Professional Documents
Culture Documents
Artificial Intelligence
Advances
Editor-in-Chief
Dr. Sergey Victorovich Ulyanov
Volume 2 | Issue 2 | October 2020 | Page 1-91
Artificial Intelligence Advances
Contents
Article
1 Intelligent Control of Mobile Robot with Redundant Manipulator & Stereovision: Quan-
tum / Soft Computing Toolkit
Kirill V. Koshelev Alena V. Nikolaeva Andrey G. Reshetnikov Sergey V. Ulyanov
32 Intelligent Robust Control of Redundant Smart Robotic Arm Pt II: Quantum Computing
KB Optimizer Supremacy
Alena V. Nikolaeva Sergey V. Ulyanov
68 Robotic Smart Prosthesis Arm with BCI and Kansei / Kawaii / Affective Engineering
Approach. Pt I: Quantum Soft Computing Supremacy
Alexey V. Nemchaninov Alena V. Nikolaeva Sergey V. Ulyanov Andrey G. Reshetnikov
Review
88 Research and Analysis of Machine Learning Algorithm in Artificial Intelligence
Yang Li Xueyuan Du Yiheng Xu
Copyright
Artificial Intelligence Advances is licensed under a Creative Commons-Non-Commercial 4.0 International Copyright
(CC BY- NC4.0). Readers shall have the right to copy and distribute articles in this journal in any form in any medium,
and may also modify, convert or create on the basis of articles. In sharing and using articles in this journal, the user must
indicate the author and source, and mark the changes made in articles. Copyright © BILINGUAL PUBLISHING CO. All
Rights Reserved.
Artificial Intelligence Advances | Volume 02 | Issue 02 | October 2020
ARTICLE
Intelligent Control of Mobile Robot with Redundant Manipulator &
Stereovision: Quantum / Soft Computing Toolkit
Kirill V. Koshelev1,2 Alena V. Nikolaeva1,2 Andrey G. Reshetnikov1,2 Sergey V. Ulyanov1,2*
1. Dubna State University, Russia
2. INESYS LLC (EFKO GROUP), Russia
Article history The task of an intelligent control system design applying soft and quan-
Received: 22 November 2019 tum computational intelligence technologies discussed. An example of a
control object as a mobile robot with redundant robotic manipulator and
Accepted: 3 June 2020 stereovision introduced. Design of robust knowledge bases is performed
Published Online: 30 June 2020 using a developed computational intelligence - quantum/soft computing
toolkit (QC/SCOptKBTM). The knowledge base self-organization process
Keywords: of fuzzy homogeneous regulators through the application of end-to-end
Quantum / Soft computing optimizer IT of quantum computing described. The coordination control between
the mobile robot and redundant manipulator with stereovision based on
Knowledge base soft computing described. The general design methodology of a gener-
Fuzzy controller alizing control unit based on the physical laws of quantum computing
Quantum fuzzy inference (quantum information-thermodynamic trade-off of control quality dis-
tribution and knowledge base self-organization goal) is considered. The
Multi-agent systems
modernization of the pattern recognition system based on stereo vision
Mobile robot stereo vision technology presented. The effectiveness of the proposed methodology is
demonstrated in comparison with the structures of control systems based
on soft computing for unforeseen control situations with sensor system.
The main objective of this article is to demonstrate the advantages of the
approach based on quantum/soft computing.
1. Introduction The prototype presented below is used in the education pro-
T
cess, the software component of each block of the control
he application of robotic systems in various prob- system is modernized through the application of soft and
lem-oriented domains gives a significant positive quantum computing IT. A significant increase in robustness
effect; moreover, not only industrial intelligent ro- [4]
when using these types of computing allows us to say
bots, but also robots for service use with various degrees of that the creation of an effective robust and adaptive control
social responsibility (medical robots, rescue robots etc.) are system for robotic devices requires an additional software
spreading [1-3]. The effectiveness of such robot’s application and algorithmic support platform in the form of sophisti-
directly depends on the quality of control systems [3], which cated toolkit based on soft and quantum computing (Quan-
are the most labor-intensive part of the robotic complex. tum computational intelligence toolkit) [4]. Specialists in
*Corresponding Author:
Sergey V. Ulyanov,
Dubna State University, Russia;
INESYS LLC (EFKO GROUP), Russia;
Email: ulyanovsv@mail.ru
the design and development of various robots rely on an in the robot artificial life. This allows us to evaluate the
intelligent platform for the application of computational in- control algorithms in online robot behavior and to reduce
telligence technologies. This approach allows unifying the difficulties connected with such troubles as robot colli-
process of intellectualization of the created hybrid industrial sions with obstacles and robot hardware damages. In this
service and special purpose robots [1-3]. Item we describe a new approach to intelligent control
In the article we are focused the attention on the fact system design using soft computing technology. A new
that the application of quantum computing in the design form of direct human - robot communications (including
process of intelligent control systems (ICS) significantly emotion, instinct and intuition) and an autonomous lo-
increases the reliability of hybrid intelligent controllers comotion control system were developed (see Figure 1).
by introducing knowledge self-organization capability. Authors considered direct human-robot communications
The basic principles of creating robust knowledge bases based on natural language (NL) and construct the simula-
(KB) are described. The paper describes an experiment, tion system of spatial scenes and robot behavior in virtual
the results of which make it clear that the effectiveness reality (VR).
of a fuzzy controller is significantly reduced in case
of emergency situations. We propose an approach that
allows to solve a similar problem by introducing a quan-
tum generalization of strategies in fuzzy inference in on-
line from a set of pre-defined fuzzy controllers (FC) by
new Quantum Fuzzy Inference (QFI). We consider a
new structure of intelligent control system (ICS) with a
quantum KB self-organization based on QFI. We espe-
cially focus on robustness of control because it is the
background for support the reliability of advanced con-
trol accuracy in uncertainty environments. The main goal
of this work is to provide a brief description of soft com-
puting tools for designing independent FC, then we will
provide the QFI methodology of quantum KB self-orga-
nization in unforeseen situations. Quantum supremacy of Figure 1. Structure of Artificial Intelligence (AI) control
classical intelligent control system design in unexpected system with distributed knowledge representation (on
control situations demonstrated. control signal levels). a - Intelligent control “in large” b
- Intelligent control “in small” c - Control on executive
1.1 State of Art of Robotic Systems Development level
The basis of the intelligent robot control system is to It was explained also the managing system which
achieve a synergistic effect from the interaction of various controls cooperatively three sub-systems of the service
on-board robot systems (a movement system, a mobile robot, as the locomotion system, the handling system for
manipulator control system, an image processing system a mobile manipulator and the image processing system
etc.). This approach was identified and supported by re- as human vision system. Also, three sub-systems which
search and development in Japan in the mid-90s of the organize the service robot system for its autonomous nav-
last century [1-3]. It should be noted that now this particular igation and these soft computing are described in [3,5]. The
approach is fundamental in robotics. locomotion control system is composed of four functions,
In particular case, in [1-3] is described how the mobile i.e. locomotion control, planning for works, learning and
robot for service use works in buildings with different recognition. These four functions are related to each other.
scenes of rooms and moves in unstructured environments By using the handling system for a mobile manipulator
in presence of many human being operators and unexpect- and the image processing system as human vision system,
ed obstacles applied cognitive graphic of path planning the robot can realize some technology operations, for ex-
and soft computing for control of different technological ample, opening a door and getting on an elevator. These
operations (open door, getting on an elevator, room con- three sub-systems are based on fuzzy control, fuzzy neural
trol operations and so on). It was suggested to construct a networks (FNN) and genetic algorithm (GA). Experimen-
simulation system for mobile service robot behavior based tal results on the developed robot show that the proposed
on cognitive graphics with virtual and complimentary re- methods are very useful for autonomous locomotion con-
ality. This system is used for possible world’s simulation trol of the robot.
Example. Intelligent control and soft computing for responses are faster and more accurate as compared to
avoidance of obstacles and execution of technology op- the conventional systems. The major drawbacks found
erations - robot is power-wheeled steering type which is while dealing with the static controllers like PI, PID
achieved by two driving wheels and a caster with passive etc. are that they are less efficient for dealing with the
suspension for stable locomotion. Thirteen ultrasonic (US) complexities and non-linear disturbances. Also, they are
sensors, nine infrared (IR) sensors, a five degree-of-free- non-adaptive and less robust. However, the computation-
dom (DOF) manipulator with a three-finger hand and a al model based on the mammalian decision system could
CCD camera are equipped on the robot for conducting deal with these complexities and non-linearity more ide-
tasks and works in buildings including human being, ally and robustly. In online process control systems, all
opening door and getting on an elevator. In process of the conditions are uncertain. Also, some complex pro-
robot’s locomotion from point to point mobile robot must cesses are very dynamic and conditionals are not known
avoid obstacles in room, and achieve starting position for with accurate precisions. Thus, in these systems, the
successful opening room door (as technology operation design of the controller is very crucial. Also, solving this
from one point to another point) and go out the room to non-linearity through conventional computational model
elevator in presence of obstacles in a corridor. Movement proves to be very slow. Also, in online scenarios the dis-
process is planned by the managing system which was de- turbance sources are unknown and thus any prior model
scribed in [5]. cannot be predicted. Adaptive models are required for
It must be said that the management of poorly formal- dealing such scenarios. Thus, various AI techniques like
ized and poorly structured control objects (CO) with a neuro-fuzzy techniques, Genetic Algorithms, PSO are
variable and complex structure, variable parameters and proved. A model based on limbic-system of mammalian
uncertain operating conditions, in a group or an autono- brain emotions has been proposed by Caro Lucas in 2004
[7]
mous object requires the use of end-to-end (in the general . He has demonstrated that emotion-based decisions
case - quantum) information technologies. Studying the are quick and more satisfying all the constraints. He has
problems of implementing control systems for such new proposed the model based on dynamic limbic system
and extremely complex CO as NICA, LHC, data process- of brain of mammalian. A BELBIC model presents the
ing of telescopes, autopilots, objects of intelligent robotics mimics of the limbic systems components - amygdala,
indicates a significant dependence of the robustness of the orbito-frontal cortex, thalamus, sensory cortex. It has
system on the embedded computing basis of the most ICS been implemented for various SISO, MIMO, and other
[6]
. It is necessary to revise the science platform that is em- non-linear systems. Various results have demonstrated
bedded in end-to-end information technology (IT) for the its very fast control action, better disturbance handling
design of ICS. capacity, and robustness (see Appendix 1).
Traditional modern automatic control systems fun- This is also due to the fact that the classical methods
damentally do not take into account the occurrence of of control theory have synthesis and design methods for
contingency management situations and do not include well-formalized, well-defined and well-described CO
the human being cognitive factor in the control loop, that operate in previously known conditions and situ-
which does not guarantee the timely achievement of the ations. However, unforeseen circumstances and their
management goal - obtaining the maximum intensity of individual characteristics determine the secrecy and
functioning and stable repetition of the required operating underdetermination of the parameters of physical and
modes with minimal cost of useful resources, such as: mathematical models, and should be taken into account
time settings, power consumption and more other. In the in the KB of intelligent FC. Quantum end-to-end IT al-
conditions of uncertainty or inaccuracy (imperfect) of the lows designing hierarchical ICS’s, which make possible
initial information, unforeseen situations or information to redistribute the degree of decision-making responsi-
risk, the traditional (using the principle of global negative bility between ICS’s depending on a dynamically chang-
feedback) and widely used in industry PID-controller of- ing situation [6,8].
ten does not cope with the management task. At the same This article also provides a brief description of the de-
time, the solution to the problem of global robustness of veloped system of pattern recognition, which is based on
the PID-controller is still unknown, despite the relevance stereovision technology. The use of stereovision allows
of this problem. obtaining data of the depth of the image, the distance to
Remark. Systems based on biological systems and objects, provides an opportunity to build a 3Dpicture of
mimicking their working conditions and behavior are the surrounding world. This article discusses the archi-
found to be more adaptive and systematically. Their tecture of convolutional neural networks and its applica-
IFICS (see Figure 2). advanced control system that all other controllers (FC1,
FC2) are failed but QFC is designed in online a new FC
with increasing robustness. This approach was applied to
other complex robotic systems [11-13].
formation interaction between KBs. Quantum computing confirmed the existence of a synergistic effect of self-orga-
technologies for creating self-organizing KB intelligent nization during the formation of robust KB from designed
regulators are considered in [16,17]. In [16] a new type of non-robust imperfect KB using quantum computing. More-
quantum search algorithm on the generalized space KB over, an additional information resource for control is based
of FC was described (KB designed on the basis of soft on the extraction of quantum information hidden in classi-
computing technologies). The QFI model implements the cal states embedded in a quantum algorithmic gate. Such
self-organization of KB, based on the physical laws of the ICS design methods allowed achieving global robustness in
theory of quantum computing [18,19] and the application of online, while using the computing resources of the embed-
four operators: superposition, quantum correlation, in- ded conventional processor of the on-board system.
terference, and measurement. The first three are unitary, The use of new quantum computing paradigm and deep
reversible quantum operators, and the fourth (measurement machine learning in the group of interaction robots pro-
operator) is classical (irreversible). vides the following advantages:
The QFI design process includes the design of a quan- (1) the potential acceleration of computing through the
tum algorithmic gate - the matrix form of three quantum use of quantum evolutionary operators, which will make it
operators: superposition, entanglement and interference, possible to fully use the powerful, but slow GA to get the
which are part of the structure of quantum search algo- possibility of training the system in online;
rithms, and a source of quantum information that is a hid- (2) the introduction of quantum superposition and
den variable in classical states. The main unit of such ICS quantum-classical correlation operators into the classical
is the quantum genetic search algorithm (QGSA). QFI algorithm entails the appearance of unique properties of
operators can be implemented both on a quantum and on the data processing process, which affects the result of the
a classical processor, which in the second case allows you work of a group of robots;
to integrate them into various control systems and built-in (3) low dependence on environmental disturbances;
intelligent controllers, taking into account the limitations (4) independence from static electricity or movement
of the computing resource of the elements of an exper- of the CO or digital video camera;
imental laboratory bench or on-board system. The very (5) registration of changes and redistribution of tasks
process of creating such an algorithmic gate requires the during changes in a group of autonomous agents.
attraction of large computational resources, in connection
with which a methodology has been developed for pro- 2.3 Control Performance Measure: Thermody-
cessing big data and conducting quantum computing on a namic Trade-off and Interrelations between Sta-
supercomputer (for example, “Govorun” LIT JINR). bility, Controllability and Robustness
The QFI algorithm (see Figure 4) for determining new According to Figure 5, one of the main tasks of designing
control parameters (a special case of implementing quan- ICS consists in providing that the developed (chosen)
tum computing on a classical processor) consists of such structure possesses the required level of control quality
stages as normalization, formation of a quantum bit, after and robustness (supports the required indices of reliability
which the optimal structure of the quantum algorithmic and accuracy of control under the conditions of informa-
gate is selected, and the state with maximum amplitude is tion uncertainty).
selected probability, and decoding the signal to obtain new
control parameters.
Note that one of the most important and hard-to-solve (b) Global unstable
problems of designing ICS’s is the design of robust KB (c) Partial unstable on generalized coordinate and
that allow the ICS to operate under the conditions of in- non-linear braces
formation uncertainty and risk. The core of technique for ② Time-dependent random structure or parametric ex-
designing robust KB of FC’s is generated by new types of citations
computing and simulation processes. ③ Type of model description
Remark. We are witnessing a rapidly growing interest (a) Mathematical model
in the field of advanced computational intelligence, a “soft (b) Physical model
computing” technique. Soft computing integrates fuzzy (c) Partial mathematical and fuzzy physical model
logic, neural networks, evolutionary computation, and (2) External random excitations
chaos. Soft computing is the most important technology ① Different probability density functions
available for designing ICSs and cognitive control. The ② Time-dependent probability functions
difficulties of fuzzy logic involve acquiring knowledge (3) Measurement system
from experts and finding knowledge for unknown tasks. ① Sensor noise
This is related to design problems in constructing fuzzy ② Time delay
rules. FNN’s and GA’s are attracting attention for their po- ③ Random time delay with sensor noise
tential in raising the efficiency of knowledge finding and (4) Different types of reference signals
acquisition. Combining the technologies of fuzzy logic, (5) Different types of traditional controllers
FNN’s and GA’s, i.e., soft computing techniques will have This problem is solved in three stages as following:
a tremendous impact on the fields of intelligent systems (1) the characteristics of stability of the controlled plant
and control design. are determined for fixed conditions of its operation in the
To explain the apparent success of soft computing, external environment;
we must determine the basic capabilities of different soft (2) a control law is formed that provides the stability of
computing frameworks. Recently, the application of ICS operation of the controlled plant for a given accuracy of
structures based on new types of computations (such as control (according to a given criteria of the optimal con-
soft, quantum computing) has drawn the ever-increasing trol);
attention of researchers. Numerous investigations con- (3) the sensitivity and robustness of the dynamic be-
ducted have shown that soft computing possess the fol- havior of the controlled plant are tested for various classes
lowing points of favor: soft computing retain the main ad- of random perturbations and noise.
vantages of conventional automatic control systems (such These design stages are considered by modern control
as stability, controllability, observable ability, etc.); soft theory as relatively independent. The main problem of
computing have an optimal (from the point of view of a designing automatic control systems is to determine an
given control objective performance) KB; soft computing optimal interaction between these three quality indices of
guarantee the attainability of the required control quality control performance. For robust structures of automatic
based on the designed KB control systems, a physical control principle can be prov-
One of the main problems of modern control theory is en that allows one to establish in an analytic form the
to develop and design automatic control systems that meet correspondence between the required level of stability,
the three main requirements: stability, controllability, and controllability, and robustness of the control. This allows
robustness. The listed quality criteria ensure the required one to determine the required intelligence level of the
accuracy of control and reliability of operation of the con- automatic control system depending on the complexity of
trolled object under the conditions of incomplete infor- the particular control problem.
mation about the external perturbations and under noise Let us briefly consider main physical principles of an
in the measurement and control channels, uncertainty in energy-based control processes that allow one to establish
either the structure or parameters of the control object, or the interrelation between the qualitative dynamic char-
under limited possibility of a formalized description of the acteristics of the controlled plant and the actuator of the
control goal. Therefore, in practice of advanced control automatic control system: stability, controllability, and
systems main sources of unpredicted control situations are robustness of control. For this purpose, we are employing
as following: the informational and thermodynamic approaches that
(1) Control object join by a homogeneous condition the criteria of dynamic
① Type of unstable dynamic behavior stability (the Lyapunov’s function), controllability, and
(a) Local unstable robustness.
Example: Thermodynamics trade-off between stability, characterized uniquely by the parameters of the nonlinear
controllability, and robustness. Consider a dynamic con- dynamic automatic control system, which results in deter-
trolled plant given by the equation mination of an optimal selective trajectory from the set of
possible trajectories in optimization problems. Thus, the
dq first condition is fulfilled automatically.
ϕ ( q, S ( t ) , t , u , ξ ( t ) ) , u =
= f ( q, qd , t ) ,(1)
dt dV
Assume that the second condition ≤ 0 holds. In
where q is the vector of generalized coordinates de- dt
scribing the dynamics of the controlled plant; S is the gen- this case, the complete derivative of the Lyapunov func-
eralized entropy of dynamic system; u is the control force tion has the form
(the output of the actuator of the automatic control sys-
dV
tem); qd(t) is reference signal, ξ(t) is random disturbance
dt
= ∑ q q
i
i i + SS = ∑ q ϕ ( q, S , t , u ) + ( S
i
i i cob ( )
− Sc ) Scob − Sc .
and t is the time. The necessary and sufficient conditions
of asymptotic stability of dynamic system with ξ(t)≡0 are Thus, taking into account (1) and the notation intro-
determined by the physical constraints (for example, as duced above, we have
for Port-controlled Hamiltonian Systems (PCHS) [20]) on
the form of the Lyapunov function, which possesses two dV
= ∑ qi ϕi ( q, ( Ψ − ϒ ) , t , u ) + ( Ψ − ϒ ) ( Ψ
− ϒ ) ≤ 0
important properties represented by the following condi- dt .
i
tions: Stability Robustness
Controllability
(I) This is a strictly positive function of generalized co- (3)
ordinates, i.e., V>0;
(II) The complete derivative in time of the Lyapunov’s Figure 6 shows the role of developed thermodynamic
function is a non-positive function, trade-off in robust control design.
dV
≤0.
dt
1 n 2 1 2
=V ∑ qi + 2 S ,(2)
2 i =1
ganization in KB of robust FC based on quantum comput- situations in an external random environment. The design
ing operations (that absent in soft computing toolkit). of private individual KB with the application of software
tools used for specified control situations is carried out
3. The Role of Robust Intelligent Control Sys- in accordance with the design technology of the ICS and
tems in Advanced Robotics discussed in detail in [26]. The main task solved by QFI is
the formation of KB with an increased level of robustness
The process of developing innovative information tech-
from a finite set of KB for workstations formed using soft
nologies for the design of embedded self-organizing ro-
computing.
bust intelligent control systems based on quantum end-
Peculiarities of quantum approach. Recall that the
to-end artificial intelligence technologies is inextricably
square of the amplitude of the probabilities of a state in
linked with the development of scientific and technolog-
quantum mechanics is equal to the classical probability
ical progress in areas such as particle physics, robotic
of finding a quantum system in a given state (Bohr’s pos-
systems (sensors, computation processors components),
tulate, which has several variants of rigorous justification
new production technologies, creation of innovative smart [27]
). From the point of view of the quantum information
materials, analytics of large experimental data, search
theory, a pure quantum state is characterized, as is known,
in unstructured databases. The development of ICS for
by the von Neumann entropy value of zero. Therefore,
traditional automated control systems with an increased
the intelligent quantum state in the considered case takes
level of robustness is of significant theoretical, practical
place for the minimum informational entropy of the Shan-
and commercial importance. It is imperative that modern
non’s quantum state. The desired minimum is achieved, in
control systems maintain the required levels of accura-
turn, with the maximum probability Pi of the state. Since
cy and reliability in unforeseen situations. The practice
Pi, by definition, is the square of the corresponding proba-
and results of modeling real objects have shown that in
bility amplitude, the principle of the maximum probability
conditions of uncertainty or inaccuracy in the initial infor-
amplitude in the correlated state can be taken as a criteri-
mation, unforeseen situations, or information risk the tra-
on for selecting the priority “intelligent” correlation (co-
ditional (using the principle of global negative feedback)
herent) state in a superposition of possible candidates [28,29].
and widely used in industry PID-controller often fails to
Thus, by calculating the amplitudes of quantum states in
cope with the management task. The application of FC
a superposition of states with mixed types of quantum
in conjunction with the PID-controller led to the creation
correlation and choosing the maximum among them, a
of hybrid fuzzy ICSs with different levels of intelligence
quantum oracle model is implemented that contains the
depending on the completeness and correctness of the de-
necessary information about the desired solution. Using
signed KB.
the standard decoding procedure (the internal product of
The use of soft computing technology (based on GA
vectors in a Hilbert space) and selecting the scaling fac-
and FNN) has expanded the field of effective use of FC
tors for the output values of the projected gain factors,
by adding new functions in the form of learning and ad-
iterative work of algorithm of QFI is carried out. Remote
aptation (see Figure 5). However, in the general case of
connection of the CO to a stationary computing system
contingency management situations, it is very difficult to
opens up the possibility of remote configuration, forma-
design a globally “good” and robust ICS structure. This
tion and self-organization of KB in online. The presented
restriction is especially typical for unforeseen control
model QFI makes it possible to solve the classical prob-
situations when the CO operates in rapidly changing con-
lems of designing robust KB in ICS structures that have
ditions (sensor failure or noise in the measuring system,
no analogues among the family of randomized classical
the presence of a delay time for control or measurement
algorithms and is characterized by polynomial computa-
signals, a sharp change in the structure of the CO or its pa-
tional complexity. The design of ICS on QFI is carried out
rameters, etc.). A solution to this kind of problems can be
using the developed software toolkit “Quantum Optimiz-
found by introducing the principle of self-organization of
er”. The technology of application of quantum fuzzy logic
KB into the design process of FC, which is implemented
allows combining several KB into a single control system,
and programmatically supported by the developed model
thereby allowing FNN to work in parallel as quantum
of QFI using the methodology of quantum soft computing
neural network.
and Intelligent System of System Engineering [25]. The
Remark. From the point of view of computational com-
proposed model of QFI uses private individual KB of FC,
plexity theories [31-33], the developed algorithm belongs to
each of which is obtained by using SCO for the corre-
the class of polynomial algorithms with a limited error
sponding operating conditions of the CO and fixed control
- the BPP class (bounded-error probabilistic polynomial
time), and its quantum generalization - to the BQP class. self-organization properties of robust knowledge bases.
