Recent Developments in Automatic Control Systems
Recent Developments in Automatic Control Systems
Control Systems
RIVER PUBLISHERS SERIES IN AUTOMATION,
CONTROL AND ROBOTICS
Series Editors:
ISHWAR K. SETHI
Oakland University, USA
TAREK SOBH
University of Bridgeport, USA
FENG QIAO
Shenyang JianZhu University, China
Editors
Yuriy P. Kondratenko
Petro Mohyla Black Sea National University of the Ministry of Education
and Science of Ukraine, Ukraine
Vsevolod M. Kuntsevich
Space Research Institute of the National Academy of Sciences of Ukraine
and State Space Agency of Ukraine, Ukraine
Arkadii A. Chikrii
Glushkov Institute of Cybernetics of the National Academy of Sciences of
Ukraine, Ukraine
Vyacheslav F. Gubarev
Space Research Institute of the National Academy of Sciences of Ukraine
and State Space Agency of Ukraine, Ukraine
River Publishers
Published 2023 by River Publishers
River Publishers
Alsbjergvej 10, 9260 Gistrup, Denmark
www.riverpublishers.com
© 2023 River Publishers. All rights reserved. No part of this publication may
be reproduced, stored in a retrieval systems, or transmitted in any form or by
any means, mechanical, photocopying, recording or otherwise, without prior
written permission of the publishers.
Preface xiii
v
vi Contents
Index 445
xiii
xiv Preface
describing the change in the value of the controlled parameter over time. As
an estimate of the future state, the conditional mathematical expectation is
used; to estimate the probability of no-failure operation, statistical modeling
of a random sequence beyond the observation area is performed. The method
can also be used to control the reliability of the objects that are characterized
by many parameters.
The chapter “Petunin Ellipsoids in Automatic Control Systems Design”,
by Klyushin D.A., Lyashko S.I. and Tymoshenko A.A., proposes a new way
of describing the optimization domain with linear constraints in the design
space using Petunin ellipsoids. The statistical properties of these ellipsoids
are given, making them an effective tool for describing data in the design
space. Authors consider a new distribution-free approach to computing ellip-
soidal conformal prediction set based on Hill’s assumption and the Petunin
ellipsoids. The resulting prediction set is a Petunin ellipsoid with a precise
confidence level depending on the number of points only.
In ”On Real-Time Calculation of the Rejected Takeoff Distance”, Vyshen-
skyy V.I., Belousov A.A. and Kuleshyn V.V. focus on a rejected takeoff in
the situation when it is decided to cancel the takeoff of an airplane. The
goal of the authors’ investigation is to obtain real-time estimations of the
distance to a full stop. The proposed real-time algorithm for the calculation
of the distance of the interrupted flight uses the step-wise interpolation of the
thrust. Using this interpolation, the Riccati equation of the aircraft motion
has explicit solutions. Besides using known time intervals of interpolation,
the time to reach a given speed can be explicitly calculated. To check and
study the proposed algorithm for estimating the rejected takeoff distance in
real time, a computer simulation system was created.
Romanenko V., Miliavskyi Y. and Kantsedal G. in the chapter “Automated
Control Problem for Dynamic Processes Applied to Cryptocurrency in Finan-
cial Markets” develop a cognitive map (CM) of the use of cryptocurrency
in the financial market and describe the dynamic model of CM impulse pro-
cesses as a difference equation system (Roberts equations). The authors chose
an external control vector for the CM impulse process provided by means of
CM nodes varying. A closed-loop CM impulse process control system was
implemented. This system includes a multidimensional discrete controller,
based on automated control theory methods, that generates a selected control
vector and directly affects respective CM nodes by varying their coordi-
nates. The authors solved three problems of a discrete controller design for
automated dynamic process control applied to cryptocurrency in financial
markets. The first problem is unstable cryptocurrency rate stabilization based
Preface xix
xxi
xxii List of Figures
Figure 8.11 First singular values of the matrix Ỹrelax for systems
of different order. Case of relative error bound
εY = 10−3 . . . . . . . . . . . . . . . . . . . . . 190
Figure 8.12 First normalized singular values of the matrix Ỹrelax
for systems of different order. Case of relative error
bound εY = 10−3 . . . . . . . . . . . . . . . . . . 191
Figure 8.13 Relative max norm of the matrix YrelaxN . . . . . . 191
Figure 8.14 Accuracy of the model, depending on the dimension
and the noise . . . . . . . . . . . . . . . . . . . . 193
Figure 9.1 Structure of the generalized MIMO fuzzy control/
decision-making system with structural optimiza-
tion unit . . . . . . . . . . . . . . . . . . . . . . . 202
Figure 9.2 Flowchart of the first modification of the informa-
tion technology for FS structural optimization based
on optimal selection of linguistic terms number . . 210
Figure 9.3 Flowchart of the second modification of the infor-
mation technology for FS structural optimization
based on optimal selection of linguistic terms
number . . . . . . . . . . . . . . . . . . . . . . . 212
Figure 9.4 Flowchart of the stage of design of the RB and
calculation of the complex objective function for the
FS with the current vector S . . . . . . . . . . . . 213
Figure 9.5 The fuzzy control system of the QUAV flight
altitude . . . . . . . . . . . . . . . . . . . . . . . 217
Figure 9.6 Complex relief of the mountainous terrain presented
by the dependence zT (x) . . . . . . . . . . . . . . 217
Figure 9.7 Changing curves of the complex objective func-
tion J C and the criteria J 1 , J 2 values during the
structural optimization process (first modification) . 220
Figure 9.8 Changing curves of the complex objective func-
tion J C and the criteria J 1 , J 2 values during the
structural optimization process (second modification) 223
Figure 9.9 QUAV’s flight trajectories over the mountainous
terrain . . . . . . . . . . . . . . . . . . . . . . . . 225
Figure 9.10 Detailed QUAV’s flight trajectories (z = –2. . . 12 m;
x = 0. . . 100 m) . . . . . . . . . . . . . . . . . . . 226
Figure 9.11 Detailed QUAV’s flight trajectories (z = 5. . . 35 m; x
= 780. . . 900 m) . . . . . . . . . . . . . . . . . . . 226
xxiv List of Figures
xxix
xxx List of Tables
xxxi
xxxii List of Contributors
xxxv
xxxvi List of Notations and Abbreviations
d˜0
dt operator of local derivative in the CXc Yc Zc coordinate
system
d˜1
dt local derivative operator in the O1 X1 Y1 Z1 coordinate
system
d
dt operator of the derivative in the inertial AXY Z
coordinate system
α rudder angle
η parametric noise
ω angular speed
ψ roll angle
θ pitch angle
ε error value
ϕ yaw (course) angle
m mass of the system
mi mass of the ith elementary particle of the AMR
platform
mj mass of the j th elementary particle of the manipulator
mM mass of the manipulator
tsij –i th switching moments for j th state coordinates
A damping coefficient matrix
AX matrix parameters that depends on the coordinates
(6×6)
Am , Bm , Cm matrices of the reference mathematical model
BX matrix of control parameters that depends on the
coordinates (6×6)
CX matrix of coefficients of perturbations that depends on
the coordinates (6×6)
E error vector
F vector of external disturbances (6×1)
Fcont measured and controlled disturbances
g gravity acceleration vector (6×1)
I unit matrix
Q, G1 , G2 matrices of weighting coefficients
R space of variables
S matrices of sensitivity functions
U control vector (6×1)
Umax maximum control constraints
Ucor corrective robust control
Uopt optimal control
List of Notations and Abbreviations xxxvii
A. Chikrii
Abstract
This chapter is devoted to an overview of the results concerning the differen-
tial games of pursuit–evasion with many participants on the opposing sides.
According to L. S. Pontryagin, such problems are the pinnacle of the theory of
conflict-controlled processes. The basic method of research is the method
of resolving functions. Here the problems with phase constraints, problems
of group and successive pursuit as well as the problems of interaction of
groups of participants are considered. Sufficient and sometimes the necessary
conditions for the completion of the game in a finite time are given. In
this case, quasi and stroboscopic strategies are used. The results are used
to solve several model examples from the book of R. Isaacs illustrating game
situations.1
1
This work was partially supported by National Research Foundation of Ukraine.
Grant No 2020.02/ 0121
3
4 Control of Moving Object Groups in a Conflict Situation
1.1 Introduction
Several effective mathematical methods have been developed in the theory of
dynamic games (differential games, conflict-controlled processes). In some
cases, they provide optimal results. From a practical point of view, the concept
of a guaranteed result is more appropriate. The game problems of pursuit–
evasion are the most interesting.
A survey of results in the theory of evasion is contained in [1]. The works
[2–6] provide an idea of the methods of pursuit. The last one describes the
method of resolving functions. This method theoretically substantiates the
rule of parallel pursuit and a more general method of pursuit along a beam,
well known to engineers engaged in rocket and space technology [7].
The mathematical apparatus for substantiating various schemes of the
method of resolving functions are the methods of applied nonlinear analysis,
in particular, the technique of set-valued mappings and their selections [8, 9].
The creation and development of the method [6] were stimulated by a special
‘minmaxmin’ function, introduced by the author in the theory of evasion,
and the chapter [10], which laid the foundation for group pursuit problems
[6, 10–13]. Subsequently, this method was used to study the traveling sales-
man problems, namely, the problems of successive pursuit [6–14], and the
problems with state constraints [2, 6, 15].
The method of resolving functions is based on the natural accumulative
principle. Reaching in the sum a certain value of the total winnings means the
game termination, that is, the fact that the trajectory of the conflict-controlled
process has hit a given terminal set.
The versatility of the method makes it possible to cover in a unified
scheme a wide range of dynamic processes which is proceeding in conditions
of conflict and uncertainty. These include, first of all, functional-differential
systems, stochastic processes, as well as processes with partial and fractional
derivatives [16–23].
n ≥ 2, ν ≥ 2,
1.2 Function ω(n, ν), Encirclement by Pshenihnyi 5
That the evader is captured by the ith pursuer means that xi = y. It was shown
that under the condition ω(n, ν) · a > 1 the straight motion of the evader in a
suitable direction provides avoiding meeting with the pursuers for any initial
states.
The next important step after the introduction of the function ω(n, ν)
was made by B.N. Pshenichnyi [10]. Let us put a = 1 and analyze the
problem from the point of view of a group of pursuers. In the expression
for the function ω(n, ν) we set
x0i − y 0
pi =
x0 − y 0 , i = 1, ..., ν,
i
M ∗ = M 0 + M, (1.4)
Q : Δ × V → 2R+ .
Its support function in the direction +1 is
α(t, τ, v) = sup{α ≥ 0 : α ∈ Q (t, τ, v)}, (t, τ ) ∈ Δ, v ∈ V. (1.6)
We call this function the resolving function [6]:
αX (p) = sup{α ≥ 0 : αp ∈ X}, p ∈ Rn , 0 ∈ X, X = X̄.
Since Pontryagin’s condition is satisfied, then the set-valued mapping
Q(t, τ, v) has nonempty closed direct images on the set Δ × V . In the
case ξ(t, g(t), γ(t, ·)) ∈ M we have Q(t, τ, v) = [0, ∞), and, therefore,
α(t, τ, v) = + ∞ for τ ∈ [0, t], v ∈ V. This corresponds to Pontryagin’s
first direct method [6]. Using the theorems on the characterization and the
inverse image [9], it can be shown that the set-valued mapping Q(t, τ, v) is
L × B-measurable for the set (τ, v), τ ∈ [0, t], v ∈ V, and the resolving
function α(t, τ, v) is joint L × B-measurable in (τ, v), by the support
function theorem [9].
Let V∗ be the set of measurable functions with values in V. We consider the
set
t
T (g(·), γ(·, ·)) = t > 0 : inf α(t, τ, v (τ )) dτ ≥ 1 . (1.7)
v (·)∈V∗ 0
8 Control of Moving Object Groups in a Conflict Situation
Then the trajectory of the process (1.3) can be brought by the pursuer to the
terminal set (1.4) at a time T using a suitable quasi-strategy.
Moreover, if the mapping Q(t, τ, v) is convex-valued, then the process
of approach can be realized in the class of stroboscopic strategies [25].
There are several modifications of the method of resolving functions
related to various aspects of the game problem. These are a scheme for
bringing the trajectory to a given point of the set M [6], a variant of the
pursuit method when the pursuer employs only stroboscopic strategies [25],
a case when the function α(t, τ, v) is t- independent that makes it feasible
to terminate the game before the estimated time at the account of the evader’s
mistakes [6], as well as the functional form of Pontryagin’s first direct method
[25, 27]. Various forms of the pursuer advantage–Pontryagin’s condition,
bring essential changes to the method scheme. For oscillatory processes, the
Pontryagin integral condition is sometimes used [6]. For the problems of soft
meeting (the instantaneous coincidence of the players’ geometric coordinates
and velocities), an effective result is achieved by using the principle of time
dilation [28], which is based on the modified Pontryagin’s condition.
The method of resolving functions makes it feasible to study in a unified
scheme the group pursuit problems, the problems with phase constraints as
the problems with static “recumbent” pursuers [6, 15], the game problems
of successive approach—traveling salesman problems [6, 14], the pursuit–
evasion problems under the interaction of groups [6], the problem of multiple
captures [11, 12], and the problem of pursuit of the formation [12]. The form
of solution representation [1] allows extending the method to the nonlinear
case.
Not so long ago, the schemes for two types of upper and lower resolving
functions were created [29, 30], focused on the case when 0 ∈ / Q(t, τ, v)
and Pontryagin’s condition does not hold. Also, the matrix resolving function
was introduced. This makes it possible not only to attract the set M −
1.3 Group Pursuit of a Moving Object 9
ξ(t, g (t), γ (t, ·)) to the intersection from (2.5) in the cone spanned on
it, but to rotate it as well [31].
Qi (t, τ, v) =
= {αi ≥ 0 : [Wi (t, τ, v) − γi (t, τ )] αi [Mi − ξi (t, gi (t), γi (t, ·))] = ∅},
Qi : Δ × V → 2R+ .
v ∈ V, i = 1, ..., ν.
Let us set
g ν (·) = (g1 (·) ... gν (·)), γ ν (·, ·) = (γ1 (·, ·)...γν (·, ·)),
Theorem 1.2. Let for the group pursuit problem (1.8), (1.9) an analog of
Pontryagin’s condition be satisfied, the sets Mi , i = 1, ..., ν, are con-
vex and suppose that for function g ν (·) there exist measurable selections,
γ ν (·, ·), γi (t, τ ) ∈ Wi (t, τ ), (t, τ ) ∈ Δ, i = 1, ..., ν, and an integer
T ν ∈ T ν (g ν (·), γ ν (·, ·)) =φ.
Then there is a quasi-strategy of the pursuers such that at least one of the
trajectories (1.8) can be brought to the corresponding set at the time instant
T ν for any admissible control of the evader.
By the way of illustration, below is given a simple example of the group
pursuit with the same control domains of the players
The encirclement by Pshenichnyi [10] for the initial states z 0 = (z10 , ..., zν0 )
can be expressed in similar forms [6], namely:
1) δ(z 0 ) = min max αi (zi0 , v) > 0;
v ≤ 1 i=1,...,ν
0
zi
2) σ(z 0 ) = min max zi0
, v > 0;
v ≤ 1 i=1,...,ν
0
z
3) 0 ∈ intco zi0 .
i
Below is given the statement that provides necessary and sufficient conditions
for the successful termination of the group pursuit–evasion and the pursuit at
12 Control of Moving Object Groups in a Conflict Situation
the moment
T0ν = min T ν (z 0 , 0).
0
z
Corollary 1.1. If 0 ∈ intco zi0 , where z 0 is the initial state of the
i
process (1.11), then the group pursuit can be terminated at the time T0ν and
the following estimate is valid: T0ν ≤ δ(zν0 ) . Herewith, if t∗ = t∗ (v(·)) is the
time of control switching, i.e. a zero of the control function
t
1 − max αi (zi0 , v (τ )) dτ,
i=1,...,ν 0
then the controls of the pursuers, realizing time T0ν on the active section
[0, t∗ ], are
and on the passive one (t∗ , T0ν ] ui (τ ) = v(τ ) for indices i, satisfying the
equality
t∗ t∗
0
αi (zi , v(τ )) dτ = max____ αi (zi0 , v (τ )) dτ,
0 i∈ 1,ν 0
and for the rest of the indices, the controls of the pursuers at the interval
(t∗ , T0ν ] can be arbitrary admissible. If
0
zi
0∈intco
¯ ,
zi0
then the escape is possible.
Corollary 1.2. Let in the group pursuit problem (1.11)
Then, if
0 ∈ int co {zi0 + εi S},
then at least one of the pursuers can catch the evader in a finite time. If
0 ∈
¯ int co {zi0 + εi S},
The goal of the pursuer’s group is the exact capture of the evader by at
least one of the pursuers (1.12). The result has the form of necessary and
sufficient conditions [10], which is similar to Corollary 1 and realizes a
parallel approach (1.13) on the active section, with the only difference that
the following estimate is fulfilled for the guaranteed time
1 a∗ · ν
T (z , 0) ≤
ν 0
ln 1 + , a∗ = − max
____ ai .
a∗ δ(z 0 ) i= 1,ν
Second-order objects.
żi1 = zi2 ,
where zi1 is the difference between the geometric coordinates of the pursuer
xi and the evader y, zi1 = xi − y, and zi2 is the difference of their speeds,
zi2 = ẋi − ẏ.
Because of the equivalence of the Equation (1.14) to the second-order
equations:
ẍi + a1 ẋi + a2 xi = ui ,
ÿ + a1 ẏ + a2 y = v,
the characteristic polynomial takes the form
λ2 + a1 λ + a2 = 0. (1.15)
We will restrict ourselves to the case of real roots and present the results in the
form of consequences of the general method scheme for classical examples
[13, 15].
14 Control of Moving Object Groups in a Conflict Situation
and set
t
ν
T (z) = min t ≥ 0 : inf max
_____ αi (t, τ, zi , v (τ )) dτ ≥ 1 .
v(·)∈V∗ i= 1,ν 0
Theorem 1.3. Suppose that for the process (1.16), (1.9) Mi = {0}, πi Ai =
Ai πi , i = 1, ..., ν, the initial position z 0 satisfies Condition 1.1, and
T ν (z 0 ) < + ∞.
Then the group pursuit can be terminated starting from the initial state z 0
no later than in time T ν (z 0 ) in the class of stroboscopic strategies.
Let us put Ai = {0}, i = 1, ..., ν. Then
żi = ϕi (ui , v), i = 1, ..., ν. (1.17)
Now we consider the set-valued mappings
____
Wi (zi , v) = πi ϕi (Ui , v) con(Mi − πi zi ),
____
W̃i (zi , v) = co {πi ϕi (Ui , v)} con(Mi − πi zi ),
and for zi ∈ Rni \Mi∗ the sets
Wi = {zi : Wi (zi , v) = φ∀v ∈ V },
W̃i = {zi : W̃i (zi , v) = φ∀v ∈ V }.
The resolving functions have the forms
ᾱi (zi , v) = max {α ≥ 0 : πi ϕi (Ui , v)
zi ∈ W̃i , v ∈ V, i = 1, ...ν.
We set
t
T̃ ν (z) = min {t ≥ 0 : inf max
_____ ᾱi (zi , v(τ ))dτ ≥ 1} .
v(·)∈V∗ i= 1,ν 0
Theorem 1.4. Let zi0 ∈ Wi , i = 1, ..., ν, for the initial state of the process
(1.17) z 0 and ᾱ(z 0 ) > 0. Then the group pursuit problem can be solved in the
class of counter-controls no later than in time T̃ ν (z 0 ) and T̃ ν (z 0 ) ≤ ᾱ(zν 0 ) .
If, otherwise, for z 0 , zi0 ∈ W̃i , i = 1, ..., ν, and α̃(z 0 ) = 0, then the
problem of evasion from a group of pursuers can be solved under the constant
control of the evader, furnishing minimum in (1.18).
Let be
Here V is a compact set from Rn , ∂coV is the boundary of the set coV. The
terminal set consists of convex compact sets M1 , ..., Mν from Rn .
Now we deduce necessary and sufficient conditions, ensuring solvability
of the group pursuit problem (1.19) for given initial states.
Let us set
____
W i (zi , ν) = [∂coV − ν] con[Mi − zi ],
W i = {zi :W i (zi , ν) = φ∀ν ∈ V }, i = 1, ..., ν,
α i (zi , ν) = max{α ≥ 0 : [∂coV − ν] [Mi − zi ] = φ},
zi ∈W i , zi ∈
¯ Mi , v ∈ V,
α (z) = inf max
_____
α i (zi , v),
v∈V i= 1,ν
t
T̂ ν (z) = min t ≥ 0 : inf max
_____
αi (zi , v (τ )) dτ ≥ 1 .
v(·) i= 1,ν 0
We now take into consideration the support function and the support set for
an arbitrary compact set X from the space Rn , respectively.
In the case the images of the mapping H(X; p) consist of single points for
any p ∈ Rn , they say that X is a strictly convex compact.
The set X is called a compact with smooth boundary if H(X; p1 )
H(X; p2 )=φ ∀p1 , p2 , p1 = p2 , p1 = p2 = 1.
Lemma 1.1. Let for the process (1.19) the set V be a strictly convex compact
with smooth boundary. Then α (z 0 ) > 0 if and only if
⎡ ⎤
0 ∈ intco ⎣ (Mi − zi0 )⎦ . (1.20)
_______
i=1,...,ν
Here Mi0 are linear subspaces from Rni , Mi are nonempty convex compact
sets from Li , the orthogonal complements to Mi0 in Rni , and Mi = {ai },
i ∈ Nν \Nk , k ≤ ν, and ai are vectors from Li , dim Li = 1.
Consequently, for i ∈ {k + 1, ..., ν}, dim Mi0 = ni − 1 and sets Mi∗ appear
as affine manifolds. We assume that the pursuers use quasi-strategies. We
denote by πi , the operator of orthogonal projecting from Rni to Li , i ∈ Nν .
Let us introduce the set-valued mappings
Wi (t, τ ) = πi eAi t ϕi (Ui , v), i ∈ Nν .
For i ∈ Nk , we set Wi (t) = Wi (t, v), t ≥ 0.
v∈V
Condition 1.2. Direct images of mappings Wi (t) are nonempty for all i ∈
∈ Nk , t ≥ 0.
We fix some measurable selections γi (t), γi (t) ∈ Wi (t), i ∈ Nk , and set
t
0 tAi 0
ξi (t, zi , γi (·)) = πi e zi + γi (τ ) dτ.
0
αi (t, τ, zi0 v, γi (·)) = sup {α ≥ 0 : [Wi (t − τ, v) − γi (t − τ )]
α[Mi − ξi (t, zi0 , γi (·))] = φ}.
They enjoy all the properties provided earlier. Consider the case i ∈ Nν \Nk .
Condition 1.3. Set-valued mappings Wi (t, v), i ∈ Nν \Nk , are continuous
single-valued functions ωi (t, v).
Let us set
ξi (t, zi0 ) = πi etAi zi0 , i ∈ Nν \Nk ,
and write the resolving functions for i = k + 1, ..., ν:
(pi , ωi (t−τ, v))
0 , pi ∈ Li , pi = 1, ai = ξi (t, zi ),
0
∗ 0 (p i , ai −ξi (t,zi ))
αi (t, τ, zi v) =
ωi (t − τ, v) + 1t , ai = ξi (t, zi0 ).
It is supposed that they are continuous in t, t > 0, γ(·) = (γ1 (·), ..., γν (·).
Let us take into consideration the function
t
ν 0
T∗ (z , γ(·)) = min t ≥ 0 : inf max max αi (t, τ, zi0 , v (τ ),
v(·) i∈Nk 0
1.5 The Group Pursuit. Linear State Constraints 19
t
γi (·)) dτ, max αi∗ (t, τ, zi0 , v (τ )) dτ ≥1 .
i∈Nν \Nk 0
Then, the trajectory of the process (1.16) can be brought from the initial state
z 0 to the corresponding set Mi∗ at the moment T∗ν (z 0 , γ(·)) for at least one
i, i ∈ Nν , using quasi-strategies.
By way of illustration, let us study the linear problem of control of a group
of pursuers when the evader cannot leave some open polyhedral set.
The dynamics of the pursuers and the evader have the forms, respectively:
ẏ = B y + v, y ∈ Rs , v ∈ V,
Ci are quadratic matrices of order ri , and B is a matrix of the order s, Ui , V
are nonempty compact sets. The motion of the evader is bounded by the
constraints
zi = y, Ai = B, ϕi (u, v) = v,
Thus, the group pursuit problem (1.22) under the state constraints (1.23)
is reduced to a group pursuit problem (1.16), (1.21) without constraints.
The above-outlined scheme can be used to study the game problem with
arbitrary state constraints being convex sets [6].
Let us analyze in detail group pursuit for ‘simple motions’
ẋi = ui , ui ≤ 1, xi ∈ Rs , s ≥ 2, i = 1, ..., k, ẏ = v v ≤ 1, y ∈ Rs ,
(1.24)
under the state constraints (1.23).
The group pursuit terminates if xi − y ≤ εi , εi ≥ 0, for some i ∈ Nk , or
the evader hits the boundary of set G, i.e. for some i ∈ Nν \Nk the equality,
(pi , y) = li takes place. The game problem (1.24), (1.23) is easily reduced
to a differential group pursuit game. Let us put zi = xi − y. i ∈ Nk . Then
żi = ui − v, ui ≤ 1, v ≤ 1, Mi∗ = {zi : zi ≤ εi }, i = 1, ..., k.
(1.25)
We set zi = y., i ∈ Nν \Nk . Then we have
żi = v, v ≤ 1, Mi∗ = {zi : (pi , zi ) = li }, i = k + 1, ..., ν. (1.26)
It is evident that the analog of Pontryagin’s condition for a group is
fulfilled, the selections γi (t) ≡ 0, i ∈ Nk , t ≥ 0. Hence, the functions
ξi (t, zi0 , 0) = zi0 , ξi (t, zi0 ) = zi0 , i ∈ Nν \Nk , ωi (t, v) = v, and the
resolving functions take the forms
αi (t, τ, zi0 , v, 0) = ρi (zi0 , v) =
2
(v, zi ) + εi + ((v, zi0 ) + εi )2 + (zi0 − ε2i )(1 − v2 )
0
= 2 , i ∈ Nk ,
z 0 − ε2
i i
(pi , v)
αi∗ (t, τ, zi0 , v) = ρi (zi0 , v) = , z 0 = ai = li pi , i ∈ Nν \Nk .
li − (pi , zi0 ) i
The termination time of the group pursuit game (1.25), (1.26) is defined as
the smallest positive root of the equation
t
inf max ρi (zi0 , v(τ )) dτ = 1. (1.27)
v(·) i∈Nν 0
Theorem 1.7. Let z 0 = (z10 , ..., zν0 ) be the initial state of the process (1.25),
(1.26). If there is a number i ∈ Nk , such that εi > 0, or k ≥ s, and
0 ∈ int {co [z10 + ε1 S, ..., zk0 + εk S] + con (pk+1 , ..., pν )}, (1.28)
Lion versus man [2]. Let k = 1, ε1 > 0, the pursuer and the evader move
according to equations (1.24), ν − 1 > s, set G be bounded and given by
inequalities (1.23).
Then con{p2 , ..., pv } = Rs and inclusion (1.28) is satisfied for all initial
positions. Hence, capture is always possible. What is more, the result holds
in the case G be an arbitrary convex compact.
Rat cornered [2]. Let motion of a cat and a rat be described by the equations
(1.24), respectively, with k = 1, ε1 > 0, a set G be a convex cone K in
space Rs . Then the barrier cone of a set G, KG = − K ∗ , where K ∗ is a cone
conjugate to K, given the form of the support function
0, p ∈ −K ∗ ,
CK (p) = ¯ − K ∗.
+∞, p∈
Consequently, if there are vectors pi , i = 2, ..., ν, pi ∈ − K ∗ ,pi = 1,
such that y 0 ∈ int{x0 + ε1 S+ +con [p2 , ..., pν ]}, then the capture can be
performed in a finite time.
Patrolling the corridor [2]. Let k = 1, s = 2, ε1 > 0. State constraints are
two parallel straight lines on the plane having the normals p2 , p3 . Without
loss of generality, we can assume that these straight lines are parallel to the
abscissa axis. We see that p2 and p3 are orthogonal to the latter. Therefore,
the condition, which is sufficient for the capture (1.28), takes the form | x01 −
y10 | < ε1 , where x01 , y10 are the first coordinates of the initial states of the
pursuer and the evader, respectively.
Death line game [2]. Let k = 1, ε1 > 0, s = 2. The game takes place on a
half-plane–below the abscissa axis, which is the line of death for the evader,
and the normal p2 coincides in direction with the axis of ordinates. The
inequalities | y10 − x01 | < ε1 , y20 >x02 separate the initial positions of the pur-
suer and the evader, starting from which the capture is possible in a finite time.
‘Simple’ group pursuit. Let k = ν ≥ s, εi ≥ 0. Then sufficient and
necessary condition for capture in the group pursuit problem (1.25) is
0 ∈ int co {z10 + ε1 S, ..., zk0 + εk S}..
with a fixed and non-fixed moment of capture of each of the evaders. The
proposed schemes [6, 14] are especially effective in the case of ‘simple
motions’ with the beforehand programmed choice of the order of capture.
Then the pursuit strategy is the parallel approach and, consequently, the total
capture time depends only on the evader’s controls. In so doing, extremum
of the functional is attained on constant controls of evaders, and the infinite-
dimensional problem of maximizing the total pursuit time is reduced to a
finite-dimensional conditional optimization problem.
Let a conflict-controlled process be described by the quasi-linear
equation
żj = Aj zj + ϕj (u, vj ), z j ∈ Rnj , u ∈ U, vj ∈ Vj , j = 1, ..., μ, (1.29)
Aj are square matrices of the order nj , u, vj are the control parameters
of the pursuer and evaders belonging to the nonempty compacts U and Vj ,
respectively, functions ϕj (u, vj ) are jointly continuous in their variables.
The terminal set consists of cylindrical sets
____
Mj∗ = Mj0 + Mj , Mj∗ ⊂ Rnj , j = 1, μ, (1.30)
where Mj0 are linear subspaces from Rnj , and Mj are convex compact sets
from the orthogonal complements Lj to Mj0 in Rnj .
The
____
goal of the pursuer is to bring in turn all the trajectories zj (t),
j = 1, μ, to the corresponding sets Mj∗ in the least total time. The pursuer
is faced with a double problem, dealing with control and enumerating, which
can in no way be divided and studied separately. Let us denote by zj a pair
(x, yj ), where x is the pursuer state and yj -the evaders. Then inclusion
zj ∈Mj∗ signifies the capture of the jth evader by the pursuer. We suppose
that the pursuer applies quasi-strategies.
We state that the successive pursuit can be completed, starting from the
initial state, z 0 = (z10 , ..., zμ0 ) no later than at the time T, if there is a quasi-
strategy of the pursuer such that for ____
any controls of evaders vT (·) there ____
exist
∗
moments tj , 0 < tj ≤ T, j = 1, μ, such zj (tj ) ∈ Mj for all j = 1, μ .
It should be noted that the values of the function vj (t) are used only on time
intervals [0, tj ]. In the case the order of bringing trajectories zj (t) to the
corresponding sets Mj∗ is chosen beforehand, i.e. is fixed at the initial moment
and then does not change, the number of possible options is μ_!
We denote by Λ the set of all possible orders for capturing the evaders.
Let us assume that at the initial moment the pursuer selects some order
l, l ∈ Λ, and then sticks it. Without loss of generality, we determine by
l = {1, 2, ..., μ} the order of bringing the trajectories to corresponding sets.
24 Control of Moving Object Groups in a Conflict Situation
j = 1, ..., μ, t ≥ 0.
Condition 1.4. The direct images of set-valued mappings W j (t) are not
empty and closed for all j = 1, ..., μ, t ≥ 0.
Since the mappings W j (t) are measurable and closed-valued, then there
exist measurable selections γj (t), γj (t) ∈ Wj (t), t ≥ 0, j = 1, ..., μ [9].
We set t
Aj t
ξj (t, zj , γj (·)) = πj e zj + γj (τ ) dτ.
0
Then the resolving functions take the forms
αj (t, τ, zj , vj , γj (·)) = sup {α ≥ 0 : [W j (t − τ, vj ) − γj (t −
τ )] α[Mj − ξj (t, zj , γj (·))]=φ}, j = 1, ..., μ, t ≥ τ ≥ 0.
We see that they enjoy all the above-listed properties.
Now we determine recursively the moments of trajectories (1.29) hitting the
corresponding sets (1.30) in order l. The following relations are true:
tj = tj (zj (tj−1 ), vj (·)) = tj (z10 , ..., zj0 , v1 (·), ..., vj (·)),
Here
tj
tj = min t ≥ tj−1 : αj (t, τ, zj (tj−1 ), vj (τ ), γj (·)) dτ ≥ 1 ,
tj−1
j = 1, ..., μ.
It is evident that the total time of capturing all the evaders coincides with the
moment of capturing the last evader and, consequently, is equal to tμ .
Let us denote z 0 = (z10 , ..., zμ0 ), v(·) = (v1 (·), . . . , vμ (·)) , and by
Tμl (z 0 , v(·)) = tμ (z10 , ..., zμ0 , v1 (·), ..., vμ (·))
–the total capture time.
This time depends on the order of capturing the evaders and its optimiza-
tion in the case of the programmed order of capture consists in maximizing
in v(·) and minimizing concerning the order, i.e.
Tμ (z 0 , γ(·)) = min sup Tμl (z 0 , v(·)),
l∈Λ v(·)
1.6 Principle of Shortest Broken Line in Successive Pursuit 25
γ(·) = (γ1 (·), ..., γμ (·)) are the fixed forehand selections.
Theorem 1.8 Let Condition 1.4 be satisfied, and there exist measurable
selections of set-valued mappings γj (t) ∈ W j (t), t ≥ 0, j = 1, ..., μ,
such thatTμ (z 0 , γ(·)) < + ∞.
Then, starting from the initial state z 0 , the successive pursuit can be
terminated at the time instant Tμ (z 0 , γ(·)), using quasi-strategies.
To apply the modified method scheme with non-fixed termination time to
solving the successive pursuit problem using stroboscopic strategies requires
significantly more burdensome constraints on the process parameters. It is
suited to the case of ‘simple motions’. For the sake of simplicity and geo-
metric descriptiveness, we restrict ourselves to the case of plane motions.
Let be
żj = u − vj , zj ∈ R2 , zj (0) = zj0 ,
u ≤ 1, vj ≤ β < 1, j = 1, ..., μ, (1.31)
zj = x − yj , ẋ = u, ẏj = vj .
It is easy to observe that each of the pursuers has an advantage over each of
the evaders, otherwise, the problem has no sense. The goal of the pursuer is to
bring in turn all trajectories zj (t) to the origin in a finite time, i.e. Mj∗ :zj = 0.
As before, it is assumed that the system with the lower index has priority in
the order of bringing. We denote
the moment of the first hitting the set Mj∗ . by the trajectory zj (t).
We have
ϕj (u, vj ) = u − vj , j = 1, ..., μ.
Then
αj (t, τ, zj , vj , 0) = αj (zj , vj )
26 Control of Moving Object Groups in a Conflict Situation
(vj , zj ) + (vj , zj )2 + zj 2 (1 − vj 2 )
= , j = 1, ..., μ.
zj 2
Therefore, the pursuer’s control, which brings the jth system (1.31) to
zero, takes the form
yj0 + βtj−1 ∂ S.
Each point of this circle and point x(tj−1 ) possesses an own Apollonius
circle with the corresponding center. Herewith, the centers of the Apollonius
circles, corresponding to the points of the attainable circle of the jth evader,
also form a circle (the circle of centers). Imagine that at each point of the
circle of centers the corresponding Apollonius circle is constructed. Then
the outer envelope of these circles is exactly the locus of the points of the
capture of player yj . The equation of this envelope in both polar and Cartesian
coordinates is presented in [6].
Hence, it is shown that if the order of captures is fixed, the pursuers
apply the law of parallel pursuit and the evaders are moving rectilinearly at
maximum speed, then the total capture time is functional depending only
on the constant controls of the evaders, and its maximization presents the
nonlinear finite-dimensional optimization problem [6].
To select the order of captures ensuring the least total time, the principle
of the shortest broken line was developed [6]. By the shortest broken line
is called a broken line of minimal length, starting from the initial position
of the pursuer and one at a time passing through all the initial positions
References 27
1.7 Conclusion
An overview of contemporary research methods of the problem of decision
making and control in conflict conditions with the participation of moving
object groups is presented. The general problem of the game interaction of
groups with the help of decomposition is divided into separate problems of
group and successive pursuit with subsequent iterative repetition.
Methods for solving the mentioned problems, based on the method of
resolving functions, are presented. Their effectiveness is demonstrated, in
particular, on test examples from the monograph by R. Isaacs.
In so doing, the presence of phase coordinates for the state of objects is
taken into account, and the case of a non-fixed time of the game termination
is also highlighted. The situation of the environment by Pshenichnyi and the
principle of the shortest broken line make the results of group interaction
geometrically descriptive.
References
[1] A. A. Chikrii, ‘Escape Problem for Control Dynamic Objects’, Jour-
nal of Automation and Information Sciences, 1997, vol. 29, No. 6,
pp. 71–82.
[2] R. F. Isaacs, ‘Differential Games: Their Scope, Nature and Future’, J.
