggered surements, IEEE Trans. Signal Process. 58 (3) (2010) 1577–1588.
approach, IEEE Trans. Cybern. 49 (5) (2019) 1979–1986. [88] N. Nasrabadi, W. Li, Object recognition by a Hopfield neural network, IEEE
[60] Q. Li, B. Shen, Z. Wang, F.E. Alsaadi, An event-triggered approach to dis- Trans. Syst. Man Cybern. 21 (6) (1991) 1523–1535.
tributed H∞ state estimation for state-saturated systems with randomly [89] X. Nie, W.X. Zheng, J. Cao, Coexistence and local mu-stability of multiple equi-
occurring mixed delays, J. Frankl. Inst.-Eng. Appl. Math. 355 (6) (2018) librium points for memristive neural networks with nonmonotonic piecewise
3104–3121. linear activation functions and unbounded time-varying delays, Neural Netw.
[61] J. Liang, Z. Wang, X. Liu, State estimation for coupled uncertain stochastic 84 (2016) 172–180.
networks with missing measurements and time-varying delays: the discrete– [90] T. Onomi, Y. Maenami, K. Nakajima, Superconducting neural network for solv-
time case, IEEE Trans. Neural Netw. 20 (5) (2009) 781–793. ing a combinatorial optimization problem, IEEE Trans. Appl. Superconduct. 21
[62] Y. Liu, Z. Wang, X. Liu, Global exponential stability of generalized recurrent (3) (2011) 701–704.
neural networks with discrete and distributed delays, Neural Netw. 19 (2006) [91] E. Painkras, L. Plana, J. Garside, SpiNNaker: A 1-w 18-core system-on-chip for
667–675. massively-parallel neural network simulation, IEEE J. Solid-State Circuits 48
[63] Y. Liu, Z. Wang, X. Liu, Robust stability of discrete-time stochastic neural net- (8) (2013) 1943–1953.
works with time-varying delays, Neurocomputing 71 (2008) 823–833. [92] Y. Pershin, M. Ventra, Experimental demonstration of associative memory
[64] Y. Liu, Z. Wang, X. Liu, On global stability of delayed BAM stochastic neu- with memristve neural networks, Neural Netw. 23 (7) (2010) 881–886.
ral networks with Markovian switching, Neural Process. Lett. 30 (1) (2009) [93] J. Qi, C. Li, T. Huang, Stability of inertial BAM neural network with time-vary-
19–35. ing delay via impulsive control, Neurocomputing 161 (2015) 162–167.
[65] Y. Liu, Z. Wang, L. Ma, F.E. Alsaadi, A partial-nodes-based information fu- [94] W. Qian, L. Wang, M.Z.Q. Chen, Local consensus of nonlinear multiagent sys-
sion approach to state estimation for discrete-time delayed stochastic com- tems with varying delay coupling, IEEE Trans. Syst. Man Cybern. Syst. 48 (12)
plex networks, Inf. Fusion 49 (2019) 240–248. (2018) 2462–2469.
[66] Y. Liu, Z. Wang, L. Ma, Y. Cui, F.E. Alsaadi, Synchronization of directed [95] W. Qian, Y. Gao, Y. Yang, Global consensus of multiagent systems with in-
switched complex networks with stochastic link perturbations and mixed ternal delays and communication delays, IEEE Trans. Syst. Man Cybern. Syst.
time-delays, Nonlinear Anal. Hybrid Syst. 27 (2018) 213–224. (2018), doi:10.1109/TSMC.2018.2883108.
[67] Y. Liu, W. Liu, M.A. Obaid, I.A. Abbas, Exponential stability of Markovian [96] R. Raja, U.K. Raja, R. Samidurai, A. Leelamani, Dynamic analysis of discrete–
jumping Cohen-Grossberg neural networks with mixed mode-dependent time BAM neural networks with stochastic perturbations and impulses, Int. J.
time-delays, Neurocomputing 177 (2016) 409–415. Mach. Learn. Cybern. 5 (1) (2014) 39–50.
[68] Y. Liu, B. Shen, Q. Li, State estimation for neural networks with Markov-based [97] R. Rakkiyappan, A. Chandrasekar, S. Laksmanan, J.H. Park, State estimation of
nonuniform sampling: the partly unknown transition probability case, Neu- memristor-based recurrent neural networks with time-varying delays based
rocomputing 357 (2019) 261–270. on passivity theory, Complexity 19 (4) (2013) 32–43.