Accordingly, to quantum algorithm theory and quantum The following levels of those shown in Figure7 were con-
Kolmogorov’s algorithmic complexity, this algorithm is sidered in detail in [21]: level 3 (physical model and objec-
effective. This means that structurally the algorithm has tive of the self-organization control process, and physical
polynomial complexity, i.e. the randomized algorithm interpretation of main operators of the quantum algorithm
is polynomial rather than exponential (classic search al- of KB self-organization control), and level 2 (dynamics of
gorithms) depending on input signals; at the same time, evolution self-organization process).
the limited probability of the accuracy of measuring the
calculation result is enough to an effective decision mak-
ing. The following actions are implemented in QFI model
[25,34,35]
:
(1) the results of the fuzzy inference of each indepen-
dent FC are processed;
(2) based on the methods of the quantum information
theory, valuable quantum information is extracted, hidden
in independent (individual) KB;
(3) in online, a generalized robust output control signal
is projected on all the sets of KB.
In this case, the output signal of QFI in online is the
optimal control signal for changing the coefficient gains
of the PID-controller, which includes the necessary (best)
quality characteristics of the output control signals of each
FC, thereby realizing the principle of self-organization. Figure 7. Hierarchical structure of quantum algorithm of
Therefore, the area of effective functioning of ICS struc- design of self-organization of robust KB’s in ICS
ture can be significantly expanded by including such an On Figure 8 is presented the layouts of robots of the
important characteristic of control quality as robustness. “mobile manipulator” type, which were designed by var-
The robustness of the control signal is the basis for main- ious research groups (HERB: Home Exploring Robotic
taining the reliability and accuracy of control in condi- Butler, STAIR: Stanford Artificial Intelligence Robot and
tions of uncertainty of information or a poorly formalized many others), which are a platform for creating new IT
description of the operating conditions and / or control and software products.
objectives [36].
In natural systems the sought robustness property is
coded in the algorithm of reproduction of the self-orga-
nization process. Therefore, such systems can autono-
mously handle an unforeseen event using different (close,
as regards the idea) approaches: adaptation (learning,
evolution) in the framework of which the system corrects
its behavior in order to handle the event variation; predic-
tion manifested in the fact that the system can “predict”
the change of situation and determine more precisely its
behavior (this property is a special case of adaptation and
does not require that the system estimates the situation
before it occurs); and robustness consisting in the fact that
the system can operate and achieve the objective if contin-
gency perturbations of a certain type occur.
Figure 8. Examples of modern mobile robots with manip-
The robustness property of the KB is achieved by
ulators based on the ROS framework
application of the quantum design algorithm of self-or-
ganization in the course of intelligent control which is One of the most striking examples of developing IT
schematically shown in Figure 7. Figure 7 show the hi- projects is the ROS framework. ROS provides libraries
erarchical levels of the design process using the quantum and tools to help software developers create robot appli-
algorithm, interconnection, and interrelation of the above cations. It provides hardware abstraction, device drivers,
pattern recognition system, Q' is a measured variable, Testing the robustness of control system models is car-
E is a control error, s(t) is a control influence limitation, ried out in experiments in standard and unforeseen control
mC(t),mMS(t) are noises in the control and measurement situations. These experiments are described in detail in [37].
system channels, U is a control action, d(t) is delay in MS, We will consider control systems with constant coef-
fCO(t) is external influences on CO, Q is adjustable value. ficients of the PID regulator and ICS based on SCO on
The creation of QFI unit is carried out using the computa- soft computing. A system of quality criteria that takes into
tional intelligence SCO toolkit based on quantum comput- account methods for assessing transients of automatic
ing [19,38]. In the selected configuration of the ICS structure control theory was introduced to evaluate and compare
are implemented seven FC’s, each of which independently the results of testing control systems. These methods have
controls one of the seven links (see Figure 13). been adapted for a specific CO (mobile robot-manipula-
tor). The following criteria were highlighted:
(1) solution of the positioning problem in standard con-
trol situations PTSKCS (Position Task Solution in known
control situations);
(2) solution of the positioning problem in external un-
foreseen control situations PTSACCS1 (Position Task Solu-
tion in the above considered control situations);
(3) solution to the problem of positioning in internal
contingency management situations PTSACCS2;
(4) IT performance;
(5) relative overshoot value σ;
(6) relative error in link positioning at the end of a giv-
en number of iterations ε;
(7) time of one iteration t;
(8) the complexity of the implementation of control P;
Figure 13. ICS of robot-manipulator with seven degrees
of freedom based on SCO (9) overall assessment of FCB management (Full Con-
trol Behavior).
Designations in Figure 13: KPi,KDi,KIi, i = 1,7 are pro- Table 1 shows a comparison of ICS based on SCO with
portional, differential and integral coefficients gain of the quantum computing using spatial correlations with the op-
PID-controller, where i is the number of the corresponding eration of ICS based on SCO with soft computing.
link of the robot-manipulator, TSi is training signal, GAi is
genetic algorithm that generates the training signal for the Table 1. Comparison of ICS based on QCO with ICS
formation of the i-th KB. The standard control situation for based on SCO
the i-th FC is a typical control situation, in the conditions
ICS based on SCO with quantum
of which the training signal TSi is received. ICS designed ICS based on computing
by SCO may contain information about seven control situa- Situation SCO with soft Using the correla- Using the
computing tion of two adjacent correlation
tions (regular or unforeseen) for each of the links. FC of seven FC
The process of integrating QFI unit into intelligent lev- 1 Regular situations 0,923 1,000 1,000
el of the control system is shown in Figure 14, where the External contingen-
2 0,744 0,821 0,821
blocks that are designed using soft and quantum comput- cies
Internal contingen-
ing technologies are highlighted. Further, ICS with QFI 3 0,923 0,846 0,923
cies
will be called SCO based on quantum computing. 4 Performance 0,092 0,477 0,459
5 Overshoot 0,969 0,973 0,971
6 Sustainability 0,911 0,962 0,960
7 Single iteration time 0,973 0,961 0,960
Implementation
8 0,946 0,957 0,986
complexity
9 General control 0,721 0,816 0,826
noise in a measurement system (an example of noise in theory of swarm artificial intelligence [40]. A multi-agent
a measurement system is shown in the Figure 17) on the system is capable of solving complex dynamic tasks for
operation of ICS based on soft and quantum computing. performing collaborative work that does not could be per-
formed by each element of the system individually in a
Noise q1 Noise q2
5
0
5
0
variety of environments without external control, control
-5
0 100 200
Noise q3
300
-5
0 100
Noise q4
200 300 or coordination. In this case, it’s described a net of weakly
5 5
0 0 interconnected robots, working together in order to solve
-5 -5
5
0 100 200
Noise q5
300
5
0 100
Noise q6
200 300
problems that go beyond individual capabilities. Different
0
-5
0 100 200 300
0
-5
0 100 200 300
nodes of such a system, as a rule, have a different level of
5
Noise q7
intellectualization (knowledge, algorithms, computational
bases) and various information resources in the design.
0
-5
0 100 200 300
solutions frameworks - add-ons, which taking into account the use of a neural network approach in the recognition
the specific features of the subject area. For example, system, significantly reduced the negative impact of exter-
the Tango Controls add-in (widely used in mega-nuclear nal factors on the quality of recognition (recognition error
projects of nuclear physics), or Robotics Operation Sys- when changing the angle of the object, changing lighting,
tem (ROS) with real time Linux. By default, parameters software sensitivity, etc.).
related to the physical features of the implementation of
the system itself, control parameters in feedback loops, 4.1 Convolutional Neural Network Architecture
are used as a custom level in such systems installed by ex- As the object of recognition was chosen a plastic cork of
perts or customizable by classical methods. Moreover, the an ordinary plastic bottle. With certain image processing,
existing tools of intellectualization have a fairly unified the module also allows tracking the object. In relation to
standard methodological approach and often do not meet the task of controlling a mobile trolley with a manipula-
modern requirements for the complexity and required tor convolutional neural network (CNN) are often used.
reliability of the developed systems. One of the main The CNN is a multi-layer sensor network, it represents a
problems of the effective use of tools and soft computing further development of the multi-layer perceptron, how-
technologies (neural networks, evolutionary algorithms, ever, unlike the latter, the convolutional network has a
fuzzy logic) in control problems was the solution of the much smaller number of weights (the principle of sharing
following problems: weights). Quite often the CNN model is divided into two
(1) an objective definition of the type of MF and its pa- main parts: the part responsible for the selection of fea-
rameters in the production rules in KB; tures, and the part with which classification is performed
(2) determination of the optimal structure of FNN in [42]
. Figure 19 shows the structure of the CNN.
deep machine learning problems (approximation of the Features of the structure of the CNN. The selection of
training signal with the required (given) error and with a features of an object occurs using layers of convolution
minimum number of production rules in KB); and subsampling. Each layer is a set of plates of neurons,
(3) the use of multicriteria GA in multicriteria control which are also called feature maps. The convolution net-
problems in the presence of discrete restrictions on the pa- work structure includes two types of layers: convolution
rameters of the CO. layers and subsampling layers. Convolution and subsam-
The listed problems were solved and tested on the pling layers are combined into macro layers (a macro lay-
basis of SCO software tools using soft computing tech- er is a convolution layer followed by a subsampling layer
nology. The developed intelligent tools made it possible [43]
). A set of several architecture-similar macro-layers is
to design robust KB based on the solution of one of the called a feature separator in a sensory neural network.
algorithmically difficult problems of the theory of artifi- Each neuron of the convolution layer and the subsample
cial intelligence - the extraction, processing and formation layer is associated with a receptive field (RF). RF is a
of objective knowledge without the use of human expert certain square area that includes neurons capable of trans-
evaluations. mitting signals to a neuron that has a given RF. Figure 20
shows a convolutional layer diagram, and also shows the
4. Object Detection and Tracking, Pattern
process of its interaction with the previous layer.
Recognition, Navigation based on Stereo Vi-
sion and Computational Intelligence
The combined use of various sensors and their combi-
nations with stereo vision technology allows for a more
qualitative and complete construction of the “world
scene” of the robot, thereby improving its interaction with
the environment. The stereo vision technology, which to
some extent repeats the features of the development of
natural vision, allows the on-board system to receive in-
formation not only about the color and brightness of the
object, but also about the distance to it, about its geomet-
ric shape, about obstacles to the object, which plays an Figure 19. General structure of the convolutional neural
extremely important role in the tasks of a mobile robot. network
The intellectualization of the control system, in particular,
input size for the classifier layer [42]. new computational paradigm that promises applications in
Remark. There are many works [45-50] that are devoted several fields, including machine learning. In the last de-
to the creation and training of the CNN. In this work the cade, deep learning, and in particular CNN, have become
recognition system based on stereo vision technology uses essential for applications in signal processing and image
the classical CNN architecture, which includes convolu- recognition. Quantum deep learning, however remains a
tion and averaging layers. Network training was done with challenging problem, as it is difficult to implement non
a teacher. In relation to the recognition problem, a teacher linearities with quantum unitaries. The quantum CNN
is the number of a class that is encoded in a vector. This (QCNN) is a shallow circuit, reproducing completely the
vector is equal to the size of the output layer of the neural classical CNN, by allowing non linearities and pooling
network. This is the desired result corresponding to this operations. The QCNN is particularly interesting for deep
input pattern. The actual response is obtained as a result networks and could allow new frontiers in image recogni-
of the reaction of the neural network with the current pa- tion, by using more or larger convolution kernels, larger
rameters on the input pattern. Error signal - the difference or deeper inputs (see Appendix 2).
between the desired signal and the current response of the The main and most important stage in the implemen-
neural network. It is on the basis of the error signal that tation of the recognition system based on the CNN is the
the tunable parameters of the neural network are corrected stage of formation of the training sample. In a specific case
[43]
. A significant minus of this training scheme is the great the training sample included 10 classes of objects (cork
difficulty in creating training samples, however with a few from a plastic bottle, ball, human face, ballpoint pen, etc.).
number of classes the negative influence of this factor can The total volume of the training sample is 50,000 images
be neglected. In the general case a deep machine learning (5,000 per class). The volume of the test sample is 10,000
quantum algorithm is used that contains the quantum su- images. A training sample was formed, partly from person-
perposition operator of all input signals with massive data ally captured images, partly from images downloaded from
processing parallelism [4]. the Internet. All images are pre-processed (32x32, RGB,
The error function depends on the system settings be- resized, mirrored) and divided according to their classifica-
ing configured. For such a function, one can construct a tion. The design, creation and training of CNN were carried
multidimensional error surface in the coordinates of free out using the Keras library. The structure of the CNN is set
parameters. In this case the real surface of the error is using this library directly in the program code, which im-
averaged over all possible examples, which are present- poses questions on the choice of the optimal structure.
ed in the form of input-output pairs. To improve system At the CNN input, the images are received in matrix
performance over time, the error value should shift to the rather than in vector form (which is necessary to save
minimum. This minimum can be both local and global [43]. information about the topology). Input image size 32x32
The most common and reliable methods for achieving a pixels, format - RGB. The first convolutional layer con-
local or global minimum on the error surface are local op- tains 32 3x3 feature maps (each with its own convolution
timization methods [51-53]. kernel), i.e. each convolutional neuron is connected to a
Remark. Many local optimization algorithms can be di- square 3x3 image. The next convolutional layer has a sim-
vided into two classes: local optimization algorithms with ilar architecture. It is known that convolution layers and
the calculation of partial derivatives of the first order (the subsampling layers are responsible for highlighting cer-
method of steepest descent, the method of one-dimen- tain attributes of various objects (borders, simple colors,
sional and two-dimensional optimization of the objective and curves) in images. Deepening into the network (going
function in the direction of the anti-gradient, the method through the following layers of convolution and subsam-
of conjugate gradients, methods that take into account the pling) allows you to define less abstract (most characteris-
direction of the anti-gradient at several steps of the algo- tic of any class of objects) features. The following depen-
rithm, etc. .) and local optimization algorithms with the dence can be traced - an increase in the number of macro
calculation of partial derivatives of the first and second or- layers (“Convolution-Subsampling”) makes it possible to
ders (Newton's method, optimization methods with sparse find more and more complex features of certain objects
Hessian matrix, quasi-Newton methods, Gauss-Newton in the image. The next step after convolution is averag-
method, Levenberg-Marquardt method). In this work, we ing (subsampling). This operation reduces the dimension
use the classical error back propagation algorithm with of feature maps obtained from the previous convolution
a further generalization to recognition algorithms based layer. This method is based on the fact that neighboring
on quantum deep machine learning using quantum neural pixels are very slightly different from each other (the so-
networks and quantum GA [54,55]. Quantum computing is a called “pixel correlation”). The averaging operation sig-
nificantly reduces the dependence of the recognition result trajectory calculated by the planning package. In this case
on the scale of the input image, and also significantly the movement correction is calculated by PID-controller,
reduces the computational load. A fully connected output while the control parameters (gain factors) are set in ac-
layer of 10 neurons contains the probability that the object cordance with the developed KB. The interaction of the
in the analyzed image belongs to a certain class. Con- intelligent control modules in the ROS system is provided
volutional neural networks are currently one of the best by the file system level. In this case it is necessary to cre-
application tools for solving recognition and classification ate a model of a controlled robot and adding this control
problems. Figure 22 shows that the system correctly rec- model to the ROS navigation system. This is necessary to
ognizes objects even when lighting is degraded. establish spatial relationships between the robot and the
objects of the surrounding world. Figure 23 shows a map
of the room obtained using the sensor system.
real time in the form of separate reactions of the corre- bots separately.
sponding individual KB optimized with various fixed
criteria of quality and control situation. Aggregation of 5. Conclusion
the obtained particular solutions in the form of a new ro-
Intensification of the production process, increasing com-
bust KB is carried out on the basis of QFI. QFI contains
plexity and the number of various technical products,
the mechanism for the formation of quantum correlation
increasing the processing power of processors indicate the
between the obtained particular solutions. As a result,
necessity for development of design technologies and the
only reactions of a finite number of individual KB’s
introduction of intelligent systems in the control loop. In
containing extremely achievable control laws in this
turn, it is necessary to have intelligent tools to configure
unforeseen situation are used. The control laws of fuzzy
similar systems in various subject areas. An increasing
PID-controller formed by the new robust KB have a sim-
in the computational basis (types of calculations) of a
pler physical implementation and as a result contain the
software product leads to an increase in the quality and
best characteristics of individual control quality criteria
reliability of control, adding the properties of adaptation
for an unforeseen control situation. It is also necessary
and learning to the feedback loop. However, increasing
to mention that according to the Thermodynamics trade-
the reliability of the control system leads to an increase in
off between stability, controllability, and robustness (see
tunable parameters and, as a result, to increasing the com-
paragraph 2.3), the property of robustness (by its phys-
plexity of tuning such systems. A control system can cope
ical nature) is an integral part of the property of self-or-
with an unpredictable environment autonomously using
ganization. The required level of robustness of ICS is
different but closely related approaches:
achieved by fulfilling the principle of minimum produc-
(1) Adaptation (learning, evolution) - the system chang-
tion of generalized entropy. The principle of minimum
es its behavior to cope with the change.
entropy production in the control system is the physical
(2) Anticipation (cognition) - the system predicts a
principle of optimal functioning with a minimum of use-
change to cope with, and adjusts its behavior accordingly.
ful work. This principle underlies the development of a
This is a special case of adaptation, where the system does
robust control system (which is very important for such
not require experiencing a situation before responding to
complex systems as, for example, a team of robots). The
it.
approach based on QFI (described in this paper) guaran-
(3) Robustness - the system is robust if it continues to
tees the necessary condition for self-organization - the
function in the face of perturbations. This can be achieved
minimum of the required initial information in the train-
with modularity, degeneracy, distributed robustness, or
ing signals. The thermodynamic criterion of the mini-
redundancy.
mum of a new measure of generalized entropy produc-
(4) Modularity: firstly, the approach to incorporating el-
tion provides a sufficient condition for self-organization
ements of intelligent computing into classical control sys-
- the robustness of control processes with a minimum
tems implies non-destruction of the lower executive level
consumption of useful resource.
(which ensures minimal interposal in the CO hardware).