Opt. Theory Appl., 1969, No 3, p. 283–295.
[3] L. S. Pontryagin, ‘Linear Differential Games of Pursuit’, Matem.
Sbornik, 1980, 112, No 3, p. 307–330. (in Russian).
[4] B. N. Pschenichnyi, ‘Structure of Differential Games’, Dokl. Akad.
Nauk. SSSR, 1969, 184, No 2. (in Russian).
[5] N. N. Krasovskii, A. I. Subbotin, ‘Positional Differential Games’,
Nauka, Moscow, 1974 (in Russian).
[6] A. A. Chikrii, ‘Conflict controlled processes’, Naukova Dumka, Kiev,
1992 (in Russian).
[7] A. S. Locke, ‘Guidance’, D. Van Nostrand Company, Inc., Princeton,
1955.
28 Control of Moving Object Groups in a Conflict Situation
Boris S. Mordukhovich
Abstract
We present recent developments and applications of variational analysis and
generalized differentiation to rather new classes of optimal control problems
governed by discontinuous differential inclusions that arise in the so-called
sweeping processes and the like. Such problems are particularly important
for numerous applications to automatics control systems, engineering design,
hysteresis, etc. To study highly challenging control problems of this type,
we develop appropriate constructions of the method of discrete approxima-
tions, which allow us to investigate various theoretical and numerical aspects
of optimal control for discontinuous differential systems. In particular, we
establish new necessary optimality conditions for such problems including
appropriate extensions of the Euler–Lagrange conditions and the Pontryagin
maximum principle along with the optimality conditions of the novel type
specific for the problems under consideration. New applications to practical
engineering and mechanical models, robotics, traffic equilibria, etc. are also
discussed.
31
32 Applications of Variational Analysis to Controlled Sweeping Processes
[36, 37], with the vast bibliographies and commentaries therein, for the gen-
esis of ideas, basic constructions and results, and applications of variational
analysis and generalized differentiation in finite and infinite dimensions. In
particular, the second volume of [36] contains a rather comprehensive account
of the methods and results of variational analysis applied to nonsmooth
controlled systems governed by ordinary differential, functional differential,
and partial differential equations and Lipschitzian differential inclusions with
a variety of applications.
New trends in variational analysis largely relate to the investigation and
solving of meaningful problems, which often come from applications. A
broad class of such problems are governed by discontinuous differential
inclusions introduced originally by Moreau [40, 41] in the form of the
sweeping process
for the continuously moving convex set C = C(t) at the point x = x(t). The
primal motivation of [41] mainly came from problems of elastoplasticity, but
over the years the sweeping process and its modifications have been devel-
oped in dynamical system theory with many applications to various areas of
mechanics, engineering, robotics, economics, traffic equilibria, etc.; see, e.g.,
the excellent recent survey [8] with the references therein. Independently,
basic theory and important engineering applications of sweeping processes
were largely developed (not under this name) in the Soviet literature by
Krasnosel’skii, Pokrovskii, and their followers for systems of hysteresis; see
the English book [30] with many references to the previous publications in
Russian.
Optimization and control problems for sweeping processes were formu-
lated much latter. A primal reason for this situation is that the Cauchy problem
for the sweeping process (1) has a unique solution, and so there is nothing to
optimize. To the best of our knowledge, optimal control problems for sweep-
ing processes were first defined with controls functions acting in additive
perturbations as in Edmond and Thibault [22]. However, the focus of [22], as
34 Applications of Variational Analysis to Controlled Sweeping Processes
where the control set U in (4) may depend also of the state U = U (x) and
thus reflects feedback laws. The sweeping inclusions (2.1) are dramatically
different from (2.3), and hence from (2.4), due to highly non-Lipschitzian
(actually discontinuous) structure of the sweeping dynamics (2.1), which
does not allow us to employ the conventional techniques developed earlier
in optimal control theory.
2.1 Introduction and Discussions 35
Another, even more essential difference between (2.1) and the dynamics
in (2.3) and (2.4) is that in the case of controlled moving sets C(t) = C(u(t))
in (2.1), the right-hand site in (2.1) is not fixed as in (2.3) and in (2.4) with
F (x) := f (U (x)). Thus optimizing (2.1) is actually a problem of shape
optimization.
The main goal of this chapter is to discuss some recent results in dynamic
optimization of controlled sweeping processes and some of their applications.
We focus on the following two types of the sweeping dynamics: with controls
acting in moving sets and with controls entering additive perturbations.
These two classes of sweeping optimal controls and the necessary optimality
conditions obtained for them are essentially different from each other.
Nevertheless, we derive necessary optimality conditions for both of these
classes of sweeping control problems by using the method of discrete approx-
imations and advanced tools of generalized differentiation in variational
analysis. Recall that the discrete approximation approach to obtain optimality
conditions for continuous-time variational problems goes back to Euler in the
classical calculus of variations. This method was used by Pshenichnyi [43]
to optimize Lipschitzian differential inclusions governed by set-valued map-
pings with convex graphs and also with convex values under rather restrictive
assumptions. Those restrictions have been overcome for convex-valued Lip-
schitzian differential inclusions, with the usage of generalized differentiation
developed by the author (see Section 2), in the books by Mordukhovich
[35] and Smirnov [46]; see also the references therein. Then the author
[36] established, by employing again the method of discrete approximations,
refined optimality conditions for Lipschitzian differential inclusions without
any convexity assumptions. Below we present major results on necessary
optimality conditions obtained by appropriate developments of the method
of discrete approximations for the aforementioned classes of optimal con-
trol problems for sweeping differential inclusions of highly discontinuous
dynamics.
The rest of the chapter is organized as follows. In Section 2 we review
basic tools of first-order and second-order generalized differentiation in
variational analysis that is broadly used in the formulations and proofs of
the main necessary optimality conditions given below.
Section 3 addresses optimal control problems for sweeping processes
with control functions entering moving sets that generate the sweeping
dynamics. Here we establish necessary optimality conditions of generalized
Euler–Lagrange and Hamiltonian types, which yield a new version of the
Weierstrass–Pontryagin maximization condition in terms of the modified
36 Applications of Variational Analysis to Controlled Sweeping Processes
and then, for (x̄, w̄) ∈ dom φ and v̄ ∈ ∂x φ(x̄, w̄), introduce the partial
second-order subdifferential of φ with respect to x at (x̄, w̄) relative to v̄ by
∂x2 φ(x̄, w̄, v̄)(u) := (D∗ ∂x φ)(x̄, w̄, v̄)(u) f or all u ∈ Rn . (2.9)
If φ is twice continuously differentiable around (x̄, w̄), we have
∂ 2 φ(x̄, w̄)(u) = {(∇2xx φ(x̄, w̄)∗ u, ∇2xw φ(x̄, w̄)∗ u)} f or all u ∈ Rn .
Consider now the following class of parametric constraint systems
S(w) := {x ∈ Rn | ψ(x, w) ∈ Θ}, w ∈ Rm , (2.10)
generated by a vector function ψ : Rn × Rm → Rs and a set Θ ⊂ Rs , and
then associate with (2.10) the normal cone mapping N : Rn × Rm ⇒ Rn
defined by
N (x, w) := N (x; S(w)) f or x ∈ S(w). (2.11)
It is easy to see that the mapping N in (2.11) admits the representation
N (x, w) = ∂x φ(x, w) with φ(x, w) := (δΘ ◦ ψ)(x, w) (2.12)
via the composition of ψ and the indicator function δΘ of the set Θ. We get
from (2.12) and the second-order subdifferential construction (2.9) that
∂x2 φ(x̄, w̄, v̄)(u) = D∗ N (x̄, w̄, v̄)(u) f or any v̄ ∈ N (x̄, w̄) and u ∈ Rn .
The following second-order calculus result can be derived from [39,
Theorem 3.1] applied to the composition in (2.12). This computation
foprmula plays an important role in the our applications to establishing
optimality conditions for controlled sweeping processes.
Theorem 2.13 (second-order subdifferential chain rule). Let ψ be C 2 -
smooth around (x̄, w̄) with the partial Jacobian matrix ∇x ψ(x̄, w̄) of full
rank. Then for each v̄ ∈ N (x̄, w̄) there exists a unique vector p̄ ∈
NΘ (ψ(x̄, w̄)) := N (ψ(x̄, w̄); Θ) satisfying
∇x ψ(x̄, w̄)∗ p̄ = v̄
and such that the coderivative of the normal cone mapping is calculated for
all u ∈ Rn by
∇ 2 p̄, ψ(x̄, w̄)
D∗ N (x̄, w̄, v̄)(u) = u + ∇ψ(x̄, w̄)∗ D∗ NΘ (ψ(x̄, w̄), p̄)
xx
∇2xw p̄, ψ(x̄, w̄)
(∇x ψ(x̄, w̄)u).
2.3 Dynamic Optimization via Controlled Moving Sets 39
over absolutely continuous control actions u(·) and the corresponding abso-
lutely continuous trajectories x(·) of the sweeping differential inclusion
ẋ(t) ∈ f (t, x(t)) − N (g(x(t)); C(t, u(t))) a.e. t ∈ [0, T ],
x(0) = x0 ∈ C(0, u(0)), (2.15)
where the controlled moving set is given by
C(t, u) := {x ∈ Rn | ψ(t, x, u) ∈ Θ}, (t, u) ∈ [0, T ] × Rm , (2.16)
with f : [0, T ] × Rn → Rn , g : Rn → Rn , ψ : [0, T ] × Rn × Rm → Rs ,
and Θ ⊂ Rs . The sets C(t, u) may not be convex, and the normal cone in
(2.15) is understood in the sense of (2.5). It follows from (2.15) due to the
normal cone definition that the formulated optimal control problem contains
the pointwise constraints on both state and control functions given by
ψ(t, g(x(t)), u(t)) ∈ Θ f or all t ∈ [0, T ].
Such constraints have been realized among the most challenging ones even in
standard optimal control theory for smooth ordinary differential equations.
Let us now formulate ourstanding assumptions in this section:
(H1) There is Lf > 0 such that f (t, x) − f (t, y) ≤ Lf x − y for all
x, y ∈ Rn , t ∈ [0, T ] and the mapping t → f (t, x) is a.e. continuous
on [0, T ] for each vectorx ∈ Rn .
(H2) There isLg > 0 such that g(x)−g(y) ≤ Lg x−y for all x, y ∈ Rn .
(H3) For each (t, u) ∈ [0, T ] × Rm , the mapping ψt,u (x) := ψ(t, x, u) is
twice continuously differentiable around the reference points with the
surjective derivative ∇ψt,u (x) satisfying
∇ψt,u (x) − ∇ψt,v (x) ≤ Lψ u − v
40 Applications of Variational Analysis to Controlled Sweeping Processes
controlled by both ui (t) and wi (t) as in [17], and also in various nonconvex
settings.
Our approach to the study of the sweeping control problem formulated in
(2.14)–(2.16) is based on the method of discrete approximations containing
the following major steps:
Step I: Construct a well-defined family of discrete approximations involving
a finite-difference replacement of the derivative ẋ in (2.15). The key point of
this step is to verify the possibility to approximate any feasible trajectory
of (2.15) by feasible trajectories of discrete-time systems in a topology
yielding the a.e. convergence of the extended discrete derivatives alonmg a
subsequence. We address not only qualitative aspects of well-posedness but
also quantitative ones with estimating error bounds, convergence rates, etc.,
which are of undoubted numerical values. Achieving these goals would lead
us then to the strong W 1,2 -norm approximation of a given local minimizer for
the continuous-time problem under consideration by a sequence of optimal
solutions to the discrete-time problems that are piecewise linearly extended
to the whole interval [0, T ].
2.3 Dynamic Optimization via Controlled Moving Sets 41
Step II: For each fixed step of discretization, the approximating discrete-time
problems can be reduced to nondynamic problems of mathematical pro-
gramming with various constraints, including increasingly many geometric
constraints. The latter problems are finite-dimensional of increasing dimen-
sions. Powerful tools of generalized differentiation in variational analysis
overviewed in Section 2 can be employed for deriving discrete-time neces-
sary optimality conditions in the approximation problems, even without any
Lipschitzian and convexity assumptions, by using appropriate calculus rules.
However, dealing with the graphical structure of the geometric constraints
in the approximation problems requires robust generalized differential con-
structions enjoying full calculus that should be subtle and small enough to be
efficiently applied to graphical sets. Note to this end, note that applying the
convexified normal cone and related constructions by Clarke [14] to graphs of
mappings often gives us the whole space or its subspace of maximal dimen-
sion; see [36, 45]. On the other hand, the nonconvex generalized differential
constructions from Section 2 satisfy all the required properties and can be
successfully used.
Step III: The concluding step in deriving necessary conditions for optimal
solutions of (2.14)–(2.16) is the passage to the limit from those for discrete-
time problems obtained in Step II with employing the strong convergence
of discrete optimal solutions obtained in Step I. This final step is highly
involved, since it requires justifying an appropriate convergence of dual arcs
as trajectories of adjoint differential inclusions. In this way we arrive at
the necessary optimality conditions for sweeping control problems presented
below.
Let us now formulate the notion of local minimizers for which we derive
necessary optimality conditions in what follows.
Definition 2.17 (local minimizers for controlled sweeping processes).
Let the pair (x̄(·), ū(·)) be feasible to problem (2.14)–(2.16) under the
imposed standing assumptions. We say that (x̄(·), ū(·)) be a local W 1,2 ×
W 1,2 -minimizer for this problem if x̄(·) ∈ W 1,2 ([0, T ]; Rn ), ū(·) ∈
W 1,2 ([0, T ]; Rm ), and
J[x̄, ū] ≤ J[x, u] f or all x(·) ∈ W 1,2 ([0, T ]; Rn ) and
u(·) ∈ W 1,2 ([0, T ]; Rm )
sufficiently close to (x̄(·), ū(·)) in the corresponding W 1,2 -norm topology.
Note that this notion, while being rather specific for the class of control
problems under consideration, occupies an intermediate position between
42 Applications of Variational Analysis to Controlled Sweeping Processes
by
q(t) = p(t) − ∇ψ(x̄(τ ), ū(τ ))∗ dγ(τ ).
[t,T ]
• MEASURED CODERIVATIVE CONDITION: Considering the
] t-dependent outer limit
γ(B)
Limsup (t) := y ∈ Rs | ∃ sequence Bk ⊂ [0, 1] with
|B|→0 |B|
γ(Bk )
t ∈ IBk , |Bk | → 0, →y
|Bk |
over Borel subsets B ⊂ [0, 1] with the Lebesgue measure |B|, for a.e.
t ∈ [0, T ] we have
Nontriviality condition:
λ + sup p(t) + γ = 0 with γ := sup x(s)dγ.
t∈[0,T ] xC([0,T ] =1 [0,T ]
Suppose in addition that η(T ) is well defined and that θ = 0 is the only vector
for which
Then the above necessary optimality conditions hold with the stronger
nontriviality
!
[ν, v] := νi αi [∇x ψ(x, u)]i
i∈I(x,u)
Observe that in the linear case h(z) := Az −b in (2.21), the full rank assump-
tion in (H3) corresponds to the classical linear independence constraint
qualification (LICQ).
The following major results can be derived from Theorem 2.20 by using
the precise computation of the second-order construction D∗ NRs− .
Theorem 3.3 (novel Hamiltonian formalism and maximum principle for
optimal control in moving sets ). Consider the sweeping control problem
(2.14)–(2.16) in the frameworks of Theorem 2.20 with the set Θ given
by (2.21), where h : Rs → Rl is C 2 -smooth around the relaxed local
46 Applications of Variational Analysis to Controlled Sweeping Processes
W 1,2 × W 1,2 -minimizer z̄(t) := (x̄(t), ū(t)) for all t ∈ [0, T ]. Suppose
that the matrix ∇h(z̄(t)) is of full rank for all t ∈ [0, T ]. Then, in addition
to the necessary optimality conditions of Theorem 2.20, the new maximum
condition
[ν(t), x̄(t)],
˙ q x (t) − λv x (t) = Hν(t) (x̄(t), ū(t), q x (t) − λv x (t)) = 0 a.e.
t ∈ [0, T ] (20)
Discussions on the Maximum Principle. Observe that our form of the new
Hamiltonian (2.22) is different from the conventional Hamiltonian function
between our form of the maximum principle for controlled sweeping pro-
cesses and the Hamiltonian formalism in models of contact and nonsmooth
mechanics; see, e.g., [7].
Here is the promised example, which shows the failure of the conven-
tional maximum principle for sweeping optimal processes with controlled
moving sets and also illustrates the fulfillment of the new one obtained in
Theorem 2.23.
Example 3.4 (failure of the maximum principle for sweeping processes
with controlled moving sets). Consider problem (2.14)–(2.16) with u =
(a, b) ∈ R6 as a = (a1 , a2 ) ∈ R4 and b = (b1 , b2 ) ∈ R2 , x = (x1 , x2 ) ∈ R2 ,
x0 = (1, 1), T = 1,
x2 1
ϕ(x) = and (t, x, u, ẋ, u̇) := (ḃ21 + ḃ22 ).
2 2
The moving sets C(t, u) are defined by
C(t, u) := {x ∈ R2 | ai (t), x ≤ bi (t), i = 1, 2}.
Fix the a-controls as ā1 ≡ (1, 0) and ā2 ≡ (0, 1), and then deduce from
Theorem 2.20 the following relationships on [0, 1] involving only the b-
components of controls together with the corresponding state and adjoint
functions:
1. w(·) = 0, v x (·) = 0, v b (·) = (ḃ1 (·), ḃ2 (·))
0 ⇒ qix (t) = 0, i = 1, 2
2. x̄˙ i (t) =
3. pb (·) is constant with nonnegative components, and −pxi (·) = λx̄(1) +
pbi (·)āi are also constant for both i = 1, 2.
˙ = pb + γ([t, 1]) for a.e. t ∈ [0, 1].
4. q x (t) = px − γ([t, 1]), q b (t) = λb̄(t)
5. λ + q(0) + p(1) = 0 with λ ≥ 0.
Observe first that the pair x̄(t) = (1, 1) and b̄(t) = 0 on [0, 1] satisfies the
necessary conditions with px1 = px2 = −1, pb1 = pb2 = γ1 = γ2 = 0, and
λ = 1. The conventional Hamiltonian function (2.25) reads now as
H(x, b, p) = sup{p, v| v ∈ −N (x; C((1, 0), (0, 1)), b)},
and we get by the direct calculation that
H(x̄(t), b̄(t), q x (t) − λv x (t))
= H((1, 1), (1, 1), (−1, −1))
= sup{(−1, −1), v| v ∈ −N ((1, 1); C(((1, 0), (0, 1)), (1, 1))}
= sup{(−1, −1), v| v1 ≤ 0, v2 ≤ 0} = ∞,
48 Applications of Variational Analysis to Controlled Sweeping Processes
while x̄(t),
˙ q x (t) − λv x (t) = 0. Thus the conventional maximum principle
fails in this example. At the same time, the new maximum condition (2.24)
from Theorem 2.23 holds trivially with the choice of ν(t) = 0 for all t ∈ [0, 1]
therein.
Observe that the sweeping differential inclusion (2.28) and the polyhedral
description (2.30) automatically yield the pointwise state constraints
x∗i , x(t) ≤ ci f or all t ∈ [0, T ].
2.4 Sweeping Processes with Controlled Dynamics 49
(A4) The vertices x∗i of (2.30) satisfy the linear independence constraint
qualification
!
[ αi x∗i = 0, αi ∈ R] ⇒ [αi = 0 f or all i ∈ I(x̄)}
i∈I(x̄)
• COMPLEMENTARITY CONDITIONS:
x∗i , x̄(t) < ci ⇒ ηi (t) = 0 and ηi (t) > 0 ⇒ x∗i , q(t) = ci
for a.e. t ∈ [0, T ] including t = T and for all i = 1, . . . , m.
• RIGHT ENDPOINT TRANSVERSALITY CONDITIONS:
!
−p(T ) = λ∇ϕ(x̄(T )) + ηi (T )x∗i with
i∈I(x̄(T ))
!
ηi (T )x∗i ∈ N (x̄(T ); C).
i∈I(x̄(T ))
where ζ(t) := Aσ(t) − A−1 (t) and C(t) := −A−1 (t) + AQ. This can
be rewritten in the frame of the sweeping optimal control problem with
controlled moving sets considered in Section 3 where x := ζ, u := ,
ψ(x, u) := x + A−1 u, and Θ := AQ. Thus we can apply Theorems 2.20
and 2.31 to this class of hysteresis operators for the general elasticity domain
Q. In this way we cover various elastoplasticity models including those with
the Drucker-Prager, Mohr-Coulomb, Tresca, and von Mises yield criteria; see
[1]. More details with precise computations, examples, and discussions can
be found in [26].
Optimal Control of Sweeping Processes in Somew Models of Robotics.
Among important dynamical models arising in robotics, the mobile robot
model with obstacles has been well recognized and investigated. It is shown
in [24] that the dynamics of this model can be adequately described as a
sweeping process. An optimal control version of the mobile robot model with
obstacles has been recently developed in [18] as a sweeping control problem
with controlled dynamics studied in Section 4. This model concerns n mobile
robots (n ≥ 2) identified with safety disks in the plane of the same radius R.
The goal of each robot is to reach the target by the shortest path during a fixed
time interval [0, T ] while avoiding the other n − 1 robots that are treated by
this robot as obstacles.
52 Applications of Variational Analysis to Controlled Sweeping Processes
x̄i (t) = (x̄i (t) cos θi (t), x̄i (t) sin θi (t)) f or i = 1, . . . , n,
g(x(t), u(t)) = (s1 u1 (t) cos θ1 (t), s1 u1 (t) sin θ1 (t), . . . , sn un (t) cos θn (t),
C := {x ∈ Rn | xj∗ , x ≤ cj , j = 1, . . . , n − 1} with
some interesting cases, while more serious practical applications require more
elaborations.
Traffic Equilibria and Crowd Motion Models. There are several practically
important models of traffic equilibria and related areas, where the dynamics
can be formulated as a sweeping process. One of the most interesting classes
of models of this type is known under the name of crowd motion models. In
what follows we consider some control versions of such problems, which can
be efficiently investigated and partly solved by using the necessary optimality
conditions discussed above.
The original motivation for crowd motion models came the modeling of
local interactions between participants in order to describe the dynamics of
pedestrian traffic. However, by now such models have been successfully used
to investigate more general classes of problems in engineering, operations
research, economics, etc.
The microscopic form of the crowd motion model is based on the follow-
ing two postulates. On one hand, each individual has a spontaneous velocity
that he/she intends to implement in the absence of other participants. On
the other hand, the actual velocity must be taken into account. The latter
one is incorporated via a projection of the spontaneous velocity into the
set of admissible/feasible velocities, i.e., those which do not violate certain
nonoverlapping constraints. A mathematical description of the uncontrolled
microscopic crowd motion model is given [49] in the sweeping process form
with the subsequent usage therein for numerical simulations and various
applications.
In the corridor version of the crowd motion model, the crucial nonover-
lapping condition is written in the form
where Nx stands for the normal cone to Q0 at x. Since all the participants
exhibit the same behavior and want to reach the exit by the shortest path,
their spontaneous velocities can be represented as follows:
U (x) = (U0 (x1 ), . . . , U0 (xn )) with U0 (x) = −s∇D(x),
where D(x) denotes the distance between the position x = (x1 , . . . , xn ) ∈
Q0 and the exit, and where s ≥ 0 stands for the speed. By taking into account
the aforementioned postulate that in the absence of other participants each
participant tends to remain his/her spontaneous velocity until reaching the
exit, the (uncontrolled) dynamic perturbations in this model are described by
g(x) = −(s1 , . . . , sn ) ∈ Rn f or all x = (x1 , . . . , xn ) ∈ Q0 ,
where si denotes the speed of the participant i ∈ {1, . . . , n}. To control
the actual speed of all the participants in the presence of the nonoverlap-
ping condition, we suggest in [11] to involve control functions u(·) =
(u1 (·), . . . , un (·)) into perturbations as follows:
g(x(t), u(t)) = (s1 u1 (t), . . . , sn un (t)), t ∈ [0, T ].
To optimize the sweeping dynamics by using controls u(·), consider the cost
functional
T
1 2
minimize J[x, u] := (x(T ) + u(t)2 dt) (2.34)
2 0
where Dij (x) := xi − xj − 2R is the signed distance between the disks i
and j of the same radius R identified with n ≥ 2 participants on the plane.
The corresponding optimal control problem formulated and investigated in
[12] is described via the sweeping dynamics as follows: minimize the cost
functional (2.34) over the controlled sweeping process
−ẋ(t) ∈ N (x(t); C(t)) + g(x(t), u(t)) f or a.e. t ∈ [0, T ],
C(t) := C + v̄(t), v̄(t) = r ∈ [r1 , r2 ] on [0, T ], x(0) = x0 ∈ C(0),
g(x(t), u(t)) := (s1 u1 (t) cos θ1 (t), s1 u1 (t) sin θ1 (t), . . . , sn un (t) cos θn (t),
r r
sn un (t) sin θn (t)), v̄i+1 (t) = v̄i (t) := √ , √ , i = 1, . . . , n − 1,
2n 2n
C := {x = (x1 , . . . , xn ) ∈ R2n | hij (x) ≥ 0 f or all i/= j as
i, j = 1, . . . , n}
with the functions hij (x) := Dij (x) = xi − xj − 2R and with
Observe that the sweeping dynamic in the described planar version of the
crowd motion model is generated by the set C, which is not a convex
polyhedron, while belonging the class of prox-regular sets in the sense of
[45]. The required extensions of necessary optimality conditions for this
type of controlled sweeping processes were obtained in [12] by using the
corresponding modification of the method of discrete approximations. The
obtained results were applied therein to establishing efficient relationships
which allowed us to determine optimal parameters, to develop numerical
algorithms, and realize them in particular settings.
2.6 Conclusion
This chapter discusses recent applications of advanced constructions of
variational analysis and generalized differentiation to challenging classes of
optimal control problems governed by discontinuous differential inclusions
that are known as sweeping processes. We present optimality conditions for
such control systems obtained by employing the method of discrete approxi-
mations and then review the usage of the obtained conditions to solving some
56 Applications of Variational Analysis to Controlled Sweeping Processes
Acknowledgments
This research was partially supported by the US National Science Foundation
under grants DMS-1007132 and DMS-1512846, by the US Air Force Office
of Scientific Research grant #15RT0462, and by the Australian Research
Council under Discovery Project DP-190100555.
References
[1] B. Acary, O. Bonnefon, B. Brogliato, Nonsmooth Modeling and Simu-
lation for Switched Circuits, Springer, Berlin, 2011.
[2] L. Adam, J.V. Outrata, ‘On optimal control of a sweeping process
coupled with an ordinary differential equation’, Discrete Contin. Dyn.
Syst. Ser. B 19 (2014), 2709–2738.
[3] S. Adly, T. Haddad, L. Thibault, ‘Convex sweeping process in the
framework of measure differential inclusions and evolution variational
inequalities’, Math. Program. 148 (2014), 5–47.
[4] C.E. Arroud and G. Colombo, ‘A maximum principle of the controlled
sweeping process’, Set-Valued Var. Anal. 26 (2018), 607–629.
[5] R.J. Aumann, ‘Integrals of set-valued functions’, J. Math. Anal. Appl.
12 (1965), 1–12.
[6] N.N. Bogolyubov, ‘Sur quelques méthodes nouvelles dans le Calculus
des Variations’, Ann. Math. Pura Appl. 7 (1929), 249–271.
[7] B. Brogliato, Nonsmooth Mechanics: Models, Dynamics and Control,
3rd edition, Springer, Cham, Switzerland, 2016.
[8] B. Brogliato, A. Tanwani, ‘Dynamical systems coupled with mono-
tone set-valued operators: formalisms, applications, well-posedness, and
stability’, SIAM Rev. 62 (2020), 3–129.
[9] M. Brokate, P. Krejčcí, ‘Optimal control of ODE systems involving a
rate independent variational inequality’, Discrete Contin. Dyn. Syst. Ser.
B 18 (2013), 331–348.
[10] T.H. Cao, G. Colombo, B.S. Mordukhovich, D. Nguyen, ‘Optimiza-
tion of fully controlled sweeping processes’, J. Diff. Eqs. 295 (2021),
138–186.
References 57
[41] J.J. Moreau, ‘On unilateral constraints, friction and plasticity’, in New
Variational Techniques in Mathematical Physics (G. Capriz and G.
Stampacchia, eds.), Proc. C.I.M.E. Summer Schools, pp. 173–322,
Cremonese, Rome, 1974.
[42] L.S. Pontryagin, V.G. Boltyanskii, R.V. Gamkrelidze, E.F. Mishchenko,
The Mathematical Theory of Optimal Processes, Wiley, New York,
1962.
[43] B.N. Pschenichnyi, Convex Analysis and Extremal Problems, Nauka,
Moscow, 1980.
[44] B.N. Pschenichnyi, V.V. Ostapenko, Differential Games, Naukova
Dumka, Kyiv, 1992.
[45] R.T. Rockafellar, R.J-B. Wets, Variational Analysis, Springer, Berlin,
1998.
[46] G. V. Smirnov, Introduction to the Theory of Differential Inclusions,
American Mathematical Society, Providence, RI, 2002.
[47] A.A. Tolstonogov, Differential Inclusions in a Banach Spaces, Kluwer,
Dordrecht, The Netherlands, 2000.
[48] A.A. Tolstonogov, ‘Control sweeping processes’, J. Convex Anal. 23
(2016), 1 099–1123.
[49] J. Venel, ‘A numerical scheme for a class of sweeping process’,
Numerische Mathematik, 118 (2011), 451–484.
[50] R.B. Vinter, Optimal Control, Birkhaüser, Boston 2000.
[51] L.C. Young, ‘Generalized curves and the existence of an attained
absolute minimum in the calculus of variations’, C. R. Soc. Sci. Lett.
Varsovie, Cl. III, 30 (1937), 212–234.
[52] V. Zeidan, C. Nour, H. Saoud, ‘A nonsmooth maximum principle for
a controlled nonconvex sweeping process’, J. Diff. Eqs. 269 (2020),
9531–9582.
3
Robust and Robustly-Adaptive Control
of Some Noninvertible Memoryless Systems
Abstract
This chapter deals with the robust control of some uncertain multivariable
memoryless (static) discrete-time systems in the presence of arbitrary unmea-
surable bounded disturbances, whose bounds may be unknown, in general.
The main feature of the systems to be controlled is that their gain matrices
are noninvertible. Again, they are of not full rank. It is assumed that the
elements of these matrices are exactly unknown. However, the bounded
intervals to which they belong are a priori known. The nonadaptive robust
control approach applicable to the case when the lengths of these intervals
are relatively short is proposed. Such an approach is based on the so-called
pseudoinverse model-based concept. The robustly-adaptive control utilizing
a modification of inverse-model concept is designed to cope with large
parametric uncertainty. The asymptotic properties of both nonadaptive and
adaptive feedback control systems are derived. To support the theoretic study,
numerical examples and simulation results are given.
61
62 Robust and Robustly-Adaptive Control of Some Noninvertible Memoryless
3.1 Introduction
The problem of a perfect performance of control systems is an important
problem from both theoretical and practical points of view. Within the frame-
work of this actual problem, new approaches have been advanced by many
researches. The latest results in this scientific area have been reported in
numerous papers and generalized in several books including [1, 2]. Among
them, the recent works [3–7] dealt with advanced multivariable control
systems.
A long-standing problem of the optimal controller design for multivari-
able system in the presence of unmeasurable disturbances [8] has been solved
using, in particular, the internal model control principle proposed in [9].
A perspective modification of this principle is the so-called model-inverse
approach. Since the pioneering work [10], the problem of inversion of linear
time-invariant multivariable systems has attracted an attention of several
authors [11–14]. Recently, a significant progress in this research are has been
achieved in [15, 16].
An inverse model-based approach to ensuring a perfect performance of
linear multivariable control systems containing the memoryless (static) plants
was first advanced in [17]. Regardless of this work, a similar approach
was proposed in [18] to deal with some memoryless plants. However,
the approach above mentioned is quite unacceptable if their gain matrices
are either square but singular or nonsquare because they are noninvertible
matrices.
It turned out that so-called pseudoinverse (generalized inverse) model-
based concept advanced in the paper [19] can be exploited to cope with the
possible noninvertibility of gain matrices. This fruitful concept was extended
in [20, 21] to robust control of a noninvertible and uncertain plant with
unmeasurable bounded disturbances.
The robust control theory gives a powerful tool to reject the unmea-
sured disturbances in the closed-loop control system with parametric and
nonparametric uncertainty. The books [22–27] et al. provide insight into the
results achieved in this theory to the beginning of the 2000s. A new esti-
mation method applicable to a nonlinear discrete-time system with bounded
disturbances has been advanced in the work [28]. Novel results with respect
3.2 Problem Statement 63
meaning that
2 ≤r≤m (3.3)
is here considered. However, in contrast with problem stated in the work [36],
the rank of B satisfies the strict inequality
rank B<r (3.4)
implying that B is a not full rank matrix.
The following basic assumptions with respect to the gain matrix B and
(i) (i)
the sequences {v (i)
n } = v0 , v1 , . . . (i = 1, . . . , m) are introduced.
A1) The elements of B are all unknown. However, there are some interval
estimates defined as
(ij)
b(ij) ≤b(ij) ≤b , i= 1, . . . , m, j= 1, . . . , r (3.5)
(ij)
where the upper and lower bounds b(ij) and b , respectively, are
assumed to be known.
A2) The rank of B satisfies Equation (3.4) giving that B is a noninvertible
matrix.
(i)
A3) vn s (i = 1, . . . , m) are all the arbitrary scalar variables bounded in
modulus according to & &
& (i) &
&vn & ≤ε(i) <∞
where ε(i) s are constant. It is assumed that these upper bounds may be
unknown, in general.
Let y 0 = [y 0(1) , . . . , y 0(m) ]T be a &desired & output vector
& 0(m) & (y 0(i) ≡
const ∀i= 1, . . . , m). Suppose that &y & +· · ·+ &y
0(1) & = 0 implying
that, at least, one y 0(i) 0(1)
of y , . . . , y 0(m) is nonzero.
Define the output error vector
en = y 0 − yn (3.6)
(i) (i) (1) (m)
with the components en = y 0(i) − yn , i.e., en = [en , . . . , en ]T .
The problem is to design the feedback controller guaranteeing the
ultimate boundedness of the sequence {en } = e1 , e2 , . . . , meaning
lim sup ||en || <∞ (3.7)
n→∞
provided
lim sup ||un || <∞ (3.8)
n→∞
3.3 Preliminaries 65
Remark 3.1. Assumptions A1) – A3) differ from similar assumptions made
in [36] only in that B is here the matrix of not full rank, and it is its essential
feature.
Remark 3.2. Equation (3.8) is here introduced additionally since it may not
be satisfied even if (3.7) takes place.
3.3 Preliminaries
Assume, for the time being, that B is a known noninvertible matrix of full
rank (rank B = r), i.e., B represents some nonsquare matrix with r < m. In
this case, the so-called pseudoinverse control
un = un−1 + B + en (3.9)
proposed by the authors in their previous works (see, e.g., [21]) ensures
the minimum of the upper bound on the Euclidean norm ||en ||2 of the
output error vector of the closed-loop control system (3.1), (3.6), (3.9)
with any bounded sequence {vn } = v1 , v2 , . . . Here the notation P +
of any pseudoinverse (generalized inverse) matrix introduced in [41] and
defined as
P + = lim (P T P + δIr )−1 P T (3.10)
δ→0
where Ir denotes the identity r × r matrix is used. Recall that if rank P = r
for some P ∈ Rm×r , then the expression of P + is simplified [42, item 7.46]
instead of (3.10) we have
P + = (P T P )−1 P T
Figure 3.2 Configuration of the pseudoinverse model-based feedback control system with
parametric uncertainty
(ij) (ij)
where δ (ij) = b0 − b̄(ij) and δ̄ (ij) = b0 − b(ij) .
3.3 Preliminaries 67
(i)
Recall that if there are no disturbances (vn ≡ 0 for all i = 1, . . . , m),
then an equilibrium state of the closed-loop control system (3.1), (3.6), (3.12)
defined by the pair (ue , y e ) in which y e = Bue , where ue represents the
solution of the equation
B0+ Bu = B0+ y 0 (3.13)
with respect to u, exist provided that rank B0 = r and the condition
q<1 (3.14)
is satisfied [36, Lemma 7.1]. In this lemma it has been established that the
condition (3.14) together with (3.15) is sufficient to achieve the ultimate
boundedness of the sequences {un } and {en } if B is a matrix of full rank.
To establish the sufficient conditions under which the equilibrium state
exist, the requirement that the r × r matrix B0+ B of Equation (3.13) is
nonsingular is utilized. ) *
It turns out that when y 0 ∈
/ ℵ B0+ and also y 0 ∈/ R (B) , where ℵ (P )
and R (P ) denote the null space and the range of a matrix P, respectively,
then the equilibrium state may exist even if det B0+ B = 0. To show this,
consider an example.
Example 3.1. Let
⎛ ⎞ ⎛ ⎞
1 −2 3 −4.5
B = ⎝ 2 −4 ⎠ , B 0 = ⎝ 1 −1.5 ⎠
−1 2 −2 3
be some nonsquare matrices of the not full rank (rank B = rank B0 = 1).