[69] Y. Liu, B. Shen, H. Shu, Finite-time resilient H∞ state estimation for discrete– [98] R. Rakkiyappan, K. Maheswari, G. Velmurugan, J.H. Park, Event-triggered h-in-
time delayed neural networks under dynamic event-triggered mechanism, finity state estimation for semi-Markov jumping discrete-time neural net-
Neural Netw. 121 (2019) 356–365. works with quantization, Neural Netw. 105 (2018) 236–248.
[70] H. Liu, Z. Wang, B. Shen, X. Liu, Event-triggered H∞ state estimation for de- [99] M. Sahebsara, T. Chen, S.L. Shah, Optimal H2 filtering with random sensor
layed stochastic memristive neural networks with missing measurements: delay, multiple packet dropout and uncertain observations, Int. J. Control 80
the discrete time case, IEEE Trans. Neural Netw. Learn. Syst. 29 (8) (2018) (2) (2007) 292–301.
3726–3737. [100] M. Sahebsara, T. Chen, S.L. Shah, Optimal H2 filtering in networked control
[71] H. Liu, Z. Wang, B. Shen, F.E. Alsaadi, H∞ state estimation for discrete-time systems with multiple packet dropout, IEEE Trans. Autom. Control 52 (8)
memristive recurrent neural networks with stochastic time-delays, Int. J. Gen. (2007) 1508–1513.
Syst. 45 (5) (2016) 633–647. [101] M. Sahebsara, T. Chen, S.L. Shah, Optimal H∞ filtering in networked con-
[72] H. Liu, Z. Wang, B. Shen, Discrete-time memristive recurrent neural networks trol systems with multiple packet dropouts, Syst. Control Lett. 57 (9) (2008)
with time-varying delays: exponential stability analysis, in: Proceedings of 696–702.
the 35th Chinese Control Conference, 2016, pp. 3584–3589. Chengdu, China. [102] R. Sakthivel, R. Anbuvithya, K. Mathiyalagan, P. Prakash, Combined H∞ and
[73] H. Liu, Z. Wang, B. Shen, X. Kan, Synchronization for discrete-time nemris- passivity state estimation of memristive neural networks with random gain
tive recurrent neural networks with time-delays, in: Proceedings of the 34th fluctuations, Neurocomputing 168 (2015) 1111–1120.
Chinese Control Conference, 2015, pp. 3478–3483. Hangzhou, China. [103] M. Sathishkumar, R. Sakthivel, C. Wang, B. Kaviarasan, S.M. Anthoni,
[74] H. Liu, Z. Wang, B. Shen, H. Dong, Delay-distribution-dependent H∞ state esti- Non-fragile filtering for singular Markovian jump systems with missing mea-
mation for discrete-time memristive neural networks with mixed time-delays surements, Signal Process. 142 (2018) 125–136.
and fading measurements, IEEE Trans. Cybern. (2018), doi:10.1109/TCYB.2018. [104] G.S. Seyboth, D.V. Dimarogonas, K.H. Johansson, Event-based broadcasting for
2862914. multi-agent average consensus, Automatica 49 (1) (2013) 245–252.
[75] H. Liu, Z. Wang, B. Shen, X. Liu, Stability analysis for discrete-time stochas- [105] B. Shen, Z. Wang, D. Ding, H. Shu, H∞ state estimation for complex networks

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx 11

with uncertain inner coupling and incomplete measurements, IEEE Trans. bination method, IEEE Trans. Neural Netw. Learn. Syst. Syst. 27 (11) (2016)
Neural Netw. Learn. Syst. Syst. 24 (12) (2013) 2027–2037. 2337–2350.
[106] B. Shen, Z. Wang, H. Qiao, Event-triggered state estimation for discrete– [136] Z. Wang, H. Liu, B. Shen, F.E. Alsaadi, A.M. Dobaie, H∞ state estimation for dis-
time multidelayed neural networks with stochastic parameters and incom- crete-time stochastic memristive BAM neural networks with mixed time-de-
plete measurements, IEEE Trans. Neural Netw. Learn. Syst. Syst. 28 (5) (2017) lays, Int. J. Mach. Learn. Cybern. 10 (4) (2019) 771–785.