The experiment with multi-agent interaction, in which
Secondly, when working with complex control objects, it
the robot manipulator participated (ICS of this robot is
is necessary: ① to separate the control blocks to reduce
described in this chapter), allows us to draw the follow-
complexity, ② to organize coordination control of the
ing conclusion - the use of SCO based on quantum com-
separated blocks to improve the quality of control, which
puting can significantly increase the robustness of the
QFI does well described on the example of the ICS based
control system. The toolkit allows to configure complex
on SCO on quantum computing of 7DoF manipulator in
control systems with many of control loops; the number
this article.
of input variables and the size of the resulting KB are
Successful self-organizing systems will use combina-
limited only by the hardware characteristics of the com-
tions of these approaches to maintain their integrity in a
puter and CO. The developed interaction options show
changing and unexpected environment. Adaptation will
the possibilities of using these systems in a wide range of
enable the system to modify itself to “fit” better within
tasks, such as automation of warehouses and production
the environment. Robustness will allow the system to
facilities, automatic catering establishments, emergency
withstand changes without losing its function or purpose,
response, etc. The main feature of a multi-agent system
and thus allowing it to adapt. Anticipation will prepare the
is the synergistic effect arising from the combination of
system for changes before these occur, adapting the sys-
several robotic systems, when the result of the functional
tem without it being perturbed. The main components and
interaction far exceeds the capabilities of each of the ro-
their interrelations in the information design technology
are based on new types of (soft and quantum) computing. limited set of sensors and in the presence of internal and
The key point of this information design technology is the external disturbing influences.
use of the method of eliciting objective knowledge about Appendix 1: Emotional Learning and Its utilization
the control process irrespective of the subjective experi- in Control Engineering [56]. The Limbic System, as part
ence of experts and the design of objective KB’s of a FC, of the mammalian creatures' brain, is mainly in charge
which is principal component of a robust ICS. The devel- of the emotional processes. The Limbic System locat-
oped SCO toolkit implements mechanisms for creating, ed in the cerebral cortex consists mainly of following
configuring, and transmitting control parameters in the components: Amygdala, Orbitofrontal Cortex, Thal-
form of control signals received from KB of FC without amus, Sensory Cortex, Hypothalamus, Hippocampus
destroying the low executive level. The use of soft com- and some other less important areas. In this section, we
puting and developed ICS design technologies reduces the try to describe briefly these main components and their
impact of expert judgment in the training and configura- tasks. Figure A1.1 illustrates the anatomy of the main
tion of ICS. components of Limbic System. The first sign of affective
From computer science viewpoint, QA of QFI model conditioning of the system appears in Amygdala which
plays the role of the information algorithmic and SW-plat- is a small almond-shaped in sub-cortical area. This com-
form support for design of self-organization process. ponent is placed in a way to communicate with all other
Quantum computing, which ensures robustness of ICS, Sensory Cortices and areas within the Limbic System.
introduces the property of self-organization, allows CO The Amygdala connections to/from other components
to function effectively in conditions of a lack of a priori are illustrated in Figure A1.2. The studies show that a
information. The use of this kind of computing can reduce stimulus and its emotional consequences are associated
the influence of the hardware (for example, reducing the in the Amygdala area. In this region, highly analyzed
number of sensors) on the effectiveness of the control stimuli in the Sensory Cortices, as well as coarsely cat-
system. Technologies for remote configuration and trans- egorized stimuli in the Thalamus are associated with an
mission of KB allow the CO to receive KB from the SCO emotional value.
unit or from other CO, which makes it possible to manage In a reciprocal connection, the Orbitofrontal Cortex,
structurally new objects, such as teams of robots, multi- as of another component of the brain system, interacts
agent systems, complex automated production, etc. In with the Amygdala. The main interrelated function of
addition, this technology gives CO ability to update and this component is: Working Memory, Preparatory Set and
adapt KB for a specific control situation, including contin- Inhibitory Control. The current and recent past events
gency. are represented in the Working Memory. The Preparatory
The experiments described above demonstrate that the Set is the priming of other structures in anticipation of
inclusion of QFI module in ICS based on SCO with soft impending action. Inhibitory Control is the selective sup-
computing made it possible to provide a complete solu- pression of areas that may be inappropriate in the current
tion to the positioning problem in standard control situa- situation. More specifically, the Orbitofrontal Cortex takes
tions and in the conditions of external unforeseen control action in omission of the expected reward or punishment
situations. The QFI module significantly improved the and control the extinction of the learning in the Amyg-
accuracy index of the positioning of the functional device dala. Another component in this area is Thalamus which
under strong internal disturbances, the performance index lies next to the basal ganglia. It is a non-homogeneous
improved 5 times, and the relative error in the positioning sub-cortical structure and a way-station between cortical
of the links decreased. It is important to note that we must structures and sub-cortical. Moreover, various parts of the
to organize the separation of control during the design of Thalamus also relay the majority of sensory information
ICS for complex CO’s (precisely such CO’s appear in the from the peripheral sensory systems to the Sensory Cor-
experiments described above). This is necessary for the tices. Particularly, the Thalamic Sensory Inputs going to
organization of coordination management without signifi- the Amygdala are believed to mediate inherently emotion-
cantly increasing the complexity of the control system. ally charged stimuli as well as coarsely resolved stimuli
However, such a management decomposition often leads in general. The Thalamus signal going to the Amygdala
to a mismatch of work and a decrease in the quality of evades the processes involved in the Sensory Cortex and
management. In the example of multi-agent interaction other components of the system. Therefore, Amygdala re-
QFI module, using data from stereo vision module, suc- ceives a non-optimal but fast stimulus from the Thalamus
cessfully generates a generalized robust control signal. which among the input stimuli is often known as a charac-
Moreover, control robustness is achieved even with a teristic signal.
Jafarzadeh et al. proposed an intelligent autopilot for a in terms of overshot, settling time, steady state error and
2-DoF helicopter model based on BELBIC. The majority of smoothness. These made the controller effective and flex-
previous systems were based on the linearization model or ible in high performance applications. Moreover, Simple
through several linearization techniques for helicopter that structure, fast auto learning and high tracking potency of
made the proposed controls unreliable. The designed model BELBIC have been made to present a new control plant
used a BELBIC controller and feedback linearization tech- that is independent of motor parameters and controls
nique to a nonlinear model of a helicopter. In this method, the speed and flux simultaneously.
states of the system have been separated into two parts, and As a demonstration of the performance of the BELBIC
each part has been controlled by one of the control inputs. controller in real-world applications beyond numerical
The performance of the two mentioned controllers simulated simulation, it will be implemented on a KUKA Light-
the in Simulink. The simulation results of controller system weight Robot (LWR), a 7DoF KUKA LWR4+ [57]. Bio-in-
by BELBIC controller and feedback linearization controller spired by the human arm and with a payload of 7 kg and
have been demonstrated in the Figure A1.5, and in Figure its 7 axes, all equipped with internal position as well as
A1.6, the control inputs of the system have been shown. force-/torque-sensors, this redundant robot offers a range
of features which are essential for the considered appli-
cation. The seven revolute joints of the industrial robot
are driven by brushless motors via harmonic drives. The
working envelope is described in Figure A1.7.
disturbances are added as time-dependent 2-dimensional The simulation verifies the tracking performance of
function. Internal uncertainties are added in the form of the BELBIC-SMC-controller for the chosen case. The
maximal +/-10% deviations from the dynamic parame- BELBIC-SMC-controller outperforms both, the PID-
ters-values. Switching constraints are implemented in the and the Computed Joint Torque-controller. The free-
form of a desired trajectory switching between different dom-shape and especially the discontinuity of the path
curved and straight segments as shown in Figure A1.8. The deteriorate the tracking performance of the conventional
desired continuous trajectory is not differentiable. controllers.
Appendix 2: quantum computing in deep convolu-
tional neural networks [58]. The growing importance of
deep learning in industry and in our society will require
extreme computational power as the dataset sizes and the
complexity of these algorithms are expected to increase.
Quantum computers are a good candidate to answer this
Figure A1.8. Trajectory to be tracked by the tool on the challenge. The recent progress in the physical realization
workpiece surface in the simulated application; the arrows of quantum processors and the advances in quantum al-
indicate the movement-direction gorithms increase more than ever the need to understand
their capabilities and limits. In particular, the field of
Table A1.1 shows the performance of the three considered
quantum machine learning has witnessed many innovative
controllers for trajectory tracking of discontinuous freeform
algorithms that offer speedups over their classical coun-
paths, exhibited on the example shown in Figure A1.8.
terparts. Quantum deep learning, the problem of creating
quantum circuit that enhance the operations of neural net-
Table A1.1. Absolute maximal, minimal and mean posi-
tion errors for both manipulator-links using PID-, Com- works, has been studied in several works. It however re-
puted Joint Torque (CJT)-control and BELBIC-SMC mains a challenging problem as it is difficult to implement
non linearities with quantum unitaries. CNN are a type of
deep learning architecture well suited for visual recogni-
tion, signal processing and time series. In this work we
propose a quantum algorithm to perform a complete con-
volutional neural network (QCNN) that offers potential
In order to demonstrate the efficiency of the suggested
speedups over classical CNNs.
controller in real-world applications beyond numerical
The adaptation of the CNNs to the quantum setting
simulation, an experimental validation is performed. The
implies some modifications that could alter the efficien-
first experiment is a goal-reaching task. The task is about
cy of the learning or classifying phases. In work [58] were
moving the robot’s joints consecutively to a specified goal
presented some experiments to show that such modified
position. Table A1.2 shows that the BELBIC-SMC-concept
CNNs can converge correctly, as the original ones. The
outperforms PID- and Computed Joint Torque-control.
experiment, using the PyTorch library, consists of train-
Table A1.2. Absolute maximal, minimal and mean posi- ing classically a small convolutional neural network for
tion errors for all manipulator-links using PID-, Comput- which was added a “quantum” sampling after each con-
ed Joint Torque (CJT)-control and BELBIC-SMC when volution. In the following results, we can see that r quan-
reaching a desired goal position tum CNN is able to learn and classify visual data from
the widely used MNIST dataset. This dataset is made
of 60.000 training images and 10.000 testing images of
handwritten digits. Each image is a 28x28 grayscale pix-
els between 0 and 255 (8 bits encoding), before normal-
ization.
Let’s first observe the “quantum” effects on an image
of the dataset. In particular, the effect of the capped non
linearity, the introduction of noise and the quantum sam-
pling (see Figure A2.1).
tensor sampling and other modifications. In particular it that a correlational measurement is made on the output
shows that only a fraction of the information is meaning- quantum state and a scalar is obtained. Gathering the scalar
ful for the neural network, and that the quantum algorithm outputs, the final output of the quantum convolutional layer
captures this information in priority. This learning can be is a 3-dimensional array. Then a pooling layer is used to
more or less efficient depending on the choice of the key reduce the dimensionality of the data. This process could be
parameters. For reasonable values of these parameters, the repeated and finally ends with a fully connected layer. The
QCNN is able to converge during the training phase. It single-qubit layer consists of Ry gates, each containing one
can then classify correctly on both training and testing set, tunable parameter. The two-qubit layer consists of CNOT
indicating that it does not overfit the data. The learning gates on nearest-neighbour pairs of qubits
curves sometimes present a late start before the conver- Concretely, the foreground pixels are represented by
gence initializes, in particular for small sampling ratio. random floating numbers ranging from 0.7 to1, whereas
This late start can be due to the random initialization of the background are small floating numbers ranging from
the kernel weights, that performs a meaningless convo- 0 to 0.1. There are 5 labels, namely S, O, I, T and L, each
lution, a case where the quantum sampling of the output of which represents a type of Tetris bricks. The dataset is
is of no interest. However it is very interesting to see that further processed by randomly splitting into a training set
despite this late start, the kernel can start converging once and a testing set that contain 80% and 20% of the images,
they have found a good combination. Overall, it is possi- respectively. This QCCNN was compared with CNN with
ble that the QCNN presents some behaviors that do not two particular structures, namely one with a single convo-
have a classical equivalent. Understanding their potential lutional layer and another with two convolutional layers. To
effects, positive or negative, is an open question, all the see the performances with a different number of labels, re-
more so as the effects of the classical CNN’s hyperparam- searchers create another dataset by only picking the two la-
eters are already a topic an active research. bels S, T out of the original training and testing data. For the
In work [59] demonstrates the potential of Quan- single-layer structure, they use a single (quantum) convo-
tum-Classical Convolutional Neural Networks (QCCNN) lutional layer with 5 filters with no padding, plus a pooling
by applying it to the Tetris dataset. Researchers create a layer also with no padding. For the two-layer structure, they
Tetris image dataset that consists of 1000 grey-scale im- use two (quantum) convolutional layers with 2 and 3 filters
ages with shape 3×3, in which each grey-scale image is a respectively, plus a pooling layer with padding 1. The win-
simulated Tetris brick. dow shape for all the layers is 2 × 2, and the stride value s =
1. Therefore, the number of qubits fed to the quantum filter
is 4, and the depth of the parametric quantum circuit is set
as 4. During 1000 iterations, it was computed the accuracy
on the testing data and store the values of the loss function,
which is chosen as mean square loss. In Figure A2.3 (a, c),
researchers plot the accuracy and loss values for the 2-label
case. While in Figure A2.3(b, d) they plot the accuracy and
loss values for the 5-label case.
CNN and the black line represents the result of two-layer CNN, the blue In particular, we can create a high pass filter for edge
dashed line represents the result of one-layer QCCNN and the black
dashed line represents the result of two-layer QCCNN. The results are
detection (Figure A2.4).
averaged over 10 random simulations. (a) Accuracy in case of 2 labels. (b)
Accuracy in case of 5 labels. (c) Loss in case of 2 labels. (d) Loss in case
of 5 labels.
We can see that QCCNN can reach almost 100% ac-
curacy for both the two structures they have used, and it
can reach much lower loss values for both cases compared
to its classical counterpart. Benefiting from the high-di-
mensional nature of the quantum system, the advantages
of QCCNN become more transparent when the number
of labels increases from 2 to 5. We can also see that the Figure A2.4. A conceptual image of the high pass filter
5-label case takes more iterations to converge than the applied to quantum multiple images
2-label case, and that QCCNN with a two-layer structure There are several issues remaining to be solved. The
converges faster than the single-layer structure, especially first one is how to effectively generate a quantum super-
in the 5-label case, which indicates that for complex prob- position of multiple data sets. Though, in principle, one
lems, better performance could be achieved by deeper can generate it using the Hadamard gate and the Pauli-X
architectures. In summary, researchers present a hybrid gate, it is still a difficult problem to superpose arbitrary
quantum-classical Convolutional Neural Network which data sets with arbitrary probability amplitudes. The second
could be used to solve real world problems with current is how to use the resultant multiple data sets obtained by
quantum computers. As a quantum machine learning al- performing the QFFT. The QFFT sustains all the informa-
gorithm inspired by classical CNN, QCCNN keeps the tion of Fourier coefficients until the moment the quantum
features of CNN such as the nonlinearity, locality of the state is measured. If the quantum state that contains the
convolutional layer, as well as extensibility to deep struc- Fourier coefficients of multiple data sets was passed on to
tures. Moreover, the generalized feature map with a para- some quantum device directly and there were some prop-
metric quantum circuit is able to explore the correlations er techniques to handle it, it would play a key role in the
of neighbouring data points in a exponentially large linear field of quantum machine learning.
space, hopefully allowing this algorithm to capture the
patterns in the dataset more efficiently or precisely with a References
quantum computer.
[1] Sergei V. Ulyanov, Kazuo Yamafuji, Valery G. Gra-
In [60], authors have discussed the implementation of
detsky, Andrea Pagni. Expert fuzzy-neuro controller
the fast Fourier transform (FFT) as a quantum circuit.
design for wall climbing robot for decontamination
The quantum version of the FFT (QFFT) is defined as
of nuclear-power station[J]. 1995, 7(1): 75-85.
a transformation of a tensor product of quantum states.
[2] Ulyanov S.V., Yamafuji K., Fukuda T. Development
The QFFT has been constructed by a combination of
of intelligent mobile robots for service use and mo-
several fundamental arithmetic operators such as an
bile automation systems including wall climbing
adder, subtractor and shift operators which have been
robots: Pt. 1: Fundamental design principles and
implemented into the quantum circuit of the QFFT
motion models[J]. Intern. J. of Intelligent Mecha-
without generating any garbage bits. One of the ad-
tronics, 1995, 1(3): 111-143.
vantages of the QFFT is due to its high versatility: the
[3] Takayuki Tanaka, Junji Ohwi, Ludmila V. Litvintse-
QFFT are applicable to all the problems that can be
va, Kazuo Yamafuji, Sergei V. Ulyanov. Intelligent
solved by the conventional FFT. The frequency domain control of a mobile robot for service use in office
filtering of digital images is one of the possible appli- buildings and its soft computing algorithms[J].
cations of the QFFT. The major advantage of using the Journal of Robotics and Mechatronics, 1996, 8(6):
QFFT lies in its quantum superposition: multiple im- 538-552.
ages are processed simultaneously (see Figure A2.4 as [4] Ulyanov S. V. Quantum fast algorithm computa-
a conceptual image of quantum parallelism). It is even tional intelligence Pt I: SW/HW smart toolkit [J].
superior to the QFT when the number of images is suf- Artificial Intelligence Advances, 2019, 1(1): 18-43.
ficiently large. Thus, the QFFT will be a great help in [5] T. Tanaka, J. Ohwi, L.V. Litvintseva, K. Yamafuji,
the industries including security, healthcare, and mar- S.V. Ulyanov. Soft computing algorithms for intel-
keting research, where numerous images are involved. ligent control of a mobile robot for service use Pt
II: Path planning, navigation and technology opera- upravlenija, 2007, 6: 71-126.
tions[J]. Soft Computing, 1997, 1: 99-106. [17] Litvintseva L. V., Ulyanov S. V. Intellektualnije
[6] Ulyanov S. V., Litvintseva L. V., Sorokin S. V. sistemy upravlenija. Kvantovije vychislenija I algo-
Robastnost’ baz znanij samoorganizujuschihsya ritm samoorganizatsii[J]. Izvestiya RAN. Teoriya I
intellectual’nih regilyatorov v nepredvidennih sit- sistemy upravlenija, 2009, 6: 102-141.
uatsiyah upravlenija[J]. Fuzzy Systems and Soft [18] Nielsen M.A., Chang I.L. Quantum Computation
Computing, 2008, 3(4): 49-73. and Quantum Information. Cambridge Univ. Press.
[7] R. Mohammadi, C. Lucas and B. N. Araabi, A novel UK, 2000.
controller for a power system based BELBIC, Pro- [19] Ulyanov S. V., Sorokin S. V., Litvintseva L. V. Opti-
ceedings of World Automation Congress, 2004, 18: mizator baz znaniy na osnove kvantovyh vychislenij
409-420. dlya proektirovanija samoorganizujuschihsja nech-
[8] Litvintseva L. V., Ulyanov I. S., Ulyanov S. V., Uly- etkih reguljatorov: programmnyj instrumentarij[J].
anov S. S. Quantum Fuzzy Inference for Knowledge System analyze in sience and education, 2012, 1.
Base Design in Robust Intelligent Controllers[J]. [20] A.J. van der Schaft. Theory of port-Hamiltonian
Journal of Computer and Systems Sciences Interna- systems (Lectures 1, 2, and 3), Network Modeling
tional, 2007, 46(6): 908-961. and Control of Physical Systems. DISC, 2005.
[9] Ulyanov S.V. System and method for control us- [21] Ulyanov S.V. “Self-organized control system,” US
ing quantum soft computing)[P]. US Patent No patent, No. 6, 411, 944, 1997; see also “Self-organi-
7,383,235 B1, 2003; EP PCT 1 083 520 A2, 2001; zation fuzzy chaos intelligent controller for robotic
Efficient simulation system of quantum algorithm unicycle: A new benchmark in AI control,” Pro-
gates on classical computer based on fast algo- ceedings 5th Intelligent System Symposium, Tokyo,
rithm[P]. US Patent No 883 2006/0224547 A1, Sep. 29 -30, 1995: 41-46.
2006. [22] Haddad W. M., Chellaboina V., Nersesov S. G.
[10] Ulyanov, S. V. Intelligent Robust Control System Thermodynamics: A Dynamical Systems Approach,
Based on Quantum KB-Self-organization: Quantum Princeton Series in Applied Mathematics Princeton.
Soft Computing and Kansei / Affective Engineering NJ: Princeton University Press, 2005.
Technologies[J]. Springer International Publishing, [23] Ulyanov S.V. System and method for stochastic
2014: 37-48. simulation of nonlinear dynamic systems with high
[11] Ohwi J., Ulyanov S.V., Yamafuji K. GA in contin- degree of freedom for soft computing application.
uous space and fuzzy classifier system for opening US patent, No 0039555 A1, Pub. Date: Feb. 26,
a door with a manipulator of mobile robot: New 2004.
Benchmark of evolutionary intelligent comput- [24] Demetriusa L. Manke T. Robustness and network
ing[J]. Robotics and Mechatronics, 1996, 8(3): 297- evolution - an entropic principle. Physica, 2005,
301. A346: 682-696.
[12] Fujii S., Ulyanov S.V. A model for motorcycle rider [25] Ulyanov S.V., Litvintseva L.V. Design of self-orga-
operation based on genetic algorithm [J]. Yamaha nized intelligent control system based on quantum
Motor Technical Review, 2004. fuzzy inference: Intelligent system of systems en-
[13] Hagiwara T., Ulyanov S.V. An application of a gineering approach[R]. Proc. of IEEE int. conf. on
smart control suspension system for a passenger car system, man and cybernetics (SMC’2005). Hawai,
based on soft computing[J]. Yamaha Motor Techni- USA, 4.
cal Review, 2003. [26] Reshetnikov A.G., Ulyanov S.V. Knowledge extract
[14] Ulyanov S.V. Intelligent self-organized robust method from physically measured teaching signal:
control design based on quantum/soft computing knowledge base design for fuzzy controller. System
technologies and Kansei Engineering[J]. Computer analyze in sience and education, 2013, 1.
Science J. of Moldova, 2013, 21, 2(62): 242 - 279. [27] Zurek W.H. Probabilities from entanglement, Born’s
[15] Ulyanov S.V. Self-organizing quantum robust con- rule from envariance. Phys. Review, 2005, A71(5).
trol methods and systems for situations with uncer- [28] Goldenblat I.I., Ulyanov S.V. Introdaction to rela-
tainty and risk [P]. Patent US 8788450 B2, 2014. tivity theory and its applications to a new technique.
[16] Litvintseva L. V., Ulyanov S. I., Ulyanov S. V., Moscow, Fizmatgiz Publ., 1979.
Ulyanov S. S. Kvantoviy nechetkiy vyvod dlya [29] Petrov B.N., Ulanov G.M., Ulyanov S.V. Control
sozdanija baz znaniy v robastnih intellektualnyh problems of relativistic and quantum dynamic sys-
regulyatorah[J]. Izvestiya RAN. Teoriya I sistemy tems. Moscow, Nauka Publ., 1982.
[30] Li Y., Ang K.H., Chong G.C.Y. Patents, software Programmnije produkti i sistemy, 2017, 4: 593-600.
and hardware for PID control: an overview and [42] Nemkov R. M. Development of neural network al-
analysis of the current art. IEEE Control Syst. Mag, gorithms for invariant pattern recognition: disserta-
2006, 26(1): 42-54. tion of the candidate of technical sciences: 05.13.18.
[31] Nielsen M.A., Chuang I.L. Quantum Computation Stavropol’, 2015: p.
and Quantum Information. - Cambridge Univ. Press, [43] Haykin S.. Neural networks: full course. Moscow,
UK, 2000. 2008: 113+281-330.
[32] Marinescu D.C., Marinescu G.M. Approaching [44] Gonzalez R. Digital image processing. Moscow,
quantum computing. Pearson Prentice Hall, USA, 2012: p.
2005. [45] Ranzato M. A., Jarrett K., Kavukcuoglu K., LeCun
[33] Ulyanov S.V., Litvintseva L.V., Ulyanov S.S. et all. Y. What is the best multi-stage architecture for ob-
Quantum information and quantum computational ject recognition? In ICCV, 2009.
intelligence: Backgrounds and applied toolkit of [46] Richard S. Sutton. Reinforcement training. Mos-
information design technologies. Milan: Note del cow, 2011: 399.
Polo (Ricerca), Universita degli Studi di Milano, [47] Ackley D. H., Hinton G. E., Sejnowski T. J. A learn-
2005, 78-86. ing algorithm for Boltzmann machines. Cognitive
[34] Ulyanov S.V., Litvintseva L.V., Ulyanov S.S. et all. Science, 1985, 9: 147-169.
Self-organization principle and robust wise control [48] Bengio Y. Learning deep architectures for AI. Foun-
design based on quantum fuzzy inference[R]. Proc. dations and Trends in Machine Learning, 2009,
of Internat. Conf. ICSCCW’2005, Antalya. Turkey, 2(1): 1-127.