By (3.10) we have
+ 1 6 2 −4
B0 =
91 −9 −3 6
) *
Put y 0 = [1, 3, 5]T . Such a choice of y 0 gives y 0 ∈
/ ℵ B0+ and y 0 ∈ /
R (B). It can be verified that in this case, Equations (3.13) will have the
infinite number of solutions ue = [ue(1) , ue(2) ]T lying on the straight line
described by
7ue(1) − 14ue(2) + 4 = 0
68 Robust and Robustly-Adaptive Control of Some Noninvertible Memoryless
(see Figure 3.3). In other words, the set Eu = {ue } of the equilibrium points
becomes here infinite.
Unfortunately, if rank B < r and y 0 ∈ / R (B), then the equilibrium state
may not exist, in principle. To demonstrate this annoying fact, consider the
following numerical example.
Example 3.2. Suppose B and B0+ are some square matrices given by
2 4 + 1 −2
B= , B0 =
1 2 2 −4
) *
and y 0 ∈/ ℵ B0+ , y 0 ∈ / R (B). Since B0+ B = 0r×r , in this case, Equation
(3.13) has no solutions. This example shows that the closed-loop control
system defined by Equations (3.1), (3.6) and (3.12) will have no equilibrium
state.
It turns out that the equilibrium state {ue , y e } of the control system above
mentioned exists iff
) * ) *
ℵ B0+ ∩ y 0 + R (B) = ∅ (3.16)
where the symbol “+” denoted the Minkowski sum of two sets of vectors
in Euclidean space. Actually, by the definition of set R (B) to which the
zero vector 0m belongs, follows that (−Bu) ∈ R (B). Thereby, the vector
y 0 − Bu belongs to the set y 0 + R (B) representing a linear manifold whose
dimension is defined as dim R (B) = rank B [41, p. 38]. Geometrically, this
set is parallel to the linear subspace R (B) and does not contain the vector 0m .
3.4 Robust Nonadaptive Control 69
is the scalar variable which depends on the elements of its kth column.
Introduce the r-dimensional vector
(11) (r1) T
g (k) (B) = [β0 , . . . , β0 ] dk (3.21)
of the variable dk (B) exist. Taking into account this fact and also Equation
(3.20) it can be written
3.4 Robust Nonadaptive Control 71
dk (B) = min b(1k) + k12 b(2k) + · · · + km−1,m b(mk)
(ik)
b(ik) ∈ [b(ik) , b ]
i = 1, . . . , m
k = 1, . . . , r (3.22)
non-zero elements of which are collinear vectors. By the definition of the null
space ℵ (P ) = {x : P x = 0r } of any r × r matrix P (see, e.g. ,[42, item
6.24]) and by virtue of the expression (3.25) it can be written
(1)T
ℵ(GT (B)) = {u : β0 u = 0r } (3.26)
where
(1) (11) (r1) T
β0 = [β0 , . . . , β0 ] (3.27)
Since the dimension of ℵ(GT (B)) is equal to dim ℵ(GT (B)) = r −
rank GT (B) = r − 1 (according to [42, item 6.25]), it is not hard to see
that, for each B, the set ℵ(GT (B)) is a hyperplane in the space Rr of some
(1)
vectors u to which β0 given by the expression (3.27) is the normal vector.
72 Robust and Robustly-Adaptive Control of Some Noninvertible Memoryless
satisfied, then the control law (3.12), (3.6) when applied to the plant (3.1)
is robust with respect to the uncertainty (3.5), and the ultimate boundedness
of the sequences {en } and {un } is guaranteed:
lim sup ||en || = ||Im − B0 B0 || ||e0 || + 2 sup ||vn ||
+
n→∞ 0<=n<∞
Figure 3.6 Choice of matrix B0 satisfying the condition rank B0 = 1 (possible cases): (a)
there are B0 s belonging to Ξ; (b) there are no B0 s belonging to Ξ (3.29)
!
s
||P ||1 := max |p(ij) | (3.30)
1≤i≤s
j=1
where
!
m
(ik) (kj)
σ (ij) = β0 δ
k=1
are the linearforms, depending on the elements of the pseudoinverse matrix
(ij) (ij)
B0+ = β0 , and δ (ij) and δ are defined by the expression (3.32).
3.4 Robust Nonadaptive Control 75
with min σ (ij) and max σ (ij) that are the solutions of LP problems given
in the expression (3.31), then the controller described in Equations (3.12)
together with (3.6) robustly stabilizes the system (3.1) for any B from the set
Ξ defined by the expression (3.11).
Proof. Follows immediately from the definition (3.30) of the 1-norm of an
arbitrary matrix.
Example 3.3. Let Ξ = {(b(ij) ) : 0.4 ≤ b(11) ≤ 1.4, −2 ≤ b(12) ≤
−1, −1.2 ≤ b(21) ≤ −0.2, 0.7 ≤ b(22) ≤ 1.7, −1.7 ≤ b(31) ≤ −0.7, 1 <=
b(32) ≤ 2}. Next, choose
⎛ ⎞
0.8 −1.6
B0 = ⎝ −0.7 1.4 ⎠
−1.0 2.0
with rank B0 = 1. Such a choice yields
+ 16/213 −14/213 −20/213
B0 =
−32/213 28/213 40/213
!
2
max{ |min σ (1j) |, |max σ (1j) |} = 70/213 < 1,
j=1
!
2
max{ |min σ (2j) |, |max σ (2j) |} = 140/213 < 1.
j=1
Figure 3.8 Robust pseudoinverse model-based controller: (a) Euclidean norm ||un ||2 of
control input vectors; (b) Euclidean norms ||yn ||2 , ||y 0 ||2 of output vectors (solid line) and
desired output vectors (dashed line), respectively
Since, at least, one of the eigenvalues λi (B) is equal to zero (due to the
singularity of B), by virtue of Equation (3.35) there are the numbers
!
r & & !
r & &
& (ij) & (i) & (ij) &
β (i) := b(ii) − &b &, β := b(ii) + &b & (3.37)
j=1 j=1
j = i j = i
such that if
|b(i1) | + · · · + |b(ir) | = 0 (3.38)
(i) (i)
then either β (i) ≤ 0 but β > 0 or β (i) < 0 but β ≥ 0. These numbers are
defined by the intersection of the ith Gerŝgorin disc with the real axis of the
complex z-plane as shown in Figures 3.9 and 3.10, respectively, left. In both
(i) (i)
cases, β (i) β ≤ 0 if the inequality (3.38) is satisfied because β (i) and β
cannot have the same sign.
(1)
Figure 3.9 The Gerŝgorin discs for r = 2 in the case |β (2) | < |β |
(2)
Figure 3.10 The Gerŝgorin discs for r=2 in the case |β | < |β (1) |
3.5 Robustly-Adaptive Control of Square Systems (Case 1) 79
Denoting
(1) (r)
β := min{ β (1) , . . . , β (r) }, β := max{ β ,...,β }, (3.39)
consider the following two cases: (a) |β| < |β|; (b) |β| > |β| (The case
when |β| = |β| can be combined with any two cases.) In order to go to the
gain matrix B̃ of the fictitious plant having the form (3.34) in the case (a), it
is sufficient to shift the Gerŝgorin disc (3.36) right taking
as shown in Figure 3.9, right. In the case b), the discs (3.36) need to be shifted
left according to
δ0 < −|β| (3.41)
See Figure 3.10, right. In both cases, the nonsingularity of B̃ is guaran-
teed. Nevertheless, the conditions (3.40) and (3.41) cannot be satisfied, as yet.
In fact, the numbers β and β given by the expressions (3.39) depend of β (i)
(i)
and β s defined by Equations (3.37). But they are unknown because b(ij) s
are all unknown.
The following operations are proposed to choose a number d0 satisfying
the reqiurement (3.35). Introduce
!
r
(ij)
β (i)
min
:=b(ii) − max{ |b(ij) |, |b |}
j=1
j = i
(i) (ii) !
r
(ij)
β max := b + max{ |b(ij) |, |b |} (3.42)
j=1
j = i
(ij)
minimizing and maximizing in b(ij) ∈ [b(ij) , b ] the right-hand sides of
(i)
Equations (3.37) for β (i) and β , respectively.
Further, the number δ0 is found to satisfy the conditions
β min := {β (1)
min
, . . . , β (r)
min
}
(1) (r)
β max := {β max , ..., β max } (3.44)
It can be clarified that if the conditions (3.43) together with Equations (3.42)
and (3.44) will be satisfied then the condition (3.35) will without fail be
ensured.
After determining the quantity δ0 we can proceed to the consideration of
(1) (r)
the fictitious plant. Since the input variables un , . . . , un and the distur-
(1) (N )
bances vn , . . . , vn of both true plant and fictitious plant are the same, this
feature makes it possible to describe our fictitious plant by the equation
It is seen from Equation (3.46) that ỹn can always be found indirectly having
un and yn to be measured.
Now, our problem reduces to the known problem of adaptive control
applicable to the fictitious plant (3.45) with the unknown gain matrix B̃ in the
(1) (r)
presence of arbitrary bounded disturbances vn , . . . , vn . Its solving follows
the steps of the section above. Namely, the adaptive control law is designed
in the form
un = un−1 + B̃n−1 ẽn , (3.47)
in which, instead of the current estimate Bn of B, another B̃n is exploited,
and the error vector en defined in Equation (3.6) is replaced by
ẽ∗(i)
(i)T
n = ∇ỹn(i) − b̃n−1 ∇un−1 (3.51)
represent the ith component of the identification error ẽ∗n given as
det B̃n = 0
Figure 3.12 Estimated parameters of the fictitious plant (solid line) and the true plant
(dashed line)
(1) (2)
Figure 3.15 Estimates εn /2 and εn /2 of Simulation experiment 3.3
It is seen from Figures 3.14 and 3.15, that the estimates of unknown
system parameters and bounds on disturbances converges as predicted by part
(a) of Theorem 3.4. We can also observe that the control inputs and outputs
remain bounded.
86 Robust and Robustly-Adaptive Control of Some Noninvertible Memoryless
(1) (2)
Figure 3.16 Control inputs un and un of Simulation experiment 3.3
(1) (2)
Figure 3.17 Outputs yn and yn of Simulation experiment 3.3
with the same un−1 and vn−1 [k]. In these equations, ỹn [k] denotes the r-
dimensional output vector related to the kth fictitious plant whose gain matrix
B̃[k] is defined as follows:
This expression shows that although as B̃[k] as B[k] remain unknown, how-
ever, the components of all N the vectors ỹn [k] can indirectly be “measured”
after measuring the components of yn and un−1 , and it is essential.
If the conditions (3.58) are satisfied, then the problem of the adaptive
stabilization of the true plant (3.1) can be reduced to the problem of simul-
taneous adaptive stabilization of all N fictitious plants (3.56) with unknown
but nonsingular r × r gain matrices B̃[k] (k = 1, . . . , N ) via forming
at each nth time instant a set of N different “potentially” possible controls
un [1], . . . , un [N ] and selecting one of them in accordance with certain
choice rule [27] given below.
Following to [39], the adaptive control law to be applicable to any
fictitious plant is designed in the form
where ẽn [k] = y 0 [k] − ỹn [k] with y 0 [k] = [y 0(i1 [k]) , . . . , y 0(ir [k]) ]T defines
the output error vector related to the kth fictitious plant at the nth time instant,
and B̃n [k] ∈ Rr×r is the current estimate of unknown r × r matrix B̃[k] at
the same time instant satisfying
ẽ∗(i)
(i) (i)T
n [k] = ỹn [k] − ỹn−1 [k] − b̃n−1 [k]∇un−1
(i)
(3.63)
represents the scalar variable making sense of the ith component of ẽ∗n [k] ∈
Rr that is the identification error vector related to the kth fictitious plant.
(i)
The coefficients γn s are chosen from the ranges [γ (i) , γ (i) ] to ensure the
requirement (3.61).
Next, add the adaptation algorithms described in the formulas (3.62)
together with (3.63) by an algorithm for estimating unknown B defined as
follows:
⎧
∗(i)
⎪ (i)
⎨ bn−1 if |en | ≤ ε0i ,
b(i)
n =
⎪ ∗(i) ∗(i)
⎩ b(i) + γn(i) en −εi sign e2n ∇un−1 otherwise,
n−1 ||∇un−1 ||2
i= 1, . . . , m,
(3.64)
(i)T
where bn represents the ith row of the estimate matrix Bn , and
e∗(i)
(i) (i)T
n = yn(i) − yn−1 − bn−1 ∇un−1 (3.65)
is the ith component of the identification error vector e∗n = yn − yn−1 −
Bn−1 ∇un−1 (εi and ε0i are given by the conditions (3.50)).
The estimation procedure defined in the algorithm (3.64) together with
Equation (3.65) makes it possible to estimate the m predicted output errors
→
− (i)
e n+1 [k] (i = 1, . . . , m) for the each ith output of true plant (3.1) at any n
using the formula
|−
→ (i)
e n+1 [k]| = |y 0(i) − b(i)T (i)
n un [k]| + ε , i = 1, . . . , m. (3.66)
The synthesis of the adaptive controller is finished by the choice of
the control un from the set {un [1], . . . , un [N ]} with un [k] given by
3.7 Robustly-Adaptive Control of Nonsquare Systems 89
Equation (3.60). This choice is implemented by the rule giving the minimum
of the 1-norm of −
→e n+1 [k] = [−
→
e n+1 [k], . . . , −
→
(1) (m)
e n+1 [k]]T according to
!
m
|−
→ (i)
un = arg min e n+1 [k]|, (3.67)
un [k]
i=1
where −→ (i)
e n+1 [k]s are specified by Equation (3.66).
The asymptotic properties of the adaptive controller described in this
section are given in theorem below (the main result).
Theorem 3.5. Consider the feedback control system containing the plant (3.1)
in which r < m, and the adaptive controller defined in Equations (3.62),
(3.67) together with Equations (3.59), (3.66) and (3.61). Using the constraints
(3.5), determine δ0 [1], . . . , δ0 [N ] to satisfy the requirement (3.58). Let
assumptions A1–A3 be valid. Then, this controller applied to the plant (3.1)
guarantees that the control objectives (3.7) and (3.8) will be achieved.
Proof. See [39].
Note that Theorem 3.5 does not guarantee that the ultimate error
limn→∞ sup ||en || will become as in the nonadaptive case when there
is no parametric uncertainty and the pseudoinverse model-based controller
proposed in [21] can by applied.
Simulation experiment 3.4. A simulation experiment was conducted to illus-
trate the performance of the proposed adaptive control in the case when
r = 2, m = 3. As the gain matrix,
⎛ ⎞
4 2
B=⎝ 2 1 ⎠
3 1.5
Figure 3.19 The behaviour of the control system: (a) the current number k of control un [k]
chosen from {un [1], un [2], un [3]} at given n; (b) the 1-norm of control vector; (c) the 1-norm
of output vector in adaptive case (solid line) and in nonadaptive optimal case (dashed line)
References 91
The quantities δ0 [1] = 1.1, δ0 [2] = 1.2 and δ0 [3] = 1.3 were taken
to satisfy the conditions (3.58) guaranteeing B̃[k] to be nonsingular were
derived from Equation (3.5). The initial B̃0 [1], B̃0 [2] and B̃0 [3] were chosen
as B̃0 [k] = B0 [k] + δ0 [k]Ir with the initial elements of B0 [k] which were
(ij)
selected from B inside the corresponding ranges [b(ij) , b ] specified as
follows: b(11) ∈ [1, 5], b(12) ∈ [0, 2], b(21) ∈ [0, 2], b(22) ∈ [1, 2],
(11) (12) (21)
b(31) ∈ [1, 4], b(32) ∈ [0, 5]. Namely, we set b0 = 1, b0 = 1 b0 = 0,
(22) (31) (32)
b0 = 1.9, b0 = 2, b0 = 2.1. The desired output vector was given as
y 0 = [1, 3, 7]T .
The performance of the simulated adaptive control system with the dis-
(i) (i) (i)
turbance sequences {vn } = v0 , v1 , . . . generated as some i.i.d. random
(1) (2)
variables taken from −0.1 ≤ vn ≤ 0.1, −0.2 ≤ vn ≤ 0.2, −0.08 ≤
(3)
vn ≤ 0.08 is presented in Figures 3.3 and 3.19.
Figures 3.19(a)–(c) demonstrate that the performance of the proposed
adaptive controller applied to the static multivariable plant having some
nonsquare gain matrix with not full rank is successful enough.
3.8 Conclusion
This chapter shed a light on the problem regarding the robust control of the
discrete-time multivariable memoryless systems with the noninvertible gain
matrices of the not full rank. It has been established that if certain condi-
tions have been satisfied then the nonadaptive pseudoinverse model-based
controller can be robust against plant uncertainty.
Again, the chapter answered the second question: Can robustly-adaptive
controller using as adjustable inverse model principle cope with the interval
parametric uncertainty of the system to be controlled? Namely, it has been
shown that such a possibility exists because these controllers are able to
guarantee the ultimate boundedness of output and control input signals.
However, this important feature is achieved via an “overparametrization” of
the uncertain systems. Nevertheless, the simulation experiments showed their
efficiency.
References
[1] V.M. Kuntsevich, et al (Eds). Control Systems: Theory and Applications.
Series in Automation, Control and Robotics, River Publishers, Gistrup,
Delft, 2018.
92 Robust and Robustly-Adaptive Control of Some Noninvertible Memoryless
[2] Y.P. Kondratenko, V.M. Kuntsevich, A.A. Chikrii, V.F. Gubarev, (Eds).
Advanced Control Systems: Theory and Applications. Series in Automa-
tion, Control and Robotics, River Publishers, Gistrup, 2021.
[3] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control,
Wiley, Chichester, 1996.
[4] P. Albertos and A. Sala, Multivariable Control Systems: an Engineering
Approach, Springer, London, 2006.
[5] T. Glad and L. Ljung, Control Theory: Multivariable and Nonlinear
Methods, Taylor & Francis, New York, 2000.
[6] L. Tan, A Generalized Framework of Linear Multivariable Control,
Elsevier, Oxford, 2017.
[7] V.F. Gubarev, M.D. Mishchenko, and B.M. Snizhko, ‘Model predic-
tive control for discrete MIMO linear systems’, In: Kondratenko, Y.P.,
Chikrii, A.A., Gubarev, V.F., Kacprzyk, J. (Eds.). Advanced Control
Techniques in Complex Engineering Systems: Theory and Applications.
Dedicated to Prof. V.M. Kuntsevich. Studies in Systems, Decision and
Control, pp. 63-81, vol. 203, Springer Nature Switzerland AG, Cham,
2019. DOI: https://doi.org/10.1007/978-3-030-21927-7_10
[8] E. Davison, ‘The output control of linear time-invariant multivariable
systems with unmeasurable arbitrary disturbances’, IEEE Trans. Autom.
Contr., pp. 621–631, no. 5, vol. AC-17, Oct. 1972.
[9] B. Francis and W. Wonham, ‘The internal model principle of control
theory’, Automatica, pp. 457–465, no. 5, vol. 12, Sep. 1976.
[10] R. Brockett, ‘The invertibility of dynamic systems with application to
control’, Ph. D. Dissertation, Case Inst. of Technology, Cleveland, Ohio,
1963.
[11] M.K. Sain and J.L. Massey, ‘Invertibility of linear time-invariant
dynamical systems’, IEEE Trans. Autom. Contr, pp. 141–149, no. 2,
vol. AC-14, Apr. 1969.
[12] L.M. Silverman, ‘Inversion of multivariable linear systems’, IEEE
Trans. Autom. Contr., pp. 270–276, no. 3, vol. AC-14, Jun. 1969.
[13] H. Seraji, ‘Minimal inversion, command tracking and disturbance
decoupling in multivariable systems’, Int. J. Control, pp. 2093–2191,
no. 6, vol. 49, Jun. 1988.
[14] G. Marro, D. Prattichizzo, and E. Zattoni, ‘Convolution profiles for
right-inversion of multivariable non-minimum phase discrete-time sys-
tems’, Automatica, pp. 1695–1703, no. 10, vol. 38, Oct. 2002.
References 93
L.P. Yugay
Dr.Sc.(Math), Professor,
Department of Natural Sciences, Uzbek State University of Physical Culture
and Sport, Chirchik City, Republic of Uzbekistan
Tel: +998-90-187-1478(m)
E-mail: [email protected]
Abstract
In the paper, a survey of the application of integral inequalities in various the-
oretical branches of mathematics and their applied fields is represented. New
integral inequalities of the Hölder type are introduced, and corresponding
estimates are proved. Also, the possibilities of applications of demonstrated
results in nonlinear differential games of avoiding encounters are indicated.
4.1 Introduction
4.1.1 On the First Integral Inequalities
Integral inequalities are widely used in the theories of differential equa-
tions, nonlinear oscillations, stability, as well as in mechanics and averaging
problems. With their help, the unique theorems of solutions of differential
equations are proved, and various approximate estimates are obtained for
solutions of differential equations that cannot be analytically solved. Let us
formulate the first variance of Grönwall–Bellman type inequality.
97
98 Nonlinear Integral Inequalities and Differential Games
Then b
v (t) ≤ Cexp u (s) ds. (4.2)
a
The inequality (4.2) was proved by T.H. Grönwall in [1] (1919) and extended
by R. Bellman in [2] (1943). We must note that Grönwall–Bellman inequality
(4.2) was the beginning point for studying many integral inequalities and their
applications. Particularly, in [3], P. Hartman writes that the result (4.2) has
as its sours the work of G. Peano [4]. Among the publications on nonlinear
Grönwall–Bellman type inequalities, the first papers were published by Bihari
[5] (1956) and A. Perov [6] (1957). Therefore integral inequalities of this
type (linear and nonlinear) are usually associated with Grönwall, Bellman,
Bihari, and others. The other names to be mentioned in connection with the
further development of the theory of integral inequalities are R.P. Agarval,
S.A. Chaplygin, N.B. Azbelev, D.D. Bainov, P.R. Beesack, S.G. Deo, U.D.
Dhongade, V. Lakshmikantham, S. Leela, B.G. Pachpatte, J. Popenda, P.S.
Simeonov, Z.B. Tsalyuk, D. Willett, J.S.W. Wong, E.C. Young. Wong, E.C.
Young. Integral inequalities are necessary for studying various equations (dif-
ferential, difference, discrete, etc.). This is why the intensive investigations of
different integral inequalities and hundreds of publications on them appear. A
part of the results can be found in the monographs of E.F. Beckenbach and R.
Bellman (1965), V. Lakshmikantham and S. Leela (1969), W. Walter (1970),
P.R. Beesack (1975), A.N. Filatov and L.V. Sharova (1976), A.A. Martynyuk
and R. Gutowski (1979), R. Rabczuk (1979), D.Ya. Mamedov, S. Ashirov, S.
Atdaev (1980), and J. Schroder (1980).
various linear and nonlinear differential equations classes. During this period,
integral inequalities constituted the central part of the proofs of existence and
uniqueness theorems for solutions of nonlinear differential equations. They
were also used in evaluating their solutions (one-sided and two-sided) [5–13].
Using integral inequalities, approximate estimates for solutions of differential
equations that do not have analytical solutions are obtained.
Primary applications of integral inequalities are outlined in the actual
work [10].
The next stage in the development of research of integral inequalities
dates back to the 1970–80s. Integral inequalities began to be used for the
solution of problems in the theory of nonlinear oscillations and the theory of
stability, as well as in mechanics, averaging problems, etc. A detailed review
of nonlinear integral inequalities can be found in [10], and applications in
nonlinear oscillations and mechanics are presented in [11–13].
As we can see from the listed sections, there are no applications of integral
inequalities in such important applied areas as the theory of optimal control
and its extension of the theory of conflict-controlled systems (differential
games). Perhaps this is due to the thematic interests of the authors because
integral inequalities are widely used in the theory of nonlinear optimal and
automatic control problems [14–18] and the theory of differential escape
games [19–59].
The theory of integral inequalities is intensively developed (third stage).
It covers new areas of applications, in particular, the theory of optimal
(automatic) control [14–18] and the theory of nonlinear differential games
[19–59].
In the theories of optimal and automatic control, as well as in differential
games, integral inequalities are applied to:
• identification (estimation) of trajectories of nonlinear controlled and
conflict controlled systems [14–30, 52, 58];
• obtaining lower estimates for the distances from the current position of
the trajectory to the terminal (target) set [31–42, 53–56];
• estimates of admissible controls that satisfy “energy” (integral) con-
straints and integral inequalities [39, 44–47, 58, 59].
Recently, there has been an active extension of new applications of
integral inequalities in differential games [48]. Thus, in [48–50], the authors,
for the first time, consider new classes of controls for the game partic-
ipants (the pursuer and the evader) based on integral inequalities of the
Grönwall–Bellman type. These controls generalize well-known controls
100 Nonlinear Integral Inequalities and Differential Games
z (ξ) = ξ − γ1 ξ α1 − γ2 ξ α2 − · · · − γn ξ αn − γn+1 ,
is hold.
Next, let’s find out the behavior of the derivative. We have, taking into
account 0 < αj < 1, j = 1, 2, . . . , n, that
!
n
lim z (ξ) = lim (1 − γi αi ξ αi −1 ) = 1,
ξ→+∞ ξ→+∞
i=1
!
n
1
lim z (ξ) = lim (1 − γi αi ) = −∞.
ξ→+0 ξ→+0 ξ 1−αi
i=1
From this and the continuity of the function z (ξ), we obtain that there is a
point, such that z (ξ∗ ) = 0.
Further investigation of the critical point ξ ∗ shows that at this
point z ( ξ ∗ ) > 0, therefore, ξ ∗ is a point of a local minimum. Further,
since z (ξ) > 0 for all ξ > 0 (including the point ξ = ξ ∗ ) then z (ξ) it is
a strictly convex function for all ξ ≥ 0, and such functions, the extremum
point ξ∗ > 0 is the only point of the global minimum of the function z (ξ)
ξ ≥ 0. in the domain ξ ≥ 0. It is easy to verify that ξ 0 > ξ∗ > 0.
Let 0 ≤ ξ ≤ ξ 0 , then from the property of the root ξ 0 we have, which
proves (4.3). Lemma 1 is proved.
Theorem 4.1 (Main result). Let the function u (t) be continuous for t0 ≤
t ≤ t0 + 1, t0 ≥ 0, and satisfies the Hölder type integral inequality (4.7):
t !
n
0 ≤ u (t) ≤ f (t) + a (s) + b (s) u (s) + ci (s)uαi (s) ds (4.7)
t0 i=1
where all functions f (t) , a (t) , b (t) ci (t) are given, nonnegative, inte-
grable, and are not equal identically to zero on the interval [t0 , t0 +1],
0 < αi < 1, i = 1, 2, . . . n.
Then there is a constant h0 ∈ (t0 ; t0 + 1], such that the function u(t) for
all t ∈ [t0 ; t0 + h0 ] satisfies the inequality
t !
n
u (t) ≤ f (t) + a (s) ds+δξ0 + ξ0αi gi (t) , (4.8)
t0 i=1
⎛ ⎞ αi ⎛ ⎞1−αi
t h !n h t
1 1
f (t)+ a (s) ds+δ u (s) ds+ ⎝ (uαi (s)) αi
ds)⎠ ⎝ ci 1−αi
(s) ds)⎠ ≤
0 0 i=1 0 0
⎛ ⎞ αi ⎛ ⎞1−αi
t h !
n h t
1
f (t) + a (s) ds + δ u (s) ds + ⎝ u(s) ds)⎠ ⎝ ci 1−αi
(s) ds)⎠
0 0 i=1 0 0
|f ( z 2 , u, v) − f ( z 1 , u, v)| ≤ Lα | z 2 − z 1 |α , (4.17)
where z 1 ∈ Rn , z 2 ∈ Rn , 0 < α ≤ 1, C ≥ 0 and Lα ≥ 0 are constants.
Terminal set M is a linear subspace of Rn .
Parameters u and v in (4.16) are chosen by opposing sides (players) as
measurable functions u = u (t) ∈ P, v = v (t) ∈ Q, t ≥ 0. Player,
choosing v = v (t) ∈ Q (evader), sets itself a task for any acceptable
behavior u (t) ∈ P to avoid encountering the corresponding trajectory z(t)
of equation (4.16), starting from any point z0 ∈ Rn \M (z0 = z(0)),
from M for any t ≥ 0. Such problem is called global avoiding encounter
problem(global evasion problem), which was formulated by L.S. Pontrjagin
and E.F. Mis̆c̆enko in [31]. At every moment t ≥ 0 the evader by forming
value of v (t) ∈ Q may use values u (s) ∈ P and z(s) with s ≤ t, but he
does not know its future values at s > t.
In addition, according to the conditions of the avoiding problem, the
evading player has a particular advantage over the pursuer for the successful
solution of the collision avoidance problem. The benefits are set in differ-
ent ways (geometric, integral, mixed, etc.) and are usually included in a
conflict-controlled process’s dynamics (4.16).
Using its advantages in dynamics and information, the evader firstly
constructs a local special control function v c (t) ∈ Q, which ensures for him,
for any admissible controls of the pursuer, to avoid on a small-time interval
0 ≤ t ≤ θ0 ( θ0 > 0 is a constant) the collision M with the corresponding
trajectory z (t) starting at point z0 = z(0)∈ Rn \M. Then, based on the local
deviation, the evader an algorithm for avoiding the trajectory z (t) from M
for all t ≥ 0 is constructed (global avoiding). At the stage of local avoidance,
the main task is to prove that the distance ρ (z (t) , M ) from the moving point
z (t) of the trajectory to the terminal set is positive for all t ≥ 0, then for
4.2 Investigation of Nonlinear Hölder Type Integral Inequality 105
Further, based on his advantage and Hölder condition (4.17), the evading
player constructs his control in such a way that provides him the inequality
t
α
|z (t) − z0 | ≤ d1 (1 + | z 0 | )t + ( d2 +Lα |z (s) − z 0 |α ) ds, (4.18)
0
2. The case when in (4.7) n= 1, a (t) ≡0, f (t) ≡f0 ≥0 (f0 is a constant
for all t≥t0 ), α1 ≥ 0, is studied in [7] (if α1 =0 or α1 =1 we have
nonlinear case). It is necessary to note, that the proof in [7] does not
extend in the general case for the proof of the inequality (4.7).
3. The case when in (4.7) n = 1, a (t) ≡ 0, b (t) ≡ 0, 0 ≤ t ≤ 1,
0 < α1 < 1, considered in [7,12,13].
4.2.4 Remarks
Remark 4.1. First, Hölder type inequality was considered in [6] and investi-
gated in [12,13,58].
Remark 4.3. Inequality (4.8) and the constant h0 are especially important
for obtaining a lower bound for the distance from a moving point of the
trajectory (4.17) of a conflict-controlled dynamical system to a given terminal
set [32–35,37,39].
4.3 Conclusion
The main aim of this work is to investigate new nonlinear Hölder type integral
inequalities, generalized well-known integral inequalities [6, 12, 13].
New integral inequalities of the Hölder type are proved, and corre-
sponding solutions of nonlinear integral inequalities of the Hölder type are
obtained, which generalized the well-known [6, 12 ,13]. The results on
integral inequalities presented in this chapter allow for solving various control
problems [14–18] and are applicable to solving new nonlinear problems in the
theory of differential avoiding games. They may also be used for constructing
new control functions(strategies) of players in differential games of avoiding
encounters [58, 59]. Further, the received results make it possible to apply
References 107
Acknowledgments
The author is grateful to the Academician of the National Academy of Sci-
ences of Ukraine, Professor A.A. Chikrii, for his kind encouragement of this
work. He also thanks to the Organizing Committee of AUTOMATION2020
for the invitation to participate in this high-level event.
References
[1] T.N. Grönwall, ‘Note on the derivatives with respect to a parameter of
the solutions of a system of differential equations,’ Ann. of Math. 20: 2
(1919), pp. 292–296. DOI 10.2307/1967124.
[2] R. Bellman. ‘The stability of solutions of linear differential equations,’
Duke Math. J. 10(1943), pp. 643–647.
[3] P. Hartman, Ordinary Differential Equations, John Wiley & Sons,
NewYork- London-Sydney, 1964, 720 p.
[4] G. Peano. Sull’ integrabilitá delle equazione differenzialli di primo
ordine, Atti. R.Accad.Torino, 21(1885/18886), 667–685 [III.2.4].
[5] I. Bihari, ‘A generalization of a lemma of Bellman and its application to
uniqueness problems of differential equations,’ Acta Math. Acad. Sci.
Hungr. 7:1(1956), pp. 71–94.
[6] A.I. Perov, ‘On Integral Inequalities,’ Proc. of the Seminar on Functional
Analysis, Voronej St. Univ. − Issue 5. − 1957.
[7] S.A. Chaplygin, ‘A new method of approximate integration of differen-
tial equations,’ Trudy TsAGI 130 (1932), pp. 5–17.
[8] E. Coddington and N. Levinson, ‘Theory of Ordinary Differential
Equations’, McGraw-Hill Book Company, Inc. New York–Toronto–
London,1955, − 474 p. (Russian Transl., Moscow: IL.1958).
[9] L.S. Pontrjagin, ‘Ordinary differential equations,’ −M.: Nauka, 1970, −
332 p.
[10] D. Bainov, P. Simeonov, ‘ Integral Inequalities and Applications.’ Dor-
drecht, Springer, Mathematics and its Applications (East European
Series, vol. 57), 1992, −256p.
108 Nonlinear Integral Inequalities and Differential Games
[23] L.A. Petrosjan, ‘On a set of differential games of survive in the space
Rn ’ Dokl. Acad.Nauk SSSR, 1965, V. 161, No. 1, –P.52-54.
[24] M.I. Zelikin, N.T. Tynjanskiĭ, ‘Determinate differential games’,
Usp.Mat. Nauk, 1965, V. 20. No. 4. −P.151–157.
[25] Yu.S. Osipov, ‘Alternative in differential-difference game,’ Dokl. Soviet
Acad.Sc., 1971, V. 197, No. 5, − P. 1022–1025.
[26] A.I. Subbotin, A.G. Chentsov, ‘Optimization of Guarantee in Control
Problems’, -M.: Nauka, 1981, − 288p.
[27] B.N. Pshenichnyĭ, V.V. Ostapenko, ‘Differential games’, Kiiv: Naukova
Dumka, 1992, −261p.
[28] A.A. C̆hikriĭ, ‘Conflict controlled processes’, Boston-London-
Dordrecht, Kluwer Acad.Publ., 1997, − 424 p.
[29] M. Sh, Mamatov, Kh. Kh. Sobirov, ‘To the theory of differential games
of pursuit on position,’ Results of Science and Technics, Modern math.
and its Appl.Subject Review, 2018, v.144, −P. 39–46.
[30] M. Sh. Mamatov, M. Tukhtasinov, ‘On a Problem of Pursuit in Dis-
tributed Control Systems,’ Cybernetics and System Analysis, Kyiv,
2009, No. 2, − P. 153–158.
[31] L.S. Pontrjagin, E.F. Mi˘s c̆enko, ‘A problem of evasion of one controlled
object from an other,’ 1969, Dokl. Acad. Nauk SSSR, V. 189, No. 4,
–P. 721–723.
[32] L.S. Pontrjagin, ‘A linear differential evasion game’, Trudy
Mat.Inst.Steklov, 19 71, V. 112, −P. 30–63.
[33] E.F. Mi˘s c̆enko, N, Satimov, ‘The evasion problem in differential games
with nonlinear controls,’ Differencialnie Uravneniya, V. 9, 10, 1973.
−P.1792-1797.
[34] E.F. Mi˘s c̆enko, M.S. Nikolskiĭ, N. Satimov, ‘The problem of avoiding
encounter in N-person differential games’, Proc. of the Steklov Institute
of Mathematics, 1980, Iss. 1, −P. 111–136.
[35] N.Satimov, ‘On a one way of avoiding encounter in differential games,’
Matematicheskiy sbornik, 99(141), 1976, −P. 280–293.
[36] F.L. Chernousko, A.A. Melikyan, ‘Game Problems of Control and
Search’, – Nauka, 1978,−270 p.
[37] P.B. Gusyatnikov, L.P.Yugay, ‘On an evasion problem in nonlinear
differential games with terminal set of compound structure,’ Izvestiya
Acad. Nauk SSSR, 1977, Techn.kibernetika, No. 2, −P. 8–13.
[38] R.V. Gamkrelidze, G.L. Kharatshvili, ‘A differential game of evasion
with nonlinear control’ SIAM J. Control, May 1974, Vol. 12, No 2,
–P. 332–349.
110 Nonlinear Integral Inequalities and Differential Games
G.Ts. Chikrii
Abstract
In this chapter, we consider the linear differential game of approaching a
cylindrical terminal set. The case when classic Pontryagin’s condition does
not hold is studied. Instead, a considerably weaker condition, incorporating
the function of time dilation, is introduced. This makes it possible to expand
the range of problems that can be solved analytically, by using information
about the enemy’s control in the past.
The study is carried out in the framework of the Method of Resolving
Functions. The gist of the method consists in constructing certain scalar
function associated with the parameters of the conflict-controlled process and
characterizing the gain of the first player at each moment. If the total gain
achieves the predetermined value then this means that the game terminates in
a certain guaranteed time. The pursuer’s control, realizing the game goal, is
constructed based on the Filippov–Castaing theorem on measurable choice.
To illustrate the developed scheme, we analyze in detail the problem of
approaching two controlled oscillation systems in geometric coordinates.
113
114 Principle of Time Dilation in Game Problems of Dynamics
5.1 Introduction
The mathematical theory of conflict-controlled processes (differential games,
dynamic games) has at its disposal a wide range of fundamental methods
to study various nature-controlled processes which function in conditions of
conflict and uncertainty.