1152–1163. [137] L. Wang, Z. Wang, T. Huang, G. Wei, An event-triggered approach to state
[107] B. Shen, Z. Wang, D. Wang, H. Liu, Distributed state-saturated recursive fil- estimation for a class of complex networks with mixed time delays and non-
tering over sensor networks under round-robin protocol, IEEE Trans. Cybern. linearities, IEEE Trans. Cybern. 46 (11) (2016) 2497–2508.
(2020), doi:10.1109/TCYB.2019.2932460. in press. [138] L. Wang, Z. Wang, Q.-L. Han, G. Wei, Event-based variance-constrained H∞ fil-
[108] B. Shen, Z. Wang, D. Wang, J. Luo, H. Pu, Y. Peng, Finite-horizon filtering for tering for stochastic parameter systems over sensor networks with successive
a class of nonlinear time-delayed systems with an energy harvesting sensor, missing measurements, IEEE Trans. Cybern. 48 (2018) 1007–1017.
Automatica 100 (2019) 144–152. [139] L. Wang, Z. Wang, G. Wei, F.E. Alsaadi, Finite-time state estimation for recur-
[109] B. Shen, Z. Wang, H. Tan, Guaranteed cost control for uncertain nonlinear sys- rent delayed neural networks with component-based event-triggering proto-
tems with mixed time-delays: the discrete-time case, Eur. J. Control 40 (2018) col, IEEE Trans. Neural Netw. Learn. Syst. 29 (4) (2018) 1046–1057.
62–67. [140] L. Wang, Y. Shen, Q. Yin, G. Zhang, Adaptive synchronization of memris-
[110] B. Shen, H. Tan, Z. Wang, T. Huang, Quantized/saturated control for sampled– tor-based neural networks with time-varying delays, IEEE Trans. Neural Netw.
data systems under noisy sampling intervals: a confluent vandermonde ma- Learn. Syst. Syst. 26 (9) (2015) 2033–2042.
trix approach, IEEE Trans. Autom. Control 62 (9) (2017) 4753–4759. [141] H. Wang, S. Duan, T. Huang, C. Li, L. Wang, Novel stability criteria for im-
[111] H. Shen, Y. Zhu, L. Zhang, J.H. Park, Extended dissipative state estimation for pulsive memristive neural networks with time-varying delays, Circuits Syst.
Markov jump neural networks with unreliable links, IEEE Trans. Neural Netw. Signal Process. 35 (11) (2016) 3935–3956.
Learn. Syst. Syst. 28 (2) (2017) 346–358. [142] H. Wang, S. Duan, T. Huang, L. Wan, C. Li, Exponential stability of complex–
[112] L. Sheng, Z. Wang, E. Tian, F.E. Alsaadi, Delay-distribution-dependent H∞ state valued memristive recurrent neural networks, IEEE Trans. Neural Netw. Learn.
estimation for delayed neural networks with (x, v)-dependent noises and fad- Syst. Syst. 28 (3) (2017) 766–771.
ing channels, Neural Netw. 84 (2016) 102–112. [143] G. Wei, Z. Wang, H. Shu, Robust filtering with stochastic nonlinearities and
[113] L. Sheng, Y. Niu, M. Gao, Distributed resilient filtering for time-varying sys- multiple missing measurements, Automatica 45 (3) (2009) 836–841.
tems over sensor networks subject to round-robin/stochastic protocol, ISA [144] H. Wei, R. Li, C. Chen, State estimation for memristor-based neural networks
Trans. 87 (2019) 55–67. with time-varying delays, Int. J. Mach. Learn. Cybern. 6 (2) (2015) 213–225.
[114] L. Sheng, Z. Wang, L. Zou, F.E. Alsaadi, Event-based H∞ state estimation for [145] S. Wen, Z. Zeng, Dynamics analysis of a class of memristor-based recurrent
time-varying stochastic dynamical networks with state- and disturbance-de- networks with time-varying delays in the presence of strong external stimuli,
pendent noises, IEEE Trans. Neural Netw. Learn. Syst. Syst. 28 (10) (2017) Neural Process. Lett. 35 (2012) 47–59.