2005. [49] Ciresan D., Meier U., Schmidhuber J. Multicolumn
[35] Litvintseva L.V., Ulyanov S.V., Takahashi K. et all. Deep Neural Networks for Image Classification[R].
Design of self-organized robust wise control sys- In Proceedings of the 2012 IEEE Conference on
tems based on quantum fuzzy inference[R] Computer Vision and Pattern Recognition (CVPR),
Proc. of World Automation Congress (WAC’2006): CVPR '12, pp.3642-3649, Washington, DC, USA,
Soft computing with Industrial Applications (ISS- 2012. IEEE Computer Society.
CI’2006). Budapest, Hungary, 2006, 5. [50] Ranzato M., Huang F., Boureau Y., LeCun Y. Un-
[36] Litvintseva L.V., Ulyanov S.V., Ulyanov S.S. Pos- supervised learning of invariant feature hierarchies
troenije robastnih baz znanij nechetkih regulyatorov with applications to object recognition[R]. In Pro-
dlya intellektual’nogo upravlenija suschestvenno ceedings of the Computer Vision and Pattern Rec-
nelinejnymi dinamicheskimy sistemamy[J]. Izvesti- ognition Conference (CVPR’07), IEEE Press, 2007.
ya RAN. Teoriya I sistemy upravlenija, 2006(5). [51] Gill F. Practical optimization. Moscow, 1985: 509.
[37] Nikolaeva A. V., Ulyanov S. V. Proektirovanije [52] Izmajlov A. F. Numerical optimization methods.
intellectual’noj sistemy upravlenija robotom manip- Moscow, 2005: 304.
ulyatorom. Ch. 3: Modelirovanije I fizicheskij ek- [53] Tarkhov D. A. Neural network models and algo-
speriment na osnove tehnologii kvantovyh vychis- rithms. Moscow, 2014: 352.
lenij[J]. System analysis in sience and education, [54] Sergey Ulyanov, Andrey Reshetnikov, Nikita Ry-
2013, 1. abov. Deep machine learning and pattern/face rec-
[38] Ulyanov S. V., Litvintseva L.V., Mishin A. A., So- ognition based on quantum neural networks and
rokin S. V. Kvantovyj optimizator robastnyh baz quantum genetic algorithm[R]. Distributed Comput-
znanij dlya proektirovanija intellektual’nyh system ing and Gridtechnologies in Science and Education.
upravlenija: RF patent №2011616869. Pub. Date Book of abstract of the 8th International Conference
5.09.2011 (Russia). Dubna, 10 - 14 September, 2018: 38.
[39] Tarasov V. B. Ot mnogoagentnyh system k intellek- [55] Ulyanov S. V., Petrov S. P. Kvantovoe raspozna-
tul’nym organisatsyjam. Moscow, 2002: 352. vanije lits i kvantovaja visual’naja kriptografija:
[40] Lebedev B. K., Lebedev V. B. Planirovanije na os- modeli i algoritmi[J]. System analyze in sience and
nove roevogo intellekta I geneticheskoj evolutsyi education, 2012(1).
[J]. Izvestija Juzhnogo federal’nogo universiteta, [56] Zahra Beheshti, Siti Zaiton Mohd Hashim. A Re-
2009, 93(4). view of Emotional Learning and It’s Utilization in
[41] Ulyanov S. V., Reshetnikov A. G. Synergetics of Control Engineering. Int. J. Advance. Soft Comput.
information-cognitive interaction in intelligent ro- Appl., 2010, 2(2): 191-208.
botic systems with remote knowledge exchange[J]. [57] Klecker S. Robotic Trajectory Tracking: Position-
ARTICLE
Intelligent Robust Control of Redundant Smart Robotic Arm Pt II:
Quantum Computing KB Optimizer Supremacy
Alena V. Nikolaeva1,2 Sergey V. Ulyanov3,4*
1. Dubna State University, Russia
2. INESYS LLC (EFKO GROUP), Russia
3. Physical and Mathematical Sciences; JINR, Russia
4. LIT JINR, INESYS LLC (EFKO GROUP), Russia
Article history This article is a continuation of the work “Intelligent robust control of
Received: 7 November 2019 redundant smart robotic arm Pt I: Soft computing KB optimizer - deep
machine learning IT”. In the first part of the paper, we examined control
Accepted: 24 April 2020 systems with constant coefficients of the conventional PID controller
Published Online: 30 June 2020 (based on genetic algorithm) and intelligent control systems based on soft
computing technologies. For demonstration, MatLab / Simulink models
Keywords: and a test benchmark of the robot manipulator demonstrated. Advantages
Quantum computing supremacy and limitations of intelligent control systems based on soft computing
technology discussed. Intelligent main element of the control system
Quantum-classical correlation based on soft computing is a fuzzy controller with a knowledge base in
Knowledge base it. In the first part of the article, two ways to implement fuzzy controllers
Fuzzy controller showed. First way applyied one controller for all links of the manipulator
and showed the best performance. However, such an implementation is
Quantum fuzzy inference
not possible in complex control objects, such as a manipulator with seven
degrees of freedom (7DOF). The second way use of separated control
when an independent fuzzy controller controls each link. The control
decomposition due to a slight decrease in the quality of management has
greatly simplified the processes of creating and placing knowledge bases.
In this Pt II, to eliminate the mismatch of the work of separate indepen-
dent fuzzy controllers, methods for organizing coordination control based
on quantum computing technologies to create robust intelligent control
systems for robotic manipulators with 3DOF and 7DOF described. Quan-
tum supremacy of developed end-to-end IT design of robust intelligent
control systems demonstrated.
*Corresponding Author:
Sergey V. Ulyanov,
Physical and Mathematical Sciences; JINR, Russia; LIT JINR, INESYS LLC (EFKO GROUP), Russia;
Email: ulyanovsv@mail.ru
1. Intelligent Control System of Manipulator (6) Decoding. The calculation of the output value of the
with Seven Degrees of Freedom corresponding signal.
(7) Denormalization. Output value calculation.
1.1 Intelligent Control System Based on Quan- The connection diagram of the QFI block in the Intel-
tum Computing Knowledge Base Optimizer ligent Control System (ICS) is shown in Figure 1 where:
(QCOptKBTM) QREF is the reference signal, Q ′ is the measured variable,
S
uch unconventional phenomena as entangled states, E is the control error, s(t) is a limitation of the control ac-
teleportation, superdense coding appear in the tion, mc(t),mMS(t) are noises in the control channels and in
quantum logic of computational intelligence and the measuring system, U is the control action, d(t) is a de-
quantum information theory, due to the failure to fulfill lay in the Measuring System (MS), fo(t) is external action
the classical law of distributivity. on the CO, Q is an adjustable value.
Quantum computing technologies for creating self-or-
Intelligent superstructure
ganizing Knowledge Bases (KB) of intelligent regulators KBO based on
soft computing
are considered in [1,2]. In [1] a new type of quantum search
FC1
algorithm described on the generalized space of the KB
ck
FC2 QFI
edba
designed for soft computing, and the quantum fuzzy al- ...
nt fe
FCi
gorithm for designing a generalized robust control signal
llige
applied. The Quantum Fuzzy Inference (QFI) [3] imple-
Inte
Quantum FC
ments the self-organization of the KBs, based on the
physical laws of the theory of quantum computing [4, 5] and Qref Ε PID- U Q
CO
the application of four operators: superposition, quantum Controller
correlation, interference, and measurement. Operators su- Q' s(t) mC (t) d(t) fCO (t)
perposition, quantum correlation and interference are uni-
tary, invertible quantum operators, and the measurement MS
Let us consider an example when, in the designed ICS for a specific CO (a 7DOF robot manipulator) and de-
based on KBO on soft computing with separate control, scribed in [12]:
FC1 contains information about the standard situation 1 (1) Position Task Solution in known control situations
for link 1, FC2 contains information about the standard PTSKCS.
situation 1 for link 2, etc. (2) Position Task Solution in the external considered
Example. The QFI block in the ICS model based on control situations PTSACCS1.
KBO on Soft Computing implements the self-organization (3) Position Task Solution in the internal considered
of the FC KB in learning situations. In this case, it is pos- control situations PTSACCS2.
sible to combine information about the standard situation (4) Performance IT.
1 for all links, and to extract additional information (by (5) Relative overshoot value σ .
methods of quantum computing and quantum information (6) Relative error in link positioning å .
theory) from the reaction of the designed KBs to imple- (7) Time of one iteration t.
ment robust control in regular and unexpected control sit- (8) The complexity of implementing control P.
uations (which are not included in existing designed KB). (9) Full Control Behavior FCB.
The KBs, whose responses used to design robust control,
in unexpected situations may not be robust. 1.4 The Quality of an Intelligent Control System
The connection diagram of the QFI block in the intelli- with QFI
gent superstructure of the control system shown in Figure Control system models of 7DOF robot manipulator in
2, on which blocks designed using soft and quantum-com- MatLab/Simulink developed to research the quality of
puting technologies highlighted. control systems.
K1′ = [K P′ 1 K D′ 1 K I′1 ] K1=[KP1 KD1 KI1]
The initial ICS on KBO on soft and quantum comput-
TS1
FC1
KB1
GA1
q1
ing contains seven FCs, each of which can correlate with
K 2′ = K P′ 2 K D′ 2 K I′ 2 ]
QFI
K2=[KP2 KD2 KI2]
FC2 TS2 q2 one or more NR systems.
GA2
... KB2
Consider six combinations:
K 7′ = [K P′ 7 K D′ 7 K I′ 7 ] K7=[KP7 KD7 KI7]
FC7 TS7
GA7
q7 (1) The correlation of 2 adjacent fuzzy KBs (the infor-
KB7
Quantum computing mation FC i and i+1 is used to control the i-th link of the
Soft computing technologies
technologies
manipulator);
Intelligent superstructure on KBO obased on soft computing and a generalizing QFI block on
KBO based on quantum computing (2) the correlation of 3 adjacent fuzzy KBs (the infor-
mation FC i, i+1 and i+2 is used to control the i-th link of
Figure 2. Intelligent add-on structure of the ICS based on
KBO on quantum and soft computing the manipulator, as shown on Figure 3);
(3) the correlation of 4 adjacent fuzzy KBs;
The ICS using the QFI called the ICS based on quan- (4) the correlation of 5 adjacent fuzzy KBs;
tum computing KBO. (5) the correlation of 6 adjacent fuzzy KBs;
(6) the correlation of all 7 fuzzy KBs.
1.3 Test Procedure and Quality Criteria
For selected six combination spatial, spatio-temporal
A series of experiments in standard and unexpected con- and temporal correlations are considered.
trol situations carried out to test the robustness of control
system models. The following configuration is taken as the
1
Quantum
FC1 correlation Link 1
initial position of the manipulator: q1 = 0 deg; q2 = 0 deg; q3 KB1 operator
2
Quantum
= 0 deg; q4 = 0 deg; q5 = 0 deg; q6 = 0 deg; q7 = 0 deg. FC2
KB2
correlation
operator
Link 2
4
Quantum
FC4 correlation Link 4
change in initial conditions; unexpected situations include KB4 operator
5
Quantum
limitation of control actions, the introduction of noise into FC5
KB5 correlation
operator
Link 5
7
Quantum
FC7 correlation Link 7
methods for assessing transients of automatic control the- KB7
operator
ory [11] was introduced to evaluate and compare the test 7DOF Robot
manipulator
results of the considered types of control systems (control Soft computing Quantum computing
systems with constant coefficients and ICS for KBO on Figure 3. The application of the correlation of three
Soft and Quantum computing). The methods were adapted neighboring FC
Performance
overshoot value
Relative
Behavior
iteration
Full Control
in link positioning
Time of one
Relative error
implementing control
The complexity of
control situations
Internal considered
control situations
External considered
situations
Known control
Spatial correlation
Spatio-temporal
correlation
correlation
Time of one iteration
Behavior
Full Control
situations
Known control
control situations
control situations
Performance
implementing control
External considered
Internal considered
The complexity of
value
Relative overshoot
Spatial
correlation
Spatio-temporal
Figure 4. The result of the work of the ICS on quantum correlation
computing KBO using the correlation of 2 FC
Temporal
correlation
Performance
overshoot value
Relative
Behavior
iteration
Full Control
in link positioning
Time of one
Relative error
implementing control
The complexity of
control situations
Internal considered
control situations
External considered
situations
Known control
Spatial correlation
Spatio-temporal
correlation
Temporal
correlation
Figure 9. The result of the work of the ICS on quantum
link positioning
Relative error in
Behavior
Full Control
situations
Known control
control situations
control situations
Performance
implementing control
External considered
Internal considered
The complexity of
value
Relative overshoot
Behavior
Full Control
situations
Known control
control situations
control situations
Performance
implementing control
External considered
Internal considered
The complexity of
value
Relative overshoot
Behavior
Full Control
situations
Known control
control situations
control situations
Performance
implementing control
External considered
Internal considered
The complexity of
value
Relative overshoot
Table 1. Comparison ICS based on quantum computing and ICS based on soft computing
ICS based on
Quantum computing
Behavior FCB
Relative
Full Control
iteration
Time of one
implementing control
control situations
The complexity of
Internal considered
situations
Known control
control situations
External considered
Relative error
in link positioning
Performance
(X − X ) + (Yref − Y ) + ( Z ref − Z ) ,
2 2 2
EEP
= ref
N EEPi
∑ ,
i =1 EEP max
Figure 14. Changing the position of the manipulator links EEPPERFORMANCE ≡
in an external control situation: a comparison of the ICS N
based on soft computing KBO and the ICS based on quan-
tum computting KBO where N is the number of experiments, EEP max is the
maximum metric error from the estimated series of exper-
The application of ICS based on quantum computing
iments.
in solving of positioning problems in internal unexpected
Updated system of quality criteria for evaluating the
situations did not help to improve the results compared
operation of control systems under conditions of strong
to ICS based on soft computing. Let us consider in more
internal interference:
detail internal unexpected situations and increase internal
(1) Metric positioning error of the manipulator end ef-
noises.
fector EEP.
2. The Quality of Work of an ICS based on (2) Relative overshoot value σ .
QFI under Conditions of Strong Internal In- (3) Relative error in link positioning ε .
(4) Time of one iteration t.
terference
(5) The complexity of implementing control P.
Earlier, the following internal unexpected situations with (6) Full Control Behavior FCB.
the introduction of interference considered:
(1) Introducing noise into the control channels (noise FCB ≡ w1 ⋅ P [ EEP ] + w2 ⋅ P [σ ] + w3 ⋅ P [ε ] + w4 ⋅ P [t ] + w5 ⋅ P ,
amplitude = 1);
(2) Making errors in the MS (1,5 degrees). where w = [ 0, 4 0,1 0,1 0, 2 0, 2] is the weight’s vec-
tor.
Let us consider the behavior of ICS under conditions
Consider an unexpected situation with the introduction
of strong interference. Previously, the solution to the po-
of a strong interference in the MS.
sitioning problem was determined by the most accurate
positioning of manipulator links: 2.1 The Quality of Control Systems in Conditions
of Strong Disturbing Interference in the MS
Previously, an internal unexpected situation was consid-
value σ
Relative overshoot
soft KBO
positioning ε
Relative error in link
and the ICS based on quantum KBO in accordance with
iteration t
Time of one
control P
The complexity
of implementing
Behavior FCB
Full Control
the introduced system of quality criteria Figure 16.
From Figure 16 can concluded that the ICS based on
quantum KBO by all criteria (except for the time of one
iteration) is better to the ICS based on soft KBO. Figure 16. The quality of the ICS based on soft and quan-
tum computing in the conditions of strong interference in
the MS
Example. Let us consider an example of the influence
of noise in a control channel of ICS based on soft and
quantum computing. In Figure 17 shows an example of
noise in a control channel. In Figure 18 shows a compari-
son of the control signals during the operation of the ICS
based on soft (a) and quantum (b) computing. In Figure 19
shows the position of the manipulator when operating the
ICS based on soft (a) and quantum computing (b).
Noise q1 Noise q2
5 5
Figure 15. Metric positioning error of the manipulator end
effector [m]: FC is result of ICS based on soft computing; 0 0
U1 U2
The error in positioning an end effector when using
40 8
only soft computing does not exceed 0,3 meters with a
30 6
20 4
length of one link of 1 m.
10 2
We will generate noise at which the metric error of po-
0
0 100 200 300
0
0 100 200 300
sitioning the functional device during the operation of the
ICS based on the KBO on soft computing will be more
U3 U4
than 0, 5 m and, with the same noise, we will evaluate the
40 40
behavior of the ICS based on quantum computing. To do
30 30
20 20
this, increase the amplitude of the noise introduced into
10 10
the control channels to 100.
0
0 100 200 300
0
0 100 200 300
Five random noises generated for 13 workspace points.
Generated noises introduced into the control channels of the
U5 U6
ICS based on KBO on soft computing with separate control
20 40
and the ICS based on KBO on quantum computing.
15 30
10 20
Error = 1,918 m Error = 0,184 m
5 10
0 0
0 100 200 300 0 100 200 300
U7
30
20 Desired
point Desired point
Reached Reached
10 point point
0
0 100 200 300
(a)
U1 U2
50 8
40
6
30
4
20
2
10
0 0
0 100 200 300 0 100 200 300
U3 U4
40 30
30
20
20
10
10
0 0
0 100 200 300 0 100 200 300
U5 U6
20 40
15 30
10 20
5 10
0 0
0 100 200 300 0 100 200 300
U7
50
40
30
20
10
0
0 100 200 300
(a) (b)
(b)
Figure 18. Control signals generated by ICS based on soft Figure 19. The position of the manipulator in space when
(a) and quantum (b) computing operating ICS based on soft (a) and quantum computing (b)
0 0
-50 -50
0 100 200 300 0 100 200 300
Noise q5 Noise q6
50 50
0 0
-50 -50
0 100 200 300 0 100 200 300
Noise q7
Figure 20. Functional device metric positioning error [m]: 50
FC is the result of ICS operation based on soft comput- 0
ing, QFC is the result of ICS operation based on quantum -50
0 100 200 300
computing
According to the results of 65 experiments, we con- Figure 22. Noise in the control channel
clude that the metric error obtained when using quantum
U1 U2
computing is much smaller than when using only soft 80 60
computing. 60
40
Let us compare the work of the ICS based on KBO on 40
soft and quantum computing in accordance with the intro- 20
20
0,745 0,859
U3 U4
100 60
0,494
0,767
40
0,683 ICS based on
0,414 50
0,976 quantum KBO
0,817 20
0,775
ICS based on
end effector EEP
error of the manipulator
Metric positioning
value σ
Relative overshoot
soft KBO
positioning ε
Relative error in link
iteration t
Time of one
0 0
control P
The complexity
of implementing
U5 U6
60 60
Figure 21. The quality of the ICM based on KBO on soft
and quantum computing in the conditions of strong noise 40 40
U1 U2
80 60
60
40
40
20
20
0 0
0 100 200 300 0 100 200 300
U3 U4
80 60
60
40
40
20
20
0 0
0 100 200 300 0 100 200 300
U5 U6 (a) (b)
60 60
Figure 24. The position of the manipulator in space when
40 40 operating ICS based on soft (a) and quantum computing (b)
20 20
In the considered unexpected situations with the intro-
duction of strong noise into the control channels, the ICS
0 0 based on KBO on quantum computing showed an advan-
0 100 200 300 0 100 200 300
tage over ICS based on KBO on soft computing.
80
U7
Consider an unexpected situation with the introduction
of a strong disturbing effect in MS.
60
soft computing. -5 -5
0 100 200 300 0 100 200 300
Example. Let us consider an example of the influence Noise q5 Noise q6
5 5
of noise in MS. In Figure 27 is an example of noise in
0 0
MS. In Figure 28 shows a comparison of control signals.
-5 -5
In Figure 29 shows the position of the manipulator when 0 100 200 300 0 100 200 300
Noise q7
operating ICS based on soft (a) and quantum computing 5
(b). 0
-5
0 100 200 300
Table 3. Comparison of the results of ICS based on KBO
on quantum and soft computing
Figure 27. Noise in the control channel
Total number of experiments 390
The number of experiments in which ICS based on quantum com-
105
puting worsen the result than when using only soft computing:
0..30% 80
31..60% 21
The number of experiments where the error
61..90% 2
increased by N%
91..120% 0
121..150% 1
The number of experiments in which ICS based on quantum com-
285
puting improve the result than using only soft computing:
0..30% 131
31..60% 79
61..90% 24
91..120% 16
The number of experiments where the error
121..150% 9
decreased by N%
150..200% 4
200..250% 4
250..300% 2
300 and more % 4
The result has not changed. 11
0,656 0,847
0,395
0,888
0,642 ICS based on
0,359
0,974 quantum KBO
0,807
0,811
ICS based on
end effector EEP
error of the manipulator
Metric positioning
value σ
Relative overshoot
soft KBO
positioning ε
Relative error in link
iteration t
Time of one
control P
The complexity
of implementing
Behavior FCB
Full Control
(a)
(a) (b)
Reached
point
Reached
Error = point
Error = Desired
point 0,262 m Desired
2,519 m
point
0,258
0,517
0,330 ICS based on
0,062
0,977 quantum KBO
0,780
0,597
ICS based on
end effector EEP
error of the manipulator
Metric positioning
value σ
soft KBO
Relative overshoot
positioning ε
Relative error in link
iteration t
Time of one
control P
The complexity
of implementing
Behavior FCB
Full Control
-5 -5
0 100 200 300 0 100 200 300
Noise q3 Noise q4
5 5
0 0
-5 -5
0 100 200 300 0 100 200 300
Noise q5 Noise q6
5 5
0 0
-5 -5
0 100 200 300 0 100 200 300
Noise q7
5
0
-5
0 100 200 300
(b)
Figure 33. Measured signals during ICS based on soft (a)
and quantum (b) computing
(a) (b) MS
Figure 34. The position of the manipulator in space when Figure 35. ICS with separated control based on KBO on
operating ICS based on soft (a) and quantum computing soft computing for 3DOF manipulator
(b)
In Figure 35 K is the matrix of proportional, differ-
In the considered unexpected situations with the intro- ential and integral coefficients of the PID controller
duction of strong interference into the MS, the ICS based
K Pi , K Di ,K Ii , i = 1,3 , i is the number of the corresponding
on quantum KBO improves quality than when using only
soft computing. link of the robot of the manipulator, TSi is the training sig-
Thus, we considered the method of organizing coordi- nal , GAi is a GA that generates a teaching signal for the
nation control using quantum computing technologies on formation of the i-th KB.
the example of a complex CO a 7DOF robot manipulator. As a standard control situation for the FCi, a typical
We have demonstrated at the simulation model (using the control situation appears in the conditions of which a
MatLab/Simulink package) that the inclusion of generalize teaching signal TSi received. Unexpected control situa-
QFI unit into the ICS based on soft computing allowed: tions divided into external and internal. Designed ICS
(1) to provide a complete solution to the problem of based on KBO on soft computing with separated divided
positioning in standard situations; control (Figure 35) contain information about three situa-
(2) to improve positioning accuracy in conditions of tions of control (standard or unexpected) for each of three
Spatio-temporal (4)
MatLab/Simulink models and manipulator Testbench ac-
ICS with separated control (1)
cording to the introduced system of quality criteria.