Attempts to construct optimal behavior of the opposing sides when solv-
ing game dynamic problems inevitably lead to the ideology of dynamic pro-
gramming, closely related to the Hamilton–Jacobi—Bellman–Isaacs equa-
tion, the main equation in the theory differential games [1]. In a slightly
different form, namely in the terms of sets, these ideas are reflected in the
Pontryagin–Pshenichnyi backward procedures [2, 3]. These are the method
of alternated integral and the method of semi-group operators Tε .
According to the aforementioned ideology, in the course of the pro-
cess of counter-acting at each moment, one of the parties maximizes and
another minimizes the payoff functional. Thus, optimization of the conflict-
controlled process is reduced to calculating continual maxmin or minmax of
the functional, depending on the current information available to each of the
parties.
The desire to find optimal solutions for counter-acting parties in the
game problems of dynamics encounters great difficulties of mathematical
nature. Therefore, a number of effective mathematical methods have been
created for deciding dynamic games, which provide a guaranteed result and
give sufficient conditions for the goal achievement without worrying about
optimality, which is quite justified from the practical point of view.
These are the First Direct Method of L.S. Pontryagin [2], the Rule of
Extremal Aiming of N.N. Krasovskii [4, 5], the Method of Programmed
Iterations [6], and the Method of Resolving Functions [7]. Adjacent to them
are the works [8, 9]. Various applications of game methods to attack applied
problems are contained in the books [10, 11].
In the course of a conflict-controlled process, information about the cur-
rent state of the process, as a rule, delays. One of the reasons is that a certain
time is required to process it. The author of this chapter developed the method
of reducing the approach problem with delay of information to the equivalent
problem with complete information, but slightly changed dynamics [12, 13],
followed by application of the known classic methods.
When performing such important tasks as, for example, interception of a
moving target, in order to achieve the goal, a certain advantage of the pursuer
in terms of control resources and (or) dynamic capabilities is necessary.
5.1 Introduction 115
Classic Pontryagin’s condition [2] plays this role in general problem state-
ments. There exist various forms of this condition [7]. One of them consists
in incorporating the solid part of the terminal set into Pontryagin’s condition.
Another way concerns the suppression of the evader’s control resource by the
way of multiplying it by a certain matrix function and subsequent subtraction
of the used resource from the terminal set. In so doing, Pontryagin’s condition
falls into two conditions. The latter technique is appropriate, for example, in
the case when the opposing objects have different inertia.
However, in many cases of conflict withstanding Pontryagin’s condition
does not hold, e.g. the problems of soft meeting and the pursuit problems for
oscillatory processes.
M.S. Nikolskij and D. Zonnevend [14, 15] suggested weakening Pon-
tryagin’s condition by the way of incorporating into it a certain function,
later on, called the function of time dilation. The establishment of the close
relation of the modified condition with the passage from the original game
with complete information to an auxiliary game with special kind information
delay [13] gave impetus to the development of an efficient approach the so-
called principle of time dilation for solving problems for which Pontryagin’s
condition does not hold.
Essentially, the transition is made from the original game with complete
information to the game with the same dynamics and the terminal set, yet
with a special kind of information delay, decreasing as the game trajectory
approaches the terminal set and vanishing as it hits the target... An important
point is that Pontryagin’s condition for the latter game involves the time
function of time dilation.
We extend the concept of the function of time dilation to the class of
functions of bounded variation. This allows applying the above-mentioned
technique to solving a wide range of game problems for the processes,
described by the systems of second-order differential equations [16, 17] and
for the problems of soft meeting [18].
In this chapter, we apply the principle of time dilation to the game
constructions based on the method of resolving functions [7, 19] This method
is closely related with the Minkowski’ inverse functionals, constructed for
special set-valued mappings. An attractive feature of the method of resolving
functions is that it provides the full theoretical substantiation of the classic
rule of parallel pursuit and the approach along the ray [20]. Moreover,
when substantiating the game constructions and obtaining meaningful results
on their basis, the modern technique of set-valued mappings is effectively
used.
116 Principle of Time Dilation in Game Problems of Dynamics
ż = Az + u − v, (5.1)
M ∗ = M0 + M, (5.2)
where
U0t = {u0 (s) : u0 (s) ∈ U, s ∈ [0, I(t) − t]}.
If the inequality in parentheses (5.4) is not fulfilled for all t ≥ 0, then
we put TI (z0 , γI (·)) = Ø. In what follows, we assume that the exact upper
bound u0 (·) is achieved in (5.4). For example, as a rule, in the case, the set
U contains 0, for simplicity’s sake, control of the first player on the initial
time interval is set identically equal to zero that providing fulfillment of the
inequality in (5.4).
Before proceeding to the formulation and proof of the main result, let
us perform some rearrangements in the representation of the system solution
(5.1), taking into account the availability of information of the first player
about his adversary control choice.
Let t be an arbitrary moment. Using the Cauchy formula, for the chosen
controls of the players, we present o the solution of system (5.1) in projection
onto the subspaceL:
t t
πz(t) = πeAt z0 + πeA(t−s) u(s)ds − πeA(t−s) v(s)ds.
0 0
Taking into account the crucial role of the function of time dilation, we
separate the section of initial control of the evader. Then we have
I(t)−t
πz(I(t)) = πeAI(t) z0 + πeA(I(t)−s) u0 (s)ds+
0
I(t) I(t)
+ πeA(I(t)−s) u(s)ds − πeA(I(t)−s) v(s)ds. (5.5)
I(t)−t 0
then the trajectory of the process (5.1) can be brought to the terminal set at
the moment I(T∗ ).
Proof. Let us study the process (5.1), (5.2) on the interval [0, I(T∗ )]. On the
initial interval [0, I(T∗ ) − T∗ ], we choose the initial control u0 (·) from the
condition for a maximum of the following expression
T∗
inf αI (T∗ , s, v(s), u0 (·))ds.
v(·)∈ΩE 0
Let us choose the pursuer control as follows. In the case N (T∗ , u0 (·)) ∈
/
M , we consider the set-valued mapping
Here
ᾱI (T∗ , s, v(I(T∗ ) − I(T∗ − s)), u0 (·)) =
αI (T∗ , s, v(I(T∗ ) − I(T∗ − s)), u0 (·)), s ∈ [0, t∗ ],
=
0, s ∈ (t∗ , T∗ ].
In view of the measurability of this mapping, there exists its selection,
which is a superposition of measurable functions [34]
If, otherwise, N (T∗ , u0 (·)) ∈ M , then control of the first player is chosen
according to a similar scheme, but with the resolving function equal to zero
on the entire interval [0, T∗ ]. This case [19] is said to correspond to the first
direct method of L.S. Pontryagin [2].
Setting t = T∗ and separating the active and the passive segments, we
finally derive the representation of the solution from formula (5.6):
I(T∗ )−T∗
AI(T∗ )
πz(I(T∗ )) = πe z0 + πeA(I(T∗ )−s) u0 (s)ds+
0
t∗ T∗
A(T∗ −s)
+ πe u(I(T∗ )−T∗ +s)ds+ πeA(T∗ −s) u(I(T∗ )−T∗ +s)ds−
0 t∗
T∗
− πeAI(T∗ −s) v(I(T∗ ) − I(T∗ − s) · I(T
˙ ∗ − s)ds.
0
122 Principle of Time Dilation in Game Problems of Dynamics
Note that in the last equation of (5.10), the convexity of the solid part of the
terminal set was taken into account.
If, otherwise, N (T∗ , u0 (·)) ∈ M , then, in view of the choice of control
(5.8), (5.9) with the zero resolving function, from (5.10) readily follows the
inclusion
πz(I(T∗ )) ∈ M.
Remark 5.1. The scheme of the method and main assertion are valid for
much more general classes of conflict-controlled processes than described by
(5.1), in particular, for generalized functional-differential systems, including
fractional, impulse, and time-delay systems.
Remark 5.2. The function of time dilation can be discontinuous, but abso-
lutely continuous on the intervals of its continuity [17].
5.5 Example
To illustrate the proposed approach, let us investigate the game problem of
convergence in geometric coordinates of two controlled systems of second-
order, describing the dynamics of the mathematical pendulum. Let their
dynamics be given by the equations
ẍ + a2 x = ρu, x ∈ Rn ,
ÿ + b2 y = σv, y ∈ Rn . (5.11)
Here x and y are geometric coordinates of the players, the control param-
eters are u v, u ≤ 1, v ≤ 1, a2 and b2 are the stiffness control factors, ρ
5.5 Example 123
where ⎛ ⎞
0 E 0 0
⎜ −a2 E 0 0 0 ⎟
A=⎜
⎝ 0
⎟.
0 0 E ⎠
2
0 0 −b E 0
Let us write Pontryagin’s condition [2]. In the case in study, it takes the form
[ρ/a |sin at| u − σ/b |sin bt| v] = ρ/a |sin at| S ∗ σ/b |sin bt| S =
v≤1u≤1
(ρ/a − σa
b2
) |sin a(t − s)|
min αI (t, s, v, 0) = .
v≤1 N (t, 0)
This can be verified using the Lagrange multiplier method [36].
Thus, by virtue of (5.4), the issue concerning the existence of a finite time
for the game termination is reduced to the existence of a positive root of the
following equation for t:
t 4
σ
a (ρ a2 − 2 ) |sin a(t − s)| ds = N (t, 0) . (5.17)
0 b
Here
N (t, 0) = cos bI(t)y0 − cos aI(t)x0 + 1/b sin bI(t)ẏ0 − 1/a sin bI(t)ẋ0 ,
5.6 Conclusion
We present a method for solving the game problems of bringing a trajectory
to the terminal set in the case when the classical Pontryagin’s condition is not
satisfied. This became feasible due to the introduction of the function of time
dilation and a weaker modified condition based on it
The method of resolving functions is used as a basis for investigation.
This method allows obtaining sufficient conditions for the approach in a finite
guaranteed time. In so doing, the first player constructs its control based on
References 127
delayed information on control of the second player, and the choice of control
itself is carried out based on the theorem on measurable choice.
The results are illustrated by the example of the game approach in geo-
metric coordinates of two second-order oscillatory systems, describing the
dynamics of mathematical pendulums.
References
[1] R.F. Isaacs, ‘Differential Games’, New York–London–Sydney: Wiley
Interscience, 479 p., 1965.
[2] L.S. Pontryagin, ‘Selected scientific papers’, 2, Moscow: Nauka, 576
p., 1988. (in Russian).
[3] B.N. Pschenitchnyi, ‘ε-strategies in Differential Games’, Topics in Dif-
ferential Games, New York, London, Amsterdam: North-Holland Publ.
Co., pp. 45–99, 1973.
[4] N.N. Krasovskii, ‘Game Problems on the Encounter of Motions,
Moscow: Nauka, 420 p., 1970 (in Russian).
[5] B.N. Pschenitchnyi, ‘Linear Differential Games’, Avtom, Telemekh.,
p. 65–78, No 1, 1968 (in Russian).
[6] A.I. Subbotin, A.G. Chentsov, ‘Optimization of Guaranteed Result in
Control Problems’, Moscow: Nauka, 286 p., 1981 (in Russian).
[7] A.A. Chikrii, ‘Conflict controlled processes’, Boston; London, Dor-
drecht: Springer Science and Business Media, 424 p., 2013.
[8] O. Hajek, ‘Pursuit games’, New York: Academic Press, 266 p., 1975.
[9] M.S. Nikolskii, ‘L.S. Pontryagin’s First Direct Method in Differential
Games’, Izdat. Gos.Univ., Moscow, 65 p., 1984 (in Russian).
[10] G. Siouris, ‘Missile Guidance and Control Systems’, New York:
Springer-Verlag, 666 p., 2004.
[11] G. Siouris, ‘Aerospace Avionics Systems: A Modern Synthesis’, San
Diego: Academic Press, 466 p., 1993.
[12] G.Ts. Chikrii, ‘An approach to solution of linear differential games with
variable information delay’, Journal of Automation and Information
Sciences: Begell House Inc., p. 163–170, No 3, 4, vol. 27, 1995.
[13] G.Ts. Chikrii, ‘Using the effect of information delay in differential
pursuit games’, Cybernetics and Systems Analysis, p. 233–245, No. 2,
vol. 43, 2007.
[14] M.S. Nikolskij, ‘Application of the first direct method in the linear
differential games’, Izvestia Acad. Nauk SSSR, p. 51–56, vol. 10, 1972
(in Russian).
128 Principle of Time Dilation in Game Problems of Dynamics
[29] A.A. Chikrii, S.D. Eidelman, ‘Game problems for fractional quasilinear
systems’, Computers and Mathematics with Applications, p. 835–851,
vol. 44, No. 7, 2002.
[30] A.G. Nakonechnyi, E.A. Kapustyan, A.A. Chikrii, ‘Control of impulse
system in conflict situation’, Journal of Automation and Information
Science, p. 54–63, No. 9, 2019.
[31] L.V. Baranovskaya, A.A. Chikrii, Al.A. Chikrii, ‘Inverse Minkowski
functional in a nonstationary problem of a group pursuit’, Journal of
Computer and Systems Sciences International, p. 101–106, vol. 36,
No. 1, 1997.
[32] L.V. Baranovskaya, Al.A. Chikrii, ‘Game problems for a class of
hereditary systems’, Journal of Automation and Information Science,
p. 87–97, vol. 29, No 2, 1997.
[33] L.A. Vlasenko, A.G. Rutkas, A.A. Chikrii, ‘On a differential game in a
stochastic system’, Proceedings of the Steklov Institute of Mathematics,
p. 185–198, vol. 309, suppl.1, 2020.
[34] J.-P. Aubin, H. Frankowska, ‘Set-valued analysis’, Boston; Basel;
Berlin: Birkhauser, 1990.
[35] N.V. Vasilenko, ‘Theory of Oscillations, Kiev: Vyshcha Shkola, 430
p., 1992 (in Russian).
[36] A.D. Joffe, V.M. Tikhomirov, ‘Theory of Extremal Problems’, North-
Holland, Amsterdam, 480 p., 1979.
[37] V.M. Kuntsevich et al (Eds). Control Systems: Theory and Applications.
Series in Automation, Control and Robotics, River Publishers, Gistrrup,
Delft, 2018.
[38] Y.P. Kondratenko, V.M. Kuntsevich, A.A. Chikrii, V.F. Gubarev
(Eds). Advanced Control Systems: Theory and Applications, Series in
Automation, Control and Robotics, River Publishers, Gistrrup, 2021.
6
Method of Upper and Lower Resolving
Functions for Pursuit Differential-difference
Games with Pure Delay
Lesia Baranovska
Abstract
A modification of the method of resolving functions for differential-
difference games with pure delay is considered. The lower and upper resolv-
ing functions are introduced. For the local game of approach, the scheme of
a method is constructed and sufficient conditions for the completion of the
game are formulated.
6.1 Introduction
Differential games is a section of mathematical control theory [1–3] studying
the manipulation of moving objects operated under conditions of uncertainty
and conflict. We consider the differential-difference games of approach. In the
theory of differential games, along with Isaacs ideology [4], several methods
have been developed that provide a guaranteed result. The most famous
methods are the first direct method of L.S. Pontryagin [5, 6], the method
of extreme aiming of M.M. Krasovskii [7, 8], and the method of resolving
functions of A.O. Chikrii [9–15]. In this chapter, the method of resolving
131
132 Method of Upper and Lower Resolving Functions
functions will be used. The first direct method of L.S. Pontryagin and the
method of resolving functions give a theoretical justification for the classical
rule of parallel pursuit. Most processes depend not only on the current state,
but they have essential knowledge of the prehistory of the state of the process,
which can’t be ignored. Today, systems of differential-difference equations
are increasingly used in control theory. The method scheme for differential-
difference games is developed in works [16–19]. In [20–22], the problem
of approach for a group of pursuers and a single evader is considered.
The scheme of the method of resolving functions for differential-difference
systems of neutral type was developed in [23]. For objects with different
inertia in [24], a modification of the Pontryagin condition by introducing the
body part of the terminal set is proposed. Differential games in case of failure
of control devices are considered in the works [25–27].
In this chapter, we first introduce the upper and lower resolving functions
of different types [28] in the method of resolving functions for the differential
game of approach with pure delay, which allows developing the method even
if the Pontryagin condition is not fulfilled. Sufficient conditions have been
obtained to complete the game.
Denote
ΩU = {u(s) : u(s) ∈ U, s ∈ [0, +∞)} ,
ΩV = {v(s) : v(s) ∈ V, s ∈ [0, +∞)} .
The function u(·) ∈ ΩU (v(·) ∈ ΩV ) ,will be called an open-loop pursuer’s
(evader’s) control. This control is chosen by the pursuer (evader) based on
knowledge of the initial condition.
The function vt (·) = {v(s) : s ∈ [0, t], v(·) ∈ ΩV } , will be called
prehistory of the control of evader at a time t, t ≥ 0 [9].
) We define * the pursuer’s “quasistrategy” [29] as a mapping
U t, z 0 (·), vt (·) . It assigns a Lebesgue measurable function to each
moment t ≥ 0, initial condition (6.2) and arbitrary prehistory vt (·) of the
control of evader, taking its values in control domain U:
) *
u(t) = U t, z 0 (·), vt (·) , t ≥ 0, (6.3)
M ∗ = M0 + M, (6.5)
Lemma 6.1 [35]. Let z(T ) be a continuous solution to the system (6.1) with
pure time-delay under the initial condition (6.2). Then,
50
z(T ) = exp τ {B, t}z 0 (−τ ) + −τ exp τ {B, t − τ − s}ż 0 (s)ds+
5t
+ 0 exp τ {B, t − τ − s}ϕ(u(s), v(s))ds.
Denote
Denote the lower and upper resolving functions [28] first type
Condition 6.2. The set-valued mapping R(t, s) has non-empty images on the
cone Δ.
Finally, denote the lower and upper resolving functions [28] second type:
respectively.
Theorem 6.1. Let the conflict-controlled process (6.1), (6.2), (6.5) satisfy
Conditions 6.1 and 6.2, M) = coM, and 0
* for a given function z (·) and a
shift functionγ(·, ·) T = T z (·), γ(·, ·) = ∅.
0
Then, the trajectory of the process (6.1), (6.2) can be under condition
α∗ (T ) < 1, brought by the pursuer (u) from z 0 (·) to the terminal set (6.3)
at the moment T by the control in the form of (6.4); and under the condition
α∗ (T ) > 1−in the counter-control class, with any permissible counteractions
of the evader.
Proof. Let v(·) be an arbitrary measurable function taking values from the
control domain V.
Suppose that α∗ (T, s, v) = +∞, s ∈ [0, T ], v ∈ V. We introduce the
controlling function
t T
h(t) = 1 − α∗ (T, s, v(s))ds − α∗ (T, s)ds, s ∈ [0, T ].
0 0
5T
If h(0) = 1 − 0 α∗ (T, s)ds = 1 − α∗ (T ) > 0,we have
T
h(T ) = 1 − α∗ (T, s, v(s))ds ≤ 0.
0
t∗
+ α∗ (T, s, v(s))[M − ξ(T )]ds+
0
T T
∗
+ α (T, s)[M − ξ(T )]ds + γ(T, s)ds =
t∗ 0
t∗ T
∗ ∗
= ξ(T ) 1 − α (T, s, v(s))ds − α (T, s)ds +
0 t∗
t∗ T
∗
+ α (T, s, v(s))M ds + α∗ (T, s)M ds =
0 t∗
T T
∗ ∗
= α (T, s)ds + α (T, s)ds M = M.
t∗ t∗
From (6.6) we have that the case α∗ (T, s, v) = +∞ for some s ∈ [0, T ], v ∈
V, is possible only if 0 ∈ M − ξ(T ), 0 ∈ exp τ {B, T }ϕ(U, v) − γ(T, s), and
in this case it is obvious that R(T, s, v) = [0, +∞), s ∈ [0, T ], v ∈ V, and
α∗ (T, s) = 0, s ∈ [0, T ].
This makes it possible to choose as the resolving functions at those points
s ∈ [0, T ] where α∗ (T, s, v(s)) = +∞, an arbitrary finite measurable
function that takes values on the interval [0, +∞) with only one condition
that the final function on the interval [0, T] provides relation h(t∗ ) = 0 for a
certain switching moment t∗ , t∗ ∈ [0, T ].
Thus, the construction of control in the active and the passive time
sections is reduced to the previous case.
6.4 Conclusion
The concept of upper and lower resolving functions of two types for
differential-difference games of approach with pure time delay is introduced
at the first time. Sufficient conditions are obtained for the trajectory of the
process (6.1), (6.2) to converge with a given cylindrical terminal set (6.5) in
the case when the Pontryagin condition does not hold.
In the future, it is planned to expand the proposed scheme for processes
that are described by systems of differential-difference equations with several
delays based on the use of the Cauchy formulas [43–45].
It is planned to apply the proposed scheme to specific examples and
to realize the visualization of the trajectory of the objects on the plane
[46, 47].
References 139
References
[1] Kuntsevich, V.M., et al (Eds). Control Systems: Theory and Applica-
tions. Series in Automation, Control, and Robotics, River Publishers,
Gistrup, Delft, 2018.
[2] Kondratenko, Y.P., Kuntsevich, V.M., Chikrii, A.A., Gubarev, V.F.
(Eds). Advanced Control Systems: Theory and Applications. Series in
Automation, Control, and Robotics, River Publishers, Gistrup, 2021.
[3] Kondratenko, Y.P., Chikrii, A.A., Gubarev, V.F., Kacprzyk, J. (Eds).
Advanced Control Techniques in Complex Engineering Systems: Theory
and Applications. Dedicated to Professor Vsevolod M. Kuntsevich.
Studies in Systems, Decision, and Control, Vol. 203. Cham: Springer
Nature Switzerland AG, 2019.
[4] R. Isaacs. Differential Games: A Mathematical Theory with Applications
to Warfare and Pursuit, Control and Optimization. John Wiley & Sons
Inc, New York, 1965.
[5] L.S. Pontryagin. Selected scientific works. Nauka, Moscow, 2, 1988.
[6] M.S. Nikol’skii. Pontryagin’s First Direct Method in Differential
Games. Mosk. Gos. Univ., Moscow, 1984 (in Russian).
[7] N.N. Krasovskii, A.I. Subbotin. Game-theoretical control problems.
Springer-Verlag, New York, 1988.
[8] V.N. Ushakov. Krasovskii’s Unification Method and the Stability Defect
of Sets in a Game Problem of Approach on a Finite Time Interval. In:
Kondratenko Y., Chikrii A., Gubarev V., Kacprzyk J. (eds) Advanced
Control Techniques in Complex Engineering Systems: Theory and
Applications. Studies in Systems, Decision, and Control, vol 203.
Springer, Cham, 2019. https://doi.org/10.1007/978-3-030-21927-7_5.
[9] A. Chikrii. Conflict-Controlled Processes. Springer Science &Business
Media, 2013.
[10] A.A. Chikrii, I.S Rappoport, K.A. Chikrii. Multivalued Mappings and
their Selectors in the Theory of Conflict-Controlled Processes. Cyber-
netics and Systems Analysis, 43(5):719–730, 2007.
[11] A.A. Chikrii, R. Petryshyn, I. Cherevko., Y. Bigun. Method of Resolving
Functions in the Theory of Conflict—Controlled Processes. In: Kon-
dratenko Y., Chikrii A., Gubarev V., Kacprzyk J. (eds) Advanced Control
140 Method of Upper and Lower Resolving Functions
[49] A.A. Chikrii, P.V. Prokopovich. Simple pursuit of one evader by a group.
Cybernetics and Systems Analysis, 28(3):438–444, 1992.
[50] Al.A. Chikrii. On nonstationary game problem of motion control. Jour-
nal of Automation and Information Sciences, 47(11):74-83, 2015.
https://doi.org/10.1615/JAutomatInfScien.v47.i11.60.
[51] I.I. Kryvonos, Al.A. Chikrii, K.A. Chikrii. On an approach scheme in
nonstationary game problems. Journal of Automation and Information
Sciences, 45(80):32–40, 2013. https://doi.org/10.1615/JAutomatInfSci
en.v45.i8.40.
[52] V.A. Pepelyaev, Al.A. Chikrii. On the game dynamics problems for non-
stationary controlled processes. Journal of Automation and Information
Sciences, 49(3):13–23, 2017. https://doi.org/10.1615/JAutomatInfScien
.v49.i3.30.
[53] Y.N. Onopchuk, Al.A. Chikrii. The analytical method to solve nonsta-
tionary differential games of pursuit. Cybernetics and Systems Analy-
sis, 49(4):603–615, 2013. https://doi.org/10.1007/s10559-013-9547-7.
7
Adaptive Method for the Variational
Inequality Problem Over the Set of
Solutions of the Equilibrium Problem
Abstract
In this chapter we consider a two-level variational problem: a variational
inequality problem over the set of solutions of the equilibrium problem. An
example of such a problem is the search for the normal Nash equilibrium.
To solve this problem, we propose a new iterative proximal algorithm that
combines the ideas of a two-stage proximal point method, adaptability, and
scheme of iterative regularization. In contrast to the previously applied rules
for choosing the step size, the proposed algorithm does not calculate any
values of the bifunction at additional points; it does not require knowledge
of information about the Lipschitz constants of the bifunction, the Lipschitz
constant and strong monotonicity constant of the operator. For monotone Lip-
schitzian bifunctions and strongly monotone Lipschitz continuous operators,
a theorem on the strong convergence of the algorithm is proved. It is shown
that the proposed algorithm is applicable to monotone two-level variational
inequalities in real Hilbert spaces.
145
146 Adaptive Method for the Variational Inequality Problem Over the Set
7.1 Introduction
In numerical optimization, and the theory of ill-posed (non-correct) problems,
the following technique for solving problems with a non-unique solution
is widespread [1–4]. The problem is associated with a family of perturbed
auxiliary problems that are uniquely and correctly (well-posed) solvable.
A particular solution to the original problem is obtained as the limit of
solutions to perturbed auxiliary problems with a decrease in perturbations.
The solutions found in this way satisfy certain additional conditions. For
example, the minimality of the norm of the normal solution to an optimization
problem obtained by the Tikhonov regularization algorithm. On the other
hand, in the operation research, optimization problems arise according to
sequentially specified criteria (lexicographic, sequential, or multi-stage opti-
mization) [5, 6]. Note that the ideas of the Tikhonov regularization method
permeate many studies in the field of automatic control [7–10].
Two-level variational inequalities arose as a natural generalization of well
studied lexicographic optimization problems with two criteria, as well as in
the analysis of ordinary optimization problems with functional constraints
in the form of a monotone variational inequality. As an independent math-
ematical object, two-level variational inequalities began to be considered in
[11]. The papers [12, 13] are devoted to the solvability of more general n-
level variational inequalities. By now, approximate methods of one-stage for
solution of all these problems are known, which are ideologically close to the
methods of penalty and regularization [6, 14].
One of the popular directions of contemporary applied nonlinear analysis
is the study of equilibrium problems (equilibrium programming problems)
[15–23], to which mathematical programming problems, variational inequal-
ities and many game theory problems can be reduced. Recently, algorithms
for solving variational inequalities and equilibrium problems have become
popular in the ML community. On their basis it is possible to design effective
learning methods for GANs [24–26].
The Chapter deals with a two-level problem: variational inequality prob-
lem on the set of solutions to the equilibrium programming problem. A
similar problem was considered in [22], where a strongly converging algo-
rithm was proposed that used the operation of calculating the value of the
resolvent of a bifunction. The latter significantly increased the complexity
of the algorithm. Below, to solve this problem, an adaptive version of the
previously studied algorithm [23] based on the two-stage proximal point
method [20, 21] and the idea of scheme of iterative regularization [1, 27]
7.2 Problem Formulation 147
will be proposed. The proposed algorithm does not calculate the values of
the bifunction at any additional points; it does not require knowledge of
information about the Lipschitzian constants of the bifunction, the constants
of Lipschitz and strong monotonicity of the operator.
For monotone bifunctions of Lipschitzian type and strongly monotone
Lipschitz continuous operators, a theorem on the strong convergence of the
algorithm is proved. It is shown that the proposed algorithm is applicable to
monotone two-level variational inequalities in real Hilbert spaces.
The results obtained are useful in the theory of optimal control. Indeed,
the problem of optimal control of the system can be formulated as a vari-
ational inequality problem. In the case of non-uniqueness of the optimal
control, an additional criterion can be introduced for the regularization of
the problem and a two-level variational in equality can be obtained. And to
solve the obtained problem, we use the algorithm proposed and studied in this
Chapter.
C6) EP (G, C) = ∅;
C7) F : C → H–μ-strongly monotone and L-Lipschitz continuous operator.
Remark 7.1. Condition C5 is provided by Giandomenico Mastroeni in [16].
For example, bifunction
G (x, y) = (T x, y − x)
x2 → min, x ∈ C : (T x, y − x) ≥ 0 ∀ y ∈ C.
ϕ (η) − ϕ (z) + (z − x, η − z) ≥ 0
Remark 7.3. For solving incorrect extremum problems that type of approxi-
mation was proposed by Tikhonov for building regularization methods. Later
Felix Browder [2, 3] applied similar regularization technique.
From results of [17] follows the existence and uniqueness of solution
xε ∈ C of problem (7.2) for all ε > 0. Auxiliary elements xε ∈ C have
several properties important for subsequent constructions.
Lemma 7.1 ([23]). The following inequalities take place:
1) xε ≤ μ1 1 + μ1 L F x̄ ∀ε > 0;
2) xε − xδ ≤ |ε − δ| εμ
1
1 + L μ1 F x̄ ∀ε, δ > 0.
Proof. Let ε > 0. For xε ∈ C and an arbitrary element w ∈ EP (G, C), we
have
(δ − ε) (F xδ , xε − xδ ) ≥ ε (F xε − F xδ , xε − xδ ) .
150 Adaptive Method for the Variational Inequality Problem Over the Set
|δ − ε| F xδ xε − xδ ≥ εμ xε − xδ 2 ,
that is,
|δ − ε|
xε − xδ ≤ F xδ .
εμ
Let us estimate from above the norm of F xδ
L
F xδ ≤ F x̄ + F xδ − F x̄ ≤ F x̄ + L xδ − x̄ ≤ 1+ F x̄ .
μ
Using the last inequality in the estimate for xε − xδ , we obtain 2).
Lemma 7.2 ([23]). Assume that conditions C1–C4 and C6, C7 are satisfied.
Then
lim xε − x̄ = 0.
ε→0
Proof. By first inequality of Lemma 7.1, from {xε }ε>0 we can extract a
sequence (xεn ) (εn → 0) that weakly converges to an element v ∈ C. Using
the weak upper semicontinuity of the functional G (·, y), we pass to the limit
in the inequality
We get
G (v, y) ≥ 0 ∀y ∈ C,
that is, v ∈ EP (G, C). And passing to the limit in inequality
we obtain
(F w, w − v) ≥ 0 ∀w ∈ EP (G, C) ,
that is, v = x̄. Let us show that
From the uniqueness of the element x̄ ∈ H we obtain limε→0 xε − x̄ = 0.
7.4 Algorithm 151
7.4 Algorithm
Following the well-known scheme of two-step proximal point method [20,
21] for solving problem (7.1) in recent work [23] the following method was
introduced: ⎧
⎨zn = xn − αn λn F xn ,
yn = proxλn G(yn−1 ,·) zn , (7.3)
⎩
xn+1 = proxλn G(yn ,·) zn ,
8 9
where λn were defined considering λn ∈ λ, λ̄ ⊆ 0, 2(2a+b) 1
and positive
sequence (αn ) was such that
∞
! αn+1 − αn
lim αn = 0, αn = +∞, lim = 0.
n→∞ n→∞ αn2
n=1
it is enough to show that the generated by Algorithm 7.1 sequence (xn ) has
the next property
lim xn − xαn = 0.
n→∞
Let us remind the well-known elementary lemma about recurrent numerical
inequalities.
Lemma 7.3 ([1]). Let the non-negative numerical sequence (ψn ) satisfies the
recurrent inequality
where numerical sequences (αn ) and (βn ) have the following conditions
αn ∈ (0, 1), αn = +∞, lim αβnn ≤ 0. Then lim ψn = 0.
n→∞ n→∞
Let us prove an important inequality for generated sequences (xn ), (yn )
and auxiliary elements xαn .
7.5 Proof of Algorithm Convergence 153
Lemma 7.4. For sequences generated by Algorithm 7.1 (xn ), (yn ) and
auxiliary elements xαn the inequality holds
2 2 λn
xn+1 − xαn ≤ (1 − αn λn μ) xn − xαn − 1 − τ xn+1 − yn 2
λn+1
λn 2 −1
− 1 − 2τ − α n λn L μ yn − xn 2
λn+1
λn
+ 2τ xn − yn−1 2 . (7.4)
λn+1
Proof. We have
xn+1 − xαn 2 = xn − xαn 2 − xn − xn+1 2
+ 2 (xn+1 − xn , xn+1 − xαn ) = xn − xαn 2 − xn − yn 2
−yn − xn+1 2 − 2 (xn − yn , yn − xn+1 )
+ 2 (xn+1 − xn , xn+1 − xαn ) . (7.5)
From the definition of points xn+1 and yn it follows that
λn G (yn , xαn ) − λn G (yn , xn+1 )
≥ (xn+1 − xn + αn λn F xn , xn+1 − xαn ) , (7.6)
λn G (yn−1 , xn+1 ) − λn G (yn−1 , yn )
≥ − (xn − αn λn F xn − yn , yn − xn+1 ) . (7.7)
Using inequalities (7.6), (7.7) for estimating inner products in (7.5),
we get
xn+1 − xαn 2 ≤ xn − xαn 2 − xn − yn 2 − yn − xn+1 2
+ 2λn (G (yn , xαn ) − G (yn , xn+1 ) + G (yn−1 , xn+1 ) − G (yn−1 , yn ))
+ 2αn λn (F xn , xαn − yn ) . (7.8)
From the rule of calculation of λn+1 it follows the inequality
G (yn−1 , xn+1 ) − G (yn−1 , yn ) − G (yn , xn+1 )
1
≤τ yn−1 − yn 2 + xn+1 − yn 2 . (7.9)
2λn+1
For estimating an expression G (yn−1 , xn+1 ) − G (yn−1 , yn ) −
G (yn , xn+1 ) in (7.8) we use (7.9). We obtain
xn+1 − xαn 2 ≤ xn − xαn 2 − xn − yn 2 − yn − xn+1 2
154 Adaptive Method for the Variational Inequality Problem Over the Set
λn λn
+τ yn−1 − yn 2 + τ xn+1 − yn 2
λn+1 λn+1
+2λn G (yn , xαn ) + 2αn λn (F xn , xαn − yn ) .
We estimate yn−1 − yn 2 in the following way
We have
Since
G (xαn , yn ) + αn (F xαn , yn − xαn ) ≥ 0
then
G (yn , xαn ) ≤ αn (F xαn , yn − xαn ) .
Taking into account the last estimation in (7.10) we obtain
× yn − xn 2 + 2τ λn λ−1 2
n+1 xn − yn−1 ,
Q.E.D.
Let us prove the estimation from which it follows the convergence to 0 of
sequences (xn − xαn ) and (xn − yn−1 ).
Lemma 7.5. For sequences (xn ), (yn ) generated by Algorithm 7.1 and
auxiliary elements xαn for large n the inequality is hold
2 2τ λn+1 λ−1
xn+1 − xα + n+2
xn+1 − yn 2
n+1
1 − αn+1 λn+1 μ
1 2
αn λ n μ 2 2τ λn λ−1
n+1 2
≤ 1− xn − xαn + xn − yn−1
2 1 − α n λn μ
2M (αn+1 − αn )2
+ , (7.13)
λn μ αn3
) *
where M = μ−1 1 + Lμ−1 F x̄.
Proof. We have
2 2
xn+1 − xαn 2 = xn+1 − xαn+1 + xαn+1 − xαn
) * 2 2
+ 2 xn+1 − xαn+1 , xαn+1 − xαn ≥ xn+1 − xαn+1 + xαn+1 − xαn
2
− 2 xn+1 − xαn+1 xαn+1 − xαn ≥ (1 − ε) xn+1 − xαn+1
ε−1 xα
2
,
+ n+1 − xαn (7.14)
ε
where ε > 0. Let in (7.14) ε = 12 αn λn μ. We get
2 − α n λn μ
xn+1 − xα 2
xn+1 − xαn 2 ≥ n+1
2
2 − α n λn μ
xα
2
.
− n+1 − xαn (7.15)
α n λn μ
156 Adaptive Method for the Variational Inequality Problem Over the Set
2M (αn+1 − αn )2
+ . (7.18)
λn μ αn3
) *
Since τ ∈ 0, 13 there exists lim λn > 0 and lim αn = 0, then starting
n→∞ n→∞
from some number n1 the inequalities
1 − 2τ λλn − αn λn L2 μ−1
n+1
> 0,
2 − α n λn μ
1 − τ λλn 2τ λλn+1
n+1
− n+2
> 0,
1 − 12 αn λn μ 1 − αn+1 λn+1 μ
2 − 2αn λn μ α n λn μ
<1− ,
2 − α n λn μ 2
will be satisfied. So for all n ≥ N = max {n0 , n1 } from (7.18) it follows
2 2τ λn+1 λ−1
xn+1 − xα + n+2
xn+1 − yn 2
n+1
1 − αn+1 λn+1 μ
1 2
αn λ n μ 2 2τ λn λ−1
n+1 2
≤ 1− xn − xαn + xn − yn−1
2 1 − α n λn μ
2M (αn+1 − αn )2
+ ,
λn μ αn3
Q.E.D.