2382–2394. [146] S. Wen, Z. Zeng, T. Huang, Exponential stability analysis of memristor-based
[115] P. Shi, M. Mahmoud, S.K. Nguang, A. Ismail, Robust filtering for jumping sys- recurrent neural networks with time-varying delays, Neurocomputing 97
tems with mode-dependent delays, Signal Process. 86 (1) (2006) 140–152. (2012) 233–240.
[116] V. Singh, A generalized LMI-based approach to the global asymptotic stability [147] S. Wen, G. Bao, Z. Zeng, Y. Chen, T. Huang, Global exponential synchroniza-
of delayed cellular neural networks, IEEE Trans. Neural Netw. 15 (1) (2004) tion of memristor-based recurrent neural networks with time-varying delays,
223–225. Neural Netw. 48 (2013) 195–203.
[117] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M.I. Jordan, S.S. Sastry, [148] C. Wen, Z. Wang, J. Hu, Q. Liu, F.E. Alsaadi, Recursive filtering for state-satu-
Kalman filtering with intermittent observations, IEEE Trans. Autom. Control rated systems with randomly occurring nonlinearities and missing measure-
49 (9) (2004) 1453–1464. ments, Int. J. Robust Nonlinear Control 28 (2018) 1715–1727.
[118] Q. Song, Z. Zhao, Y. Li, Global exponential stability of BAM neural networks [149] A. Wu, J. Zhang, Z. Zeng, Dynamics behaviors of a class of memristor-based
with distributed delays and reaction-diffusion terms, Phys. Lett. A 335 (2005) Hopfield networks, Phys. Lett. A 375 (11) (2011) 1661–1665.
213–225. [150] A. Wu, S. Wen, Z. Zeng, Synchronization control of a class of memristor-based
[119] Q. Song, Z. Zhao, Y. Liu, Stability analysis of complex-valued neural networks recurrent neural networks, Inf. Sci. 183 (1) (2012) 106–116.
with probabilistic time-varying delays, Neurocomputing 159 (2015) 96–104. [151] A. Wu, Z. Zeng, Lagrange stability of memristive neural networks with dis-
[120] Q. Song, H. Yan, Z. Zhao, Y. Liu, Global exponential stability of complex-valued crete and distributed delays, IEEE Trans. Neural Netw. Learn. Syst. Syst. 25 (4)
neural networks with both time-varying delays and impulsive effects, Neural (2014) 690–703.
Netw. 79 (2016) 108–116. [152] A. Wu, Z. Zeng, Global Mittag-Leffler stabilization of fractional-order mem-
[121] D. Strukov, G. Snider, G. Stewart, R. Williams, The missing memristor found, ristive neural networks, IEEE Trans. Neural Netw. Learn. Syst. 28 (1) (2017)
Nature 453 (7191) (2008) 80–83. 206–217.
[122] S. Sun, L. Xie, W. Xiao, Optimal full-order and reduced-order estimators for [153] Y. Xia, J. Wang, A general projection neural network for solving monotone
discrete-time systems with multiple packet dropouts, IEEE Trans. Signal Pro- variational inequalities and related optimization problems, IEEE Trans. Neural
cess. 56 (8) (2008) 4031–4038. Netw. 15 (2) (2004) 318–328.
[123] M. Tabbara, D. Nešić, Input–output stability of networked control systems [154] J. Xiao, S. Zhong, Y. Li, F. Xu, Finite-time Mittag-Leffler synchronization of frac-
with stochastic protocols and channels, IEEE Trans. Autom. Control 53 (5) tional-order memristive BAM neural networks with time delays, Neurocom-
(2008) 1160–1175. puting 219 (2017) 431–439.
[124] P. Tabuada, Event-triggered real-time scheduling of stabilizing control tasks, [155] Y. Xu, H. Su, Y. Pan, Z. Wu, W. Xu, Stability analysis of network control sys-
IEEE Trans. Autom. Control 52 (9) (2007) 1680–1685. tems with round-robin scheduling and packet dropouts, J. Frankl. Inst. 350 (8)
[125] D. Tank, J. Hopfield, Simple neural optimization networks: an A/D converter, (2013) 2013–2027.
signal decision circuit, and a linear programming circuit, IEEE Trans. Circuits [156] Y. Xu, R. Lu, P. Shi, H. Li, S. Xie, Finite-time distributed state estimation over
Syst. 33 (5) (1986) 533–541. sensor networks with round-robin protocol and fading channels, IEEE Trans.