Position Task
Solution in standard
control situations
Position Task
Solution in
unexpected control
situations
Performance
Relative
overshoot value
Relative error in
link positioning
One iteration
time
Implementation
complexity
Full Control
Behavior
Relative error in
link positioning
One iteration
time
Implementation
complexity
Position Task
Solution in standard
control situations
ICS based on KBO on soft
Position Task
Solution in
unexpected control
situations
computing with one FC
Performance
Relative
overshoot value
Relative error in
link positioning
One iteration
time
Implementation
complexity
Full Control
Behavior
Figure 41. Comparison of the results of ICS based on
KBO on soft computing with one FC and ICS based on
quantum computing when testing Matlab / Simulink mod-
els
Performance
Relative error in
link positioning
Implementation
complexity
Full Control
Behavior
separated control in unexpected control situation (a); ICS
based on KBO on quantum computing (b)
The inability of ICS based on soft computing to solve Figure 42. Comparison of the results of ICS based on
the problem of exact position control in Figure 40 also KBO on soft computing with one FC and ICS based on
illustrated. FC responsible for managing the second link quantum computing when testing manipulator Testbench
for the allotted time was not able to rehabilitate after a
It can see from the comparison results that ICS based
powerful external impact. Because of which the position-
on KBO on soft computing and ICS based on quantum
ing error of the second link was more than 50 degrees, the
computing solves the positioning problems for standard
control goal has not achieved and the control system as a
and external unexpected situations. ICS based on soft
whole was not robust.
computing with one FC not always coping with internal
unexpected control situations. Full Control Behavior for
External impact on the second link
ICS on quantum computing is higher, both for MatLab/
Current position
Simulink models and for the manipulator Testbench.
Reference signal
Let us demonstrate the operation of ICS based on KBO
on soft computing with one FC in the conditions of an
internal unexpected control situation (Figure 43) in com-
parison with ICS based on KBO on quantum computing.
Figure 40. The change in the position of the second link
under the control of ICS based on KBO on soft computing
with separated control
(3) Intelligent agents are used different types of quan- to so called QPB-class) that can solve classical algorith-
tum correlations (as behavior control coordinator with mically unsolved problems;
quantum computation by communication) and information (4) In QFI the self-organization principle is realized
transport (value information); using the on-line responses in a dynamic behavior of clas-
(4) The (hidden) quantum value information extracted sical FC’s on new control errors in unpredicted control
from classical states of control signal classical trajectories situations for the design of robust intelligent control;
(with minimum entropy in “intelligent states” of designed (5) Model of QFI supports the thermodynamic inter-
robust control signals). relations between stability, controllability and robustness
B. Features for design of self-organization processes.
(1) Developed QFI model is based on thermodynamic Specific features of QFI applications in design of ro-
and information-theoretic measures of intelligent agent in- bust KB. Let us stress the fundamentally important spe-
teractions in communication space between macro- and mi- cific feature of operation of the QA (in the QFI model) in
cro-levels (the entanglement-assisted correlations in an active the design process of robust laws for the coefficient gain
system represented by a collection of intelligent agents); schedules of fuzzy PID controllers based on the individual
(2) From computer science viewpoint, QA of QFI KB that designed on Soft Computing Optimizer (SCO)
model plays the role of the information-algorithmic and with soft computing.
SW-platform support for design of self-organization pro-
cess; A1.2 Quantum Information Resources in QFI Al-
(3) Physically, QFI supports optimally a new developed gorithm
thermodynamic trade-off of control performance (between In this section we introduce briefly the particularities of
stability, controllability and robustness) in self-organiza- quantum computing and quantum information theory that
tion KB process. are used in the quantum block, QFI supporting a self-or-
From quantum information theory viewpoint, QFI ganizing capability of FC in robust ICS. According to
reduces the redundant information in classical control described below algorithm the input to the QFI gate is
signals using four facts[4] from quantum information for
considered as a superposed quantum state K1 (t ) ⊗ K 2 (t ) ,
data compression in quantum information processing:1)
where K1,2 (t ) are the outputs from fuzzy controllers FC1
efficient quantum data compression; 2) coupling (separa- and FC2 designed by SCO (see, below Figure A1) for the
tion) of information in the quantum state in the form of given control task in different control situations (for ex-
classical and quantum components; 3) amount of total, ample, in the presence of different stochastic noises).
classical, and quantum correlation; and 4) hidden (observ-
able) classical correlation in the quantum state. A1.2.1 Quantum Hidden Information Extraction
The gate structure of QFI model with self-organization in QFI
KB properties that includes all of these QA features (see,
below Figure A1) based on above-mentioned proposals. Using the four facts from quantum information theory
Let us discuss the following question. QFI extracts the hidden quantum value information from
Q. What is a difference between our approach and Nat- classical KB1 and KB2 (see Figure A1).
ural (or man-made) models of self-organization?
A. Main differences and features are as followings:
(1) In our approach a self-organization process is de-
scribed as a logical algorithmic process of value informa-
tion extraction from hidden layers (possibilities) in classi-
cal control laws using quantum decision-making logic of
QFI-models based on main facts of quantum information,
quantum computing and QA’s theories;
(2) Structure of QFI includes all of natural elements of
self-organization (templating, self-assembly, and self-or-
ganization structure) with corresponding quantum opera-
tors (superposition of initial states, selection of quantum
correlation types and classes, quantum oracles, interfer-
ence, and measurements);
(3) QFI is a new quantum search algorithm (belonging (a)
1 d −1 1
=ρ
2d =∑∑
k 0=t 0
(k k ⊗t t ) ⊗ (U
A t k k U t† ) B
.(A1)
To realize QFI process we developed new tools called “QC Remark. On Figure A6, the first internal layer of QC
Optimizer” that are the next generation of SCO tools. Optimizer is shown (inputs/output). On Fig/ A7, the quan-
tum block realizing QFI process based on three KB is
A1.4.1 QC Optimizer Toolkit described. The “delay time = 20 (sec)” on the Figure A7,
corresponds to the parameter “Δt” given in temporal
QC Optimizer Toolkit is based on Quantum & Soft Com-
quantum correlations description (see Table A1); the knob
puting and includes the following:
named “correlation parameters” call other block (see
(1) Soft computing and stochastic fuzzy simulation
Figure A7) where a chosen type of quantum correlations
with information-thermodynamic criteria for robust KBs
(Table A1) is described.
design in the case of a few teaching control situations;
On Figure A8 description of temporal quantum correla-
(2) QFI-Model and its application to a self-organization
process based on two or more KBs for robust control in tions is shown. Here “kp1_r” means state 0 for kp(t) of
the case of unpredicted control situations. FC1 (or KB1); “kp1_r_t” means state 0 for kp(t+Δt) of
Internal structure of QC Optimizer is shown on Fig-
ures A6 and A7. FC1 (or KB1); “kp1_v” means state 1 for kp(t) of FC1 (or
KB1); “kp1_v_t” means state 1 for kp(t+Δt) of FC1 (or
KB1); and so on for other FC2 (KB2) and FC3(KB3).
K dnew Probability
KiFC 0
QFI (Self-Organization)
function
Kinew challenge of QFI using in problems of CO with weakly
Self-
Assembly
formalized (ill-defined) structure and a large dimension of
K pFC 2
Robust QFC the phase space of controlled parameters.
K dFC 2
Appendix 2: Quantum Computing Optimizer of
KiFC 2 Knowledge Base (QCOptKBTM)
Scale coefficients:
max K p ( d ,i ) ⋅Q_A_params Quantum Computing Optimizer of KB (QCOptimizer)
Description of correlation type
is a software toolkit designed for modeling, creation and
Figure A9. QFI-process by using QC Optimizer (QFI optimization of complex control structures, including
kernel) fuzzy and quantum control modules. Working with
Main module of Quantum Computing Optimizer is a QCOptimizer also has standard window menu, displayed
Win32 application, called qcowin.exe. It can be used in on the top of program window. To choose a command
Windows 95, 98, NT, Me, 2000 and XP environments. In with the mouse or keyboard, first select the menu and then
order to start the program, you should type “QCOptimizer” choose command you want. Each underlined character in
in command prompt or double-click QCOptimizer icon in menu and command names corresponds to the key you
windows explorer. During process of model optimization can press to select a menu or to choose a command.
QCOptimizer creates temporary files in current directo- To choose a command by using the mouse you should
ry. Those files have “mms” extensions. Before starting first click the menu name containing the command you
QCOptimizer check that current directory does not con- want and then click the command name in the pop-up
tain files with same extension. menu.
To choose a command by using the keyboard first press
A2.2.1 SCOptimizer Program Window the ALT key. The File menu appears selected. Use left and
right arrow keys to select required menu and then press
Main program window of QCOptimizer is shown on Fig- down arrow or ENTER to open pop-up menu. Use up and
ure A10. down arrows to select desired command and press ENTER
QCOptimizer window is mostly occupied with model to activate it. You can also press one of the underlined
display. It displays current control system model as a set characters to quickly select desired menu or command.
of connected blocks. Each block corresponds to some A2.2.3.1 FILE Menu
data-processing module in control system. Lines show QCOptimizer allows project files to be managed by the
how inputs and outputs of modules are connected. Model means of commands in the FILE Menu.
display allows access to each module configuration and The following items are available in this menu:
allows manipulation of modules and connections. New: closes actual model and starts creation of new
model.
Open: open existing model from disk file.
Import: import SCOptimizer file to the current model.
Save: save actual model to the file.
Close: close current model without saving. situation use Teaching Signal Manager to point QCOpti-
Signal Manager: activate learning signal manager. mizer to new location of teaching signal file.
Number Format: set number format conventions for
teaching signal and model files. A2.3 Working with QCOptimizer
Exit: closes QCOptimizer.
A2.3.1 Creating New Model
A2.2.3.2 VIEW Menu
View menu can be used to open signal view windows: Select File/New from menu to create new model. An emp-
Open signal view: displays new signal window. ty model will be created and you can than add various
A2.2.3.3 HELP Menu blocks to it.
Help menu currently have only one command About,
which displays version information of QCOptimizer. A2.3.2 Loading Model from File
If you have a file with previously saved model you should
A2.2.4 Dialog Boxes
select File/Open command. Standard windows File Open
When you choose a command having options, QCOp- dialog box will appear. Select file with you model and
timizer shows a dialog box. A dialog box may contain press Open button.
fields in which you can enter text, numbers or select some
items. Typical dialog box will have OK button, which you A2.3.3 Saving Model files
should press after filling all fields in order to activate the You can use File/Save command to save you model files.
command, and CANCEL button, which will abort com- Standard windows File Save dialog box will appear. Se-
mand execution. You can also use ENTER key on the key- lect file with you model and press Save button.
board to activate the same function as OK button and ESC We recommend saving you model regularly, in order
as CANCEL. to avoid loss of data due to possible program/computer
If an operation requires many parameters Wizard-style failure or undesired model change. Note that most
dialog box will be used. Wizard box consists of a sequence operations of QCOptimizer do not have “undo” option, so
of dialog boxes that will guide you through the steps of an the only way to restore to previous state is to load model
operation. Three buttons will be available for navigation from saved file.
of the wizard box. NEXT>> button tells wizard that you
have successfully completed filling of current page and A2.3.4. Importing SCOptimizer Files
want to switch to the next page or perform the command
if no additional parameters are required. <<BACK button You can import SCOptimizer model files (.sco exten-
should be used to return to one of the previous pages, if sion) in order to use fuzzy control systems developed in
you want to change some parameters. CANCEL button SCOptimzier in QCOptimizer projects. Select File/Import
aborts command execution. command to start model input. Standard windows File
Open dialog box will appear. Select file with SCOptimizer
A2.2.5 QCOptimizer Files model and press Open button.
Import options dialog will than appear as shown in Fig-
SCOptimizer saves project in special files: model files ure A11.
(with *.sc2 extension). Model file is used to store all mod-
el data, including model blocks, their properties, connec-
tions and positions, as well as most optimization and other
settings.
QCOptimizer can save genetic algorithm optimization
state to disk file, so you can continue optimization process
interrupted some time ago. Those files have *.st extension
and will be called state file. Figure A11. Import options dialog
If you wish to copy QCOptimizer files to new location Check option checkboxes as required and press OK to
(for example to another computer) you should copy model import model. The options are following:
file, teaching signal files and state files, if it is required. If Convert input blocks to scale: by default imported
your teaching signal file was originally located in directo- SCOptimizer model includes new input blocks that are
ry other than one with project file, than QCOptimizer may used to feed input data to fuzzy control system. If this
be unable to load it automatically in new location. In this option is selected the model will be imported with scale
blocks instead, which will use input normalization param- You can change any parameter by entering desired
eters to scale signal. This option can be useful when creat- symbol(s) to the corresponding field. There are also
ing models, which includes several fuzzy control modules two buttons, which you can use to switch to one of the
using the same input. predefined configurations:
Convert output blocks to scale: same as before, but ap- (1) System: switch to system defaults, as defined by
plied to fuzzy control output. Windows locale settings.
Load active database only: allow import of only one (2) Matlab: switch to format, used by Matlab for text
fuzzy rule database from SCOptimizer model, the active files. Following settings are default for the Matlab:
one. By default, all databases are imported and active da- ① List separator: “;”
tabase is connected to output blocks. ② Decimal separator: “.”
③ Negative sign: “-”
A2.3.5 Changing Locale Settings ④ Positive sign: “+”
This operation is available from File/Number Format ⑤ Grouping method: “3;”
menu and from New Model creation Wizard. After selec- ⑥ Thousand’s separator: “ ”
tion of this operation QCOptimizer will show you dialog When entering locale parameters be careful not to:
box, containing current settings (Figure A12). (3) set different parameters to the same symbol
(4) make list separator, decimal separator or negative
sign an empty string
(5) do not set any parameter to one of the following
symbols: (,),{,},:,’.
Breaking these rules may cause program not to save
and/or load files correctly.
After you set all options to desired characters press
OK to apply changes or CANCEL if you want to reset you
changes.
blocks by holding CTRL key and clicking on blocks to be Matlab file should contain an array of real numbers
selected. To deselect all blocks click on a free space. with number of columns equal to sum of number of input
Selected block[s] can be moved across display. To move, and output variables of the model. If version file contains
click and hold left mouse on selected block, and drag it to several variables then first one will be loaded.
desired location. Release mouse button to stop moving. Text files are processed using locale settings. Please set
You can also access block properties in this mode. locale settings as described in section 4.2.5 before loading
Double click left mouse button over block to display cor- text files. If locale settings do not match the file format
responding option dialog box. than signal will be loaded incorrectly and no error mes-
(2) Move window tool (Table A2): select this button sage will typically be displayed. Text file should contain
to move display window if the model does not fit inside learning signal data separated by any separators. This will
screen. Click toolbar button and then click and hold left include text files produced by Matlab, CSV files and prac-
mouse button on model view and drag mouse to change tically any other file of this kind.
display position. Release mouse to stop moving. To manage learning signals, select File/Signal Manag-
(3) Create block (Table A2): click this button to create er command from QCOptimizer menu. Signal manager
new block. After you click it menu with list of available window will be displayed on Figure A13.
block types will be displayed. Select required block type
from this menu. Depending on block type dialog box may
appear, where you should enter initial block settings. After
finishing with block settings, you will see new block dis-
played in model view. Drag block to it’s desired location
and click mouse to finish block creation.
(4) Create link tool (Table A2): click this button to create
link between blocks. Then click to ports (one input and one
output, in any order) to create link between ports. Any output
port can have as many connections as required. Input ports
can have only one connection. If you select input port which
already have connection it will be reconnected as instructed. Figure A13. Signal manager dialog box
(5) Delete block tool (Table A2): To delete block click To load new signal file press Load signal button. Stan-
this button and then block to be deleted. dard windows file open dialog will be displayed. Select
(6) Delete link tool (Table A2): To delete link click this file with signal and click Open to load it. Loaded files will
button and then input port, which should be disconnected. appear in list box labeled Loaded signals.
You cannot delete links by clicking on output ports. Click on a file name in this list box and it signal to vari-
able assignments will be displayed below. Those define
Table A2. The icons
how loaded signal and model variables are related. By
Select tool default, table show “not connected” to indicate that sig-
nal variable is not assigned to model variable. To set the
Move window tool
assignment (or change it) click Change button. Another
Create block dialog will appear as shown on Figure A14.
Create link tool
Loaded and assigned signals can be used during opti- In this dialog you can change block name (initially
mization, if teaching signal is selected as a source of fit- equals block index) and normalization parameters. Signal
ness function in optimization manager parameters. In this range corresponding to current not-normalized signal are
case for model evaluation loaded signals will be presented shown below.
on model input and block outputs will be compared to as- Press OK to accept changes and close window. Press
signed signals. Apply to change block parameters without closing
File with learning signal can be unloaded by selecting window. Press Cancel to close window without applying
it in the list and clicking Unload button. Unloaded signals modifications.
will not be used during optimization. A2.3.8.2 Model Output Blocks
Output blocks are used to pass results of calculation to
A2.3.8. Available blocks external world. Data arriving at output blocks is then sent
Blocks are building parts of a control system models. Fol- over external link or compared to assigned learning sig-
lowing block types are supported: nals.
(1) Model input: feeds input data to model; Output blocks are numbered from 0, and the index of
(2) Model output: used to pass calculated output from block is shown after block name. Output signal sent over
model to external world; external connection are compiled from signals of output
(3) Fuzzification: performs fuzzification of signal; blocks in order of their indexes. Block index depend on
(4) Inference: fuzzy rule database, performing Sugeno the order in which blocks where created and can only be
inference; changed by deleting blocks with lower indexes.
(5) Constant: generates constant signal; Output blocks support signal denormalization. To alter
(6) Quantum generalization: perform quantum general- signal normalization parameters double click block. Fol-
ization algorithm; lowing dialog will be displayed on Figure 16.
(7) File: injects data from external file to model;
(8) Sum: calculates weighted sum of its inputs;
(9) Scale: perform scale and offset on signal.
Blocks are described in the following sections.
A2.3.8.1 Model Input Blocks
Input blocks are used to pass incoming data to model.
They do not have input ports and receive data from exter-
nal connection or assigned learning signal file.
Input blocks are numbered from 0, and the index of
block is shown after block name. Vector signals from ex-
ternal connection are separated to components and send to
input blocks according to their index. Block index depend
on the order in which blocks where created and can only
be changed by deleting blocks with lower indexes.
Input blocks also support signal normalization. To alter
signal normalization parameters double click block. Fol- Figure A16. Input block parameter dialog box
lowing dialog will be displayed on Figure A15.
In this dialog you can change block name (initially
equals block index) and normalization parameters. Signal
range corresponding to current not-normalized signal are
shown below.
Note that output block perform demoralization on the
signal, so that directly connected input and output blocks
with equal parameters will leave signal intact. Input block
offset signal to offset parameter and than scale it, output
block perform reverse operation.
Press OK to accept changes and close window. Press
Apply to change block parameters without closing
window. Press Cancel to close window without applying
Figure A15. Input block parameter dialog box
eters of newly added membership function. If you press Maximal number of rules: Maximal number of rules for
OK MF with those parameters will be added, if you press current model.
CANCEL it won’t. Since rule database structure is high- Total rules in db: Number of rules stored in the data-
ly dependant on number of MF’s, rule database will be base. For complete database this is always equal to max-
cleared and recreated if you add a MF. imal number of rules. For LBRW database it can be less
A2.3.8.4 Inference Blocks then maximal.
Inference block support fuzzy inference with Sugeno Show rules for output: Rule database is displayed for
model. When the block is being created following dialog one of the outputs. This list selects output variable for
will appear, where you should enter initial block parame- which database will be displayed.
ters (Figure A20) Selected rule: Displays textual representation of select-
ed rule, if any.
Rule Database editor display database as network with
four layers. First layer is input variable layer. Each circle
in this layer represents input variable. Second layer is in-
put MF layer. Circles of this layer represent membership
functions of variables. Circles in the third layer represent
rules of the database. Number written inside this circle is
a rule number of the rule in the database. Last layer is the
output layer. For Mamdani model output layer is com-
posed of circles, corresponding to membership functions
of selected output variable. For Sugeno models output lay-
er displays numerical parameters of the rule.
Database structure is shown with lines, connecting
different layers. Each node in the rule level is linked with
those MF’s in input MF layer, which are included in the
Figure A20. Inference block creation page
if-part of the rule. It is also linked with output MF or nu-
To view or edit inference block properties double click merical parameter of the then-part.
the block. Dialog box as shown on Figure A21 will ap- Since all rules of the model won’t fit on display, scroll
pear. bar is implemented which scrolls nodes of the rule level.
You can select a rule from the database by clicking on
the node of the rule level. Textual representation of the
rule will be shown in the Selected rule field and you will
be able to edit or delete this rule. When you select a rule
in the database it activation level is displayed as red lines
on the Rule Activation Level page.
If you wish to delete the rule, select it and press delete
button. Rule will be removed from the database. Note that
rules from complete database can not be deleted.
By pressing Edit button, you can change selected rule.
The dialog window as shown on Figure A22 will appear.
The left part of the dialog represents if-part of the rule, search for block optimization
and the right part corresponds to the then-part. You can (3) Cyclic is indicate that value have cyclic property
change parameters of any part by selecting items from the which is used during optimization
list and changing values in the drop-down box below the Number of block outputs cannot be changed after the
list. block is created. Other parameters can be changed by acti-
For Sugeno 0 model this dialog will have a slightly dif- vating parameter dialog by double clicking on the block.
ferent appearance (Figure A23). A2.3.8.6 Quantum Generalization Blocks
During creation of quantum generalization block you
should enter initial parameters in the dialog box shown on
Figure A25.
(2) Record enable: check to enable histogram recording tion vectors. Each to columns of the table correspond to
for selected channel; real (one with “_r” appended) and virtual (“_v”) parts of a
(3) Decode: uncheck to skip this input during decoding vector for each output.