Let us formulate the main result of Chapter.
Theorem 7.1. Let the conditions C1–C7 and D1–D3 are satisfied. Then for
two sequences (xn ), (yn ) generated by iterative Algorithm 7.1 takes place
From inequality
xn − x̄ ≤ xn − xαn + xαn − x̄ ,
Lemma 7.2 and (7.20) we obtain (7.19).
Remark 7.5. Easy to see that sequence (zn ) (zn = xn − αn λn F xn ) also
converges strongly to point x̄ ∈ C.
where
Cn = {z ∈ H : (zn − λn F1 yn−1 − yn , z − yn ) ≤ 0} .
7.7 Conclusion
In this chapter we consider a two-level variational problem: a variational
inequality problem over the set of solutions of the equilibrium problem
[22, 23]. To solve this two-level problem, an iterative proximal method is pro-
posed that combines the ideas of a two-stage proximal point method [20, 21],
adaptation and scheme of iterative regularization [1, 27]. The algorithm has
the structure:
⎧
⎪
⎪ zn = xn − αn λn Fxn ,
⎨
yn = arg miny∈C λn G (yn−1 , y) + 12 y − zn 2 ,
⎪
⎪
⎩ xn+1 = arg miny∈C λn G (yn , y) + 1 y − zn 2 ,
2
160 Adaptive Method for Variational Inequality Problem Over the Set
Acknowlegdments
This work was supported by Ministry of Education and Science of Ukraine
(project “Mathematical modelling and optimization of dynamical systems for
defense, medicine and ecology”, 0119U100337) and the National Academy
of Sciences of Ukraine (project “New methods of research of correctness and
solution search for discrete optimization problems, variational inequalities
and their applications”, 0119U101608).
References
[1] A.B. Bakushinskii, A.V. Goncharskii, Iterative Methods for Solving Ill-
Posed Problems, Moscow: Nauka, 1989.
[2] F. Browder, “Existence and approximation of solutions of nonlinear
variational inequalities,” Proc. Nat. Acad. Sci. USA, vol. 56, no. 4,
pp. 1080–1086, 1966.
References 161
[27] L.D. Popov, “On schemes for the formation of a master sequence in a
regularized extragradient method for solving variational inequalities,”
Russian Mathematics, vol. 48, iss. 1, pp. 67–76, 2004.
[28] H.H. Bauschke, P.L. Combettes, Convex Analysis and Monotone Oper-
ator Theory in Hilbert Spaces. Berlin, Heidelberg, New York: Springer,
2011.
[29] Y.I. Vedel, S.V. Denisov, V.V. Semenov, “An adaptive algorithm for the
variational inequality over the set of solutions of the equilibrium prob-
lem,” Cybernetics and Systems Analysis, vol. 57, issue 1, pp. 91–100,
2021.
Part II
Advances in Control Systems Application
8
Identification of Complex Systems in the
Class of Linear Regression Models
Abstract
Most methods of analysis and synthesis of control assume the known mathe-
matical model of the control object. To obtain the desired control result, the
model must accurately describe the real processes [1]. It is often not possible
to obtain quantitative relations between the parameters and variables of the
system under consideration on the basis of theoretical laws. For many causal
systems, these relations can only be assessed experimentally, during identifi-
cation. The most developed approach is the so-called stochastic identification,
which consists in obtaining unbiased estimates of the parameters for certain
model. However, the possibilities of such identification are significantly
limited in the case of complex systems. This chapter discusses a more general
approach to identification within the regression model class, which searches
for approximate solutions that are consistent in accuracy with errors in the
data. Since identification problems in complex systems are often incorrectly
posed, the proposed method includes regularization that provides practically
suitable solutions.
167
168 Identification of Complex Systems in the Class
8.1 Introduction
System identification, i.e. the problem of constructing its mathematical model
from experimental data has a long history. It is conducted by scientists who
study the behavior of various systems, as well as control specialists who
need mathematical models to solve problems of analysis and synthesis. Many
results on this problem within the framework of a discrete description have
been published in well-known books [2, 3], including monographs [4, 5].
The most interesting problems from a practical point of view are associ-
ated with uncertainty in the output and input variables. The corresponding
type of identification problem is called “errors in variables” (EIV) and is
considered by many scientists [6–9]. Recent publications can be found in
[10–13]. The main focus there was on obtaining unbiased parameter estimates
and the issue of consistency.
Different approaches were considered within EIV identification, for
example, for describing an uncertain dynamic system it was proposed to use a
graph metric introduced in [14] with the representation of model uncertainties
by the graph subspace consisting of the input and output spaces of the system.
Moreover, it was assumed everywhere that the errors are independent and
identically distributed (i.i.d.) measurement noises. Recently graph subspace
approach was proposed and developed in [15] for EIV identification problem
and a new estimation algorithm was derived that is more general than total
last square algorithm used in preceding works. It is aimed at estimation of the
model parameters based on finite measurement samples of the graph subspace
of a given system.
This chapter explores a different approach to address the EIV identifica-
tion problem with a different interpretation of the uncertainty in the data. We
assume that the errors are randomly distributed but limited in magnitude. As
in [15], the data series can be large, but always finite. This initial information
about the system is more realistic than in the stochastic case.
The approach proposed here is focused on the identification of controlled
systems and allows the use of a fairly arbitrary input signal. This allows data
collection in active experiments with specially designed informative input
signals. The problems of identifying systems of large dimension, including
infinite ones, based on inaccurate data, are usually ill-posed. Obtaining unbi-
ased estimates of the model parameters in these cases becomes impossible.
Moreover, a regularized solution in ill-posed case can give an approximate
model of a lower dimension than that of the original system. Thus, the
proposed identification method involves determining the size of the model
that guarantees stability, followed by parameter estimation.
8.2 Active Experiments and Informative Data 169
The resulting models were tested when simulating the identification pro-
cess. This chapter is actually developing an approach and results obtained
earlier in [16, 17].
where ξu (t) and ξy (t) are errors in data. According to the non-stochastic
approach, we assume that ξu (t) and ξy (t) are unknown arbitrary random
sequences satisfying condition
where εu and εy are sufficiently small compared to the magnitude of u(t) and
y(t).
The kind of input signal affects the information contained in the observa-
tions. The most important characteristic of an input is its ability to excite all
modes of the system. Insufficiently excited modes cannot be identified.
For stochastic identification, the persistently exiting input sequence of a
given order [18] is used. An ergodic sequence u(t) is persistently exiting of
order l if and only if
1
rank lim UlN · UlN = l,
T
(8.3)
N →∞ N
where
⎡ ⎤
u(t) u(t + 1) . . . u(t + N − 1)
⎢ u(t + 1) u(t + 2) . . . u(t + N ) ⎥
⎢ ⎥
UlN = ⎢ .. .. .. .. ⎥.
⎣ . . . . ⎦
u(t + l − 1) u(t + l) . . . u(t + N + l − 2)
When u(t) is a zero-mean discrete white noise sequence it has the statistical
property
1
UlN · UlN = σu2 · Il + ε2N · EN , (8.4)
N
8.2 Active Experiments and Informative Data 171
!
t−1 !
P
y(t) = h0p (t − j)u(j), (8.6)
j=0 p=1
172 Identification of Complex Systems in the Class
where h0p (k) = ρkp [fpc cos ωp (k) + fps sin ωp (k)], and
ρp = |λp | , ωp = arg(λp ), fpc = ccp bcp + csp bsp , fps = ccp bsp − csp bcp .
Ỹrelax = U ΣV T , (8.11)
where · F means Frobenius norm. The value σn̂+1 can be used to judge
the proximity of the matrix Ỹrelax to the matrix of rank n̂. Sometimes, the
true dimension n can be determined by the rate of decrease of the singular
values. This can be done if there is a pronounced gap between the values
of σi and σi+1 . However, there is often no clear gap. It all depends on
8.3 Method of Identification 175
the signal-to-noise ratio of individual modes (8.5). Modes with a low SNR
becomes indistinguishable from background noise, making impossible in the
non-stochastic case to establish the exact dimension of the system. In such
cases, we can find a model of reduced dimension. For this, the principle of
regularization is used. The required approximate solution must be consistent
in accuracy with the uncertainty of the data.
Consider how the quantities σn̂+1 and εy are related. Let’s represent the
matrix Ỹrelax in the form of the exact and noise part, omitting here the “relax”
index
Ỹ = Y + ΔY. (8.15)
Inequality (8.2) gives ΔY max < εY . Let us estimate from above the
quantity σn̂+1 . From (8.14) we have σn̂+1 = Y N (n̂) = Ỹ − Y S (n̂) .
F F
Substituting (8.15) and using the triangle inequality we derive
σn̂+1 ≤ Y − Y S (n̂)F + ΔY F .
Since the quantity Y − Y S (n̂)F can not be larger then Y − Ỹ , we
F
have
σn̂+1 ≤ 2 ΔY F .
N M 2
From norm equivalence AF = j ai,j , Amax =
& & i
& &
max &ai,j &, we have estimation
i = 1, M
j = 1, N
√
σn̂+1 ≤ 2 Rl · εY . (8.16)
Relation (8.16) means that the modes of the system, starting with such n̂,
cannot be guaranteed to be identified. The possibility of their identification is
determined only by the case of noise realization. This property will be used to
determine the dimension of the model, namely the dimension of the vector a.
In certain cases, by reducing Δ0 and Δ1 , and, consequently, the magni-
tude of R, it is possible to achieve an approximation to the exact dimension.
Everything will depend on the dynamic properties of the system, determined
by its expansion invariants (8.5). Low SNR values appear for modes with
close ωp or modes with small fpc and fps . There exist those εy , starting from
which their signals are not distinguishable against the background of noise.
The presence of such modes in the system will lead to the reduction of the
model order in the proposed approach.
176 Identification of Complex Systems in the Class
Then the basic equation for finding the vector b will be the following equation
˜ = b0 u(t) + b1 u(t − 1) + . . . + bm u(t), d = 0
ỹ(t)
˜ = b1 u(t − 1) + . . . + bm u(t), , (8.20)
ỹ(t) d=0
for any moment of time t where we have nonzero excitation signals. From
the data on the intervals of excitation of a suitable duration, we form an
overdetermined system
˜
W · b = ỹ, (8.21)
where b is a vector whose components are b0 , b1 , . . . , bm or
b1 , b2 , . . . , bm , the vector ỹ˜ is composed of ỹ(t)
˜ for the time moments
corresponding to (8.20), and W is a matrix consisting of the input signals for
the appropriate time moments. In (8.21), both the matrix W and vector ỹ(t) ˜
are given approximately. Taking into account the specified errors, we will
find an approximate solution of (8.21), using the generalized LS. In this case,
you should make sure that the problem being solved is correctly posed. If it
is ill-conditioned, a regularized solution should be found in accordance with
the well-known methods for solving incorrect linear systems. In this case, it
is possible to obtain stable approximate solutions, like with a vector a, by
lowering the dimension m of the vector b.
Bad conditionality of the matrix W indicates that the dimension of the
vector b is less than n. In such cases, the further actions are similar to those
described in Section 2.1, when determining the dimension of the vector a. To
do this, we carry out the SVD decomposition of the matrix W , and check for
singular values. After establishing the dimension of the vector b its value is
found from the redefined system (8.21).
Solving the system of equations (8.21) by generalized LS completes the
system identification, resulting in an approximate model consistent with the
EIV.
where a(n) and b(n) are the values of the corresponding vectors for the
dimension n. Various behaviors of the discrete function defined by (8.22)
are possible. At first, it decreases, and then begins to grow, while local
minima are possible. Then the last local minimum determines the quasi-
optimal dimension. It is possible when this function grows continuously,
starting from the smallest n. It is also possible a noticeable scatter in the
behavior of this function, starting from some point. Then the value n before
the scatter determines the desired dimension.
The parameters a and b corresponding to the quasi-optimal n determine a
regularized approximate solution to the identification problem.
λp for all modes, ensuring stability, slow decay and sufficient distance from
each other. We also set invariants of |fpc | and |fps | to be equal for all modes, to
provide better observability. The definition of the system in the form (8.5) in
canonical observable representation (ccp = 1, csp = 0) makes it easy to pass
to equivalent description in a normal observable realization in the state space
⎡ ⎤
0 0 . . . 0 −an ⎡ ⎤ ⎡ ⎤
⎢ 1 0 . . . 0 −an−1 ⎥ bn 0
⎢ ⎥ ⎢ bn−1 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
A = ⎢ 0 1 . . . 0 −an−2 ⎥ , b = ⎢ . ⎥ , c = ⎢ .. ⎥ .
⎢ .. .. . . .. .. ⎥ ⎣ .. ⎦ ⎣ . ⎦
⎣ . . . . . ⎦
b1 1
0 0 . . . 1 −a1
where k1 = 100, k2 = 157, l = 1000. The noisy outputs are stored line
by line in the matrix Ỹ , the inputs are stored line by line in the matrix Ũ .
2. From Ỹ on the relaxation, a matrix Ỹrelax ∈ RR×l is formed. R =
k1 + k 2 .
3. Using the SVD, dimension n is determined.
4. S
Matrix Ỹrelax is splitted by n into signal Yrelax N
(n) and noise Yrelax (n)
parts.
5. A linear system for the regression coefficients a is formed from
S
Yrelax (n). Equation
included if its coefficients exceed the SNR thresh-
old KN S · ε · Ỹ .
max
6. The regression coefficients of free motion are found.
7. We define forced movement Ỹf orced by subtracting free movement from
Ỹ . ' (
8. (optional). We build compound matrix Ỹf orced |Ũ and apply SVD.
If it is not possible to determine dimension m by 'singular values, (
then we accept m = n. Extract signal part of Ỹf orced |Ũ =
' (S ' (N
Ỹf orced |Ũ (m̂) + Ỹf orced |Ũ (m̂), and then break it back:
' (S ' (
Ỹf orced |Ũ = ỸfSorced |Ũ S .
9. From the matrices Ỹf orced and Ũ (or ỸfSorced and Ũ S ) we form a system
for the control coefficients b of regression. The system includes all
equations whose coefficients exceed the threshold value, similarly to
item 5.
10. By singular values we check whether the chosen dimension m̂ is true. If
not, we rebuild the system for the specified dimension.
11. We solve the system and find the regression coefficients for control.
Then we check regression model:
1. Check if the resulting model is stable.
2. Calculate the deviation of the model output from the real system when
testing with a rectangular pulse followed by relaxation.
8.4 Simulation Results 181
where x0 is the state vector at the first relaxation step. Therefore, the best
conditions for identifying the mode correspond to maximization x2 for any
implementation of A.
Conclusion: the magnitude of the mode excitation will be defined as the norm
of the state vector.
182 Identification of Complex Systems in the Class
Figure 8.1 State of a complex mode under harmonic excitation with different initial phases
Let us consider the behavior of the state vector of a complex mode upon
excitation by a harmonic signal with a different initial phase. Figure 8.1 shows
trajectory (xc , xs ) of mode 0.95e0.5i for excitations u (t) = sin (0.2t) (left),
and u (t) = cos (0.2t) (right). In both cases, the limiting trajectories coincide.
Conclusion: the steady-state forced motion of the complex mode does not
depend on the initial phase of the harmonic signal.
Consider the excitation of a complex mode by harmonic signals of differ-
ent frequencies. Figure 8.2 shows trajectory (xc , xs ) of mode 0.95e0.5i for
excitations u (t) = sin (0.5t) (left), and u (t) = cos (0.8t) (right). The first
case is resonant.
The steady-state forced motion of the complex mode under harmonic
excitation has trajectory of the form of an ellipse and depends on the ratio
of the natural frequency and the input frequency. In the resonant case, the
trajectory is close to circular.
Let us consider the trajectories of free motion after the excitation of
a mode by a harmonic signal. We use harmonic excitation and turn off
control at times corresponding to different points of the elliptical trajectory.
Figure 8.3 shows trajectory (xc , xs ) of mode 0.95e0.5i for input u (t) =
−→
sin (0.2t) , t = 1, τ
for τ = 196 (solid) and τ = 204 (dashed).
0, t>τ
Conclusion: in the case of nonresonant harmonic excitation of the complex
mode, the value of the relaxation signal can significantly depend on the initial
moment.
8.4 Simulation Results 183
Figure 8.2 State of a complex mode under harmonic excitation of different frequencies
Figure 8.3 State of a complex mode during relaxation after harmonic excitation
Figure 8.4 State of a complex mode during constant excitation and relaxation
Figure 8.5 The norm of the state vector of modes under harmonic excitation
Let us consider the change in the norm of the state vector of various
modes.
6 We will take a system of7 two real and two complex modes: {λ} =
0.5, 0.9, 0.95e0.5i , 0.95e1.2i . Figure 8.5 shows dependency xp and
|xp | on time for input u (t) = sin (0.2t).
Excitation of the system by a harmonic input signal leads all modes to
forced oscillations with the same frequency, but out of phase.
Figure 8.6 shows dependency xp and |xp | on time for input u (t) ≡ 1.
8.4 Simulation Results 185
Figure 8.6 The norm of the state vector of modes under constant excitation
Figure 8.7 The norm of the state vector of modes under rectangular pulse and on relaxation
186 Identification of Complex Systems in the Class
Figure 8.8 The maximum of norm of the modes state vectors under harmonical input
Table 8.1 Maximum and average errors of models for εU = 10−4 and εY = 10−4
Algorithm configuration Error in a Error in b Error by output
A 0.02263 (0.01980) 0.19602 (0.16965) 0.01854 (0.01625)
B 0.00281 (0.00091) 0.02635 (0.00853) 0.00234 (0.00074)
C 0.00281 (0.00091) 0.03752 (0.02049) 0.00332 (0.00225)
Table 8.2 Maximum and average errors of models for εU = 10−6 and εY = 10−4
Algorithm configuration Error in a Error in b Error by output
A 0.02246 (0.01985) 0.19539 (0.17013) 0.01835 (0.01631)
B 0.00283 (0.00092) 0.02718 (0.00820) 0.00219 (0.00074)
C 0.00283 (0.00092) 0.02239 (0.00677) 0.00216 (0.00068)
S
The extraction of a matrix Yrelax (n) of a given rank from the matrix
Ỹrelax can significantly improve the quality of the resulting model. The
selection of a matrices ỸfSorced and Ũ S from matrices Ỹf orced and Ũ has little
effect on the quality of the resulting model. When εU is relatively small,
we can achieve some improvement. In other case we risk worsening the
result.
Figure 8.9 Average number of rows in overdetermined systems for a and b (left), the quality
of the resulting model for different implementations of random noise (right)
Figure 8.9 shows the number of not discarded equations (left) and model
quality (right) for varied coefficient KN S .
The right figure shows the dependencies for four different variants of
the implementation of random measurement noise. In one of these cases,
dropping a large number of equations did not improve the result. An increase
in the threshold SNR coefficient can significantly reduce the number of lines
in the systems of equations and, in most cases, slightly increase the accuracy
of the model.
0.2, 0.5, 0.85, 0.9 0.00682 (0.00227) 0.00629 (0.00208) 0.00363 (0.00117)
0.2, 0.5, 0.70, 0.9 0.01780 (0.00495) 0.01673 (0.00474) 0.00425 (0.00132)
0.2, 0.5, 0.89, 0.9 0.07777 (0.05265) 0.07756 (0.05248) 0.13216 (0.09800)
Figure 8.10 First singular values of the matrix Yrelax for systems of different order
Figure 8.11 First singular values of the matrix Ỹrelax for systems of different order. Case of
relative error bound εY = 10−3
√
σ +1
Let us put that Y max ≈ Ỹrelax , then 2Pσ 1
≤ 2 Rl · εY . Therefore,
max
let’s build a graph of the singular values normalized to the first of them.
Figure 8.12 shows normalized singular values for εY = 10−3 and a threshold.
8.4 Simulation Results 191
Figure 8.12 First normalized singular values of the matrix Ỹrelax for systems of different
order. Case of relative error bound εY = 10−3
√
As we can see, the guaranteed estimate 2 Rl · εY is overestimated due
to the ratio of matrix norms · F and · max . To get rid of this we
check the possibility of determining the dimension directly in max norm.
Yrelax
N (n̂)max
Figure 8.13 shows values σ̃n̂+1 = for εY = 10−3 and a
Ỹrelax max
threshold 2εY .
N
Figure 8.13 Relative max norm of the matrix Yrelax
192 Identification of Complex Systems in the Class
The threshold 2εY is much more accurate, but values σ̃n̂+1 are not
monotonically decreasing. As we can see, this method can also be used to
determine the dimension.
Thanks to a good selection of input signals, it is possible to accurately
determine the order of the system up to high dimensions.
Figure 8.14 Accuracy of the model, depending on the dimension and the noise
8.5 Conclusion
In the classical statistical method, it is assumed that the implemented errors
have the specified statistical properties, which ensure the consistency of the
estimation, i.e. convergence to unbiased estimates of system parameters. In
case of non-stochastic identification, arbitrary random errors in the data,
including unfavorable ones, are allowed. In these cases, in order to ensure
the quality of the models, during active experiments, it is necessary to pay
serious attention to the information content of the system input. In this
chapter, this issue is focused on. It is substantiated how the experiment should
be planned so that the SNR for each of the modes of the system was the
greatest. Numerical modeling has been carried out, which makes it possi-
ble to evaluate the effectiveness of the proposed approach to identification
and interactively construct a mathematical model of the system, taking into
account its structural features and experimental capabilities, on which the
quality of the resulting models significantly depends.
194 Identification of Complex Systems in the Class
References
[1] Yuriy P. Kondratenko, Arkadii A. Chikri, Vyacheslav F. Gubarev, Janusz
Kacprzyk (Eds.), ‘Advanced Control Techniques in Complex Engi-
neering Systems: Theory and Applications’, Dedicated to Professor
Vsevolod M. Kuntsevich, Studies in Systems, Decision and Control, vol.
203, Springer Nature Switzerland AG, 337 pp., 2019.
[2] T.Söderström, & P.Stocia, ‘System identification’ U.K.: Prentice-Hall
International, Hemel Hempstead, 612 pp., 1989.
[3] L.Ljung, ‘System Identification: Theory for the user’, Prentice-Hall,
1999.
[4] T. Söderström, ‘Errors-in-variables methods in system identification’,
Springer, 2018.
[5] Van Huffel, S., & Lemmerling, P. (Eds.), ‘Total least square and errors-
in-variables modeling: Analysis, algorithms and applications’, Springer,
2002.
[6] S. Beghelli, R.P. Guidorzi, U. Soverini, ‘The frisch scheme in dynamic
system identification’, Automatica, vol. 26, pp. 171–176, 1990.
[7] M. Hong, T. Söderström, W.X. Zheng, ‘Accuracy analysis of bias-
eliminating least squares estimates for errors-in-variables systems’,
Automatica, vol. 43, pp. 1590–1596, 2007.
[8] T. Söderström, ‘Accuracy analysis of the Frisch scheme for identifying
errors-in-variables systems’, IEEE Transactions on Automatic Control,
vol. 52, no 6, pp. 985–997, 2007.
[9] T. Söderström, ‘Errors-in-variables methods in system identification’,
Automatica, vol. 43, pp. 939–958, 2007.
[10] T. Söderström, U. Soverini, ‘Errors-in-variables identification using
maximum likelihood estimation in the frequency domain’, Automatica,
vol. 79, pp. 131–143, 2017.
[11] E. Zhang, R. Pintelon, ‘Identification of multivariable dynamic
errors-in-variables system with arbitrary inputs’, Automatica, vol. 82,
pp. 69–78, 2017.
[12] E. Zhang, R. Pintelon, ‘Nonparametric identification of linear dynamic
errors-in-variables systems’, Automatica, vol. 94, pp. 416–425, 2018.
[13] Zheng, W.X. & Feng, C.B., ‘Unbiased parameter estimation of linear
systems in presence of input and output noise’, International Journal of
Adaptive Control and Signal Processing, vol. 3, pp. 231–251, 1989.
[14] Georgiou, T.T., & Smith, M.C., ‘Optimal robustness in the gap metric’,
IEEE Transactions on Automatic Control, 35, pp. 673–686, 1990.
References 195
[15] Hyundeok Kang, Guoxiang Gu, Wei Xing Zheng, ‘A graph subspace
approach to system identification based on errors-in-variables system
models’, Automatica, Volume 109, 2019, 108535.
[16] Kuntsevich, V.M., Gubarev, V.F., Kondratenko, Y.P., Lebedev, D.V.,
Lysenko, V.P. (Eds), ‘Control Systems: Theory and Applications’. River
Publishers Series in Automation, Control and Robotics, 329 pp., 2018.
[17] Kondratenko, Y.P., Kuntsevich, V.M., Chikrii, A.A., Gubarev, V.F (Eds),
‘Advanced Control Systems: Theory and Applications’, River Publish-
ers Series in Automation, Control and Robotics, 441 p., 2021.
[18] Verhaegen M. and Dewilde P, ‘Subspace model identification. Part 1:
The output-error state space model identification class of algorithms’,
International Journal of Control, Vol. 56, No 5, pp. 1187–1210, 1992.
[19] V.F. Gubarev, ‘Modeling and Identification of Complex Systems’,
Naukova Dumka, 246 p. 2019 (in Ukrainian).
[20] G.H. Golub, Ch.F. Loan, ‘Matrix Computations’, The John Hopkins
University Press, Baltimore, 780 p., 2013.
[21] A.N. Tikhonov and V.Y Arsenin, ‘Solution of ill-posed problems’, V.H.
Winston & Sons, Washington, D.C.: John Wiley & Sons, New York, 258
p., 1977.
9
Fuzzy Systems Design: Optimal Selection
of Linguistic Terms Number
Rzeszow, Poland
E-mail: [email protected]; [email protected];
[email protected]; [email protected]
Abstract
This chapter considers developing and researching advanced information
technology for fuzzy systems (FSs) design and structural optimization with
optimal linguistic terms (LTs) numbers. The application of the proposed
information technology allows increasing the FS accuracy and performance,
reducing the computational costs spent on the generation of rule base (RB)
and optimization of parameters, and simplifying its software and hardware
implementation. The effectiveness study of the developed information tech-
nology is carried out on a specific example, particularly when designing and
optimizing the fuzzy control system of a quadrotor unmanned aerial vehicle
(QUAV). The research results obtained confirm the high effectiveness of the
proposed information technology and the feasibility of its use in creating
various types of fuzzy decision-making and control systems.
197
198 Fuzzy Systems Design: Optimal Selection
9.1 Introduction
In today’s world, artificial intelligence (AI) is the engine of overall progress.
AI technologies can help address global challenges such as combating cli-
mate change and optimizing results in transport, medicine, agriculture, etc.
[1, 2]. In particular, modern AI technologies are used for trading financial
instruments on stock exchanges, carrying out crewless vehicles, establish-
ing medical diagnoses, analyzing large amounts of data, recognizing and
generating images, creating industrial and household robots, as well as high-
precision autonomous weapons [3, 4]. AI processes vast amounts of data,
making conclusions with the help of specific algorithms, which become the
basis for independent decision-making.
Fuzzy logic is a relatively widespread and powerful AI tool with great
potential [5–7]. It allows to operate effectively with expert knowledge,
simulate human decision-making processes, and form linguistic models of
complicated dynamic plants [8–10]. The most expedient is the application of
methods and means of fuzzy logic to construct intelligent decision-making
and control systems of different types. In particular, expert systems for
decision-making based on fuzzy logic techniques are successfully used in
conditions of uncertainty in many areas: financial management, technical
and medical diagnostics, stock market forecasting, transport logistics, and
others [11–13]. Also, fuzzy devices and inference systems are developed and
frequently implemented as observers and adaptive devices, controllers and
identifiers, units of tactical and strategic control for automation of complex
nonstationary and nonlinear objects, such as power plants, robotic production
lines, chemical reactors, spacecraft, mobile robots, satellites, marine ships,
and floating structures, crewless underwater vehicles, drones and others
[14–16].
A specific feature of units and systems based on fuzzy logic is a consider-
able number of structural elements and parameters that significantly influence
their efficiency and should be tuned by the experts or automatically using spe-
cific progressive techniques during their design process [17–19]. In particular,
the main adjustable parameters and elements are: (a) the shapes and parame-
ters of the linguistic terms membership functions, (b) the number of LTs for
input and output signals, (c) the types of operations of aggregation, activation,
accumulation, and the defuzzification procedures, (d) the number of rules of
the fuzzy rule base and its antecedents and consequents, (e) the adjustable
input gains, etc. [20–22]. It gives the ability to create effective, flexible,
and complicated strategies of decision-making and control but, at the same
9.2 Related Works and Problem Statement 199
time, significantly complicates the development process of the given units and
systems. Thus, in many cases, using expert assessments, fuzzy logic systems
do not enormously increase the main quality indicators compared to similar
systems based on conventional principles. Moreover, they may require more
computational costs and more powerful hardware [23–25]. However, recently,
the development and implementation of perspective progressive algorithms,
approaches, and information technologies for fuzzy systems design based on
specific optimization procedures are one of the leading directions of research
in the FSs modern theory [26–28]. Modern fuzzy systems developed using
these advanced optimization-based techniques can be successfully trained
like neural networks, using training samples or objective functions for high
accuracy attainment, and still maintain good interpretability and transparency
[29–31].
This chapter focuses on developing and researching an advanced infor-
mation technology for designing and optimizing fuzzy systems based on the
optimal selection of linguistic terms number. The structure of the paper is
organized as follows. Section 9.2 presents the brief literature survey, the
statement of the research problem, and the primary purpose of this study.
Section 9.3 sets out in detail the proposed information technology of the
FS structural optimization based on the optimal selection of linguistic terms
number. Section 9.4 presents the results of studying the effectiveness of the
developed information technology on a specific example of a fuzzy control
system for the quadrotor unmanned aerial vehicle with a detailed discussion
of the results of computer simulation. Finally, Section 9.5 concludes the work
and suggests the directions for future studies.
Figure 9.1 Structure of the generalized MIMO fuzzy control/decision-making system with
structural optimization unit
Y = fFS (X),
(9.1)
Y = (y1 , y2 , ..., yj , ..., ym ), X = (x1 , x2 , ..., xi , ..., xn ),
where Sxi and Syj are selected numbers of LTs for ith input and jth output FS
variables.
For example for fuzzy system with 3 input variables (n = 3) and 2 output
variables (m = 2) the following numbers of LTs are selected for the given
variables: Sx1 = 5, Sx2 = 3, Sx3 = 3, Sy1 = 5, Sy2 = 7. Then, the vector S has
the following form
S = {5, 3, 3, 5, 7}. (9.4)
In turn, for each found variant of the vector S by the unit of LTs number
optimal selection, the new variant of the rule base (vector R) should be syn-
thesized. Only then, the objective function J can be calculated for evaluation
FS with obtained configuration. Thus, the task of the RB synthesis is the sub-
task of the LTs number optimal selection, which should be solved each time
at every iteration during the implementation of the LTs number optimization
process. The given subtask should be solved in automatic mode by the RB
synthesis unit. In addition, in the optimization process, both units of the LTs
204 Fuzzy Systems Design: Optimal Selection
values, the operating ranges should be set from –1 to 1 [80, 81]. In addition, if
some inputs or outputs can be only positive by nature (electric heating power,
water or gas consumption, and others), then their changing ranges should be
from 0 to 1.
Stage 3. Choosing the shapes of LT membership functions for FS inputs and
outputs. The shapes of linguistic terms MFs for all n input and m output
variables are chosen. Often, the same shapes are selected for all LTMFs of the
system’s variables at the initial stage of the designing process. For instance,
Gaussian, trapezoidal or triangular shapes may be chosen [82, 83]. Also, the
initial values of their parameters for each FS input and output should be
set in an automatic mode so that the LTs would be evenly distributed over
the operating ranges (between the minimum and maximum values) for any
specified LTs number.
Stage 4. Determination of the vector S initial structure. The vector S that
defines LTs numbers for FS input and output variables is formed at this stage,
taking into account the chosen at the Stage 1 n inputs and m outputs. It has
the following form
S = {Sx1 , Sx2 , ..., Sxi , ..., Sxn , Sy1 , Sy2 , ..., Syj , ..., Sym }. (9.5)
The given vector includes variables Sxi,yj that define the following num-
bers of the LTs for the fuzzy system’s variables. These variables are located
in order, starting from the first input x1 and ending with the final output ym .
Stage 5. The setting of vector S constraints. The constraints Smin and Smax
on the LTs number are selected for each FS input and output. For example,
the minimum value of the number of the terms may be selected equal to 2
(Smin = 2), and the maximum value can be chosen equal to 7 (Smax = 7)
for each FS input [28, 39]. In turn, for the FS outputs, the minimum and
maximum values can be equal to 3 (Smin = 3) and 9 (Smax = 9), respectively
[82]. Moreover, at this stage, some additional constraints may be selected. For
instance, for fuzzy control systems, if their variables can vary from positive
to negative values within the symmetric operating ranges, the additional
constraints may be set, that the number of the linguistic terms should be only
odd [39]. Therefore, in this case, in addition to basic constraints (Smin and
Smax ), the LTs number maybe only be as follows: 3, 5, and 7 – for FS inputs
(Sxi ∈ {3, 5, 7}); 3, 5, 7, and 9 – for output variables (Syj ∈ {3, 5, 7, 9}).
Stage 6. Selection of the FS complex objective function J C . The main
parameters and boundary (optimal) value of the complex objective function
206 Fuzzy Systems Design: Optimal Selection
JC = J1 + kJ2 J2 , (9.6)
where rmax is the total number of the RB rules, defined by the number of all
possible combinations of the LTs of the system’s inputs.
Moreover, the boundary (optimal) values of the functions J 1opt and J Copt
and the corresponding coefficient kJ 2 are preliminarily selected at this stage.
It is necessary to consider the main requirements and peculiarities of the
problem of fuzzy system design. For example, the optimal value of the
objective function J Copt and the scaling factor kJ 2 can be selected based on
the following conditions
JCopt JCopt JCopt
< 2, > 2, = 2. (9.8)
J1opt J1opt J1opt
m –1. If the current vector of the LTs number S is already recorded to the
Checklist, then transition to Stage 13 is performed. In the opposite case, go
to Stage 10.
Stage 10. Development of the rule base for the system with the current vector
S and calculation of its complex objective function. The rule base generation
is conducted at this stage through the ACO-based method or technique of
the sequential search, designed and approbated in papers [29] and [21],
respectively. Also, the complex objective function J C value is calculated
for this designed rule base. As the previous comprehensive research [45]
shows, it is expedient to choose the sequential search technique or the
nature-inspired ACO-based method for performing an effective automatic
RB design, depending on its size and complexity, defined with the help of
criterion J 2 (9.7). Herewith, if the criterion J 2 ≤ 600, the sequential search
approach should be used. If J 2 > 600, then the nature-inspired ACO-based
technique will be more efficacious [45]. Consequently, at this stage, the
criterion J 2 (9.7) for the current vector S is first calculated, and only then,
based on its defined value, a suitable method for RB design is chosen (ACO-
based method or sequential search). Further, the rule base development is
performed with the calculation of the criterion J 1 using the chosen technique.
After that, the value of the complex objective function J C is finally calculated
by the equation (9.6) using values of the criteria J 1 and J 2 . In turn, the
adjustable parameters of the ACO-based and sequential search algorithms
(number of rounds of sequential search l, number of agents in the population
Z max , number of elite agents e, maximum number of iterations N*max , etc.)
are previously set based on conducted experimental studies and obtained
recommendations in the previous research [29, 45]. The antecedents of the
rule base for the current vector S are synthesized in an automatic mode as all
possible combinations of LTs of the system’s inputs. As for their consequents,
they are determined during the design process through the above sequential
search method or ACO-based approach.
Stage 11. Check for the attainment of the complex objective function optimal
value. At this stage, the achievement of the optimal value of the complex
objective function J Copt is checked for the obtained vector S of the LTs
number. If the given check is positive, then transition to Stage 16 is carried
out. Otherwise, go to Stage 12.
Stage 12. Record the current vector S, the corresponding rule base, and the
system’s value of the complex objective function to the Checklist. The current
vector S, the corresponding developed rule base, and the system’s current
9.3 Information Technology for Fuzzy Systems Design 209
Figure 9.2 Flowchart of the first modification of the information technology for FS structural
optimization based on optimal selection of linguistic terms number
where
ΔSxi = Sxi max − Sxi min , i = 1, 2, ..., n; (9.11)
ΔSyj = Syj max − Syj min , j = 1, 2, ..., m. (9.12)
Moreover, each N-th iteration of the presented information technology
includes conducting N*max iterations of the ACO-based method or sequential
search technique at Stage 10.
The second modification based on the random search approach uses the
concept of creating H random variants of vectors of the LTs number and
9.3 Information Technology for Fuzzy Systems Design 211
choosing the best from them. The first six stages of this modification are
the same as for the first modification of the information technology. Next,
consider the rest of the stages of this modification, starting from Stage 7.
Stage 7. Generation of H different variants of vector S in a random way.
At this stage, H different variants of the vector of LTs number are randomly
generated. In turn, each variable Sxi,yj of each h-th vector Sh is generated ran-
domly within the constraints Sxi,yj min and Sxi,yj max , that are set at Sage 5.
Moreover, the number of random variants H is previously selected before
implementing this modification of the information technology. Herewith, if
there are some identical vectors among the generated variants, then they are
deleted, and new ones are generated in their place until all H vectors S are
different. The rule bases will be synthesized, and the complex objective func-
tion J C values will be calculated for each h-th vector of the given generated
vectors at the further stages of the given modification of the information
technology.