[126] E. Tian, Z. Wang, L. Zou, D. Yue, Probability-constrained filtering for a class of Cybern. 48 (1) (2016) 336–345.
nonlinear systems with improved static event-triggered communication, Int. [157] X. Yang, W.C. Ho, Synchronization of delayed memristive neural networks:
J. Robust Nonlinear Control 29 (5) (2019) 1484–1498. Robust analysis approach, IEEE Trans. Cybern. 42 (16) (2016) 3377–3387.
[127] V. Ugrinovskii, E. Fridman, A round-robin type protocol for distributed esti- [158] X. Yang, J. Cao, J. Liang, Exponential synchronization of memristive neu-
mation with H∞ consensus, Syst. Control Lett. 69 (2014) 103–110. ral networks with delays: interval matrix method, IEEE Trans. Neural Netw.
[128] G.C. Walsh, H. Ye, L.G. Bushnell, Stability analysis of networked control sys- Learn. Syst. 39 (2) (2017) 1878–1888.
tems, IEEE Trans. Control Syst. Technol. 10 (3) (2002) 438–446. [159] Y. Yang, J. Hu, N. Yang, J. Du, Non-fragile set-membership state estimation for
[129] X. Wan, Z. Wang, M. Wu, X. Liu, H∞ state estimation for discrete-time nonlin- memristive neural networks with incomplete measurements via round-robin
ear singularly perturbed complex networks under the round-robin protocol, protocol, in: Proceedings of the 2019 Chinese Control and Decision Confer-
IEEE Trans. Neural Netw. Learn. Syst. Syst. 30 (2) (2019) 415–426. ence, Nanchang, China, 2019, pp. 3395–40 0 0.
[130] X. Wan, Z. Wang, Q.-L. Han, M. Wu, Finite-time H∞ state estimation for dis- [160] Y. Yuan, Z. Wang, P. Zhang, H. Dong, Nonfragile near-optimal control of
crete time-delayed genetic regulatory networks under stochastic communica- stochastic time-varying multiagent systems with control-and state-dependent
tion protocols, IEEE Trans. Circuits Syst. – Part I 65 (10) (2018) 3481–3491. noises, IEEE Trans. Cybern. 49 (7) (2018) 2605–2617.
[131] X. Wan, L. Xu, H. Fang, G. Ling, Robust non-fragile H∞ state estimation for [161] Y. Yuan, Z. Wang, L. Guo, Event-triggered strategy design for discrete-time
discrete-time genetic regulatory networks with Markov jump delays and un- nonlinear quadratic games with disturbance compensations: the noncooper-
certain transition probabilities, Neurocomputing 154 (2015) 162–173. ative case, IEEE Trans. Syst. Man Cybern. Syst. 48 (11) (2017) 1885–1896.
[132] Z. Wang, Y. Liu, X. Liu, On global asymptotic stability of neural networks with [162] D. Yue, E. Tian, Y. Zhang, C. Peng, Delay-distribution-dependent stability and
discrete and distributed delays, Phys. Lett. A 345 (4–6) (2005) 299–308. stabilization of t-s fuzzy systems with probabilistic interval delay, IEEE Trans.
[133] Z. Wang, D.W.C. Ho, X. Liu, State estimation for delayed neural networks, IEEE Syst. Man Cybern. Part B: Cybern. 39 (2) (2009) 503–516.