(required by QPID algorithm). Entries in the table show which value will be used at
To the left of table there is a histogram review window. which position of a vector. Those values can be changed
It shows histogram for the channel selected in the drop- by clicking on them and selecting new value from menu.
down list above it. Clear All button can be used to erase Name of the component is composed of input name (or
all recorded histograms for all channels. input number if name is not specified), and modifier, as
Delay time control in the top part of the section deter- follows:
mine amount of delay introduced for temporal correlation. (1) _r: real component is used;
In the output section there is a table with output param- (2) _r_t: real delayed component used;
eters: (3) _v: virtual component used;
(1) Name: output name, to be used in correlation con- (4) _v_d: virtual delayed.
figuration; Quantum algorithm requires the histogram to be load-
(2) Scale coef.: output is scaled to this constant. ed for each channel. Histograms can be recorded during
Output scale coefficients can be optimized. Optimiza- block operation or by processing file from disk. In the first
tion settings can be accessed in dialog box activated by case check Record enable checkbox in input table of cor-
Optimization Parameters button (Figure A27). responding inputs and run signal through the model. Data
appeared on inputs corresponding to checked table lines
will be recorded in histogram. To load histogram press
“Load from file button” (Figure A29).
Figure A30. File block parameter dialog Figure A32. Scale block parameter dialog
Select file you wish to use by clicking on “…” button Set desired parameters in dialog and press OK or Apply
and selecting file name using standard windows file open to change block parameters.
dialog. Data from file can be normalized before sending to
A2.3.9 External Link
block output. Normalization parameters should be entered
to the corresponding fields of the dialog. Press OK to cre- QCOptimizer support data exchange with external pro-
ate block. grams, which can be used to pass data from external pro-
During operation file block output one record from the gram as input to QCOptimizer model and return resulting
file for every model calculation step. Current record index model output. This link uses the same interface as one
is shown in block properties dialog and can be reset to available for external inference in SCOptimizer, and can
zero by pressing Rewind button. Block automatically re- be accessed from SCLib library functions.
starts from first record when all records from file are read. If current fitness calculation mode is set to “teaching
A2.3.8.8 Sum Blocks signal” external link is not used and cannot be activated.
Sum block calculate weighted average of its inputs. Change fitness calculation mode to Matlab/external signal
During creation of a block dialog window will appear to enable external link access.
where you should select required number of inputs and To connect via external link QCOptimizer and anoth-
initial weights (Figure A31). er program should use the same link name. Default link
name is SC2. Link name in SCOptimizer can be changed
by pressing toolbar button. Dialog box will appear
where you can enter desired link name (Figure A33).
Figure A31. Sum block parameter dialog Figure A33. Link name configuration dialog
Enter desired number of inputs in input count field, Current link name is also displayed in the status line in
weight factors in scale column in input configuration table the left-bottom corner of main QCOptimizer window pre-
and press OK to create the block. ceding with “IPC:”.
You cannot change number of inputs of sum block, but To enable external link press toolbar button. To indi-
input weights can be changed. To do it double click the cate that QCOptimizer is ready for external connection an
block. The same dialog window will appear, where you ‘*’ symbol will be shown after link name in status line. To
can set weight factors as desired. disable external connection press , ‘*” will disappear.
A2.3.8.9 Scale Blocks Always enable external link first from QCOptimizer
Scale blocks transform input signal by scaling and off- and than in external program. Close link in reverse order –
setting it. To set scale and offset parameters double click first in external program, then in QCOptimizer.
A2.3.10 Viewing Data Passing through the Model First select desired fitness calculation mode:
(1) Teaching signal: model is evaluated using teaching
You can view signals passing through the model in the signal loaded in Teaching Signal Manager and fitness is
graphical representation. To view signal select View/Open computed by comparing actual data from model with data
signal view menu item. Signal view window will appear, selected as teaching signal.
showing signals on model input. To change signal shown (2) Fitness calculation in matlab: during optimization
in window click right mouse button above window. In the QCOptimizer set model to the state being evaluated and
menu appear select block (or model input/output) and than calls matlab function specified in configuration. Function
block output. should evaluate model quality (typically using external
You can open as many data view windows as you like, link to do some calculations) and return fitness value. To
by selecting the same menu item several times. work in this mode matlab should be available and one
should create three functions and set it names to following
A2.3.11 Testing Model on Input Signal from File
parameters:
You can make QCOptimizer perform calculations using ① Initialize session command: called by QCOptimizer
input signal recorded in file. To do so you should: when optimization is started. Should perform initializa-
(1) Load signal using Teaching Signal Manager and tion, for example load some simulink model.
associate it with model inputs. ② Fitness calculation command: called by QCOpti-
(2) Select “teaching signal” fitness mode calculation in mizer to calculate fitness for current state.
optimization configuration. ③ Close session command: called when optimization
Once those two actions are done, you should see “file” is finished. May free resources allocated for calculations.
displayed in the status line typically used to show external (3) Fitness calculation from external signal. Signal
link name. Press toolbar button to perform a full run coming from external link is used for model input and
through data available in loaded files. You can view re- fitness calculation. Concrete fitness function is configured
sulting signals using signal view windows, or you can use during optimization start, but here you can enter initial
this mode to record histograms in quantum block. conditions and calculation time. Initial conditions are set
in Initial state table where you can enter desired initial
A2.3.12 Optimization value for some variables and possible delta. In example
QCOptimizer support various optimization algorithms shown on figure above optimization will start when output
and modes. It can perform optimization of single block, or 1 of quantum block 432 will equal 1±2. Once initial con-
several blocks at a time, including optimization with dif- dition is met QCOptimizer will wait for delay time set in
ferent algorithms, running at the same time and evaluated wait before averaging field and than calculate fitness func-
over same link. tion as average over averaging time selected.
A2.3.12.1 Selecting Optimization Options In this window you can also select what QCOptimizer
will do if it detects that the same model state evaluation
To set optimization settings press toolbar button (Fig-
ure A34). is requested during single optimization process. If do not
check for duplicate chromosomes is selected than dupli-
cate detection will be disabled and all states will be eval-
uated. You can greatly improve optimization time by se-
lecting single evaluation mode. In this case you may also
want to set hash table size used to store processed states.
A2.3.12.2 Starting Optimization
To start optimization press button. Optimization start
dialog will appear (Figure A35).
References
[1] Ulyanov, S. V. System and method for control using
quantum soft computing [P]. US Patent No 7,383,235
B1, 2003; EP PCT 1 083 520 A2, 2001; Efficient
simulation system of quantum algorithm gates on
classical computer based on fast algorithm [P]. US
Patent No 2006/0224547 A1, 2006.
[2] Ulyanov, S. V. Quantum fast algorithm computation-
al intelligence PT I: SW / HW smart toolkit [J]. Arti-
Figure A36. Optimization progress dialog ficial Intelligence Advances, 2019, 1(1): 18-43.
ARTICLE
Robotic Smart Prosthesis Arm with BCI and Kansei / Kawaii / Affective
Engineering Approach. Pt I: Quantum Soft Computing Supremacy
Alexey V. Nemchaninov1 Alena V. Nikolaeva1 Sergey V. Ulyanov1* Andrey G. Reshetnikov1,2
1. INESYS LLC (EFKO GROUP), Russia
2. Dubna State University, Russia
Article history A description of the design stage and results of the development of the
Received: 22 November 2019 conceptual structure of a robotic prosthesis arm is given. As a result, a
prototype of man-made smart prosthesis on a 3D printer as well as a foun-
Accepted: 3 June 2020 dation for computational intelligence presented. The application of soft
Published Online: 30 June 2020 computing technology (the first step of IT) allows to extract knowledge
directly from the physical signal of the electroencephalogram, as well as
Keywords: to form knowledge-based intelligent robust control of the lower perform-
Robotic prosthetic arm ing level taking into account the assessment of the patient’s emotional
state. The possibilities of applying quantum soft computing technologies
Cognitive computational intelligence (the second step of IT) in the processes of robust filtering of electroen-
Brain-computer-device neurointerface cephalogram signals for the formation of mental commands of robotic
Mental commands prosthetic arm discussed. Quantum supremacy benchmark of intelligent
control simulation demonstrated.
Quantum soft computing
Fuzzy cognitive controller
Quantum supremacy benchmark
1. Introduction ligent information technologies in the field of intelligent
T
control using cognitive technologies (“brain-computer-de-
he development of robotic human limbs prostheses vice” interface - BCI) are considered. Neurointerface of
and the production of human-like electromechani-
this type (manmade interface) allow to restore and expand
cal devices - anthropomorphic robots are receiving
the capabilities of a person with physical (for example,
more, but not sufficient, attention in both the scientific,
disability with loss of limbs) or mental disorders in var-
technical and socio-economic plants [1-4]. It should also
be noted that in the strategy for the development of arti- ious activities (for example, autistic children or patients
ficial intelligence (AI) in the Russian Federation [5] in the with impaired mental activity - dementia). Cognitive
Healthcare section, the application of AI in such an im- interfaces provide the ability to communicate, evaluate
portant socio-technical domain as intelligent prosthetics emotions, transfer and control devices with mental com-
and smart cognitive control systems, as well as rehabilita- mands.
tion of disabled people, is not indicated at all. The process of developing a cognitive intelligent simu-
Principles of the development and application of intel- lator toolkit discussed. The application of developed tool-
*Corresponding Author:
Sergey V. Ulyanov,
INESYS LLC (EFKO GROUP), Russia;
Email: ulyanovsv@mail.ru
kit provides developers with the ability to control robotic apply such interfaces to solve everyday problems and
devices at the lower (executive) level (using the so-called control various devices. Interfaces are actively used for
“mental commands”), and at the upper level - the intelligent rehabilitation and diagnostic procedures, helping to im-
level (to develop cognitive intelligent control technologies prove interaction with the human environment, including
with the possibility of application in applied tasks). with robotic devices [10-14,16,17]. Technologies are actively
Let us consider briefly main principles, peculiarities involved quantum end-to-end technologies in EEG data
and features of cognitive intelligent control applied in processing and educational processes at state levels [18].
biomechanical products and presented the description Research of this kind has been funded by states since the
of a hierarchical intelligent control system based on early 70’s. There are a number of research collaborations
QSCOptKBTM (knowledge base optimizer on quantum soft on the creation and development of man - machine inter-
computing). faces associated with all three areas [10-14,19,20] etc. In par-
ticular, research in this area can divided into the following
2. Cognitive Intelligent Control: Design Prin- groups:
ciples and Features of Biomechanical Product (1) Recognition research - development of devices for
the diagnosis, modeling, simplification and reduction of
Sources of technological and breakthrough innovations in
threats to the interaction of the brain with the system.
these areas are following: (1) new technologies for creating
(2) Simulation of the brain mechanism - the use of neu-
intelligent materials; (2) technologies for creating an intelli-
ral network effects and the phenomena of the functioning
gent software product integrated into devices and applied at
of the brain in applied problems of information technolo-
all stages of interaction with devices; (3) new human-ma-
gy, for example, analysis and synthesis of information.
chine interfaces, the principles of which are based primarily
(3) Restorative medicine - restoration of behavioral
on the method of reading the activity of the functioning of
cognitive functions lost as result of damage to the brain or
the brain and nerve endings of the body [2,5].
body [8].
So, the first direction allows a person to restore (and in
(4) Elaboration - development of brain-computer sys-
the future, to exceed) the functional state of limbs dam-
tems in the feedback loop to accelerate and improve the
aged as a result of any injuries due to the creation of more
functional behavior of the system [1, 21].
advanced alloys, material structures, nano coatings. The
The development of these researches made it possi-
second, innovative direction is associated with the cre-
ble to create new technologies for the neural interface to
ation of sophisticated software that allows the biomechan-
detect fundamental and interregional brain functions in
ical device to learn and adapt to individual physiological
online, as well as to develop complex mathematical algo-
and psychological qualities and characteristics of the
rithms to model brain activity and the resulting behavioral
human operator. Through the application of deep machine
functions and reactions.
learning and medical recommendation systems with a
deep knowledge representation, it is possible to recognize 2.2 Architecture of the Limbic System
complex human commands, read and recognize the emo-
tional state of the operator [6 - 8]. At the same time, a certain The Limbic System, as part of the mammalian creatures’
computing basis in the form of embedded end-to-end brain, is mainly in charge of the emotional processes. (see
technologies of cognitive computing and computational Appendix 1). The Limbic System located in the cerebral
intelligence must correspond to software of this level. The cortex consists mainly of following components: Amygda-
development of the third direction based on the sources of la, Orbitofrontal Cortex, Thalamus, Sensory Cortex, Hy-
new human-machine interfaces that can effectively com- pothalamus, Hippocampus and some other less important
plement and expand human information capabilities. Such areas. We try to describe briefly these main components
interfaces include infrared - spectrometers, electroenceph- and their tasks.
alographs, magnetoencephalographs, cognitive helmets Figure 1 illustrates the anatomy of the main compo-
and equipment for virtual and complementary (augmented) nents of Limbic System [22].
3D - and 4D - reality, invasive and non-invasive sensors The first sign of affective conditioning of the system
and beacons, for example, mounted on the wrist, or other appears in Amygdala which is a small almond-shaped in
parts of the human body [9 - 15]. sub-cortical area. This component placed in a way to com-
municate with all other Sensory Cortices and areas within
2.1 Related Works the Limbic System.
Many researches known in which patients routinely
anism based on sensory inputs and emotional cues (Reward amputation, send signals (red arrows) or the cortical re-
signals) (see Appendix 1). organization (red star in the brain) generates the phantom
limp pain. Other sensations that can felt involve tingling,
cramping, heat, and cold.
Figure 10. The structure of cognitive intelligent control Let us consider the redistribution problem of the level of
system responsibility between the cognitive and fuzzy controller.
The basis of prostheses and robotic manipulators is spatial
In the considered part of the work cycle, the existing mechanisms with many degrees of freedom. According-
hardware and software research basis and information ly, when designing a cognitive intelligent controller for
technology of sophisticated class of cognitive intelligent a prosthesis, the intelligent control system of the robotic
control system (see Figure 11) for supporting the design arm can take as the basis. In this example, three fuzzy
and operation of a new class of devices are presented. controllers are implemented in the structure of the in-
telligent control system of the robot manipulator with 3
{K ( t ) , K ( t ) , K ( t ) , K ( t ) , K ( t ) , K ( t )} → K ( t )
1
P
2
P
3
P
1
D
2
D
3
D
new
P
{K ( t ) , K ( t ) , K ( t ) , K ( t ) , K ( t ) , K ( t )} → K ( t ) ,
1
D
2
D
3
D
1
I
2
I
3
I
new
D
{K ( t ) , K ( t ) , K ( t ) , K ( t ) , K ( t ) , K ( t )} → K ( t )
1
I
2
I
3
I
1
P
2
P
3
P
new
I
Spatio - temporal: solves the problem of the exact positioning of the manipu-
lator robot in contrast to the intelligent control systems on
{K ( t ) , K ( t − ∆t ) , K ( t ) , K ( t − ∆t ) , K ( t ) , K ( t − ∆t )} → K ( t )
1
P
1
D
2
P
2
D
3
P
3
D
new
P the SCO on soft computing.
{K ( t ) , K ( t − ∆t ) , K ( t ) , K ( t − ∆t ) , K ( t ) , K ( t − ∆t )} → K ( t ).
1
D
1
I
2
D
2
I
3
D
3
I
new
D
Let us demonstrate the operation of the intelligent con-
trol systems on the SCO on soft computing with divided
{K ( t ) , K ( t − ∆t ) , K ( t ) , K ( t − ∆t ) , K ( t ) , K ( t − ∆t )} → K ( t )
1
I
1
P
2
I
2
P
3
I
3
P
new
I
control in the conditions of an external unforeseen control
Temporal: situation (Figure 17) in comparison with the intelligent
control systems in the SCO on quantum computing. As an
{K ( t ) , K ( t ) , K ( t ) , K ( t − ∆t ) , K ( t − ∆t ) , K ( t − ∆t )} → K ( t )
1
P
2
P
3
P
1
P
2
P
3
P
new
P
unforeseen situation is the forced displacement of the sec-
{K ( t ) , K ( t ) , K ( t ) , K ( t − ∆t ) , K ( t − ∆t ) , K ( t − ∆t )} → K ( t ).
1 2 3 1 2 3 new ond link of 3DOF manipulator.
D D D D D D D
{K ( t ) , K ( t ) , K ( t ) , K ( t − ∆t ) , K ( t − ∆t ) , K ( t − ∆t )} → K ( t )
1
I
2
I
3
I
1
I
2
I
3
I
new
I
foreseen situation, the control of the intelligent control FC is created and the amount of memory in the system in
systems on the SCO on quantum computing copes with which the FC is located.
the task of positioning with a given accuracy, in contrast Moreover, for complex control objects, such as a robot
to the intelligent control systems on the SCO on soft com- manipulator with 7DoF, the organization of a single FC is
puting with a divided control. impossible.
In general, the possibility of decomposing the control
(dividing one knowledge base into several identical inde-
pendent knowledge bases) and organizing coordination
management by introducing the quantum fuzzy factor link
significantly increases the possible number of input vari-
ables and thereby expands the possibilities of accounting
for the parameters of the system and the control object.
dt
(
= ∑qi ⋅ ϕ q, t , k ( H O − ( H Тc + H Cc ) ) , u )
i =1
Stability Controllability
(2)
( )
+ ( H O − ( H Тc + H Cc ) ) ⋅ H O − ( H Тc + H Cc ) ≤ 0
Robustness
ensure the guaranteed achievement of the control goal. its behavior depending on the circumstances, as well as
One of the key tasks of modern robotics is the devel- to plan its communication and cooperation strategies with
opment of technologies for the cognitive interaction of ro- other nodes. Here the indicators of the level of coopera-
botic systems, which allow solving the tasks of intelligent tion are the nature of the distribution of tasks, the unifica-
hierarchical management by redistributing knowledge and tion of various information resources and, of course, the
control functions, for example, traditionally between a possibility of solving a common problem in a bounded
leader and a subordinate (“master - slave” system). Mod- time interval.
ern approaches to solving this problem based on the theo- Example. As noted above, if microscopic degrees of
ry of multi - agent systems, the theory of swarm artificial freedom are available to the observer in the form of the
intelligence, and many others [33 - 35]. Maxwell demon, then the second law of thermodynam-
Due to the synergetic effect, an additional information ics can violated. Szilard showed from an analysis of the
resource created and the multi-agent system is able to Maxwell demon model that work extracted from the ther-
solve complex dynamic tasks for performing team mutu- modynamic cycle in the form as kT ln 2 . Moreover, it
al work. The given task may not fulfill by each element was shown [38] that recoverable work WextS from the system
(agent) of the system separately in various environments
is determined by the amount of information (or quan-
without external management, control or coordination,
tum-classical mutual information) I, which estimate the
however, exchange of knowledge and information allows
extracted knowledge of the system from measurements.
performing useful teamwork to achieve the management
At the same time, a similar ratio in the form of a lower
goal under the conditions of uncertainty of the initial in- M
formation and limited consumption of useful resources. boundary exists for the full cost Wcost measuring and eras-
In particular, it is known that for closed-loop control M
ing information WextS ≤ −∆F S + kTI and Wcost ≥ kTI , where
systems, the amount of useful work W that is extracted
∆F S determines the free energy of the system. Then it is
t
=
satisfies the inequality Wmax ( t ) k Tmin Ic dt ′ ≤ kTI c ,
∫ easy to see the speed of the recoverable work Wext limited
0
by Wext ≤ kTI , i.e. limited by the speed of information re-
where k is the Boltzmann constant, Tmin ( t ) is interpret-
trieved [39].
ed as the lowest achievable temperature by the system in
Model of self-organized intelligent control system pro-
time t for feedback control, assuming Tmin ( 0 ) = T and Ic posed based on the principles of minimum informational
determines the amount of Shannon information (entropy entropy (in the “intelligent” state of control signals) and
transfer) , extracted by the system from the measurement the generalized thermodynamic measure of entropy pro-
process [36]. duction (in the system “control object + regulator”). The
Physically, the synergetic effect means the design of main result of the application of the self-organization pro-
knowledge self-organization and the extraction of addi- cess is the acquisition of the necessary level of robustness
tional information that allows the multi-agent system to and the flexibility (adaptability) of the reproducible struc-
perform the most useful work with a minimum loss of ture.
useful resource and with a minimum of the required initial It has noted that the property of robustness (by its phys-
information, and without destroying the lower executive ical nature) acts as an integral part of self-organization,
level of the control system. and the required level of robustness in intelligent control
Together with the information-thermodynamic law of system achieved by fulfilling the principle of minimum
intelligent control (optimal distribution of the manage- production of generalized entropy noted above. The prin-
ment qualities as “stability - controllability - robustness”), ciple of minimum entropy production in object control
an intelligent control system (ICS) is designed with multi- and system control serves as the physical principle of op-
agent systems, ensuring the achievement of the manage- timal functioning with a minimum consumption of useful
ment goal under the conditions of uncertain initial infor- work and underlies the development of robust intelligent
mation and limited useful resource [37]. control system.