Stage 8. Selection of the first variant of the generated vectors S. The first
variant (h = 1) of the generated vectors of the LTs number S1 is selected at
this stage to start the procedures of the optimization process.
Stage 9. Development of the rule base for the system with the current vector
S and calculation of its complex objective function. The given stage of
this modification is the same as Stage 10 for the first modification of the
information technology.
Stage 10. Check for the attainment of the complex objective function optimal
value. The given stage of this modification is the same as Stage 11 for the first
modification of the information technology. If this check is positive, then the
transition to Stage 13 is carried out. In the opposite case, go to Stage 11.
Stage 11. Check for the optimization process completion of the LTs number
for the FS. The calculations of structural optimization are considered to be
complete at this stage if the design of the rule base and determination of
the complex objective function J C is conducted for all H vectors of the LTs
numbers generated at Stage 7 (h = H). If the given check is positive, then
the transition to Stage 12 is performed. Otherwise, go to the next (h + 1)
generated vector Sh+1 and return to Stage 9.
Stage 12. Selection of the best variant of the generated vectors S. At this
stage, the best variant of the H generated vectors S is selected based on their
values of the complex objective function J C calculated at Stage 9.
212 Fuzzy Systems Design: Optimal Selection
Stage 13. Completion of the LTs number optimization process. The given
stage of this modification is the same as Stage 16 for the first modification of
the information technology.
Figure 9.3 shows a flowchart of the second modification of the pro-
posed information technology for fuzzy systems development and structural
optimization with optimal selection of LTs number.
At the implementation of the given modification of the proposed infor-
mation technology, it is advisable to set the number of generated vectors H
the same as the maximum number of iterations N max for the first modifica-
tion for the objectivity of comparing the computational costs of these two
modifications.
Figure 9.3 Flowchart of the second modification of the information technology for FS
structural optimization based on optimal selection of linguistic terms number
9.3 Information Technology for Fuzzy Systems Design 213
Figure 9.4 Flowchart of the stage of design of the RB and calculation of the complex
objective function for the FS with the current vector S
route length that a specific ant has traveled in the current iteration depends on
the value of the FS objective function (first criterion) J1Z (N ∗).
During the movement of all ants along the graph, the probabil-
ity PrγjZ (N ∗) of the Zth agent passing along the rγjth edge from
the node with number r(j–1) to the rj-th node (r ∈ {1, ..., s} ; γ ∈
{1, ..., Syj } ; j ∈ {1, ..., m}) in the N*-th iteration (N* = 1, . . . , N max ) is
calculated based on the equation [88]
where r is the current rule number of the RB; γ is the current consequent
number of the jth output variable; τrγj (N*) is the pheromone intensity on the
rγjth edge at the iteration N*; ηrγj is an inverse of the relative length Drγj of
the rγjth edge; α is an adjustable parameter which determines the importance
of the pheromone trace considering on edge; β is the parameter that regulates
importance of the relative length Drγj of the rγjth edge.
Other equations and main steps of the ACO-based method and sequential
search method for automatic rule base design of fuzzy control and decision-
making systems are considered in detail in [29, 89].
The effectiveness study of the developed information technology is per-
formed in this work on a specific example, particularly when designing the
fuzzy control system for the quadrotor unmanned aerial vehicle.
and inspection tasks in various areas of the civil and military sector [90–92].
QUAVs have many advantages over other types of crewless aerial vehicles,
namely: high maneuverability, high efficiency, the ability to take off and land
vertically and hovering in space, ease of maintenance, and others. However,
quadcopters are complicated control plants, for automation of which it is
advisable to use intelligent automatic control systems, particularly fuzzy con-
trol systems [93–95]. The stabilization and automatic control of the QUAV
flight altitude when moving over terrain with a complex relief is one of the
most important tasks in automating its spatial motion. In this work, the fuzzy
control system of the QUAV flight altitude is designed to test the effectiveness
of the proposed information technology for FSs structural optimization with
optimal selection of linguistic terms number.
The dynamics of spatial motion of the quadrotor unmanned aerial vehicle
is described by the following system of basic equations [90]:
⎧
⎪
⎪ ẍ = (F1 + F2 + F3 + F4 ) (cos φsinθcosψ + sinφ sin ψ)
⎪
⎪
⎪
⎪
⎪
⎪ −kx1 (ẋ)2 − kx2 Fx /m;
⎪
⎪
⎪
⎪
⎪
⎪ ÿ = (F1 + F2 + F3 + F4 ) (sin φsinθcosψ + cosφ sin ψ)
⎪
⎪
⎪
⎪
⎪
⎪ −k ( ẏ) 2
− k F
y2 y /m;
⎪
⎨
y1
z̈ = (F1 + F2 + F3 + F4 ) cos θcosψ − Fg − kz1 (ż)2 − kz2 Fz /m;
⎪
⎪ 2
⎪
⎪
⎪
⎪ ψ̈ = l −F1 + F2 + F3 − F4 − kψ1 ψ̇ − kψ2 Fψ /Iψ ;
⎪
⎪
⎪
⎪ 2
⎪
⎪
⎪
⎪ θ̈ = l −F1 − F2 + F3 + F4 − kθ1 θ̇ − kθ2 Fθ /Iθ ;
⎪
⎪
⎪
⎪ 2
⎪
⎪
⎪
⎩ φ̈ = Cφ F1 − F2 + F3 − F4 − kφ1 φ̇ − kφ2 Fφ /Iφ ,
(9.14)
where x, y and z are QUAV’s longitudinal-horizontal and transverse-
horizontal coordinates, as well as flight altitude; ψ, θ and ϕ are the roll,
pitch and yaw angles; m is the QUAV mass; F 1 , F 2 , F 3 and F 4 are the
corresponding values of the rotors’ lifting force; Fx , Fy , Fz , F ψ , F θ, and
F ϕ are the corresponding values of disturbing effects of wind; Fg is the
gravity force; I ψ , I,θ and I ϕ are the QUAV’s moments of inertia about the
longitudinal, transverse and vertical axes; l is the distance from the rotor to
the center of mass; Cϕ is the yaw moment coefficient; kx1 , kx2 , ky1 , ky2 , kz 1 ,
kz 2 , kψ1 , kψ2 , kθ1 , kθ2 , kϕ1 and kϕ2 are the QUAV’s model coefficients.
216 Fuzzy Systems Design: Optimal Selection
Since the time constants of the quadcopter drives are significantly less
than the time constants of the quadcopter itself, then as an assumption,
the dynamics of the rotor drives can be neglected. Thus, the control of the
quadcopter is carried out using four control forces: u1 , u2 , u3 , and u4 [90]. In
turn, these forces are calculated by the equations
u1 = F1 + F2 + F3 + F4 ;
u2 = −F1 − F2 + F3 + F4 ;
(9.15)
u3 = −F1 + F2 + F3 − F4 ;
u4 = F1 − F2 + F3 − F4 .
Figure 9.5 The fuzzy control system of the QUAV flight altitude
Figure 9.6 Complex relief of the mountainous terrain presented by the dependence zT (x)
Figure 9.7 Changing curves of the complex objective function J C and the criteria J 1 ,
J 2 values during the structural optimization process (first modification)
Sopt = {5, 3, 3, 7} .
In turn, the vector S, obtained at the ninth iteration, has less value of the
criterion J 1 (J 1 = 0.103) than Sopt . However, its criterion J 2 is significantly
bigger (J 2 = 405), which ultimately gives a bigger value of the complex
objective function J C (J C = 0.236). In addition, for the vector S obtained
at the first iteration, the value of criterion J 2 is the smallest (J 2 = 135), but its
criterion J 1 has the worst value (J 1 = 0.671), which as a result gives the worst
value of the complex objective function J C (J C = 0.715).
9.4 Design and Structural Optimization of the Fuzzy Control System 221
Table 9.2 RB fragment for the vector Sopt obtained using the first modification
Input and Output Variables
Rule Number 5
εz dεz
dt
εz dt u1
1 BN N N BN
5 BN Z Z BN
12 SN N P N
18 SN P P SN
23 Z Z Z Z
27 Z P P P
30 SP N P SP
36 SP P P BP
41 BP Z Z BP
45 BP P P BP
Also, a particular feature of these calculations (carried out using the first
modification) is that the controller RB was developed using the sequential
search technique without using the ACO-based method during all iterations.
Since, for all obtained values of the vector S, criterion J 2 was less than 600.
Table 9.2 presents the RB fragment developed during the optimization
process for the found optimal vector Sopt . The given RB has 45 rules.
Also, the entire vector R of the RB consequents has the following form:
R = (BN, BN, BN, BN, BN, BN, BN, BN, BN, BN, BN, N, N, N, SN, SN,
SN, SN, N, N, SN, SN, Z, Z, SP, SP, P, SP, SP, SP, SP, P, P, SP, P, BP, BP, BP,
BP, BP, BP, BP, BP, BP, BP).
Further, the remaining stages of the second modification of the proposed
information technology are carried out (from 7th to 13th ).
At Stage 7 of the second modification, H different variants of vector S
of LTs number are randomly generated, considering constraints selected at
Stage 5. In this case, the number of generated random variants H is equal
to 9 (H = 9) to be the same as the maximum number of iterations N max for
the first modification for the objectivity of comparing the computational costs
of these two modifications. The rule bases are synthesized, and the complex
objective function J C values are calculated for each h-th vector Sh (h = 1,
2, . . . , H) of the given generated vectors at the further stages of the given
modification. In turn, the RBs are developed through sequential search and
ACO-based methods, depending on the current value of the criterion J 2 . The
authors set the same adjustable parameters of the sequential search method
and ACO-based technique as for the first modification.
222 Fuzzy Systems Design: Optimal Selection
Table 9.3 presents obtained results during the optimization of the vector
S for the QUAV altitude fuzzy controller using the second modification.
Figure 9.8 shows the changing curves of the complex objective function
J C and separately the criteria J 1 and J 2 at implementing the second modifica-
tion of the information technology during the structural optimization process
of the QUAV’s altitude FC.
As shown from Table 9.3 and Figure 9.8, the optimal value of the complex
objective function J Copt has not been achieved for any generated variant dur-
ing the implementation of the given (second) modification of the information
technology. The best variant (7th variant) of the nine generated vectors is
selected that has the lowest value of the complex objective function (J C =
0.376). It has the following form
Sbest = {3, 5, 5, 7} .
Also, for most of the generated variants of vector S, the synthesis of the
rule base was carried out using the ACO-based approach since the value of
criterion J 2 was more than 600. In turn, the sequential search method was
used only for the sixth and seventh variants.
In general, with the help of this modification, more complex configura-
tions of fuzzy systems with a larger number of LTs for inputs and output were
found in comparison with the variants obtained using the first modification of
the information technology.
The found more complex variants made it possible to achieve lower values
of criterion J 1 (especially 1, 3, 4, 5, 8). However, they have sufficiently
large values of criterion J 2 , which gave unsatisfactory values of the complex
objective function J C for all obtained vectors S. We can assume that with
an increase in the number H of randomly generated variants of vectors of
the number of the linguistic terms, it would be possible to find the optimal
9.4 Design and Structural Optimization of the Fuzzy Control System 223
Figure 9.8 Changing curves of the complex objective function J C and the criteria J 1 ,
J 2 values during the structural optimization process (second modification)
vector Sopt (with the optimal value of the complex objective function J Copt )
using this second modification. Nevertheless, this would require more com-
putational costs in comparison with the first modification of the information
technology.
Thus, the studies carried out show that the first modification, in general,
has a higher efficiency than the second since it allows to find the optimal
configurations of fuzzy systems (by the LTs number of input and output
variables) using lower computational costs. Also, we can conclude that when
creating fuzzy systems, for which simplicity of configuration, the small size
of the rule base, and simplicity of software and hardware implementation are
essential, it is more expedient to use the first modification of this information
technology. Since it allows synthesizing the simplest and enough effective
fuzzy systems, and, at the same time, uses low computational costs. In turn,
the second modification can be quite effectively used to design fuzzy systems,
for which high accuracy and efficiency of solving the assigned problem are
paramount, and the configuration complexity, the RB size, and the complexity
of the software and hardware implementation are of less importance.
224 Fuzzy Systems Design: Optimal Selection
Table 9.4 RB fragment for the vector Sbest obtained by the second modification
Input and Output Variables
Rule Number 5
εz dεz
dt
εz dt u1
1 N BN BN BN
12 N Z SN BN
20 N SP BP BN
28 Z BN Z SN
38 Z Z Z Z
46 Z BP BN SP
54 P BN SP BP
61 P Z BN BP
68 P SP Z BP
75 P BP BP BP
Also, Figure 9.12 presents the transients of the flight control system of the
quadcopter (changing of the real values of flight altitude relative to the terrain
surface) with developed fuzzy controllers at flying over the mountainous
terrain with complex relief zT (x) (Figure 9.6). In turn, the detailed transients
of the flight control system are presented in Figure 9.13 (t = 0. . . 40 s) and
Figure 9.14 (zR = 3.5. . . 6.5 m; t = 100. . . 220 s). The simulation results
presented in Figures 9.9–9.14 are obtained for the QUAV with a flight altitude
control system based on the FCs: 1 – developed by the first modification;
2 – developed by the second modification. The flight altitude set value
zS (relative to the terrain surface) is 5 m, the QUAV speed ẋs along the
longitudinal-horizontal axis x over the mountainous terrain equals 4.3 m/s.
As Figure 9.9–9.14 shows, the QUAV’s flight control system based on
the FC developed using the first modification has higher accuracy and qual-
ity indicators of control compared to the same system based on the fuzzy
controller designed by the second modification. Moreover, the FC developed
using the first modification has a more straightforward structure, fewer rules
in the RB, and less computational costs were spent during its development,
which in general confirms the high effectiveness of this modification of the
proposed information technology.
226 Fuzzy Systems Design: Optimal Selection
Figure 9.12 Transients of the altitude control system of the QUAV with developed FCs
Figure 9.14 Detailed transients of the altitude control system (zR = 3.5. . . 6.5 m;
t = 100. . . 220 s)
9.5 Conclusion
The authors presented the development and study of the advanced infor-
mation technology for design and structural optimization of fuzzy systems
with optimal selection of linguistic terms number in this chapter. The appli-
cation of the developed information technology increases the FS accuracy
and performance, reduces the computational costs spent on the rule base
generation and optimization of parameters, and simplifies its software and
hardware realization. In particular, in this chapter, two modifications of the
given information technology for fuzzy systems development and structural
optimization with optimal selection of LTs number are developed and studied:
1) based on the sequential search approach; 2) based on the random search
approach. In turn, each modification uses a bioinspired ACO-based method
References 229
and sequential search technique for automatic rule base design for each new
obtained variant of the vector S that defines numbers of linguistic terms. The
first modification uses the concept of sequential selection of the best values
of the number of LTs for the FS, starting from the first input x1 and ending
with the last output ym . The second modification uses the concept of creating
H random variants of vectors of the LTs number and choosing the best from
them.
The effectiveness study of both modifications of the developed informa-
tion technology is performed on a specific example, namely when designing
the fuzzy control system for the quadrotor unmanned aerial vehicle. In
particular, the development and structural optimization with optimal selection
of LTs number of the fuzzy controller for QUAV’s altitude control system is
carried out alternately using first and second modifications of the proposed
information technology. The conducted studies confirm, in general, the high
effectiveness of the presented information technology since its application
made it possible to develop a highly efficient QUAV’s fuzzy control system
with a fairly simple configuration and RB. The high efficiency is defined by
the obtained flight trajectories of the QUAV over the mountainous terrain with
complicated relief and control transients and low computational costs spent
at developing the studied fuzzy control system.
Moreover, when comparing the first and the second modifications with
each other, it is evident that the first modification has a higher efficiency than
the second since it allows to find the optimal configurations of fuzzy systems
(by the number of LTs for input and output variables) using lower computa-
tional costs. Also, we can conclude that when designing fuzzy systems, for
which simplicity of configuration, the small size of the RB, and simplicity
of software and hardware implementation are more important, it is more
advisable to use the first modification of this information technology. Since it
allows developing enough effective fuzzy systems with the simplest structure
and uses low computational costs at the same time. In turn, the second
modification can be quite effectively used to design FSs, for which high
accuracy and efficiency of solving the assigned problem (control or decision-
making) are paramount. Also, the configuration complexity, the RB size, and
the complexity of these systems’ software and hardware implementation are
of less importance.
Further research should be conducted towards developing and studying a
new specific modification of the proposed information technology for design
and structural optimization with optimal selection of LTs number for fuzzy
decision-making and control systems of the Takagi–Sugeno type.
230 Fuzzy Systems Design: Optimal Selection
References
[1] K. Knight, C. Zhang, G. Holmes, M.-L. Zhang (Eds.), ‘Artificial Intel-
ligence,’ Second CCF International Conference, ICAI 2019, Xuzhou,
China, August 22–23, 2019, Proceedings, Springer Singapore, 2019.
DOI 10.1007/978-981-32-9298-7.
[2] S. N. Vassilyev, A. Yu. Kelina, Y. I. Kudinov, F. F. Pashchenko, ‘Intel-
ligent Control Systems,’ in Procedia Computer Science, Vol. 103, pp.
623–628, 2017.
[3] C.C. Aggarwal, ‘Artificial Intelligence,’ Springer International Publish-
ing, 2021. DOI 10.1007/978-3-030-72357-6.
[4] G. Krivulya, I. Skarga-Bandurova, Z. Tatarchenko, O. Seredina, M.
Shcherbakova, E. Shcherbakov, ‘An intelligent functional diagnostics
of wireless sensor network,’ in Proceedings of 2019 International Con-
ference on Future Internet of Things and Cloud Workshops, FiCloudW
2019, pp. 135–139, 2019.
[5] L. A. Zadeh, ‘Fuzzy Sets,’ in Information & Control, 8, pp. 338–353,
1965.
[6] L. A. Zadeh, ‘The role of fuzzy logic in modeling, identification and
control,’ in Modeling Identification and Control, 15 (3), pp. 191–203,
1994.
[7] L. A. Zadeh, A. M. Abbasov, R. R. Yager, S. N. Shahbazova, M. Z.
Reformat, Eds., ‘Recent Developments and New Directions in Soft
Computing,’ STUDFUZ 317, Cham: Springer, 2014.
[8] J. Kacprzyk, ‘Multistage Fuzzy Control: A Prescriptive Approach,’ John
Wiley & Sons, Inc., New York, NY, USA, 1997.
[9] B. Kosko, ‘Fuzzy Systems as Universal Approximators,’ in IEEE Trans.
on Computers, Vol. 43, Iss. 11, pp. 1329–1333, 1994.
[10] Y. P. Kondratenko, O. V. Korobko, O. V. Kozlov, ‘Synthesis and Opti-
mization of Fuzzy Controller for Thermoacoustic Plant,’ in Recent
Developments and New Direction in Soft-Computing Foundations and
Applications, Studies in Fuzziness and Soft Computing 342, Lotfi A.
Zadeh et al. (Eds.), Berlin, Heidelberg: Springer-Verlag, pp. 453–467,
2016.
[11] J. M. Mendel, ‘Uncertain Rule-Based Fuzzy Systems, Introduction and
New Directions,’ Second Edition, Springer International Publishing,
2017.
[12] M. Solesvik, Y. Kondratenko, G. Kondratenko, I. Sidenko, V.
Kharchenko, A. Boyarchuk, ‘Fuzzy decision support systems in marine
References 231
[89] R.-M. Chen, Y.-M. Shen, C.-T. Wang, ‘Ant Colony Optimization
Inspired Swarm Optimization for Grid Task Scheduling,’ in 2016 Inter-
national Symposium on Computer, Consumer and Control (IS3C),
pp. 461–464, 2016.
[90] V.L. Timchenko, D.O. Lebedev, ‘Optimization of Processes of Robust
Control of Quadcopter for Monitoring of Sea Waters,’ Journal of
Automation and Information Sciences, NY., Begell House, Vol. 51, Issue
2, pp. 1–10, 2019.
[91] N. Ben, S. Bouallègue, J. Haggège, ‘Fuzzy gains-scheduling of an inte-
gral sliding mode controller for a quadrotor unmanned aerial vehicle,’
Int. J. Adv. Comput. Sci. Appl., Vol. 9, no. 3, pp. 132–141, 2018.
[92] Y. Kondratenko, ‘Robotics, Automation and information systems:
Future perspectives and correlation with culture, Sport and life science,’
Lecture Notes in Economics and Mathematical Systems, Vol. 675,
Springer Verlag, pp. 43–55, 2014. DOI: 10.1007/978-3-319-03907-7_6.
[93] A. Eltayeb, M. F. Rahmat, M. A. M. Basri, M. A. M. Eltoum, S. El-
Ferik, ‘An Improved Design of an Adaptive Sliding Mode Controller
for Chattering Attenuation and Trajectory Tracking of the Quadcopter
UAV,’ IEEE Access, Vol. 8, pp. 205968–205979, 2020.
[94] R. Duro et al. (Eds). ‘Advances in Intelligent Robotics and Collaborative
Automation.’ River Publishers, Gistrup, 2015. DOI: https://doi.org/10.1
3052/rp-9788793237049.
[95] Y.P. Kondratenko, et al (Eds). ‘Advanced Control Systems: Theory and
Applications.’ River Publishers, Gistrup, 2021.
[96] I.P. Atamanyuk, V.Y. Kondratenko, O.V. Kozlov, Y.P. Kondratenko,
‘The Algorithm of Optimal Polynomial Extrapolation of Random Pro-
cesses,’ Lecture Notes in Business Information Processing: Modeling
and Simulation in Engineering, Economics and Management., K. J.
Engemann, A. M. Gil-Lafuente, J. M. Merigo (Eds.), Berlin, Heidelberg:
Springer-Verlag, Vol. 115, pp. 78–87, 2012.
[97] D. Simon, ‘Design and rule base reduction of a fuzzy filter for the
estimation of motor currents,’ in International Journal of Approximate
Reasoning, 25, pp. 145–167, 2000.
10
Analysis of the Dynamics
and Controllability of an Autonomous
Mobile Robot with a Manipulator
Abstract
The results of the study of the dynamics and control of an autonomous mobile
robot with a manipulator are presented. An all-wheel drive model with a
four-wheeled chassis layout and equipped with a four-link manipulator is
considered. The design of the manipulator consists of a docking ring rotating
around a vertical axis and three rod links of the arm, connected by rotary
kinematic pairs of the fifth class. The mathematical model of controlled
movement of autonomous mobile robot with a manipulator is proposed. With
relative motion of the manipulator, the tensor of inertia of an autonomous
mobile robot in the coordinate system associated with the chassis is non-
diagonality and non-stationarity. Simplifications of the mathematical model
for several driving modes are substantiated. An algorithm for integrating the
equations of dynamics is developed, taking into account certain features.
241
242 Analysis of the Dynamics and Controllability of an Autonomous Mobile Robot
10.1 Introduction
The development of new technologies makes it possible to use manipulators
(M) for work in extreme conditions for humans. For example, the use of M
and autonomous mobile robots (AMR) for eliminating the consequences of
environmental and man-made disasters, the operation of remotely controlled
M for working with radioactive substances, the use of M for assembling
structures in space or on the seabed, etc. For work in extreme conditions, as
a rule, remote-controlled robots and M. are used. In many cases, a remote-
controlled arm-manipulator is installed on a controlled mobile platform.
This expands the service characteristics of M, but imposes more stringent
requirements on the design of control and monitoring systems. An example
of such structures is: a manipulator arm is placed on a spacecraft, on a ship,
on a wheeled chassis, etc. In all of the above cases, the movement of the
base affects the kinematics and dynamics of the manipulator arm, increasing
the positioning error and deviation of the gripper pole from the specified
trajectory. Dynamic processes for M on a mobile platform can be divided
into three modes of movement:
• controlled movement of the platform along a given route with “check-
in” M;
• performing technological operations M with the platform stopped;
• M performs technological operations with a controlled movement of the
platform.
The controlled movement of a platform with a “locked” manipulator has been
well studied for various chassis designs. In [1], the dynamics of an all-wheel
drive 4-wheeled bogie when turning is investigated. The synthesis of the law
of control of the plane motion of a transport robot is a typical task: several
methods of stabilization of motion along a given trajectory are known. In [2],
the authors propose a modification of the system of dynamic equations with
the change of coordinates and the transformation of outputs to chain forms.
In [3], a transformation of the system of equations to the Cauchy normal form
was applied with subsequent differentiation with respect to a new independent
variable [3–5]. In [2, 6], the equations of the bogie dynamics are relative to
road coordinates, which well fix deviations from a given trajectory.
10.1 Introduction 243
+Ω̄ × (w̄
) 1 × p̄o1M )* + (Ω̄ × w̄1 ) × p̄o1M +
+w̄1 × Ω̄ × p̄o1M +w̄1 × (w̄1 × p̄o1M )] = v F̄v .
(10.1)
For the sections of the trajectory on which the AMR moves with an
immovable manipulator, let’s write formula (10.1) in the form:
1 2
d2 r̄c d˜0 Ω̄ ) * !
m· 2
+ × p̄co1 + mM Ω̄ × Ω̄ × p̄CM = F̄v . (10.2)
dt dt v
If the AMR moves with a slow angular motion of the manipulator, from
formula (10.1) let’s obtain:
10.2 Results of Study of Dynamics and Control of Autonomous Mobile Robots 249
1 2 1
d2 r̄c d˜0 Ω̄ d˜21 p̄o1M ) * d˜1 p̄o1M
m· 2
+ × p̄ co1 + m M · 2
+ 2 w̄1 + Ω̄ × +
dt dt dt dt
2
d˜1 w̄1 !
+ × p̄o1M = F̄v . (10.3)
dt v
the indicated features at the switching points jof f , jon , the control functions
χi have several values (Figure 10.2).
It is proposed to solve the system of equations by the Runge–Kutta–
Fehlberg method [27]. In this case, on the sections of motion corresponding
to χi =const, numerical integration is carried out with a certain constant step
h. If |j-jof f |≤E (or |j-jon |≤E , where E = const – the vicinity of the switching
point of the relay characteristic), then the integration step decreases and is
equal to h1 =h/C. The most common is the so-called double step method with
C=2n , n = const, n>0.
If the condition |j-jof f |≤E/C (or |j-jon |≤E/C) let’s assume that the switch-
ing of the relay characteristic takes place and χi changes the value in
accordance with the control law. The values of the constants h1 , h, E, C, n are
chosen based on the initial conditions. The structural diagram of the algorithm
for solving the system of equations by the Runge–Kutta–Fehlberg method is
shown in Figure 10.3. The use of the Runge–Kutta–Fehlberg method makes it
possible to create software for identifying the current dynamic state of AMR
with M.
The results of mathematical modeling of the AMR dynamics without
taking into account the non-diagonality and non-stationarity of the tensor of
inertia are given in Table. 10.1 (t – time; q1 (t) – the generalized coordinate of
the relative motion of the manipulator; αx (t), αy (t), αz (t) – the angular coor-
dinates of the system of bodies relative to the coordinate system CXC YC ZC ;
ω x (t), ω y (t), ω z (t) – the angular velocities of the system of bodies relative to
the coordinate system CXC YC ZC ; χx , χy , χz – control functions).
The results of mathematical modeling of the AMR dynamics taking into
account the non-diagonality and non-stationarity of the tensor of inertia are
252 Analysis of the Dynamics and Controllability of an Autonomous Mobile Robot
BEGIN
Yes Yes
No No
Integration with step ,
Figure 10.3 Algorithm for solving the problem of dynamics of AMR with M by the Runge–
Kutta–Fehlberg method
Table 10.1 Results of mathematical modeling of AMR dynamics without taking into account
the non-diagonality and non-stationarity of the tensor of inertia
t, c q1 , αx , ω x .10−3 , χx αy , ω y .10−3 , χy αz .10−3 , ω z .10−3 , χz
rad rad rad /s rad rad /s rad rad /s
0 0 0 0 0 0 0 0 15,71 17,45 −1
0,03 0 0 −0,1 1 0 0,12 −1 15,71 −17,45 1
18,03 0,05 −0,24 −0,1 1 −0,18 −0,1 1 −15,71 −17,45 1
18,05 0,05 −0,24 1,0 −1 −0,18 0,05 0 −15,71 17,45 −1
36,05 0,11 1,47 1,0 −1 −0,05 0,09 0 15,71 17,45 −1
36,08 0,11 1,47 −0,1 1 −0,05 −0,1 1 15,71 −17,45 1
54,08 0,16 1,24 −0,1 1 −0,01 0,02 0 −15,71 −17,45 1
54,11 0,16 1,37 1,0 −1 −0,01 0,14 −1 −15,71 17,45 −1
72,11 0,22 3,08 1,0 −1 0,23 0,18 −1 15,71 17,45 −1
72,14 0,22 3,08 −0,1 1 0,23 0,06 0 15,71 −17,45 1
90,14 0,27 2,85 −0,1 1 0,38 0,11 −1 −15,71 −17,45 1
Table 10.2 Results of mathematical modeling of AMR dynamics taking into account the
non-diagonality and non-stationarity of the tensor of inertia
t, c q1 , αx , ω x .10−3 , χx αy , ω y .10−3 , χy αz .10−3 , ω z .10−3 , χz
rad rad rad /s rad rad /s rad rad /s
0 0 0 0 0 0 0 0 15,71 17,45 −1
0,03 0 0 −0,1 1 0 −0,1 1 15,71 −17,45 1
18,03 0,05 −0,24 −0,1 1 −0,02 −0,1 1 −15,71 −17,45 1
18,05 0,05 −0,24 0,1 −1 −0,02 0,1 −1 −15,71 17,45 −1
36,05 0,11 0,14 0,1 −1 0,01 0,1 −1 15,71 17,45 −1
36,08 0,11 0,14 −0,1 1 0,01 −0,1 1 15,71 −17,45 1
54,08 0,16 −0,14 −0,1 1 −0,01 −0,1 1 −15,71 −17,45 1
54,11 0,16 −0,14 1,0 −1 −0,01 1,0 −1 −15,71 17,45 −1
72,11 0,22 0,13 1,0 −1 0,01 1,0 −1 15,71 17,45 −1
72,14 0,22 0,13 −0,1 1 0,01 −0,1 1 15,71 −17,45 1
90,14 0,27 −0,12 −0,1 1 −0,01 −0,1 1 −15,71 −17,45 1
the direction of motion. The control system tries to work out the disturbances
and prevent the growth of angular velocities in the directions perpendicular
to the direction of motion. Despite the operation of the control system, due
to the non-diagonality and non-stationarity of the tensor of inertia and cross-
coupling of control channels, after 90 s of the AMR movement along the
trajectory, the angular coordinates reach the values αx = 2.85 rad, αy = 0.38
rad. Thus, AMR loses its vertical position, in the conditions of autonomous
operation it leads to the impossibility of performing technological
operations.
If the magnitude and direction of action of the control actions is deter-
mined taking into account the cross-connection of control channels due to the
non-diagonality and non-stationarity of the AMR inertia tensor relative to the
coordinate system CXC YC ZC , then the angular coordinates reach the values
αx = −0.12 rad, αy = −0.01 rad. In this case, the growth of q1 (t) from 0
to 0.27 rad also causes the occurrence of angular coordinates and velocities
in the directions perpendicular to the direction of motion, but AMR does not
lose its vertical position and continues to perform technological operations.
The effectiveness of manipulator control is largely determined by teach-
ing methods, applied methods and means of adaptation. The control of the
grab of the AMR manipulator is carried out by the operator remotely. The
design provides for three data transmission channels (fiber optic wire, Wi-fi
and protected radio communication). Correction of movements and control
of the results of the manipulator’s activity is carried out using a video
surveillance system.
The operating conditions of AMR with a manipulator are often not only
not known a priori, but can also vary unpredictably over a wide range. The
reasons for the uncertainty and nonstationarity of these conditions are: 1) the
lack of information about the properties of the external environment; 2) work
in technical limitations, natural spread and drift of parameters of sensory and
motor systems; 3) in the occurrence of obstacles and computational errors in
communication and control channels.
Usually an adaptive control system has to adapt to the conditions of the
external environment. That is why work with adaptive control and elements
of artificial intelligence is significantly superior in terms of the possibilities of
working with programmed control: they can adequately respond to changes
in the external environment, adapt to the drift of environmental parameters,
recognize and avoid obstacles, identify target objects, and determine their
characteristics.
10.2 Results of Study of Dynamics and Control of Autonomous Mobile Robots 255
Figure 10.4 Block diagram of the AMR adaptive control algorithm with M
256 Analysis of the Dynamics and Controllability of an Autonomous Mobile Robot
10.3 Conclusion
The peculiarities of the dynamics of AMR with M are the non-diagonality
and non-stationarity of the tensor of inertia of bodies relative to the basic
coordinate system, which determine:
• occurrence of inconsistency of the main central axes of inertia of the
system of bodies with the axes associated with the center of mass of the
AMR platform by the basic CX c Y c Z c coordinate system,
• inconsistency of control actions with the directions of the main central
axes of inertia of the system,
• interdependence of control channels.
A mathematical model of controlled motion of AMR with M and an algo-
rithm for solving the problem of dynamics by the Runge–Kutta–Fehlberg
method are presented. Software has been developed for identifying the cur-
rent dynamic state of AMR with M and assessing the mutual connection of
control channels.
References 257
Acknowledgments
The results presented in the article were obtained by the authors in the
course of research on the topic 0120U103294 “Universal anthropomorphic
manipulator” and on the topic 0119U101151 “Applied research in mechanics
and mechatronics.”
References
[1] N. Ashchepkova, S. Ashchepkov, S. Kapera, ‘Dynamics of transport
robot model during the turns’, Science and Education in New Dimen-
sion, pp. 26–29, VI(19), Issue 171, 2018. (in Ukrainian)
[2] B. Siciliano, O. Khatib, ‘Handbook of Robotics’, pp. 799–825, Springer,
2008.
[3] S. Tkachev, ‘Stabilization of nonminimum-phase multi-input affine
system’, Science and Education. Scientific periodical of the Bauman
MSTU, #8, 2012.
[4] O. Andrianova, ‘Path following simulation of wheleed vehicle’, Science
and Education. Scientific periodical of the Bauman MSTU, #10, 2011.
[5] R. Gilimyanov, A. Pesterev, L. Rapoport, ‘Motion control for a wheeled
robot following a curvilinear path’, Journal of Computer and Systems
Science International, pp. 987–994, Vol. 47, #6, 2008.
[6] A. Kanatikov, T. Kasatkina, ‘Features of transition to path coordinates
in a problem of path stabilization’, Science and Education. Scientific
periodical of the Bauman MSTU, #7, 2012.
258 Analysis of the Dynamics and Controllability of an Autonomous Mobile Robot
[29] V.M. Kuntsevich, et. al (Eds). Control Systems: Theory and Applica-
tions. Series in Automation, Control and Robotics, River Publishers,
Gistrup, Delft, 2018.
[30] Y.P. Kondratenko, V.M. Kuntsevich, A.A. Chikrii, V.F. Gubarev,
(Eds). Advanced Control Systems: Theory and Applications. Series in
Automation, Control and Robotics, River Publishers, Gistrup, 2021.
11
Safe Navigation of an Autonomous Robot
in Dynamic and Unknown Environments
Abstract
For robots to become an efficient means for performing tasks, they need to
navigate safely and effectively without supervision. This chapter is devoted
to the problem of fuzzy control system design for mobile robots operating in
dynamic environments with obstacles. The decision-making method, based
on the transformation of radar-sensor information to fuzzy set “Obstacles”
and robot’s heading course to fuzzy set “Direction to Target Point,” is
considered. We also address the safety aspects of the proposed approach.
The different scenarios of the working environment (with static and
dynamic obstacles), the adjusted speed of the autonomous robot, and pecu-
liarities of “robot-obstacle” interaction are the issues for discussion. The
simulation results confirm the efficiency and safety of the robot’s autonomous
navigation in dynamic and unknown environments.
261
262 Safe Navigation of an Autonomous Robot in Dynamic
11.1 Introduction
Service robots acting in uncertain environments have become very popular in
the last few years [1–5]. Many robots and robotic systems successfully work
in hospitals, office buildings, department stores, museums, enterprises, etc.
[6–8].
For robots to become capable team members and helpful assistants,
especially in dynamic human-populated environments, they need to navigate
efficiently and safely [9–12] in the target area.
In many cases, the efficiency of moving vehicles or robot missions
depends on the properties of the sensor and control systems in providing the
obstacles avoiding [13–20] (spacecraft, ships, crewless underwater vehicles,
etc.). The target area, especially for mobile and underwater robots, often
consists of static and dynamic obstacles. Thus, it is necessary to consider
various environmental uncertainties. The obstacle avoidance with a priori
unknown parameters is the peculiar feature of the control systems for above
mentioned autonomous vehicles and robots.
Intelligent systems based on fuzzy sets and fuzzy logic [3, 7, 11, 12, 21-
34] can be a good solution for this problem, especially when a system’s
mathematical model is either unavailable or too complex. Many publications
[1, 3, 9, 11, 12, 35–45] are devoted to robot trajectory planning in the
working environment with obstacles. However, this problem still requires
developing new models and algorithms that optimize and improve the process
of automatic path planning in real-time.
The authors of this chapter develop and investigate intelligent algorithms,
providing the efficient planning and optimization of the robot’s trajectory in
uncertain and dynamic environments. The proposed algorithm is based on the
representation of radar-distance observations and direction to the target point
by specially constructed fuzzy sets with adjusted parameters of membership
functions according to the current “robot-obstacle” situations.