Trans. Neural Netw. 16 (1) (2005) 279–284. [163] Z. Zeng, D. Huang, Z. Wang, Pattern memory analysis based on stability the-
[134] Z. Wang, Y. Liu, K. Fraser, X. Liu, Stochastic stability of uncertain Hopfield ory of cellular neural networks, Appl. Math. Model. 32 (1) (2008) 112–121.
neural networks with discrete and distributed delays, Phys. Lett. A 354 (2006) [164] C. Zhang, Y. He, L. Jiang, W. Lin, M. Wu, Delay-dependent stability analysis of
288–297. neural networks with time-varying delay: A generalized free-weighting-ma-
[135] Z. Wang, S. Ding, Z. Huang, H. Zhuang, Exponential stability and stabiliza- trix approach, Appl. Math. Comput. 294 (2017) 102–120.
tion of delayed memristive neural networks based on quadratic convex com- [165] G. Zhang, Y. Shen, Exponential synchronization of delayed memristor-based

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066
JID: NEUCOM
ARTICLE IN PRESS [m5G;February 4, 2020;9:59]

12 H. Liu, L. Ma and Z. Wang et al. / Neurocomputing xxx (xxxx) xxx

chaotic neural networks via periodically intermittent control, Neural Netw. Fellow at the King’s College London, U.K. His current research interests include non-
55 (2014) 1–10. linear control and signal processing, variable structure control, distributed control
[166] J. Zhang, L. Ma, Y. Liu, Passivity analysis for discrete-time neural networks and filtering, time-varying systems and multi-agent systems. He has published more
with mixed time-delays and randomly occurring quantization effects, Neuro- than 20 papers in refereed international journals. He serves as an Associate Editor
computing 216 (2016) 657–665. for International Journal of Systems Science and Neurocomputing. He is a very active
[167] J. Zhang, L. Ma, Y. Liu, M. Lyu, F.E. Alsaadi, Y. Bo, H∞ and l2 − l∞ finite-hori- reviewer for many international journals.
zon filtering with randomly occurring gain variations and quantization ef-
fects, Applied Math. Comput. 298 (2017) 171–187. Zidong Wang was born in Yangzhou, Jiangsu, China, in
[168] W. Zhang, L. Yu, G. Feng, Optimal linear estimation for networked systems 1966. He received the B.Sc. degree in mathematics from
with communication constraints, Automatica 47 (9) (2011) 1992–20 0 0. Suzhou University, Suzhou, China, in 1986, and the M.Sc.
[169] W. Zhang, Y. Tang, W. Wong, Q. Miao, Stochastic stability of delayed neural degree in applied mathematics and the Ph.D. degree in
networks with local impulsive effects, IEEE Trans. Neural Netw. Learn. Syst. electrical engineering from the Nanjing University of Sci-
26 (10) (2015) 2336–2345. ence and Technology, Nanjing, China, in 1990 and 1994,
[170] H. Zhang, Z. Wang, D. Liu, A comprehensive review of stability analysis of respectively. He is currently a Professor of dynamical sys-
continuous-time recurrent neural networks, IEEE Trans. Neural Netw. Learn. tems and computing with the Department of Computer
Syst. 25 (7) (2014) 1229–1262. Science, Brunel University London, Uxbridge, U.K. From
[171] Y. Zhang, Z. Wang, L. Ma, F.E. Alsaadi, Annulus-event-based fault detection, 1990 to 2002, he held teaching and research appoint-
isolation and estimation for multirate time-varying systems: applications to a ments in universities in China, Germany, and the U.K. He
three-tank system, J. Process Control 75 (2019) 48–58. has published over 220 papers in the IEEE transactions
[172] Y. Zhang, Z. Wang, L. Ma, Variance-constrained state estimation for networked and over 60 papers in Automatica. His current research
multi-rate systems with measurement quantization and probabilistic sensor interests include dynamical systems, signal processing, bioinformatics, and control
failures, Int. J. Robust Nonlinear Control 26 (16) (2016) 3507–3523. theory and applications. Dr. Wang is a fellow of the Royal Statistical Society and
[173] X.-M. Zhang, Q.L. Han, Global asymptotic stability analysis for delayed neural a member of the program committee for many international conferences. He re-
networks using a matrix-based quadratic convex approach, Neural Netw. 54 ceived the Alexander von Humboldt Research Fellowship of Germany, the JSPS Re-
(2014) 57–69. search Fellowship of Japan, and the William Mong Visiting Research Fellowship of
[174] X.-M. Zhang, Q.-L. Han, Z. Wang, B.L. Zhang, Neuronal state estimation for Hong Kong. He serves (or has served) as the Editor-in-Chief of Neurocomputing,
neural networks with two additive time-varying delay components, IEEE the Deputy Editor-in-Chief of the International Journal of Systems Science, and an
Trans. Cybern. 47 (10) (2017) 3184–3194. Associate Editor for 12 international journals, including the IEEE Transactions on
[175] Z. Zhao, Z. Wang, L. Zou, H. Liu, Finite-horizon h-infinity state estimation Automatic Control, the IEEE Transactions on Control Systems Technology, the IEEE
for artificial neural networks with component-based distributed delays and Transactions on Neural Networks, the IEEE Transactions on Signal Processing, and
stochastic protocol, Neurocomputing 321 (2018) 169–177. the IEEE Transactions on Systems, Man, and Cybernetics-Part C.