A network of loosely coupled groups of robots working This statement is based on the fact that for the general
together to solve tasks that go beyond individual capa- case of controlling dynamic objects, the optimal solu-
bilities. Different nodes of such a system, as a rule, have tion to the finite variational problem of determining the
a different intelligent level (knowledge, algorithms, and maximum useful work W according to, it is equivalent
computational bases) and various information resources to solving the final variational problem of finding the
for design process. Each node should be able to modify minimum of entropy production S. Thus, the study of the
The supplied software allows receiving, recognizing the operator develops his skill in working with a helmet,
and recording the EEG signal from the helmet. To ana- which subsequently makes it possible to control real tech-
lyze the received signal, the so - called EEG frequency nical devices (manipulators, wheelchairs, and other de-
rhythms are distinguished. vices). Training can be either active, with external stimuli
Remark. The term “frequency rhythm” means a certain (for example, light of a certain frequency, video and audio
type of electrical activity corresponding to a certain state material, pictures, etc.), or passive if the operator presents
of the brain, for which the boundaries of the frequency it without external influence during the generation of a
range are defined. In the process of cognitive activity, mental command.
characteristic rhythms of the beta, alpha, theta and delta Also, additional hardware, such as a virtual reality hel-
ranges appear [47] (Figure 26). met, which will visualize the user’s actions in the virtual
The set of simultaneously present rhythms forms a world, thereby increasing the effectiveness of training,
specific spatial-frequency EEG pattern. Patterns are char- can be used to form a user with a more complete picture
acteristic of different types of cognitive activity and are of learning. The duration of training and the intensity
highly individually specific [15, 19].The ability of an individ- of training affects the quality and number of recognized
ual to establish rhythmic EEG patterns when performing mental teams. If the first time a person can learn how to
certain cognitive tasks makes up an “encephalographic” generate signals in the brain for one team, then for good
portrait of his personality. recognition of two or more teams, several exercises are
Comment. One of the main components of cognitive necessary.
neurointerface technology is gaming simulators. It is Accordingly, both the psychophysiological character-
important to note that during training, the control model istics of a person’s state (including his position in space)
acts as a system model on the monitor screen. This, in and his level of training will influence the management
turn, allows not only to train the brain to generate men- system and the achievement of the management goal.
tal commands, but also to tune the control system of the Usually (according to information from the manufacturer)
control object itself, adjusting it to the operator, thereby for the system to work well using four teams. It is neces-
ensuring increase of overall performance in the system sary to conduct regular training for 2-3 weeks, and after
“brain - computer”. This kind of feature is because when a training the operator experiences fatigue. And accordingly,
person works with a cognitive helmet, the program adapts time is required to restore strength. The learning process
to the characteristics of the operator, adjusting the control described in more detail in [37, 45]. A program based on the
system. Developers of such equipment with the level of Epoc cognitive helmet SDK written to control the prosthe-
operator training to generate various mental commands sis servomotors.
usually associate the quality of team recognition. The learning process consists of two main blocks: (1)
analysis of electroencephalographic indications and iden-
tification of frequency patterns specific to a particular op-
erator; (2) generating and recording mental commands.
At the training stage, using the wet non-immersed
brain-computer interface with passive electrodes of Emo-
tivEpoc+, the activity of certain parts of the operator’s
brain recorded. Initially, it is necessary to obtain a fre-
quency slice of the operator’s neutral state (the operator
is at rest), this frequency slice is necessary to improve
detection of the nervous excitation of the operator during
further analysis.
In the future, to generate mental commands, the op-
erator for a certain period of time (8 seconds) mentally
Figure 26. EEG frequency rhythms models the necessary action, which will serve as the basis
of the mental command (contraction, relaxation of various
The software bundle provides specialized games - sim-
muscle groups, etc.). And also at this moment performs
ulators, in which the learning and training process takes
this action with «Live» hand. An individual signal record-
place. Typically, computer games used for this, where the
ed for each mental command.
operator must perform an action that can matched with
From the viewpoint of the control system, the operator
some command used to control the object. When playing,
must be able to re-generate the recorded signals, which
interpreted by the system to control the movement of the helmet, it is necessary to mark the “Connect” field, then
device. The frequency rhythm shown in Figure 27. select a mental command from the drop-down list or cre-
ate a new one. Now of pressing the “Train” button, the
operator begins active cognitive activity associated with
the selected team. After completing the training process,
it is necessary to save the obtained data, connect the pros-
thesis to the computer, set the boundaries for activating
the mental command. And then put the application into
the prosthesis control mode (by marking the “Rсactive”
field). When click on the “Train” button, information sent
to the helmet about the currently selected mental team
for training, after which some time activity indicators of
the human brain recorded. When the helmet fixes similar
Figure 27. Frequency rhythm of the neutral state (a) and parameters of mental activity, the key of the current team
the state of excitation (b) will change from neutral to the team associated with these
parameters.
Typically, statistical methods used to evaluate and rec-
In addition to connecting directly to the helmet and
ognize signals, including the calculation of variance and
prosthesis, it is possible to test the program using the
mathematical expectation.
third-party program Emotiv Xavier Composer, which
Figure 28 demonstrate highlights the most different sig-
simulates the work of a cognitive helmet and allows send-
nals from each other, which the operator associates with a
ing virtual mental commands to the application. In the
mental command and a neutral state.
program code, this implemented as a call to the Remote
Connector Connect methods.
If the user sends a mental command to the program,
then the spatial-frequency pattern evaluated, after the
scaling of the received signal relative to the set minimum
and maximum, the corresponding signal sent to the COM
port. The following event occurs when the timer is started,
Figure 28. Dispersion of mental command signals
which starts to control the prosthesis during the connec-
Dispersion of signals generated by the operator is a tion of the cognitive helmet, which is shown on the main
fairly simple and effective method for comparing signals screen.
of mental commands. Additional various methods de- Comment. The variable “power” is the power of mental
scribed in [47]. effort. Since the prosthesis has physical limitations during
Figure 29 shows the interface of software tools that al- operation (it is not possible to send servomotors an an-
lows you to receive and process data from the brain-com- gle of rotation that extends beyond the interval [0.180]), the
puter interface of EmotivEpoc+ (the program code created variable “a” trimmed in possible values to the boundaries
in the C# programming language using the Windows of this interval. In the program code, the second possi-
Forms graphical interface). ble value is another such construct. Thus, a binary tree
implemented. After the operations performed, a line of
the form “a A” is sent to the COM port, where “A” is the
angle from 0 to 180, by which all five servomotors will be
rotated, taking into account the physical restrictions previ-
ously established in the Arduino program for each one in
particular. Such methods prevent possible damage to the
prosthesis if the commands coming from the helmet are
misinterpreted.
Figure 30 shows a graph showing the frequency rhythm
Figure 29. Program interface of the operator. The trained cognitive control system, ana-
Data exchange between the cognitive helmet and the lyzing this frequency pattern, recognizes the mental com-
computing center done using the Bluetooth protocol stack, mands present in the frequency rhythm.
as well as the UART protocol. To initialize a cognitive
5. Conclusion
(1) This work (the first part) presents the development
of several high-tech areas of robotics, which have practi-
cal scientific and technical interest, both in separate and in
joint developments.
Figure 30. Execution of mental commands (2) It has been shown that the prospect of developing
Figure 31 shows a graph showing the registration of cognitive intellectual control using soft and quantum
mental commands at discrete time intervals. computing technologies is one of the important tasks in
creating a robotic prosthetic arm, such as a simple case of
a robot avatar, and is integral to the development of infor-
mation technology in the framework of the concept of an
intelligent simulator [41].
(3) The use of expert recommendation systems with a
deep representation of knowledge [6,7] and quantum end-to-
end technologies of deep machine learning with quantum
EEG processing [17,49] allows the appointment, selection
of control of robotic prostheses of the hand, taking into
account the individual psychophysiological characteristics
of the patient and the operating environment.
Figure 31. Regulatory signaling for the actuator
(4) On the one hand, these products if properly devel-
Figure 32 shows the process of the experiment. oped can presented on the market of commercially attrac-
tive products, on the other hand, technologies for using
new types of intelligent information technologies and
human-machine cognitive interfaces.
(5) The next stage of development is the creation of a
cognitive intellectual control system for a robotic arm-pros-
thesis for maintenance based on IT quantum soft comput-
ing, quantum EEG processing filters [17] and Kansei / Affec-
tive Engineering intelligent computing technologies with an
assessment of the user’s emotional state [49 - 51].
Figure 32. Conducting an experiment: execution of the (6) The work, in its essence, reflects the completeness
“unload” command (a), the execution of the command of the formation of a new educational approach in intel-
“compression” (b) ligent robotics [32, 40, 41, 52] - a hybrid cognitive intelligent
The operator, using mental commands that read by the robotics based on neural interfaces with new types of IT
EPOC+ cognitive interface, controls the compression-un- data processing.
loading process of the prosthesis brush. Appendix 1. Contemporary models of affect, motiva-
In accordance with the scheme shown in Figure 10, tion, emotion and cognitive control for Kansei /Kawaii /
EmotivEpoc in on line captures the excitations of the Affective Engineering
cerebral cortex at the time of a certain cognitive activity Let us briefly consider the models of negative affect,
(generation of mental command). The received data goes pain and emotion that play important role in Kansei engi-
to the computer center, where they are recognized. Based neering and cognitive control.
on the analysis of the obtained frequency pattern and Example. In humans and other primates, the cingulate
estimation of the signal level at a discrete time instant, a - a thick belt of cortex encircling the corpus callosum - is
control action generated for the actuator. one of the most prominent features on the mesial surface
of the brain (Figure 33a). Early research suggested that cortex; (b) - Negative affect, pain and cognitive control
the rostral cingulate cortex (Brodmann’s ‘precingulate’; activate a common region within the anterior subdivision
architectonic areas) plays a key part in affect and motiva- of the midcingulate cortex - aMCC (The map depicts the
tion (Figure 33b). More recent research has enlarged the results of a coordinate-based meta-analysis (CBMA) of
breadth of functions ascribed to this region; in addition to 380 activation foci derived from 192 experiments and
involving more than 3,000 participants); (c) - Activation
emotion, the rostral cingulate cortex has a central role in
likelihood estimate (AlE) maps of the three behavioral do-
contemporary models of pain and cognitive control. The
mains (left) and pairwise ALE minimum conjuction maps;
most basic question is whether emotion, pain and cogni- (d) - Activation foci maps [22]
tive control are segregated into distinct subdivisions of
the rostral cingulate or are instead integrated in a common Example: Emotional learning occurs mainly in the
region. There is a growing recognition that aMCC might amygdala. The system operation consists of two levels:
implement a domain-general process that is integral to the amygdale learns to predict and react to a given rein-
negative affect, pain and cognitive control [22]. forcement signal. This subsystem cannot unlearn a con-
Cognitive control is a range of elementary processes (such nection. The incompatibility between predictions and the
as attention, inhibition and learning) that are engaged when actual reinforcement signals causes inappropriate respons-
automatic or habitual responses are insufficient to sustain es from the amygdala. As depicted in Figure 34, the sys-
goal-directed behavior. Control can be engaged proactively tem on Figure 34a, consists of four main parts [54]. Sensory
or reactively. Activation foci maps between negative affect, input signals first enter the thalamus. Since the thalamus
pain and cognitive control is shown on Figure 33d. must provide a fast response to stimuli, in this model the
maximum over all stimuli S is sent directly to the amyg-
dala as another input. The amygdala receives inputs from
the thalamus and sensory cortex, while the orbitofrontal
cortex (OFC) part receives inputs from the sensory cortex
and the amygdala. The system also receives a reinforcing
signal (REW - emotional signal).
For each Anode in the amygdala, there is a plastic con-
nection weight Vi. Any input is multiplied by this weight
to provide the output of the node. The O nodes show sim-
ilar behavior, with a connection weight Wi applied to the
input signal to create an output. There is one output node
in common for all outputs of the model, called E (see Fig-
ure 34a). The E node sums the outputs from A except for
the Ath and then subtracts the inhibitory outputs from the O
nodes. The result is the output from the model.
Learning Based Intelligent Controller (BELBIC) structure according to the end application. For example in Lucas
[53]
. In a biological system, emotional reactions are utilized et al. the reward signal r(t) is obtained as a weighted sum
for fast decision‐making in complex environments or of the error signal and the control effort and the external
emergency situations. It is thought that the amygdala and stimulus signal i(t) is defined as a linear combination of
the orbitofrontal cortex are the most important parts of the the system output and its first derivative.
brain involved in emotional reactions and learning. The It should be observed that it essentially converts two
amygdala is a small structure in the medial temporal lobe sets of inputs (sensory inputs and emotional cues or rein-
of the brain that is thought to be responsible for the emo- forcing signals) into the decision signal (the emotional re-
tional evaluation of stimuli. This evaluation is in turn used action) as its output. Closed loop configurations using this
as a basis for emotional states and reactions and is used for block (BELBIC) in the feedforward control loop of the to-
attention signals and laying down long‐term memories. tal system in an appropriate manner have been implement-
The amygdala and the orbitofrontal cortex compute their ed so that the input signals have the proper interpretations.
outputs based on the emotional signal (the reinforcing The block implicitly implemented the critic, the learning
signal) received from the environment. The final output algorithm and the action selection mechanism used in the
(the emotional reaction) is calculated by subtracting the functional implementations of emotionally based (or, gen-
amygdala’s output from the orbitofrontal cortex’s (OFC) erally, reinforcement learning based) controllers, all at the
output (see Figure 34b). A control system strategy, based same time.
on brain emotional learning, was proposed by Caro Lucas Reza Keramat et al. [54] consider that in practical sys-
in the early 2000s (Lucas et al., 2004). The limbic model tems important information originates from two sourc-
used was based on the neural link, between the amygdale es. One of the sources is the experts, who define their
and the orbitofrontal cortex, proposed by Balkenius and knowledge of the system using the natural language. The
Moren. This control paradigm is commonly designated by other consists of measurements and mathematical models
BELBIC which stands for brain emotional learning based derived from physical laws. Hence, what matters is how
intelligent control. The reasoning behind the integration of to incorporate these types of information into the design
the limbic model into a closed loop control system can be of systems. The question is that how it is possible to for-
tracked down to the seemingly robust way that the brain mulate human knowledge within a framework similar to
performs decision making. Actually, control has all to mathematical models.
do with decision making: the controller goal is to devise Basically, the main function of a fuzzy system is to
the best input actions based on the incoming information make such a conversion possible. Fuzzy systems are
according to the system states. These actions can be tak- based on knowledge or rules. The core of a fuzzy system
en considering the past, the present or even forecasts on is a knowledge base following the IF-THEN fuzzy rules.
the future system states. Hence the controller produces a If a fuzzy system is used by a controller, the controller is
mapping between its input signals and the output control called “fuzzy controller”.
signals by means of an arbitrary decision function which The fuzzy controller designed for the full bridge DC-
can be described by means of differential equations, as DC converter takes two inputs: error (e) and error varia-
in PID-controllers, or by an inference mechanism such tions (Δe). The membership input functions for this con-
as in Fuzzy or Neuro-Fuzzy control. Alternatively, it can troller are in the range of (-5, 5). Membership functions
be based on the result of the optimization of a cost func- for each of the two input components of error and error
tion such as linear quadratic regulators (LQR) or model variations are shown in Figure 35.
predictive control (MPC). In the BELBIC control system
architecture this input-to-output transformation is imposed
by means of the limbic system model. In this case, both
the external stimuli and reward signals are generated in
such a way as to produce a closed loop system response
according to some target characteristics. In addition, due
to the recursive nature of the weights update law, this con-
troller is able to gradually learn how to handle changes in Figure 35. The membership functions for e and Δe
the system dynamics. A key point in BELBIC is the exter-
Seven membership functions are used here for the in-
nal stimulus and reward signals definition.
put: negative big (NB), negative medium (NM), negative
Notice that there are not universal rules to carry out this
small (NS), zero (Z), positive big (PB), positive medium
task. This choice is flexible and must be custom defined
(PM), and positive small (PS). The controller output is as- automata feature. Based on this deduction, we felt no need
sumed to be equal to the duty cycle, which varies between to emphasize the comparison of these controllers after set-
0 and 0.5. Figure 36 displays the controller output func- ting optimization.
tions.
References
[1] Robbin A. Miranda, Casebeer William D., Heinc
Amy M., etc.. DARPA-funded efforts in the devel-
opment of novel brain-computer interface technolo-
gies[J]. Journal of Neuroscience Methods, 2015, 244:
52-67.
[2] An analytical review of the global robotics market
Figure 36. The output membership functions
2019[M]. Sberbank Robotic Lab., Sberbank, 2019,
Reza Keramat et al. show the effectiveness of BELBIC 272.
for the full bridge DC-DC converter. Figure 37 shows the [3] Guangzhong Yang, Jim Bellingham, Pierre E. Du-
system output voltage after the application of the fuzzy pont, etc. The grand challenges of Science Robot-
and BELBIC controllers. ics[J]. Sci. Robot. 3, 2018, eaar7650: 1-14.
[4] An analytical review of the global robotics mar-
ket[R]. Laboratory of Robotics, Sberbank. 2018: 79.
[5] Artificial Intelligence: Approaches to Formation. AI
development strategies in the Russian Federation[R].
Sberbank, 2019.
[6] Amirova E.K., Efimov V.A., Ulyanov S.V., etc. Ex-
pert system for selecting lower-extremity (thigh)
prostheses and diagnosis of the quality of artificial
replacement. Pt 1[J]. ISSN 0006-3398, Consultants
bureau, New York, J. of Biomedical Engineering,
May-June, 1991, 3: 26-31.
Figure 37. The comparison of the fuzzy controller with [7] Amirova E.K., Efimov V.A., Kuzhekin A.P., Ulyanov
the BELBIC S.V., etc. An expert system for selecting lower-ex-
As seen in this figure, the BELBIC has a slight over- tremity (thigh) prosthesis and evaluation of prosthet-
shoot while the fuzzy controller has a considerable over- ic quality. Part 2[J]. ISSN 0006-3398, Consultants
shoot (the initial overshoot is 35 V). The time required for bureau, New York, J. of Biomedical Engineering,
stabilization is almost the same in both cases. It is there- November-December, 1991, 6: 5-12.
fore concluded that the BELBIC outperforms the fuzzy [8] Lupina I.V., Slepchenko A.N., Ulyanov S.V., etc. Hy-
controller. brid expert system with an in-depth representation of
In Reza Keramat’s experiment the BELBIC controller knowledge for the design and diagnosis of biotechno-
outperforms the fuzzy controller. Considering the uncer- logical products[J]. Izv. USSR Academy of Sciences.
tainty of system parameters (including inductance, ca- Ser. Technical cybernetics, 1991, 5(152-175).
pacitance, and input voltage and acceptable variations of [9] Kaplan A., Shishkin S., Ganin I., etc. The prospects
load), the BELBIC presents better performance than the of the P300-based brain computer interface in game
fuzzy controller. Although fuzzy control is a robust and control[M]. IEEE Transactions on Computational In-
telligence and AI in Games, 2013
effective method for a large number of engineering sys-
[10] Shishkin S.L., Ganin I.P., Kaplan A.Y. Event-related
tems, but its design (and consequently its performance) is
potentials in a moving matrix modification of the
almost depend on the experience and tact of the designer.
P300 brain-computer interface paradigm[J]. Neuro-
Furthermore, after design and installation, its performance
science Letters, 2011, 496(2): 95-99.
is not improved, and in other word, it is not a learn-
[11] Leigh R. Hochberg, Mijail D. Serruya, Gerhard M.
ing-based or intelligent controller.
Friehs, etc. Neuronal ensemble control of prosthetic
It can be stated that a Learning Based Intelligent Con-
devices by a human with tetraplegia[J]. Nature, 2006,
trol which “may” in the first step, act not as satisfying as
442: 164-171.
any another modern controller, any straight-forward-de-
[12] Höller Y., Bergmann J., Kronbichler M., etc. E. Real
signed controller after few iterations thanks to its learning
movement vs. motor imagery in healthy subjects[J].
Int. J. Psychophysiol, 2012: S0167-8760. Emotional Learning in the Amygdala [J], Cybernetics
[13] Salvaris M., Cinel C., Citi L., etc. Novel protocols and Systems, 2000, 32(6): 611-636.
for P300-based brain-computer interfaces[J]. IEEE [25] C. Lucas, D. Shahmirzadi, N. Sheikholeslami. Intro-
Trans. Neural. Syst. Rehabil. Eng., 2012, 20: 8-17. ducing BELBIC: Brain Emotional Learning Based
[14] Musallam S., Corneil B.D., Greger B., etc. Cognitive Intelligent Controller[J]. International Journal of
Control Signals for Neural Prosthetics[J]. Science, Intelligent Automation and Soft Computing, 2004,
2004, 305: 258-262. 10(1): 11-22.
[15] Ivanitsky G. A., Determining the nature of mental ac- [26] H. Rouhani, M. Jalili, B. N. Araabi, W. Eppler, C.
tivity according to the rhythmic pattern of the EEG] Lucas. Brain emotional learning based intelligent
[A]. in Russia. Available URL: controller applied to neuro-fuzzy model of mi-
https://www.rfbr.ru/rffi/portal/project_search/ cro-heat exchanger[J]. Expert Systems with Applica-
o_53385 tions, 2007, 32: 911-924.
[16] Sunny T.D., Aparna T., Neethu P., etc. Robotic Arm [27] Rajesh P. N. Rao, Paul G. Allen. Towards Neural
with Brain Computer Interfacing[M]. International Co-Processors for the Brain: Combining Decoding
Conference on Emerging Trends in Engineering, Sci- and Encoding in Brain-Computer Interfaces[J]. In-
ence and Technology, ICETEST, 2015. vited submission to the journal Current Opinion in
[17] Gandhi V. Brain-Computer Interfacing for Assistive Neurobiology, 2018.
Robotics 1st Edition: Electroencephalograms, Re- [28] José del R. Millán, Pierre W. Ferrez, Anna Buttfield.
current Quantum Neural Networks, and User-Centric Non Invasive Brain-Machine Interfaces [A]. IDIAP
Graphical Interfaces[J]. N.Y. Academic Press, 2015. Research Institute. Available URL:
[18] Meng Jing. Brain Co CEO says his “mind-reading” https://www.esa.int/gsp/ACT/doc/ARI/ARI%20
tech is here to improve concentration, not surveil- Study%20Report/ACT-RPT-BIO-ARI-056402-Non_
lance[M]. South China Morning Post May, 2019. invasive_brain-machine_interfaces_-_Martigny_IDI-
Available URL: AP.pdf
https://www.scmp.com/tech/innovation/arti - [29] Maksimenko V. A., Kurkin S. A., Pitsik E. N.. Artifi-
cle/3008439/brainco-ceo-says-his-mind-reading- cial Neural Network Classification of Motor-Related
tech-here-improve-concentration EEG: An Increase in Classification Accuracy by Re-
[19] Kazuki Yanagisawa, Hitoshi Tsunashima, Kaoru ducing Signal Complexity[A]. Hindawi Complexity,
Sakatani. Brain-Computer Interface Using Near 2018, Article ID 9385947. Available URL:
Infrared Spectroscopy for Rehabilitation, Infrared https://doi.org/10.1155/2018/9385947
Spectroscopy[J]. Life and Biomedical Sciences, Prof. [30] US Patent No 7,219,087B2, Soft computing optimiz-
Theophile (Ed.), ISBN: 978-953-51-0538-1, InTech. er of intelligent control system structures[P]. (Inven-
Available URL: tor: S. V. Ulyanov), Date of patent: May 15, 2007
https://www.intechopen.com/books/infrared-spec- [WO 2005/013019 A3, 2005].
troscopy-life-and-biomedical-sciences/brain-comput- [31] US Patent No 2006, 0218 A1, System for soft com-
er-interface-using-near-infrared-spectroscopy-for-re- puting simulation[P]. (Inventor: S. V. Ulyanov), Date
habilitation of patent: Sept. 2006, 2006.