This chapter pays special attention to an autonomous robot’s efficient and
safe navigation in a dynamic or a priori unknown environment. Efficient
navigation means the ability of the robot to reach a target point from a
starting position in a priori unknown environment, and the safety measure
corresponds to the number of the robot collisions with dynamic obstacles
during its navigation to the target point.
11.2 Related Works 263
for the mobile robot, which uses a Microsoft Xbox Kinect sensor. The main
task of the fuzzy controller is providing control of the robot in obstacle avoid-
ance and target following. The designed multi-robot system [60] can work
autonomously in an outdoor environment. In [61], the various robot motion
planning schemes, including a genetic-fuzzy system, genetic-neural system,
and a conventional potential field approach, have been compared in terms
of traveling time taken by the robot, robustness, adaptability, goal-reaching
capability, and repeatability. Neural networks, genetic algorithms and genetic
programming are augmented with fuzzy logic-based schemes to enhance
artificial intelligence of mobile robots [14, 16, 29, 30, 34, 35, 39, 42, 62–67].
In [68], the authors propose a neuro-fuzzy system architecture for behavior-
based control of a mobile robot in unknown environments. The methodology
of the behavior-based control approach includes two steps: (a) to analyze and
to decompose a complex task based on stimulus-response behavior, (b) to
formulate (quantitatively) each type of behavior with a simple feature by
fuzzy sets and fuzzy rules, and (c) to coordinate conflicts and competition
among multiple types of behavior by fuzzy reasoning.
An integrated representation of the environment with an approximation
of obstacles’ shapes by discs or polygons [69] is the base for collision-free
navigation of a mobile robot in complex dynamic environments with moving
obstacles. The navigation algorithm presented in [69] provides a short robot
path planning through the crowd of moving or steady obstacles.
An efficient stereovision-based motion compensation method for moving
robots, discussed in [70], uses the disparity map and three modules: seg-
mentation, feature extraction, and estimation. In the segmentation module,
the authors propose using extended type-2 fuzzy sets to extract the objects.
Fuzzy logic is used for the robot in [71] to implement the behavior design
and coordination during the realization of the “memory grid” and “minimum
risk method.” The robot can choose the safest region to avoid colliding with
obstacles in different scenarios, particularly in the long wall, large concave,
recursive UU-shaped, unstructured, cluttered, maze-like, and dynamic indoor
environments. This chapter [72] presents the fuzzy-logic controller based on
the Mamdani-type fuzzy inference engine for robot navigation and obstacles
avoidance in a cluttered environment. The fuzzy controller with three inputs
and single output provides safe navigation of the robot in a static environment
considering the accuracy of absolute measurements of its position, obstacle
distances, goal distance, velocity, orientation, and rate of change in its head-
ing angle. This chapter [73] describes a fast and reliable obstacle avoidance
method for ground mobile robots in both outdoor and indoor navigation. The
11.2 Related Works 265
• the time of the fuzzy information processing for the calculation and cor-
rection of the robot motion’s course providing the obstacle avoidance;
• the time for calculating and correcting the robot motion’s velocity to
secure the save motion near the obstacle.
Therefore, the actual research task is to modify the existing and create new,
more efficient navigation and control systems algorithms for their successful
implementation in mobile robots and autonomous vehicles.
As discussed above, the artificial intelligence methods, including the
application of fuzzy sets and fuzzy logic, are powerful tools for solving such
problems as planning efficient trajectories for the safe movement of mobile
robots in an uncertain environment with obstacles. Choosing a corresponding
artificial intelligence method for implementation in a robot navigation-control
system (providing efficient fuzzy information processing) requires taking into
account: (a) all specific robot’s environment conditions and peculiarities;
(b) an extensive spectrum of the possible missions for a mobile robot; (c)
the universality for different real working scenarios with static and dynamic
obstacles and other uncertainties.
Thus, it is reasonable to develop new design methods, algorithms, and
models that provide efficient fuzzy information processing. It will simul-
taneously improve the design processes for robots’ navigation-control sys-
tems and enhance the control indicators of their functioning in uncertain
environments with obstacles.
The main idea of this chapter is an investigation of the fuzzy approach for
mobile robots’ trajectory planning in an environment with uncertain obsta-
cles, using proposed by authors fuzzy reasoning algorithm. This algorithm
will provide: (a) the increasing levels of universality for different uncertain
scenarios with obstacles, and (b) simplicity by decreasing the complexity
of the fuzzy information processing and by reducing intricacy of the cor-
responding device’s synthesis process for hardware realization, as well as, (c)
increasing the reliability and (d) decreasing the time for decision making in
different conflict situations.
Finally, let us formulate the aims of this chapter as:
• introducing the fuzzy reasoning algorithm [21–25] instead of fuzzy-rule-
base inference engine for increasing efficiency of robot’s navigation in
the uncertain environment with unknown obstacles;
• the investigation of proposed by authors “radar-fuzzy-set” fuzzy
processing approach, concerning the continuous adjusting of robot
movement parameters and parameters of fuzzy reasoning algorithm,
based on the current information of the radar-sensory system;
11.3 Problem Statement 267
Figure 11.1 The scenario of mobile robot’s working environment with obstacles
2π
ϕi+1 = ϕi + Δϕ = (i + 1) , i = 0, 1, 2, ...n − 1,
n
where n is the number of discrete steps for a scanning beam; Li is a distance
between the robot and the nearest obstacle at every ith step ϕi of the scanning
beam, Li /Lmax , Li ∈ [0, 1].
Figure 11.2 illustrates the circular scan (radar-beam-diagram) for the
configuration of the obstacles Obs2 and Obs3 at the mobile robot’s trajectory
point A (Figure 11.1). The distance Li , (i = 1, 2, ..., n) between the robot
Figure 11.2 The radar-beam-diagram of the robot sensor system in point A on the motion
trajectory, presented in Figure 11.1
270 Safe Navigation of an Autonomous Robot in Dynamic
(point 0 in the center of the circle diagram, Figure 11.2) and any obstacle on
the corresponding beam is presented in the relative form Li /Lmax ∈ [0, 1].
The proposed algorithm for fuzzy information representation and process-
ing consists of the following stages:
Stage 1. Scanning radar area. The interpretation of the scanned results of
the robot’s visible zone (domain) {L0 , L1 , ..., Li , , ..., Ln } for the working
environment with detected obstacles as the fuzzy set E. ∼
Such fuzzy set E∼
(for fuzzy interpretation of the radar information shown
in Figure 11.2) is shown in Figure 11.3, considering the following dependence
Li
μE
∼
(ϕi ) = , (11.1)
Lmax
where μE∼
(ϕi ) is a value of the membership function μE ∼
E,
of the fuzzy set ∼
which corresponds to the value of beam’s angle position ϕi , included to the
support S (E)
∼
[87] of the fuzzy set E:
∼
∀ϕi : ϕi ∈ S (∼ E) = supp (∼ E) = ϕi : μE ∼
(ϕ i ) > 0, ϕi ∈ [0, 2π] .
If any obstacle is absent on the ith scanning beam, the value of member-
Li
ship function μ∼E (ϕi ) = Lmax = 1, because Li = Lmax . If the scanning beam
E (ϕi ) = Lmax ∈ [0, 1] , because Li < Lmax . If
Li
meets the obstacle, then μ∼
the robot contacts a corresponding obstacle or robot situated at the minimal
admissible distance Lmin to the obstacle (and the further movement in the
current direction is impossible), then μE ∼
(ϕi ) = LLmaxi
= 0, as Li = 0 or
Li ≤ Lmin .
It is evident that the membership function of the fuzzy set E ∼
has a
rectangular form for the case if no obstacle is in the radar’s visible zone
where is a fuzzy intersection operator, for example, t-norm operator [87,
90, 91, 92].
The membership function μR∼
(ϕi ) of the fuzzy set R
∼
can be calculated
using the algorithm
R (ϕi ) = μ∼
μ∼ G (ϕi ) μ∼
E (ϕi ) . (11.4)
Usually, a fuzzy set with a membership function μR ∼
(ϕi ) has a one-
interval or multi-intervals core [87] if an obstacle exists in the radar range
of the robot.
Figure 11.5 illustrates the resulting fuzzy set R ∼
with membership function
μR∼
(ϕi ) for interpreting the robot’s heading course ψ as an intersection of the
two fuzzy sets G ∼
and E ∼
according to (11.4).
The shape of the membership function μ∼ R (ϕi ) mostly depends on the
shapes of the membershipfunctions μG ∼
(ϕi ) and μE∼
(ϕi ), but the choice of
the intersection operator ( ) also influences the final form.
It is possible to use different types of t-norm operators ( ) as intersection
operators from two families [87] of non-parameterized and parameterized
t-norm operators. All t-norm operators satisfy the special conditions of
t-norm: mapping space, zeroing, the identity of unity, commutativity, union,
monotonicity.
For example, we can use the non-parameterized minimum intersection
operator (MIN-operator) for the calculation of the values μ∼ R (ϕi ) according
to (11.4) in the following way:
μR∼
(ϕ i ) = μ G
∼
(ϕ i ) μ E
∼
(ϕ i ) = M IN μ G
∼
(ϕ i ) , μE
∼
(ϕ i ) . (11.5)
11.4 Fuzzy based Path Planning 273
First, the robot finds all sections where the radar distance value is more
than a threshold for finding a segment with less probability. Since there are
measurement noises, radar may not have the exact value of 1.
a) No Obstacles, if
μ∼ (ϕi ) ≤ threshold
Determining Obstacles in μR
∼
(ϕi ) = R
b) Obstacles, if
R (ϕi ) > threshold
μ∼
Then the robot chooses the segment which is the longest. Figure 11.6 shows
that segment (1) is longer than segment (2), so the robot will select a part (1).
Defuzzification of this segment will return the most appropriate direction
for the robot. In some cases, the choice of the defuzzification method is
limited by the necessity to take into account the multi-intervals core of the
resulting fuzzy set ∼R and its membership function μR ∼
(ϕi ) that practically
does not meet in the majority of the Mamdani-type fuzzy inference systems.
The multimodality presence explains that if the direct way to the target point
is blocked for the mobile robot by an obstacle, there can be some alternate
route concerning the shape of the specified obstacle.
Stage 4. Defuzzification. The last stage is the determination of the crisp value
of the robot’s direction ψ by defuzzification of the resulting fuzzy set R
∼
with
membership function μR ∼
(ϕi ), presented as an example in Figure 11.5:
' (
ψ = Arg Def uzzif ication μR ∼
(ϕi ) (11.6)
11.7 Conclusion
We propose a fuzzy-based approach enabling robots to navigate efficiently
and safely in uncertain environments.
Experimental results show the applicability of the proposed approach
for autonomous robot navigation in dynamic environments where mov-
ing obstacles are avoided efficiently, and the robot manages to reach its
destination.
In our experiments, the autonomous robot chooses a safe path for naviga-
tion. In situations where it needs to avoid collisions with dynamic obstacles,
the robot can adjust its speed to escape from dangerous situations as safely
as possible. We showed that changing rate resulted in having no collisions
between the robot and obstacles.
278 Safe Navigation of an Autonomous Robot in Dynamic
For future work, we would like to extend this work to a multi-robot setting
and investigate the robustness of sensor-based multi-robot control in an
uncertain environment [95–99] using hybrid artificial intelligence methods.
Acknowledgments
The authors thank the Fulbright Program (USA) for supporting Prof. Y. P.
Kondratenko with a Fulbright scholarship and making it possible for this team
to conduct research together in the USA at the Cleveland State University,
Electrical Engineering, and Computer Science Department.
References
[1] M. Derkach, D. Matiuk, I. Skarga-Bandurova, ‘Obstacle Avoidance
Algorithm for Small Autonomous Mobile Robot Equipped with Ultra-
sonic Sensors,’ in IEEE 11th Int. Conf. on Dependable Systems,
Services, and Technologies, DESSERT 2020, pp. 236–241, 2020.
[2] A.N. Tkachenko, N.M. Brovinskaya, Y.P. Kondratenko, ‘Evolutionary
adaptation of control processes in robots operating in non-stationary
environments,’ Mechanism and Machine Theory, Vol. 18, No. 4, pp.
275–278, 1983. DOI: 10.1016/0094-114X(83)90118-0
[3] Y. Kondratenko, G. Khademi, V. Azimi, D. Ebeigbe, M. Abdelhady,
S. Fakoorian, T. Barto, A. Roshanineshat, I. Atamanyuk, D. Simon,
‘Robotics and Prosthetics at Cleveland State University: Modern Infor-
mation, Communication, and Modeling Technologies,’ in Information
and Communication Technologies in Education, Research, and Indus-
trial Applications, ICTERI 2016. Communications in Computer and
Information Science, Vol. 783. Springer, Cham, pp. 133–155, 2017.
DOI: https://doi.org/10.1007/978-3-319-69965-3_8
[4] Y. Kondratenko, L. Klymenko, V. Kondratenko, G. Kondratenko,
E. Shvets, ‘Slip Displacement Sensors for Intelligent Robots: Solu-
tions and Models’, in Proc. of the 2013 IEEE 7th Int. Conf.
on Intelligent Data Acquisition and Advanced Computing Systems
(IDAACS), Berlin, vol. 2, Sept. 12–14, pp. 861–866, 2013. DOI:
10.1109/IDAACS.2013.6663050
[5] U. Patel et al., ‘Beam: A Collaborative Autonomous Mobile Service
Robot,’ in Proceedings of the AAAI Fall Symposium on Artificial
Intelligence for Human-Robot Interaction (AI-HRI), 2017.
References 279
New York: Begel House Inc., pp. 16–29, 2011. DOI: 10.1615/JAutomat-
InfScien.v43.i6.20
[16] R. Duro, Y. Kondratenko (Eds.), ‘Advances in Intelligent Robotics and
Collaborative Automation’, in Book Series: River Publishers Series in
Automation Control and Robotics, 2015. DOI: https://doi.org/10.13052
/rp-9788793237049
[17] O.S. Gerasin, et al. ‘Remote IoT-based control system of the mobile
caterpillar robot,’ in 16th International Conference on ICT in Educa-
tion, Research and Industrial Applications. Integration, Harmonization,
and Knowledge Transfer, Volume I: Main Conference, ICTERI 2020,
Kharkiv, CEUR Workshop Proceedings, Vol. 2740, pp. 129–136, 2020.
CEUR-WS.org/Vol-2740/20200129.pdf
[18] T. Huntsberger, A. Hrand, E. Baumgartner, S. Paul, ‘Behavior-based
control systems for planetary autonomous robot outposts,’ in Aerospace
Conference Proceedings, 2000 IEEE, vol. 7, IEEE, pp. 679–686, 2000.
[19] D. Simon, ‘Neural Networks for Optimal Robot Trajectory Planning,’ in
E. Fiesler, R. Beale (Eds), Handbook of Neural Computation, Institute
of Physics Publishing, pp. G2.5:1–8, 1997.
[20] T. Huntsberger, ‘Biologically inspired autonomous rover control,’ in
Autonomous Robots 11, No. 3, pp. 341–346, 2001.
[21] L.A. Zadeh, ‘Fuzzy sets. Information and control’, No. 8, pp. 338–353,
1965.
[22] L.A. Zadeh, ‘Outline of a new approach to the analysis of complex
systems and decision processes,’ in IEEE Transactions on Systems, Man
and Cybernetics, Vol. 3, No. 1, pp. 28–44, 1973.
[23] L.A. Zadeh, ‘Inference in fuzzy logic,’ in IEEE Proceedings, Vol. 68,
pp. 124–131, 1980.
[24] L.A. Zadeh, ‘From computing with number to computing with words –
from manipulations of measurements to manipulation with perceptions,’
in IEEE Transactions on Circuits and Systems Part I, Vol. 4, no. 3,
pp. 105–109, 1999.
[25] J. Kacprzyk, ‘Multistage Fuzzy Control: A Model-Based Approach to
Control and Decision-Making’, in Wiley, Chichester, 1997.
[26] Y.P. Kondratenko, N.Y. Kondratenko, ‘Reduced library of the soft com-
puting analytic models for arithmetic operations with asymmetrical
fuzzy numbers,’ in Alan Casey (Ed), Soft Computing: Developments,
Methods, and Applications. Series: Computer Science, technology
and applications, NOVA Science Publishers, Hauppauge, New York,
pp. 1–38, 2016.
References 281
[69] A.V. Savkin, C. Wang, ‘Seeking a path through the crowd: Robot naviga-
tion in unknown dynamic environments with moving obstacles based on
an integrated environment representation,’ in Robotics and Autonomous
Systems, Vol. 62, Issue 10, pp. 1568–1580, October 2014.
[70] T.K. Kang, et al., ‘Ego-motion-compensated object recognition using
type-2 fuzzy set for a moving robot’, in Neurocomputing, Vol. 120,
pp. 130–140, 2013.
[71] M. Wang, J.N.K. Liu, ‘Fuzzy logic-based real-time robot navigation in
an unknown environment with dead ends,’ in Robotics and Autonomous
Systems, Vol. 56, Issue 7, pp. 625–643, 2008.
[72] A. Pandey, D.R. Parhi, ‘MATLAB Simulation for mobile robot naviga-
tion with hurdles in cluttered environment using minimum rule based
fuzzy logic controller,’ in Procedia Technology, Vol. 14, pp. 28–34,
2014.
[73] P. Bigai, J. Bartoszek, ‘Low time complexity collision avoidance method
for autonomous mobile robot,’ in Intelligent Systems 2014, Proceed-
ings of the 7th Intern. Conference IS’2014, September 24–26, Warsaw,
Poland, Vol. 2, Springer, Cham – London, pp.141–152, 2014.
[74] P.G. Zavlangas, S.G. Tzafestas, ‘Industrial robot navigation and obstacle
avoidance employing fuzzy logic,’ in Journal of Intelligent and Robotic
Systems, 27(1–2), pp. 85–97, 2000.
[75] D. Simon, H. El-Sherief, ‘Real-Time Navigation Using the Global Posi-
tioning System,’ in IEEE Aerospace and Electronic Systems Magazine,
vol. 10, no. 1, pp. 31–37, 1995.
[76] A.V. Palagin, V. N. Opanasenko, ‘Reconfigurable-computing technol-
ogy,’ in Cybernetics and Systems Analysis 43, no. 5, pp. 675–686,
2007.
[77] Lin, Mingjie, et al., ‘Performance benefits of monolithically stacked 3-D
FPGA’, in Computer-Aided Design of Integrated Circuits and Systems,
IEEE Transactions on 26, no. 2, pp. 216–229, 2007.
[78] A. Drozd, et al. ‘Green Experiments with FPGA,’ in Green IT Engi-
neering: Components, Networks, and Systems Implementation, V.
Kharchenko, et al. (Eds), Vol. 105. Berlin, Heidelberg: Springer Interna-
tional Publishing, pp. 219–239, 2017. DOI: 10.1007/978-3-319-55595-
9_11
[79] A. Drozd, et al. ‘Checkable FPGA Design: Energy Consumption,
Throughput, and Trustworthiness,’ in Green IT Engineering: Social,
Business and Industrial Applications, Studies in Systems, Decision,
and Control, V. Kharchenko, et al. (Eds), Vol. 171. Berlin, Heidelberg:
286 Safe Navigation of an Autonomous Robot in Dynamic
[97] R. Zhao, H.-K. Lee, ‘Fuzzy-based Path Planning for Multiple Mobile
Robots in Unknown Dynamic Environment,’ Journal of Electrical Engi-
neering and Technology, vol. 12, no. 2, pp. 918–925, Mar. 2017.
[98] Z. Gomolka, et al., ‘From homogeneous network to neural nets with
fractional derivative mechanism.’ In: Rutkowski, L. et al. (Eds), Int.
Conf. on Artificial Intelligence and Soft Computing, ICAISC-2017, Part
I, Zakopane, Poland, 11–15 June 2017, LNAI 10245, Springer, Cham,
2017, pp. 52–63. DOI: https://doi.org/10.1007/978-3-319-59063-9_5
[99] G. Khademi, et al. ‘Hybrid invasive weed / biogeography-based opti-
mization,’ Engineering Applications of Artificial Intelligence, Vol. 64,
pp. 213–231, July 2017.
12
Algorithmic Procedures Synthesis
of Robust-Optimal Control
for Moving Objects
Abstract
The transient process control of a variety of moving objects described by
ordinary nonlinear differential equations requires development of robust-
optimal systems with variable structure. Using general algorithmic procedure
for constructing optimal trajectories, determining switching moments and
synthesizing control functions for multidimensional systems are suggested.
The control of the mismatch between the physical object trajectory and
the optimal model calculated trajectory enables to take the values of opti-
mal control into account and to form a robust subsystem, which provides
invariance to incomplete information about the moving object. Simulating
quadrotor UAV stabilization and sea vessel maneuvering under conditions of
external disturbances and parametric noise demonstrates the required control
accuracy in the vicinity of the formed optimal trajectory of the controlled
coordinates. The algorithms and circuit solutions for the synthesis of robust-
optimal variable-structure systems for multidimensional nonlinear moving
objects have been developed as a practical basis of automated procedures
of the synthesis of robust-optimal systems and development of software tools
for various types of moving object control systems.
289
290 Algorithmic Procedures Synthesis of Robust-Optimal Control
12.1 Introduction
The original classical works of R. Bellman, R. Kalman, N. Krasovsky, A.
Feldbaum, L. Pontryagin and other well-known scientists demonstrate the
rapid development of the foundations of modern control theory related to the
problems of analyzing desired (optimal) trajectories of object movement.
The basic requirements for the tasks of the optimal control synthesis
for quadratic criteria of optimality and maximum speed were mathemati-
cally formalized, supported by algorithms for the determination of switching
moments. It was a breakthrough in control systems, especially in the space
sector, where the highest accuracy and efficiency were required for well-
defined objects and given properties of the environment [1, 2]. It is well
known there are many control problems, e.g. motion under conditions of
aerodynamic and/or hydrodynamic resistance of a stochastic medium with
incompletely known parameters. However, the deviation of a mathematical
model from a moving physical object had so significant impact on control
errors that it stimulated the rapid development of methods of adaptive, and
later, robust control. Later, summing up the development of these methods,
V. Kuntsevich [3] classified them as adaptive-optimal methods requiring
identification of parameters and adjustment of control (which is rather dif-
ficult to provide in real-time), and robust-optimal methods that create control
algorithms with known stable solutions concerning bounded uncertainties.
These methods were further developed in subsequent works [4–8].
Other widely used methods based on robust properties of sliding modes
(considering the additional power of control for their implementation), but
frequent switching of control devices negatively effects their efficiency
[8–11]. Robust properties are also inherent in PID controllers, since they use
approximate tuning for the parameters of the plant model, respond to an error,
and compensate uncertain disturbances [12].
The complexity of the synthesis of multidimensional objects control was
revealed and caused the development of a new approach for forming the
desired trajectories in solving inverse problems of dynamics. The symmetry
of solutions of the motion modeling and control synthesis made it possible
to focus on developing the programmed trajectories of moving objects at
the initial stage of control synthesis. These methods simplify the synthesis
of optimal control since they do not require solving Riccati equations or
12.1 Introduction 291
water areas proves that about 70% of accidents occur due to human factor. It
is partly due to the lack of sufficient traffic safety data for operators to make
decisions in extreme operational conditions.
The use of UAV (quadrotor UAV) for marine environment monitoring
makes it possible to expand the navigation safety information assessment
and to increase the control of the marine moving objects improving in
this way the marine emergency prevention. It is necessary to improve the
quality control systems to ensure the optimal trajectory of stabilization of
the quadrotor UAV along all given controlled coordinates (the dynamics are
described by a system of six differential equations of the second-order [37,
38]). Quadrotor UAV functions under conditions of limited rotors power and
incomplete information of uncontrolled external disturbances (the complexity
of a complete description of the physical impact of wind perturbation on a
UAV involves the selection of the following components of airflow: high-
frequency random pulsations; low-frequency piecewise constant components
in horizontal (constant wind) and vertical directions [39]).
The LQR method is widely used for dynamic system stabilization as a
modern and effective method to synthesize optimal controllers for UAVs at
a minimal energy cost [40–45]. The comparison between LQR and the PID
algorithm shows that LQR has better energy efficiency than PID. At the same
time, LQR algorithms working well in conditions of noisy control processes
do not ensure accurate control in conditions of uncontrolled disturbance
compared to the control systems with the sliding modes [46–48]. The use
of control synthesis based on quadratic functionals and Lyapunov functions
[49–51] is limited to linear systems. Accordingly, the main problem of the
QFT method consist in the formation of feedback under the conditions of
uncertainty of disturbance parameters [52–54]. H 2 /H control synthesis is
increasingly used for the synthesis of robust controllers. This optimization-
based method is associated with the mathematical description of the expected
behavior of a closed system and its stability [55–60]. It has a strictly math-
ematical basis for the optimization process thus having been applied to both
classical optimal and robust control. There are some advantages of the linear
approach due to its powerful tools for the design and analysis of controllers.
This method considers the resistance to incoming perturbations (aerodynamic
moments and forces of wind) as the main source of errors in orientation
[61–64]. Neural networks are used to synthesize both robust and optimal
controllers [65, 66]. Wind disturbances are compensated by an artificial
neural network, the parameters of which are configured in the current mode of
operation. The controller is complemented by multilayer neural networks and
294 Algorithmic Procedures Synthesis of Robust-Optimal Control
the control object. The structure of the interaction between the object and the
environment is determined by the type of differential equations and is suffi-
ciently definite. The structure can be changed due to the correct simplification
of the system of equations, e.g. by replacing the nonlinear model with a linear
model. However, the values of the parameters (coefficients) of the model
of the controlled object are rather undefined values, and they describe an
incompletely informative model. While designing automatic control systems
for dynamic objects, the parameters of the mathematical model of the object,
which are the basis of synthesis, differ from the calculated ones in real
systems due to the impossibility to physically determine and mathematically
describe all factors that effect the value or function of the parameter. For
example, hydrodynamic drag coefficient and the attached mass of the vessel
depend, among other things, on the shape of the vessel hull, which can be
changed during operation (including overgrowing with marine organisms),
and the density of the marine environment, the value of which is also not
constant for different areas of operation of the vessel.
The disturbances under which the controlled object functions can be
divided into controlled (measured, physically measurable) and uncontrolled
(unmeasured), e.g. wind speed or wind direction that can be accurately
measured using an anemometric system. It should be noted that if the control
object does not allow the placement of the anemometric system on the
hull, e.g. a quadrotor UAV, then, in this case, the wind effect is uncontrol-
lable. Irregular sea waves that determine vessel pitching are uncontrollable
disturbances.
The system of equations for the generalized model of the nonlinear non-
stationary dynamic system of a moving object is described by Newtonian
mechanics in vector-matrix form
perturbations
Ẍ(t) = AẊ(t) + BU(t). (12.7)
Work during the movement of the object (in the absence of external
perturbations) along the trajectories of stabilization is performed by control
and damping (aerohydrodynamic resistance) forces. To estimate the energy
losses due to damping forces, the Rayleigh dissipative function must be
introduced T
T
D(Ẋ) = −0.5 Ẋ AẊdt. (12.8)
0
To optimize the problem of minimum energy consumption, lets consider
the Equations (12.7, 12.8), the minimized functional in the form of Lagrange
function can be written as
T
L = −0.5Ẋ (t)AẊ(t) + 0.5UT (t)QU(t) + λT [Ẍ(t) − AẊ(t) − BU(t)].
(12.9)
The extremum of functional (12.9) is determined by the following Euler–
Lagrange equation considering Ẋ(t) = V(t)
∂L d ∂L ∂L d ∂L
− = − = −AT Ẋ(t) − AT λ(t) − λ̇(t) = 0,
∂V(t) dt ∂ V̇(t) ∂ Ẋ(t) dt ∂ Ẍ(t)
(12.10)
and for control
∂L
= QU(t) − BT λ(t) = 0. (12.11)
∂U(t)
From the Equation (12.11), optimal control can be expressed as
Uopt (t) = Q−1 BT λ(t). (12.12)
When solving the Equation (12.10), it is assumed that polynomial rep-
resentation of the moving object trajectories is used. In this case, the object
can be at rest or perform a uniform (steady-state relative to the speed Ẋ(t)
with zero acceleration), or uniformly accelerated (steady-state relative to
acceleration Ẍ(t) with zero third derivative), or established for the third
...
derivative X (t) (with zero fourth derivative) and further, established for mth
(m)
derivative X (t) (with zero (m+1)-th derivative), motion.
(m) (m)
In general, for X (t) = X (0) = const the equation of one segment of
the trajectory for phase coordinate vector will take the form
! (i)
m−1
ti (m) tm
X(t) = X(0) ± X (0) . (12.13)
i! m!
i=0
12.3 Optimal Synthesis 299
Together with the equation for the controlled system and control (12.12)
Equation (12.15) forms a boundary value problem for optimal control, the
solution of which for a trajectory with zero third and fourth derivative can
also be described in the polynomial form
! (i)
m−2 s
s t
U(t) = U(ti ) i , (12.17)
i!
i=0
(i)
where order m and values U(tsi ) are limited by the physical constraints of the
controls.
For the polynomial representation of trajectories, each coordinate is inde-
pendent of the others (although in the general case considered below, they
300 Algorithmic Procedures Synthesis of Robust-Optimal Control
have different end times of the transient process). It should also be noted
that for a zero matrix A, the polynomial and exponential representations
of the trajectories coincide. Thus, the polynomial representation of the tra-
jectories, in contrast to the exponential one, makes it possible to construct
the moving object stabilization trajectories in advance (before the synthesis
of the control) rather simply. The polynomial representation of trajectories
is applicable for a complete mathematical description of the motion of the
moving object, i.e. considering nonlinearity and non-stationarity. With the
exponential representation of the trajectories, it is necessary to numerically
solve nonlinear non-stationary differential equations of a higher order which
describe the dynamics of the object.
The use of the polynomial form of trajectories makes it possible to
estimate the dependence of the optimality functional for the minimum energy
consumption on the time of the transient process. Consider the motion of the
moving object along the trajectory segments (12.4) for the 3rd and 4th zero
derivatives for the coordinate vector X(t), as well as the final conditions (for
simplicity, X0 = Ẋ0 = 0)
...
X(t) = 0.5Ẍ0 t2 ; Ẍ(t) = Ẍ0 = const; X (t) = 0; X(T ) = Xk ;
(12.18)
1 1 ... 1 ... ...
X(t) = Ẍ0 t2 + X0 t3 ; Ẋ(t) = Ẍ0 t + X0 t ; Ẍ(t) = Ẍ0 + X0 t;
2
2 6 2
... ... (IV )
X (t) =X0 = const; X (t) = 0; X(T ) = Xk ; Ẋ(T ) = Ẋk .
(12.19)
Then it is possible to obtain the necessary initial conditions, respectively,
for the 1st
Ẍ0 = 2T −2 Xk ,
and for the 2nd segment
...
Ẍ0 = T −2 (6Xk − 2Ẋk T ); X0 = T
−3
(6Ẋk T − 12Xk ).
Further, for the 1st segment, we obtain the equation for the initial conditions
for the control
Ẍ0 = AẊ0 + BU0 ⇒ U0 = B−1 Ẍ0 ;
...
−1
X0 = AẌ0 + BU̇0 = 0 ⇒ U̇0 = −B AẌ0 ,
and the general equation for the control in the form of
12.3 Optimal Synthesis 301
For the tasks of maximum speed, the trajectories are optimal relative to
the criterion (12.3) when a moving object moves with the maximum possible
number of maximum possible values of the derivatives of the coordinate
vector, taking into account the constraints on the control action
(i)
maxmax{Ẋ X (t0 ) , t },
i (i)
X (t0 )
(i)
and the fulfillment of all specified boundary conditions X(t0 ) =
(i−1)
f [X(t0 ), U max (t0 )].
Figure 12.1 Algorithmic procedure A for constructing the optimal trajectory in terms of
maximum speed
in general form (12.1), the parameters and external disturbances are precisely
known and specified.
The equations (12.17) define the representation of control functions in
the form of time polynomials, which are switched by certain algorithmic
procedures A and B, and is described by the block diagram shown in
Figure 12.3.
An alternative is to represent controls in the form of integrating func-
tions in the feedback loops of a moving object (Figure 12.4). Therefore,
we use the differential transformation of the vector-matrix Equation (12.1)
considering the above assumptions regarding disturbances, for example, the
third and fourth zero derivatives of the coordinate vector. This differential
304 Algorithmic Procedures Synthesis of Robust-Optimal Control
Figure 12.2 Algorithmic procedure B for the problem of minimum energy consumption
Figure 12.3 Block diagram of the implementation of control in the form of time polynomials
(F(t) is not considered; the index X for the matrices A, B is omitted)
(IV)
and for the condition X (t) = 0
...
AX X (t) + 2ȦX Ẍ(t) + ÄX Ẋ(t) + BX Ü(t) + 2ḂX U̇(t)+
+B̈X U(t) + CX F̈(t) + ĊX Ḟ(t) + C̈X F(t) = 0. (12.23)
U(tsi ) = B−1
X [Ẍ(ti ) − AX Ẋ(ti ) + g − CX F(ti )],
s s s
...
U̇(tsi ) = B−1
X [X (ti ) − AX Ẍ(ti ) − ȦX Ẋ(ti ) − ḂX U(ti ) − ĊX F(ti )+
s s s s s
Figure 12.4 Block diagram of the implementation of control in the form of integrating
functions in the feedback
model of the moving object motion and the corrective control signal formed
by a robust loop based on the mismatch between the optimal and physical
signal output. The integral signal of optimal and robust control is transferred
to a moving object propellers. The general constraint on the control action
should limit the values of the optimal and corrective control
|Uopt | + |Ucor | ≤ |Umax | .
The proposed approach for an admissible simplification of the synthesis of
robust control includes the linearization of the dynamics equations in the
vicinity of the nominal values of the parameters, which makes it possible
to apply the superposition principle to analyze control errors. The optimal
control and the specified trajectories of the moving object are formed, as
described above, taking into account the fact that the nonlinearity of the
model, and the discrepancy arising from the application of linearization can
be attributed to additional uncertainty, which is compensated by a robust
control.
The external disturbances F, some of which Fcont can be measured and
controlled, the parametric measurement noise η, the mismatch between the
parameters of the physical moving object, described by the matrix transfer
12.4 Robust Correction 307
function Wo (p), and the optimal mathematical model Wopt (p) are considered
to be uncertainties that are compensated by the robust loop.
Based on the structural diagram of a robust-optimal system (Figure 12.5),
the error vector can be determined as
E(p) = [I + Wo (p)Wcor (p)]−1 {[Wopt (p) − Wo (p)]Uopt (p)+
+ Wopt (p)Fcont (p) − Wo (p)F(p) − η(p)} = EU (p)+
+ EF (p) + Eη (p). (12.27)
In a more simplified form, the error vector can be represented in an
approximate form due to the unmeasurable part of the perturbations ΔF(p)
EF (p) ≈ [I + Wo (p)Wcor (p)]−1 Wopt (p)ΔF(p).
The requirements for the structure and parameters of a robust correction in the
form of the minimum H 2 norm [71] of the matrices of sensitivity functions
can be written for:
• input programmed optimal control based on the mismatch between the
parameters of the optimal model and the physical object
SU = [I + Wo (jω)Wcor (jω)]−1 [Wopt (jω) − Wo (jω)];
• the difference between the values of the vector of external disturbances
F(t) and their controlled part Fcont
SF = [I + Wo (jω)Wcor (jω)]−1 Wopt (jω);
308 Algorithmic Procedures Synthesis of Robust-Optimal Control
12.5 Experiments
12.5.1 Sea Vessel Maneuvering
To simulate the process of stabilizing the course angle ϕ (t) = x1 (t) and
angular speed ω (t) = x2 (t) of the vessel, controlled by changing the angle
of the rudder blade α(t), considering the values of the reduced aerohydro-
dynamic coefficients and external disturbance f (t) [27,31], let’s consider the
310 Algorithmic Procedures Synthesis of Robust-Optimal Control
U4 N1 − N2 + N3 − N4
C = diag (0.08, 0.036, 1.14, 0.008, 0.008, 0.004) .
Figure 12.14 Comparison of control accuracy between optimal control with variable
structure and PID control over time
Figure 12.15 Comparison of energy consumption between optimal control with variable
structure and PID control over time
and control, respectively, are shown by means of the index xyz), and system
with optimization by two higher coordinates x, y and a block of four PID
controllers for the remaining four coordinates (the graphs shown by mean
of the index xy), are given. The simulation results demonstrate a decrease in
energy consumption (values are given in arbitrary units) of control for the
proposed control system.
The simulation results (Figures 12.16, 12.17, the red line inside the "tube"
shows the optimal trajectory) demonstrates the errors change in the form of
ńtubesż for different values of external uncontrolled wind disturbance.
316 Algorithmic Procedures Synthesis of Robust-Optimal Control
12.6 Conclusion
Optimal control systems provide moving objects with the highest control
quality indicators in terms of speed, accuracy, and energy consumption. In the
case of multidimensional nonlinear systems, there are significant restrictions
on the applicability of classical optimal control methods associated with the
computational difficulties due to boundary value problems complex solutions,
the correctness of the formation of weight matrices of optimality criteria and
others.