[176] L. Zou, Z. Wang, H. Gao, Set-membership filtering for time-varying systems
with mixed time-delays under round-robin and weighted try-once-discard
protocols, Automatica 74 (2016) 341–348.
[177] L. Zou, Z. Wang, H. Gao, X. Liu, State estimation for discrete-time dynam-
ical networks with time-varying delays and stochastic disturbances under Yurong Liu received the B.Sc. degree in mathematics
the round-robin protocol, IEEE Trans. Neural Netw. Learn. Syst. 28 (5) (2016) from Suzhou University, Suzhou, China, in 1986, the M.Sc.
1139–1151. degree in applied mathematics from the Nanjing Uni-
versity of Science and Technology, Nanjing, China, in
Hongjian Liu received his B.Sc. degree in applied mathe- 1989, and the Ph.D. degree in applied mathematics from
matics in 2003 from Anhui University, Hefei, China and Suzhou University in 20 0 0. He is currently a Professor
the M.Sc. degree in detection technology and automa- with the Department of Mathematics, Yangzhou Univer-
tion equipments in 2009 from Anhui Polytechnic Univer- sity, Yangzhou, China. He has published over 100 pa-
sity, Wuhu, China, and the Ph.D. degree in control sci- pers in international journal. His current interests include
ence and engineering in 2018 from Donghua University, neural networks, nonlinear dynamics, time-delay systems,
Shanghai, China. In 2016, he was a Research Assistant and chaotic dynamics. He serves as an editorial board
with the Department of Mathematics, Texas A&M Univer- member for Neurocomputing.
sity at Qatar, Doha, Qatar, for two months. From March
2017 to March 2018, he was a Visiting Scholar in the De-
partment of Information Systems and Computing, Brunel
University London, UK. He is currently an Associate Pro-
fessor in the School of Mathematics and Physics, Anhui Fuad E. Alsaadi received the B.S. and M.Sc. degrees in
Polytechnic University, Wuhu, China. His current research interests include filtering electronic and communication from King AbdulAziz Uni-
theory, memristive neural networks and network communication systems. He is a versity, Jeddah, Saudi Arabia, in 1996 and 2002, respec-
very active reviewer for many international journals. tively, and the Ph.D. degree in optical wireless communi-
cation systems from the University of Leeds, Leeds, U.K.,
Lifeng Ma received the B.Sc. degree in Automation from in 2011. From 1996 to 2005, he was a Communication In-
Jiangsu University, Zhenjiang, China, in 2004 and the structor with the College of Electronics and Communica-
Ph.D. degree in Control Science and Engineering from tion, Jeddah. He is currently an Associate Professor with
Nanjing University of Science and Technology, Nanjing, the Department of Electrical and Computer Engineering,
China, in 2010. From August 2008 to February 2009, he Faculty of Engineering, King Abdulaziz University. He has
was a Visiting Ph.D. Student in the Department of Infor- published widely in the top IEEE communications confer-
mation Systems and Computing, Brunel University Lon- ences and journals. His current research interests include
don, U.K. From January 2010 to April 2010 and May 2011 optical systems and networks, signal processing, and syn-
to September 2011, he was a Research Associate in the chronization and systems design. He was a recipient of the Carter Award, University
Department of Mechanical Engineering, the University of of Leeds for the best Ph.D.
Hong Kong. He is currently a Professor in the School of
Automation, Nanjing University of Science and Technol-
ogy, Nanjing, China, and is currently a Visiting Research

Please cite this article as: H. Liu, L. Ma and Z. Wang et al., An overview of stability analysis and state estimation for memristive neural
networks, Neurocomputing, https://doi.org/10.1016/j.neucom.2020.01.066

You might also like