[20] Vaughan TM, McFarland DJ, Schalk G, etc.. The [32] Nikolaeva A. V., Ulyanov S.V. Intelligent robust con-
Wadsworth BCI Research and Development Pro- trol of an autonomous robot-manipulator[A]. Soft-
gram: at home with BCI[J]. IEEE Trans Neural Syst. ware systems and computational methods. 2014(1):
Eng., 2006, 14(2): 229-233. 34-62.
[21] Ulyanov S. V., Shevchenko A. V., Mamaeva A. A., DOI: 10.7256/2305-6061.2014.1.11466
etc. Soft Computing Optimizer as Deep Machine Available URL: https://en.nbpublish.com/library_
Learning in Intelligent Cognitive Estimation of Hu- get_pdf.php?id=28019
man Being Emotion for Robotic Control[J]. Robots [33] Tarasov V. B. From multi-agent systems to intelli-
and Engineering Cybernetics, 2019. gent organizations: philosophy, psychology, com-
[22] D. Purves, G. J. Augustine, D. Fitzpatrick, etc.. Neu- puter science[J]. M.: URSS editorial, 2002: 352. (in
roscience[M]. Sinauer Associates, 2001. Russ).
[23] J. Moren. Emotion and Learning - A Computational [34] Zaitsev A.A., Kureichik V.V., Polupanov A.A. Over-
Model of the Amygdala[D]. Lund University, Lund, view of evolutionary optimization techniques based
Sweden, 2002. on swarm intelligence[J]. News SFU. Technical sci-
[24] J. Moren, C. Balkenius, A Computational Model of ence, 2010(12). (in Russian).
[35] Kureichik V.M., Kazharov A.A.. The use of swarm 224. (in Russia).
intelligence in solving difficult problems[J]. News [47] Sotnokov P., I., Overview of EEG Signal Processing
SFU. Technicalscience, 2011(7). (in Russ). Techniques at Brain Computer Interfaces[J]. Engi-
[36] Sandberg H. et al. Maximum work extraction and neering Bulletin, Electronic Journal, 2014(10). (in
implementation costs for nonequilibrium Maxwell’s Russia).
demon[J]. Physical Review E. 2014, 4: 042119. [48] Ulyanov S.V., etc. Intelligent self-organizing cogni-
[37] Chatzis S. P., Korkinof D., Demiris Y.. A quan- tive regulators. Part 2: models of cognitive interfaces
tum-statistical approach toward robot learning by “brain - device”[J]. Journal of Systems Analysis in
demonstration[J]. IEEE Transactions on Robotics, Science and Education, 2015, 1. (in Russia).
2012, 28(6): 1371-1381. Available URL: http://sanse.ru/download/236
[38] Sagawa T, Ueda M. Minimal Energy Cost for Ther- [49] Ulyanov S.V.. Intelligent self-organized robust con-
modynamic Information Processing: Measurement trol design based on quantum / soft computing tech-
and Information Erasure[J]. Phys. Rev. Lett., 2009, nologies and Kansei engineering[J]. Computer Sci-
102(25): 250602. Erratum. Phys. Rev. Lett. 106, ence Journal of Moldova. 2013, 21, 2(62): 242-279.
189901, 2011. [50] Ulyanov S.V., Yamafuji K.. Intelligent self-orga-
[39] Horowitz J. M., Sandberg H. Second-law-like nized cognitive controllers. Pt. 1: Kansei / affective
inequalities with information and their interpreta- engineering and quantum / soft computing technol-
tions[J]. New Journal of Physics, 2014, 16: 125007. ogies[J]. System Analysis in Science and Education,
[40] Ulyanov S. V., Reshetnikov A. G., Mamaeva A. A., 2014, 4. Available URL: http:/www.sanse.ru/ar-
etc.. Hybrid cognitive control systems on the exam- chive/48
ple of driving[J]. Journal of Systems Analysis in Sci- [51] Ulyanov S.V.. Quantum fast algorithm computational
ence and Education, 2010, 3. (in Russia) intelligence Pt I: SW / HW smart toolkit. Artificial
Available URL: http://sanse.ru/download/261 Intelligence Advances, 2019, 1(1): 18-43. Available
[41] Ulyanov S. V., Reshetnikov A. G. Basis of cognitive URL:
computer training in robotics. Intellectual simulator https://doi.org/10.30564/aia.v1i1.619
for the formation of active knowledge[J]. Journal of [52] Ulyanov S.V., Reshetnikov A.G., Tyatyushkina O.Yu.
Systems Analysis in Science and Education, 2016(4). Intelligent robotics, part 2: socio-economic-techni-
(in Russia) cal platform of the cognitive educational process[J].
Available URL: http://sanse.ru/download/277 Journal of Systems Analysis in Science and Educa-
[42] Gael Langevin, InMoov is the first Open Source 3D tion, 2019. No 4. (in Russia) Available URL:
printed life-size robot[R]. Available URL: http://in- http://sanse.ru/download/277
moov.fr [53] D. Shahmirzadi. Computational Modeling of the
[43] Hand robot InMoov, opensource - project[R]. Avail- Brain Limbic System and its Application In Control
able URL: Engineering[D]. Texas A&M University, U.S.A.,
https://www.thingiverse.com/thing:17773 2005.
[44] EmotivEpoc, Neurocomputer interface[R]. Available [54] Reza Keramat, Mohammad Hosein Ershadi,
URL: https://www.emotiv.com Shahrokh Shojaeian. A Comparison of Fuzzy and
[45] Arduino, electronic device design tool[R]. Available Brain Emotional Learning-Based Intelligent Control
URL: http://arduino.ru/ Approaches for a Full Bridge DC-DC Converter[J].
[46] Petrov B.N., Ulanov G.M., Ulyanov S.V. etc. The International Journal of Industrial Electronics, Con-
theory of models in control processes: information trol and Optimization, 2019, 2(3): 197-206.
and thermodynamic aspects[J]. M.: Nauka, 1978:
REVIEW
Research and Analysis of Machine Learning Algorithm in Artificial In-
telligence
Yang Li1 Xueyuan Du1* Yiheng Xu2
1. University of Liverpool, Liverpool, L69 3BX, England
2. Xi’an Jiaotong-Liverpool University, Suzhou, Jiangsu, 215213, China
Article history This article firstly explains the concepts of artificial intelligence and
Received: 14 April 2020 algorithm separately, then determines the research status of artificial in-
telligence and machine learning in the background of the increasing pop-
Accepted: 20 April 2020 ularity of artificial intelligence, and finally briefly describes the machine
Published Online:15 July 2020 learning algorithm in the field of artificial intelligence, as well as puts for-
ward appropriate development prospects, in order to provide theoretical
Keywords: reference for industry insider.
artificial intelligence
machine learning algorithm
application of science and technology
1. Introduction 2. An Overview of Artificial Intelligence and
W
Machine Learning algorithm
ith the continuous development of science and
technology, artificial intelligence technology 2.1 Concept of Artificial Intelligence
has been widely used in various fields with
great effects. As the core technology of artificial intelli- In the year of 1950, Allen Matheson Turing, the father
of modern computer science, named a test with his own
gence, the initial research goal of machine learning is to
name. The test will randomly ask questions to the testees
enable computer systems to have the ability to “learn” just
using some devices when the tester is separated from the
like humans, and to lay the foundation for the realization
testees. If more than 30% of all testees cannot determine
of artificial intelligence. With the unremitting efforts of
whether the tester is a person or a machine, then it shows
growing experts and scholars, the development of artifi- that the machine is intelligent. This is exactly the famous
cial intelligence technology can be described as progress Turing test, which also has its shortcomings: it takes hu-
with each passing day. In some complicated but ordinary mans--which have much uncertainty because of influence
fields, machines can gradually replace manual work and of subjective consciousness--as the actual test standard of
exert extremely efficient work. the entire test, so the test results of the Turing test will be
*Corresponding Author:
Sergey V. Ulyanov,
INESYS LLC (EFKO GROUP), Russia;
Email: ulyanovsv@mail.ru
lack of some imprecision [1] pace of human development. In other words, learning is
With the continuous development of information tech- the basis for the continuous development of artificial in-
nology, the definition of artificial intelligence also devel- telligence [4]
ops continuously. Professor Nielsen of the United States Compared with traditional algorithm, learning algo-
believes: “Artificial intelligence is a discipline about rithm emphasizes the attribute of “learning”, which means
knowledge--how to express, obtain and use knowledge.”; the algorithm can develop itself, find new laws, and even
Professor Winston, who is also from the United States, improve the logic and behavior of algorithm program
believes: “ Artificial intelligence is to study how to enable through “learning”. In today’s academics, a large num-
computers to do intelligent work that could only be done ber of scholars believe that a computer system cannot be
by humans in the past.” These two definitions of artificial called an intelligent system if it is not able to learn.
intelligence can show basic ideas and contents of it to a Nowadays, the machine learning algorithm is usually
certain extent: to study the characteristics of human in- divided into three types: supervised learning, unsuper-
telligence activities and construct artificial systems with vised learning, and reinforcement learning. Supervised
certain intelligence, in other words, a basic theory, method learning means that learning algorithm works out predic-
and technology to study how to use computer technology tion models from the labeled training data, so as to effec-
to simulate human intelligent behavior[2]. tively predict unknown situations in the future; Unsuper-
vised learning means that learning algorithm performs
2.2 Concept of Algorithm the exploration of data structures and extracts meaningful
The academics have not yet reached a unification on the information or results [5] without the guidance of known
concept of the algorithm, among which philosophers variables or reward functions; Reinforcement learning is
mostly study algorithms just in terms of the learning algo- mainly used as a trial and error method, which is able to
rithm. These philosophers believe that the algorithm has help learning algorithm improve prediction effect of pre-
a strong correlation with human cognition, but they rarely diction models through interaction with the environment.
study algorithm [3]. To summarize the research results of Compared with traditional algorithm, computer systems
today’s academics, there are following four views on the can accomplish different things by using the same learn-
definition of the algorithm: ing algorithm, and the operation process is simpler and
① Algorithm is the method to determine a certain pro- more convenient, even a single learning algorithm can
cess route. derive countless different programs by itself. To achieve
② Take the algorithm as a program. After determining the same effect, the traditional algorithm requires more
a problem, the program will tell us to perform steps con- process.
tinuously, and analyze the problem, finally get the corre-
3. Research status of artificial intelligence and
sponding result.
machine learning algorithm
③ The third view is similar to the second one, both of
which take the algorithm as a program. There are some 3.1 Research status of artificial intelligence
differences in details: for the third view, the program can
always tell us how to deal with a series of rules systemati- Since its birth in the 1950s, the annual investment of sci-
cally. entific research in the field of artificial intelligence has
④ Algorithms are limited rules that present a series of also increased year by year with strong support from glob-
operations to solve specific types of problems. And algo- al governments. As a first-class power in the century and
rithm has the characteristics of finiteness, certainty, input, a major R & D country, many public funds in the United
output and effectiveness [3]. States have entered into the field of artificial intelligence
under the guidance of government departments, and arti-
2.3 Machine learning algorithm ficial intelligence has gained rapid development. In order
to further speed up the R & D of artificial intelligence, the
As the core technology of artificial intelligence, the initial
United States released the report of “2019 R & D Strategic
research goal of a machine learning algorithm is to enable
Planning of National Artificial Intelligence “in June 2019.
computer systems to have the ability to “learn” just like
The report not only effectively updates the first edition of
humans and to complete tasks that humans do now. To
“ R & D Strategic Planning of National Artificial Intelli-
achieve this goal, learning will be an indispensable part.
gence” issued by the US government in 2016, but also
Without learning, artificial intelligence cannot complete
proposes research goals for problems like human-machine
tasks that humans do now, neither can it keep up with the
collaboration methods, artificial intelligence safety, ethics,
legal and social impacts, which lays a planning foundation D achievements in the field of artificial intelligence, and
for the development of country artificial intelligence. applied these achievements to industrial development. [7]
At present, with the support of advanced technologies
such as information technology and computer technology, 4. Classic learning algorithm in artificial in-
two key technical breakthroughs have emerged in the field telligence
of artificial intelligence, namely cloud robot technology and
human brain bionic computing technology. As cloud robot
4.1 regression algorithm
technology related aspects have gradually matured, there The so-called regression algorithm is an algorithm that
are much more relevant existing researches of cloud robot attempts to explore the relationship among variables by
technology than that of human brain bionic computing measuring the error. In the field of artificial intelligence,
technology. With the rapid development and popularization it means a regression algorithm when people say “regres-
of high-tech technologies such as science and technology, sion”. The common regression algorithm includes the
computer networks, cloud computing, and big data, the least square method, logistic regression, stepwise regres-
costs of R & D and production of robot technology have sion, multivariate adaptive regression splines, local scatter
decreased year by year, also it is possible for robots to do smoothing estimation, etc.
cloud control and cloud computing as well as obtain and
screen network data through big data technology. At this 4.2 Instance-based Algorithm
stage, the main research aspects of cloud robot technology
An instance-based algorithm is often used to build game
are building a networked robot system platform, an open
problem models. In the actual application process, after
system robot architecture, robot network platform algo-
the model is established, it is often necessary to select a
rithm, image processing system development, and cloud
batch of sample data, then the computer system will ana-
robot related network infrastructure construction.[6]
lyze and compare new data with sample data according to
In China, Baidu is the company with the deepest re-
certain similarity characteristics, and find the best match-
search in the field of artificial intelligence. As an Internet
ing data from new data. Therefore, the instance-based
company, Baidu’s most significant research achievement
algorithm is also called memory-based learning. Common
in recent years is the Deep Speech 2 speech recognition
instance-based algorithms include KNN, learning vector
system, which has not only been recognized by the world
quantization, self-organizing mapping algorithms, etc.
but also successfully been chosen by MIT Technology
Review as one of 2016 Top Ten Progressive Technolo- 4.3 Regularization algorithm
gies. As an Internet company, Alibaba has developed an
intelligent customer service system based on the actual Generally speaking, the regularization algorithm can be
needs of its own e-commerce platform, which effectively regarded as an extension of the regression algorithm. This
deals with current restriction problem on the sharp rise in algorithm will adjust algorithm reasonably according to
customer service demand for e-commerce platforms, and the complexity of it. In the adjustment process, the regu-
promotes basic customer service tasks such as automatic larization algorithm usually rewards simple models, and
voice transfer, emotion recognition and keyword recog- punishes complex algorithm correspondingly. Common
nition to be replaced by intelligent customer service. This regularization algorithm includes Ridge Regression-LAS-
intelligent customer service system not only effectively SO-and resilient network.
reduces the workload of manual customers, but also re-
4.4 Decision-Tree Learning Algorithm
duces the company’s labor demand costs, generates more
economic benefits, and lays the foundation for further In the actual application process, the decision-tree al-
development. Tencent has also achieved certain achieve- gorithm will reward the tree-structure decision model
ments in the field of artificial intelligence, among which according to data attributes. This decision model is often
the most advanced artificial intelligence technology is used to solve classification and regression problems. The
the face recognition technology developed by “Tencent common decision-tree algorithm includes regression trees,
Youtu”. This technology is at the leading level even in the ID3, C4.5, CHAID, Decision Stump, random forest, mul-
whole world, it also provides strong image and pattern tivariate adaptive regression splines and gradient boosting
recognition technology support for Tencent subordinate machines.
industries such as QQ, Tenpay, Webank, etc. In addition to
the above three major companies, other companies such 4.5 Bayesian method
as iFLYTEK and AISpeech have also achieved great R & Bayesian method algorithm is a kind of algorithm de-
veloped on the basis of Bayesian positioning, which is are semi-supervised learning algorithms, which are mostly
often used to solve classification and regression problems. used to process large data sets with a small amount of un-
Common Bayesian method algorithm includes the Bayes- identified data. [8]
ian naive algorithm, averaged one-dependence estimator
and BBN. 5. Conclusion
4.6 Kernel-based algorithm With the rapid development of computer technology and
information technology, artificial intelligence technol-
Among all kernel-based algorithms, the support vector ogy has also been deeply researched and popularized.
machine algorithm is the most famous. In practical appli- Although existing artificial intelligence technology has
cations, the kernel-based algorithm will map input data to learning ability to some extent, it is difficult to meet the
a higher-order vector space, in which data is processed. In ever-changing social demand due to limitations of theory,
high-order vector space, some classification or regression algorithm and technology. Therefore, it is necessary for
problems can be better solved and data resolution can be academics from home and abroad to further study and
improved. The common kernel-based algorithm includes develop artificial intelligence, to promote the further im-
support vector machines, radial basis functions, and linear provement of machine learning algorithms in the field of
discriminant analysis. artificial intelligence, to ensure the effective growth of
the learning ability of artificial intelligence technology, to
4.7 Clustering algorithm meet social demand, and to contribute more to social de-
In application, the clustering algorithm will classify and velopment.
summarize all input data in a layered manner according to
References
the central point process, then find out internal structures
of data and determine the biggest common point, based [1] Xie Juanying, Liu Ran. Research progress of target
on which input data will be classified. The common clus- detection algorithm based on deep learning[J]. Jour-
tering algorithm includes the k-Means algorithm and the nal of Shaanxi Normal University (Natural Science
expectation maximization algorithm. Edition), 2019(5):1-9.
[2] Sun Shaojing, Chen Changfeng, Li Shigang. Devel-
4.8 Association rules algorithm opment and Challenge of ‘Algorithm Recommen-
Computer systems can choose rules that can explain the dation and Artificial Intelligence’[J]. Journalism
relationship among variables through association rules Bimonthly, 2019(6):40-47.
[3] Tian Qichuan, Wang Manli. Research progress of
learning, then find out relevant ones from a large amount
deep learning algorithms[J]. Computer Engineering
of data. Common association rules algorithm includes the
and Applications, 2019(22):25-33.
Apriori algorithm and the Eclat algorithm.
[4] Chang Yingchun. Anomaly detection of game artifi-
4.8 Artificial Neural Network cial intelligence based on layered time series memory
algorithm[D]. Journal of Hebei University of tech-
In application, an artificial neural network algorithm will nology, 2015.
simulate biological neural networks and solve classifica- [5] Gan Jingzhong, Yang Xiulan, Lv Jie. Overview of
tion and regression problems. Nowadays, as the focus of Unsupervised Learning Algorithms in Artificial In-
academic research, an artificial neural network algorithm telligence[J]. Technology and Industry Across the
has covered hundreds of different types of algorithms, Straits, 2019(1):134-135.
among which deep learning algorithm is the most com- [6] Klaus Mainzell. Artificial Intelligence and Machine
monly used one. Common used artificial neural network Learning: Algorithm Basics and Philosophical Per-
algorithm includes perceptron neural network, reverse spectives[J]. Journal of Shanghai normal university
transfer, Hopfield network, self-organizing mapping and (Philosophy and Social Sciences), 2018(3):13-24.
learning vector quantization. As an important algorithm [7] Zhou Jin. Deep learning algorithms on artificial intel-
category derived from the development process of artifi- ligence[J]. Computer Programming Skills & Mainte-
cial neural network algorithms, deep learning algorithm nance, 2019, (7):115-117.
has now received extensive attention from home and [8] Huang Fayang. Face Recognition Technology Based
abroad and experienced in-depth research and develop- on Neural Network Deep Learning Algorithm[J]. In-
ment. However, most of today’s deep learning algorithms telligent building, 2018(10):65-67.
Ⅰ. Format
Ⅱ. Cover Letter
Ⅲ. Abstract
A general introduction to the research topic of the paper should be provided, along with a brief summary of its main
results and implications. Kindly ensure the abstract is self-contained and remains readable to a wider audience. The
abstract should also be kept to a maximum of 200 words.
Authors should also include 5-8 keywords after the abstract, separated by a semi-colon, avoiding the words already used
in the title of the article.
Abstract and keywords should be reflected as font size 14.
Ⅳ. Title
The title should not exceed 50 words. Authors are encouraged to keep their titles succinct and relevant.
Titles should be reflected as font size 26, and in bold type.
Ⅳ. Section Headings
Ⅴ. Introduction
The introduction should highlight the significance of the research conducted, in particular, in relation to current state of
research in the field. A clear research objective should be conveyed within a single sentence.
Ⅵ. Methodology/Methods
In this section, the methods used to obtain the results in the paper should be clearly elucidated. This allows readers to be
able to replicate the study in the future. Authors should ensure that any references made to other research or experiments
should be clearly cited.
Ⅶ. Results
In this section, the results of experiments conducted should be detailed. The results should not be discussed at length in
this section. Alternatively, Results and Discussion can also be combined to a single section.
Ⅷ. Discussion
In this section, the results of the experiments conducted can be discussed in detail. Authors should discuss the direct and
indirect implications of their findings, and also discuss if the results obtain reflect the current state of research in the field.
Applications for the research should be discussed in this section. Suggestions for future research can also be discussed in
this section.
Ⅸ. Conclusion
This section offers closure for the paper. An effective conclusion will need to sum up the principal findings of the papers,
and its implications for further research.
Ⅹ. References
References should be included as a separate page from the main manuscript. For parts of the manuscript that have
referenced a particular source, a superscript (ie. [x]) should be included next to the referenced text.
[x] refers to the allocated number of the source under the Reference List (eg. [1], [2], [3])
In the References section, the corresponding source should be referenced as:
[x] Author(s). Article Title [Publication Type]. Journal Name, Vol. No., Issue No.: Page numbers. (DOI number)
J = Journal/Magazine
M = Monograph/Book
C = (Article) Collection
D = Dissertation/Thesis
P = Patent
S = Standards
N = Newspapers
R = Reports
Kindly note that the order of appearance of the referenced source should follow its order of appearance in the main manu-
script.
Graphs, Figures, Tables, and Equations
Graphs, figures and tables should be labelled closely below it and aligned to the center. Each data presentation type
should be labelled as Graph, Figure, or Table, and its sequence should be in running order, separate from each other.
Equations should be aligned to the left, and numbered with in running order with its number in parenthesis (aligned
right).
Ⅻ. Others
Conflicts of interest, acknowledgements, and publication ethics should also be declared in the final version of the manu-
script. Instructions have been provided as its counterpart under Cover Letter.
About the Publisher
Bilingual Publishing Co. (BPC) is an international publisher of online, open access and scholarly peer-reviewed
journals covering a wide range of academic disciplines including science, technology, medicine, engineering, education
and social science. Reflecting the latest research from a broad sweep of subjects, our content is accessible world-
wide——both in print and online.
BPC aims to provide an analytics as well as platform for information exchange and discussion that help organizations
and professionals in advancing society for the betterment of mankind. BPC hopes to be indexed by well-known
databases in order to expand its reach to the science community, and eventually grow to be a reputable publisher recognized
by scholars and researchers around the world.
Database Inclusion