To optimize nonlinear systems up to the sixth order, the proposed algorith-
mic procedure of the synthesis of variable structure system for different types
of moving objects based on preliminary construction of optimal trajectories of
stabilization of dynamic processes has been applied. The generated minimax
optimality conditions for various criteria allow to consider the constraints on
the control actions and solve the synthesis problem for the corresponding
functional control problems: maximum speed (for critical modes) and mini-
mum energy consumption (for technological modes). The uncertainty of the
external environment and the dynamic object requires robust correction of the
optimal system. The generated optimal trajectories and control determine the
formation of the additional robust control loop to compensate incomplete a
priori definiteness of the mathematical model and uncontrolled (unmeasured)
disturbances and noises.
The proposed procedure for the synthesis of robust-optimal systems has
been tested for several models of a nonlinear moving object and demonstrated
the versatility and engineering practicality, as well as efficiency under the
influence of external limited disturbances in terms of accuracy and con-
trol energy consumption. A number of algorithmic procedures for optimal
trajectories construction, switching moments determination, and synthesis
of control functions for multidimensional systems have been developed.
Controlling the mismatch between the trajectory of a physical object and
the calculated optimal trajectory of the reference model allows to form a
robust subsystem that provides invariance to incomplete information about
the dynamic system taking into account the values of the optimal control from
the reference system. The presented results of the simulation of quadrotor
UAV stabilization and maneuvering of a vessel under the conditions of
external disturbance and parametric noise demonstrate the provision of the
318 Algorithmic Procedures Synthesis of Robust-Optimal Control
References
[1] Krasovsky, N.N.: Some problems of the theory of stability of motion,
Phizmathlit, Moscow (1959) (in Russian)
[2] Feldbaum, A.A.: Theory Of Optimal Systems, Phizmathlit, Moscow
(1963) (in Russian)
[3] Kuntsevich, V.M.: Synthesis of Robust Optimal Adaptive Control Sys-
tems for Nonstationary Objects under Bounded Disturbances. In: Jour-
nal of Automation and Information Sciences, NY, Begell house inc., vol.
36, issue 3, pp.14–24 (2004)
[4] Kuntsevich, V.M., et al (Eds): Control Systems: Theory and Applica-
tions. Series in Automation, Control and Robotics, River Publishers,
Gistrup, Delft, (2018)
[5] Kondratenko, Y.P., Kuntsevich, V.M., Chikrii, A.A., Gubarev, V.F.
(Eds): Advanced Control Systems: Theory and Applications. Series in
Automation, Control and Robotics, River Publishers, Gistrup, (2021)
[6] Kondratenko, Y.P., Chikrii, A.A., Gubarev, V.F., Kacprzyk, J. (Eds):
Advanced Control Techniques in Complex Engineering Systems: The-
ory and Applications. Dedicated to Professor Vsevolod M. Kuntsevich.
Studies in Systems, Decision and Control, Vol. 203. Cham: Springer
Nature Switzerland AG (2019)
[7] Kuntsevich, V.M.: Estimation of Impact of Bounded Perturbations on
Nonlinear Descrete Systems. In: Kuntsevich, V.M. et al. (Eds). Control
Systems: Theory and Applications. Series in Automation, Control and
Robotics, River Publishers, Gistrup, Delft, pp. 3–15 (2018)
[8] Emelynov, S.V., Korovin, S.K.: New types of feedback, Nauka, Phiz-
mathlit, Moscow, (1997) (in Russian)
[9] Emelynov, S.V.: Automatic control systems of variable structure: syn-
thesis of scalar and vector systems by state and by output. In: Nonlinear
dynamics and control, Phizmathlit, Moscow, vol. 5, pp. 5–24 (2007) (in
Russian)
[10] Utkin, V.I.: Sliding Modes in Control and Optimization, Springer-
Verlag, (1992)
[11] Shtessel, Y.B., Moreno, J.A., Fridman, L.M.: Twisting sliding mode
control with adaptation: Lyapunov design, methodology and applica-
tion. In: Automatica, vol. 75, pp. 229–235 (2017)
References 319
[12] Dorf, R.,. Bishop, R.: Modern Control Systems, LBZ, Moscow (2002)
(in Russian)
[13] Boychuk, L.M.: Method of structural synthesis of nonlinear automatic
control systems, Energia, Moscow (1971) (in Russian)
[14] Galiulin, A.S.: Methods for solving inverse problems of dynamics,
Moscow, Nauka, (1986) (in Russian)
[15] Krutko, P.D.: Robust Stable Structures of Controlled Systems of High
Dynamic Accuracy. Algorithms and dynamics of motion control of
model objects. In: Control theory and systems, no. 2, pp. 120–140
(2005) (in Russian)
[16] Krutko, P.D.: Inverse problems of dynamics in the theory of automatic
control, Mashinostroenie, Moscow, (2004) (in Russian)
[17] Kondratenko, Y.P., Timchenko,V.L.: Optimal feedback switching
method for linear control system. In: Intern. System and Networks:
Mathematical theory and applications, Berlin, Academie Verlag, Vol. 2,
pp. 291–292 (1994)
[18] Timchenko, V.L., Lebedev, D.O., Robust-optimal stabilization of non-
linear dynamic systems. In: Radio Electronics, Computer Science,
Control, no. 3., pp. 185–197 (2019)
[19] Gabasov, R., Kirillova, F.M., Ruzhitskay, E.A.: Implementation of
Bounded Feedback in a Nonlinear ontrol Problem. In: Cybernetics and
Systems Analysis, no. 1, pp. 108–116 (2009) (in Russian)
[20] Balashevich, N.V., Gabasov, R., Kalinin, A.I. and Kirillova, F.M.: Opti-
mal Control of Nonlinear Systems. In: Computational Mathematics and
Mathematical Physics, vol. 42, no. 7, pp. 931–956 (2002)
[21] Horowitz, I.M.: Survey of quantitative feedback theory (QFT). In: Int.
Journal of Robust and Non-Linear Control, vol. 11., no. 10., pp. 887–921
(2001)
[22] Larin, V.B.: On the reversal of the problem of analytical design of
controllers. In: Control and Information Promblems, no. 1, pp. 17–25
(2004) (in Russian)
[23] Yaesh, I., Boyarski, S., Shaked U.: Probability-guaranteed robust H∞
performance analysis and state-feedback design. In: Systems & Control
Letters, vol, 48, issue 5, pp. 351–364 (2003)
[24] Chesnov, V.N.: Synthesis of H∞ - controllers for Multidimensional
Systems of Given Accuracy and Stability. In: Autom. Remote Control,
vol. 72, no. 10, pp. 2161–2175 (2011)
[25] Roberts, G.N.: Trends Marine Control Systems. Im: IFAC, CAMS’07,
Zagreb, Croatia (2007)
320 Algorithmic Procedures Synthesis of Robust-Optimal Control
[26] Ornerdi, c E., Robert,s G.N.: A fuzzy track-keeping autopilot for ship
steering. In: Journal of Marine Engineering and technology, London,
no. A2, pp. 23–35 (2003)
[27] Lukomsky, Y.K., Chugunov, V.S.: Marine mobile objects control sys-
tems, L., Sudostroenie, (1988) (in Russian)
[28] Fossen, T.I., Perez, T.: Kinematic models for maneuvering and seakeep-
ing of marine vessels. In. Journal of modeling, identification and control,
vol. 28, issue 1, pp. 19–30 (2007)
[29] Johansen, T., Fossen, T.I.: Control allocation—A survey. In. Automat-
ica, vol.49, pp. 1087–1103 (2013)
[30] Comasòlivas, R., Escobet, T. and Quevedo, J.: Automatic design of
robust PID controllers based on QFT specifications. In: Proceeding of
IFAC Conference on Advances in PID Control, pp. 715–720 (2012)
[31] Timchenko, V.L., Kondratenko, Y.P.: Robust stabilization of marine
mobile objects on the basis of systems with variable structure of feed-
backs. In: Intern. Journal of Automation and Information Sciences, NY,
Begell house inc., vol. 43, issue 6, pp. 16–29 (2011)
[32] Timchenko, V.L.: Synthesis of variable structure systems for stabi-
lization of ships at incomplete controllability. In: Intern. Journal of
Automation and Information Sciences, NY, Begell house inc., vol. 44,
issue 6, pp. 8–19 (2012)
[33] Timchenko, V.L., Ukhin,O.A.: Optimization of Stabilization Processes
of Marine Mobile Object in Dynamic Positioning Mode. In: Intern.
Journal of Automation and Information Sciences, NY, Begell house inc.,
vol. 46, issue 7, pp. 40–52 (2014)
[34] Timchenko, V.L., Ukhin, O.A.: Variable Structure Robust-Optimal Sys-
tems for Control of Marine Vehicles. In: Proceeding of the IEEE 4th
International Conference MSNMC, Kyiv, Ukraine, pp. 151–155 (2016)
[35] Timchenko, V.L., Lebedev, D.O.: Automated Algorithmic Procedure of
Robust-Optimal Control System’s Synthesis for Marine Vehicles. In:
Proceeding of the IEEE 4th International Conference MSNMC, Kyiv,
Ukraine„ pp. 98–101 (2018)
[36] Timchenko, V.L., Ukhin, O.A. and Lebedev, D.O.: Optimization of
nonlinear systems of variable structure for control of marine moving
vehicles. In: Intern. Journal of Automation and Information Sciences,
NY, Begell house inc., vol. 49, issue 7, pp. 33–47 (2017)
[37] Altug, E., Ostrowski, J.P. and Taylor. C.J.: Control of a quadrotor heli-
copter using dual cameravisual feedback. In: The International Journal
of Robotics Research, vol. 24, no. 5, pp. 329–341 (2005)
References 321
[38] Pyrkin, A.A., Maltseva, T.A., Labadin, D.V., Surov, M.O. and Bobtsov,
A.A.: Sintez sistemy upravleniya kvadrokopterom s ispol’zovaniyem
uproshchennoy matematicheskoy modeli. In: Izv. Vuzov. Priborostroye-
niye, RF, SPb, vol. 56, no. 4, pp. 47–51 (2012) (in Russian).
[39] Andreev, M.A., Miller, B., Miller, B.M. and Stepanyan, K.V.: Path
planning for unmanned aerial vehicle under complicated conditions and
hazards. In: International Journal of Computer and Systems Sciences,
vol. 51, no. 2, pp. 328–338 (2012)
[40] Younes, Y.A., Drak, A., Noura, H. et al.: Robust Model-Free Control
Applied to a Quadrotor. In: J. Intell. Robotics Syst., 84, pp. 37–52 (2016)
[41] Larin, V.B., Tunik, A.A.: Synthesis of the quad-rotor flight control sys-
tem. In: Proceeding of the IEEE 4th International Conference MSNMC,
Kyiv, Ukraine, pp. 12–17 (2016)
[42] Wang, D., Liu, D., Li, H., et al.: An Approximate Optimal Control
Approach for Robust Stabilization of a Class of Discrete-Time Non-
linear Systems With Uncertainties. In: IEEE Transactions on Systems,
Man, and Cybernetics: Systems, vol. 46, no. 5, pp. 713–717 (2016)
[43] Lu, T., Wen, P.: Time Optimal and Robust Control of Twin Rotor
System. In: Proceeding of the IEEE International Conference on Control
and Automation, Guangzhou, China, pp. 862–866 (2007)
[44] Faruque, I.A., Muijres, F.T., Macfarlane, K.M. et al.: Identification of
optimal feedback control rules from micro-quadrotor and insect flight
trajectories. In: Biol Cybern., 112, pp. 165–179 (2018)
[45] Kamran, J.M., Ahmed, S.F., Bakar, M.I., Ali, A.: Horizontal Motion
Control of Underactuated Quadrotor Under Disturbed and Noisy
Circumoyostances, Information and Communication Technology. In:
Advances in Intelligent Systems and Computing, Springer, vol. 625.,
pp. 63–79 (2018)
[46] Jayakrishnan, H.J.Z: Position and Attitude control of a Quadrotor UAV
using Super Twisting Sliding Mode. In: IFAC-PapersOnLine, vol. 49,
issue 1, pp. 284–289 (2016)
[47] Moreno, J.A., Osorio, M.: A Lyapunov approach to second-order sliding
mode controller and observers. In: Proceeding of the 47th Conference on
Decision and Control, pp. 2856–2861 (2008)
[48] Matveev, A.S., Teimoori, H., Savkin, A.V.: A method for guidance
and control of an autonomous vehicle in problems of border patrolling
and obstacle avoidance. In: Automatica, vol. 47, issue 3, pp. 515–524
(2011)
322 Algorithmic Procedures Synthesis of Robust-Optimal Control
[49] Leira, F.S., Johansen, T.A. and Fossen,T.I.: Automatic detection, classi-
fication and tracking of objects in the ocean surface from UAVs using a
thermal camera. In: Proceeding of Aerospace Conf., pp. 1–10(2015)
[50] Liu, Y., Chen, C., Wu, H. et al.: Structural stability analysis and opti-
mization of the quadrotor unmanned aerial vehicles via the concept of
Lyapunov exponents. In: Int J Adv Manuf Technol 94, pp. 3217–3227
(2018)
[51] Liu, Y., Li X., Wang, T., Zhang, Y., Mei, P.: The Stability Analysis
of Quadrotor Unmanned Aerial Vechicles. In: Wearable Sensors and
Robots. Lecture Notes in Electrical Engineering, Springer, vol. 399,
pp. 383–394 (2017)
[52] Liu, Y., Li, X., Wang, T., et al.: Quantitative stability of quadrotor
unmanned aerial vehicles. In: Nonlinear Dyn, pp. 1819–1833 (2017)
[53] Gharib, M. R., and Moavenian, M.: Full dynamics and control of a
quadrotor using quantitative feedback theory. In:Int. J. Numer. Model,
29, pp. 501–519 (2017)
[54] Xu, Y., Tong, Ch.: Quantitative feedback control of a quadrotor. In:
Proceedings of the IEEE International Symposium on Industrial Elec-
tronics, Hangzhou, China, pp. 1309–1314 (2012)
[55] Raffo, G. V., Ortega M. G., Rubio F. R.: An integral predictive/nonlinear
H∞ control structure for a quadrotor helicopter. In: Automatica, vol. 46,
issue 1, pp. 29–39 (2010)
[56] Viegas, D., Batista, P., Oliveir,a.P., Silvestre, C.: Decentralized H2
observers for position and velocity estimation in vehicle formations
with fixed topologies. In: Systems & Control Letters, vol. 61, issue 3,
pp. 443–453 (2012)
[57] Falkenberg, O., Witt, J., Pilz, U., Weltin, U., Werner, H.: Model Identi-
cation and H∞ Attitude Control for Quadrotor MAV’s. In: ICIRA, Part
II, pp. 460–471 (2012)
[58] Sushchenko, O.A., Shyrokyi, O.V.: H 2 /H-Optimization of system for
stabilization and control by line-of-sight orientation of devices operated
at UAV. In: Proceedings of the IEEE 3rd International Conference
APUAVD, pp. 235–238 (2015)
[59] Zhang, X., Kamgarpou,r M., et al.: Robust optimal control with
adjustable uncertainty sets. In: Automatica, vol. 75, pp. 249–259 (2017)
[60] Sushchenko, O.A.: Robust control of angular motion of platform with
payload based on H∞ -synthesis. In: Journal of Automation and Infor-
mation Sciences, vol. 48, issue 12, pp. 13–26 (2016)
References 323
[61] Satici, A.C., Poonawala, H., Spong, M.W.: Robust Optimal Control of
Quadrotor UAVs. In: IEEE Access, vol. 1, pp. 79–93 (2013)
[62] Vossen, G., Maurer, H.: On L1-minimization in optimal control
and applications to robotics. In: Optim. Control Appl. Meth., 27,
pp. 301–321 (2006)
[63] Chen, Y., Han, J., & Zhao, X., Three-dimensional path planning for
unmanned aerial vehicle based on linear programming. In: Robotica,
30(5), pp. 773–781 (2012).
[64] Jiang, B., Bishop, A.N., Anderson, B. D.., Drake S.P.: Optimal path
planning and sensor placement for mobile target detection. In: Auto-
matica, Vol. 60, pp. 127–139 (2015)
[65] Lewis, F.L., Vrabie, D., Vamvoudakis, K. G.: Reinforcement Learning
and Feedback Control: Using Natural Decision Methods to Design.
Optimal Adaptive Controllers. In: IEEE Control Systems Magazine,
vol. 32, no. 6, pp. 76–105 (2012)
[66] Bisheban, M., Lee T.: Geometric Adaptive Control With Neural Net-
works for a Quadrotor in Wind Fields. In: Proceeding of the IEEE
Transactions on Control Systems Technology, pp. 1–19 (2019)
[67] Timchenko, V.L., Lebedev, D.O., Kuklina, K.A. and Timchenko, I.V.:
Robust-optimal control system of quad for maritime traffic’s monitoring.
In: Proceeding of the IEEE 4th international conference APUAVD,
Kyiv, Ukraine„ pp. 192–196 (2017)
[68] Timchenko, V.L., Lebedev, D.O., Optimization of Processes of Robust
Control of Quadcopter for Monitoring of Sea Waters. In: Journal of
Automation and Information Sciences, NY., Begell house, vol. 51,
issue. 2, pp. 1–10 (2019)
[69] Timchenko, V.L., Lebedev, D.O.: Robust control systems with subopti-
mal models for stabilization of UAV. In: Proceedings of the 2019 IEEE
5th International Conference APUAVD, Kyiv, Ukraine, pp. 124–129
(2019)
[70] Timchenko, V.L., Lebedev, D.O.: Synthesis of Optimal Phase Trajec-
tory for Quadrotor UAV Stabilization. In: Proceedings of the IEEE 6th
International Conference MSNMC, Kyiv, Ukraine, pp. 20–24 (2020)
[71] Sumacheva, V.A., Kharitonov, V.L.: Calculation of the H2 Norm of the
Transfer Matrix of a System of Neutral Type. In: Differential Equations
And Control Processes, vol. 4, pp. 22–32 (2014)
Part III
Recent Developments in Collaborative Automation
13
Modeling of Cyber-Physical Systems
Abstract
The purpose is to review current research in the field of constructing cyber–
physical systems. The structure of the chapter is as follows: a review of
the most effective approaches and methods of constructing cyber–physical
systems; using these methods—design cyber–physical–logistical system; per-
form the review of the obtained results and come up with a plan for further
researches based on the reviewed methods. For validation of the designed
system we came up with a live scenario from the logistical systems: the
autonomous object delivers goods from one warehouse to another. We also
will model obstacles along the way. The system should proactively react to
this and recalculate the optimal path. This will lead to the timely delivery of
goods to the destination.
13.1 Introduction
Cyber–physical systems are the integration of computational and physical
processes. Controllers and computer networks control physical processes,
usually through feedback, which is also affected by physical processes and
vice versa.
327
328 Modeling of Cyber-Physical Systems
The economic and social potential of such systems is much greater than
that of their counterparts, which is why the main global investment goes to the
development of this technology. Nevertheless, there are significant problems,
especially because the physical components of such systems require certain
security and reliability requirements that are qualitatively different from those
for general purpose computing.
Logistics is a set of entities involved in the delivery of goods and
services to the end customer. Logistics management is the effective end-to-
end process management, which includes: 1) planning the optimal delivery
route; 2) optimal appointment of “performers” (autonomous objects that carry
out delivery); 3) coordination of system subjects among themselves; 4) risk
forecasting; 5) timely response to emergency situations. The logistics system
faces problems that can potentially lead to inefficiencies and waste in supply
chains, such as delayed deliveries, rising fuel costs, inconsistent suppliers,
and ever-increasing end-user expectations.
that capture information flows between this element and the elements with
which it is associated with the help of the above relations.
The definition of the system, in this case, will be as follows. Let there be
a system
{P, C, CPS, Rp , Rc } (13.1)
where P is the set of physical components that are related by the relation-
ship Rp ;
1) C—cyber components that are related by the relationship Rc ;
2) CPS—a cyber–physical system, including as a child, having the same
properties as the parent.
Figure 13.2 shows an example of building a composite cyber–physical sys-
tem. It includes three child cyber–physical systems (CPS1, CPS2, CPS3) as
part of one parent (CPS4); each of the systems has at least one physical (P)
and one cyber component (C); communication occurs only between com-
ponents of the same type (physical component physical component, cyber
component ⇔ cyber component).
The main focus is on the “form a single CPS” relationship between the
“Cyber Component” and “Physical Component” classes, which relate to each
other as “one to one” (one cyber component to one physical component).
As stated in the metamodel, the “Cyber Component” and “Physical Com-
ponent” classes are subclasses of the “CPS Component” base class, which
contains the fundamental properties. Each “CPS Component” can either act
as a communicator between two components, or be an independent unit and
communicate with other homogeneous components. Communication is done
via input and output interfaces.
The next approach to the description of cyber–physical systems is the use
of Business Process Modeling Notation (BPMN 2.0), taking into account the
specifics of cyber–physical systems. Classifying new metaclasses requires a
deep understanding of BPMN metaclasses [2, 11–14].
For this, new abstract classes of cyber–physical systems are introduced as
a set of BPMN subclasses. This concept will allow the developer of a cyber–
physical system to describe the logic of internal processes, various types of
actions inside and real objects.
In a cyber–physical system, business processes must be defined accord-
ing to a centralized approach to execute business logic. BPMN defines the
behavior of different participants using different pools.
The first pool is for physical processes. The second pool is for cyber pro-
cesses. The third pool is a central controller that directly coordinates the inter-
action between cyber and physical processes. Therefore, the BPMN4CPS
process must consist of at least three pools: physical, controller, and cyber.
The separation into the processing logic, executed at the physical level,
the controller and the cyber part, forces the modeler to explicitly indicate the
interaction between them using messages. Data objects provide information
about what physical actions need to be performed and/or what information
they generate. Data objects are the temporarily stored data of a running
process instance.
Figure 13.4 shows an example illustrating the execution of a closed loop
of a cyber–physical task in the context of three pools: physical, controller,
and cyber process pools.
In cyber–physical systems, there are three types of tasks: cyber, man-
ual, and physical. These types of actions are also covered in the BPMN
metamodel.
Cybertask is an extension of the standard BPMN service task. This task is
performed at the cyber level and, in essence, is the support for the controller
in the formation of the control action. For example, solving optimization
332 Modeling of Cyber-Physical Systems
problems, the cyber level finds a solution to a certain problem, which it sends
to the controller.
A physical task is an atomic activity within a process. The physical
process can be divided into two parts:
1) retrieve sensor data and send them to the controller level;
2) receipt by executive devices of control actions from the controller.
In this model, the controller plays the role of a kind of “mediator” and
“aggregator”. The role of the “mediator” is that the controller receives data
from the sensors and the cyber layer and sends it to the physical layer.
13.2 Review of Modeling Methods for Cyber–Physical Systems 333
The role of the “aggregator” is that having stabilized the system in the event
of a disturbance and having received data from the cyber level, the controller
brings this data into a single form that is understandable for executive devices.
The third approach to modeling cyber–physical systems is the model-
based design method for constructing dynamic systems [3, 15–17].
Designing a complex cyber–physical system, especially with heteroge-
neous subsystems distributed over networks, is challenging. Commonly used
design methods are complex and include mathematical modeling of physical
systems, formal computation models, modeling of heterogeneous systems,
software synthesis, verification, validation, and testing.
The proposed algorithm consists of 10 steps:
1) problem statement;
2) description of physical processes by equations;
3) description of minimum/maximum parameters and/or restrictions;
4) synthesis of control;
5) selection of computing software;
6) choice of hardware;
7) modeling of the process;
8) design;
9) software development;
10) performance check and testing.
Let’s briefly describe each step of the algorithm.
Problem statement.
A preparatory step that requires maximum attention. In this step, a basic
understanding of the problem is formed. For the most part, this step involves a
written description of the problem, without the use of complex mathematics.
Description of physical processes by equations.
In this step, an understanding of the relevant physical systems is formed,
such as the environment in which the cyber–physical system is located, or
the physical processes that need to be controlled.
Models of physical processes are simplified representations of real
systems and usually take the form of systems of differential equations.
Description of minimum/maximum parameters and/or restrictions;
At this stage, need to highlight the constants, adjustable parameters, and
variables that need to be controlled.
334 Modeling of Cyber-Physical Systems
Also need to determine the values that characterize the system and/or
physical processes, such as minimum and maximum values, initial states,
final states.
Synthesis of control.
This step requires determining the conditions under which the physical pro-
cesses are controllable and synthesizing the control law. It is also necessary
to define the parameters of time delays, sampling rate, and quantization
so that the physical dynamics of interest can be accurately measured and
appropriately controlled.
Selection of computing software.
The formal computation model defines semantics that often leads to greater
analyzability and the ability to model cyber–physical systems using hetero-
geneous modeling tools.
Models described by formal computation models are easier to analyze in
terms of determinism, runtime, state reachability, and memory usage.
Choice of hardware.
When having a basic understanding of all the components of the system, it is
necessary to select the components from which the production model will be
built.
The selection of equipment should be based on the fact that the finished
model should adequately interact with the simulated physical systems and
implement the control law in a high-precision measure.
For each component, consider its input and output bandwidth, input-to-
output latency, power consumption, resolution, and measurement speed, as
well as mechanical parameters such as form factor, electrical noise rejection,
durability, and lifespan.
Modeling of the process.
This step allows us to verify the adequacy of the developed model of the
cyber–physical system. It is necessary to collect quantitative and qualitative
parameters to make sure that the optimal control law takes the system from
the initial state to the final one. Should also check that all limits and ranges
for each module are met.
Design.
This involves assembling the production model itself according to the spec-
ifications. It is best to build a model so that tests can be run at each step
13.3 Problem Formulation 335
half indicates the cyber component. Accordingly, the “form a single CPS”
relation is represented by combining the half ellipses into a whole ellipse.
Some properties of a cyber–physical system can be directly deduced from
Figure 13.6 by the type of connection between cyber–physical components.
For example, a cyber–physical component that has no physical inputs and
outputs and only interacts in the cybernetic dimension can be thought of as a
computing node. In this notation, “System” is a computational node.
In addition, a component that receives only physical input and provides
only cyber output can be called a sensor. A radar that interacts with ODG
through physical signals and provides a cyber response on the state of the
road. In the metamodel, it can be referred to as sensors.
Figure 13.7 shows the implementation of a model of a logistics cyber–
physical system in BPMN4CPS terminology.
Here, the extended BPMN notation divides the processing logic into three
parts: the cyber part, the controller, and the physical part.
These parts interact with each other through the exchange of messages.
The set of components here is as follows
1) ODG;
2) radar installed on the ODG;
338 Modeling of Cyber-Physical Systems
3) controller;
4) the cyber component of the system responsible for data processing
(cloud computing resources).
This notation clearly shows the relationship between components. ODG is a
basic component.
The level of the physical process is a radar as a sensor device and a
steering wheel system as an executive component. The radar tracks objects
in the environment and the distance to them. Having received information
about the next object from the radar, the controller sends data to the cloud
13.4 Building Models of a Logistic Cyber–Physical System 339
where
e(t) = z(t) − y(t) (13.19)
e(t) – difference between desirable z(t) and real y(t) outputs. Rewrite e(t)
as a function of z(t) and x(t):
Substitute the error (13.19) into the optimality criterion (13.18) and find
the Hamiltonian H of the system:
1 1
H = [z(t) − C(t)x(t)]T Q(t)[z(t) − C(t)x(t)] + uT (t)R(t)u(t)p0 (t)
2 2
+ A(t)T x(t)T p(t) + B(t)T u(t)T p(t). (13.21)
The vector g(t) (with n components) is the result of the solution of the
differential equation:
g(t) = −[A(t) − B(t)R−1 (t)BT (t)K(t)]T g(t) − CT (t)Q(t)z(t) (13.40)
Figure 13.8 Primary trajectory obtained by ODG at the initial moment of time
346 Modeling of Cyber-Physical Systems
Figure 13.10 The state of the cyber–physical system at the moment the radar detects an
obstacle (the state of the radar (a), the state of the ODG (b).
348 Modeling of Cyber-Physical Systems
Figure 13.11 MATLAB script responsible for the operation of the radar and movement of
the ODG.
Figure 13.12 New logistics route recalculated due to the occurrence of an obstacle.
13.5 Simulation Results 349
Figure 13.13 MATLAB script showing the characteristics of the radar and the obstacle
tracking algorithm.
350 Modeling of Cyber-Physical Systems
Figure 13.14 The state of the cyber–physical system after receiving a new logistic route (the
state of the radar (a), the state of the ODG (b).
13.6 Conclusion 351
13.6 Conclusion
The chapter discusses the most promising in terms of completeness of
description approaches to modeling cyber–physical systems. Not only the
analysis of modeling approaches to cyber–physical systems is carried out but
also the results of building a model for the selected object are presented.
The FCA metamodel approach is excellent for building a high-level
view of the cyber–physical system under development. The result gives a
clear understanding of what systems/functional blocks the cyber–physical
system consists of and how these nodes are interconnected (using cyber or
physical links). Based on the types of incoming and/or outgoing links, you
can determine the type of object. For example, a cyber–physical component
that has no physical inputs and outputs and only interacts in the cybernetic
dimension can be thought of as a computing node. In addition, a component
that receives only physical input and provides only cyber output can be called
a sensor.
The BPMN approach is great for the next design phase—describing
the internal interactions between systems/functional blocks. If the previous
approach showed which components in the system are physical and which are
cyber, then this approach illustrates in detail how these components interact
with each other. What is the nature of the original physical signal from the
steering wheel system to the ODG? Or, what type of cyber signal comes from
the radar to the level of controller interaction. The advantage of this approach
is precisely in detailing the “hidden” interactions between components and
how these interactions form a closed loop of data flow from cyber elements
to physical ones and vice versa.
The last so-called model-based approach, perhaps, in part, is a continu-
ation of the previous one. However, it should be performed most effectively
in conjunction with it. There are many model-based steps worth completing
before starting to use the BPMN-based method. For example, “problem
statement,” “description of physical processes by equations,” “Description of
minimum/maximum parameters and/or restrictions.” Completing these steps
will adequately describe the internal interactions between the components.
The main disadvantage of this approach is the absence of any graphical
representations of intermediate results in the process of building a cyber–
physical system. In turn, its great advantage is algorithmic. Performing each
step, you can get the entire set of data, characteristics, details of technical
implementation for adequate testing of the performance of the developed
cyber–physical system.
352 Modeling of Cyber-Physical Systems
Each approach has its advantages and disadvantages. The authors see the
sense in some symbiosis of these approaches. The first approach is ideal
for the stage of initial prototyping of the problem that has arisen. Follow-
ing the first steps of the third approach “description of physical processes
by equations” and “Description of minimum/maximum parameters and/or
restrictions”, it is possible to collect the necessary set of values/data to
describe the internal interactions between the components of a cyber–physical
system following the second approach. Further, the development of a cyber–
physical system should be continued according to the model-based approach
algorithm.
References
[1] D. Morozov, et al. ‘Multi-paradigm modeling of cyber–physical
Systems’, IFAC-PapersOnLine, 51(11), 2018, pp. 1385–1390. DOI:
https://doi.org/10.1016/j.ifacol.2018.08.334
[2] Graja, et al. ‘BPMN4CPS: A BPMN Extension for Modeling cyber–
physical Systems’, 2016 IEEE 25th International Conference on
Enabling Technologies: Infrastructure for Collaborative Enterprises
(WETICE), 2016, DOI: https://doi.org/10.1109/wetice.2016.41
[3] J. C. Jensen, et al. ‘A model-based design methodology for cyber–
physical systems’, 2011 7th International Wireless Communications and
Mobile Computing Conference, 2011, DOI: https://doi.org/10.1109/iw
cmc.2011.5982785
[4] A. Holovatenko, et al. ‘Energy-Efficient Path-Following Control System
of Automated Guided Vehicles’, J Control Autom Electr Syst 32, 2021,
pp. 390–403. DOI: https://doi.org/10.1007/s40313-020-00668-8
[5] A. Chikrii. Control of Moving Objects in Condition of Conflict. In:
Kuntsevich, V.M. et al. (Eds). Control Systems: Theory and Applica-
tions. Series in Automation, Control and Robotics, River Publishers,
Gistrup, Delft, 2018, pp. 17–43.
[6] V. F. Gubarev, et al. Model Predictive Control for Discrete MIMO
Linear Systems. In: Kondratenko, Y.P., Chikrii, A.A., Gubarev, V.F.,
Kacprzyk, J. (Eds) Advanced Control Techniques in Complex Engi-
neering Systems: Theory and Applications. Dedicated to Prof. V.M.
Kuntsevich. Studies in Systems, Decision and Control, Vol. 203.
Cham: Springer Nature Switzerland AG, 2019, pp. 63-81. DOI:
https://doi.org/10.1007/978-3-030-21927-7_4
References 353
Abstract
A method for determining the suitability of a technical object for further
operation based on the information about the current state and the history
of its functioning during an arbitrary period of time is obtained. The check
is carried out on the basis of the analysis of the state of the object under
study at future points in time and the value of the posterior probability
of no-failure operation. The method for predicting the individual reliability
of objects is based on the canonical decomposition of a random sequence
describing the change in the value of the controlled parameter over time. As
an estimate of the future state, the conditional mathematical expectation is
used; to estimate the probability of no-failure operation, statistical modeling
of a random sequence beyond the observation area is performed.
355
356 Reliability Control of Technical Systems based on Canonical
The method can also be used to control the reliability of the objects
that are characterized by many parameters (in this case, the vector canon-
ical decomposition is used). The proposed approach does not impose any
restrictions on the random sequence of change of the controlled parameters
and, thus, makes it possible to fully take into account the peculiarities of the
functioning of the object under study.
The work presents block diagrams that characterize the features of using
the proposed method. Expressions for assessing the quality of predicting the
state of the controlled object are obtained.
14.1 Introduction
The main features of the development of modern technology [1] are increas-
ing the degree of automation [2, 3]; increasing loads [4, 5], speeds, tem-
peratures, pressures; reduction in size and weight; increasing requirements
for accuracy and efficiency of functioning [6, 7]; combining individual units
into systems with a single control [8, 9], etc. The increase in complexity
and requirements for performance indicators inevitably leads to the need
to increase requirements for the reliability and durability of equipment
(especially of critical infrastructure systems [10–12]).
The problem of individual prediction [13–16] of state and reliability is of
particular interest. Its solution allows not only to obtain an estimate of the
reliability of each specific sample of products, but also, in the presence of
developed diagnostic support, to move from maintenance by time or resource
to planning operation based on the actual state.
Technical objects are operated under conditions of continuous influence
of different kinds of disturbances, the impact of which has random nature
from the point of view of the intensity, duration and moment of occurrence.
Accordingly, the resulting changes in the state of the operated objects also
turn out to be random [17–19] and form a certain random process in time.
Thus, the problem of predicting the future state of an object from the known
past and present is a problem of a probabilistic-statistical nature, and its
solution is always associated with the study of random processes occurring in
the devices, the state of which is to be predicted. Such studies are constantly
carried out in practice, both in the process of testing products for reliability,
and directly in the process of their operation. This information constitutes
14.3 Solution 357
If the parameter S crosses the boundaries of the admissible region [c, d], a
refusal is recorded. Based on the measurements s (μ) , μ = 1, k, k < I at
discrete moments in time tμ , μ = 1, k, it is required to draw a conclusion
about the suitability of an object for operation at future moments in time
ti , i = k + 1, I.
14.3 Solution
14.3.1 Forecasting the State of Control Objects
The most general approach for solving the problem of assessing the future
state of the control object is based on the Kolmogorov-Gabor polynomial
[20–22]. However, finding its parameters for a large number of known
values and the used order of nonlinear relation is a very complicated and
time-consuming procedure (for example, for 11 known values and order of
nonlinearity 4, it is necessary to obtain and solve 1819 partial differential
equations of the mean square of the extrapolation error). In this regard,
when forming the forecast algorithms realized in practice, various simpli-
fications and restrictions on the properties of a random sequence are used.
For example, Pugachev [23, 24] proposed a number of suboptimal methods
of nonlinear extrapolation with a limited order of stochastic relation based
on the approximation of the posterior probability density of the estimated
vector by the Hermite orthogonal polynomial expansion or in the form of
an Edgeworth series. Solution of the non-stationary equation of Kolmogorov
[25] (a particular case of Stratanovich [26] differential equation for describing
Markov processes) is obtained under the conditions that (a) the diffusion
coefficient is a constant value, and (b) the drift coefficient can be presented as
linear dependence from the state. There are exhaustive solutions of optimal
358 Reliability Control of Technical Systems based on Canonical
linear extrapolation problem for different types of random sequences and dif-
ferent levels of data support in considered forecasting task. Among examples
are (a) Kolmogorov’s equation for random sequences with stationary char-
acter which are measured without errors; Kalman’s technique [27] for noisy
random sequences with Markov properties; Wiener-Hopf’s filter-extrapolator
which is efficient for noisy sequences with stationary character; Kudritskiy’s
algorithms of optimal linear extrapolation [28–30] synthesized using the
Pugachev’s canonical expansion, and others. Nevertheless, we should remem-
ber that only for Gaussian random sequences it is possible to achieve the
highest forecast accuracy using methods of linear extrapolation.
The universal mathematical model (taking into account the restrictions,
which are imposed on the random sequence under consideration) is the
canonical decomposition of such type [31–33]:
!
i !
N
S (i) = M [S (i)] + L(λ)
ν ρ (λ, ν; 1, i) , i = 1, I, (14.2)
ν=1 λ=1
(λ)
where the elements of the canonical expansion Lν , ρ (λ, ν; h, i) are deter-
mined by the recurrent relations:
(λ) 8 9 N (j)
Lν = S λ (ν) − M S λ (ν) − ν−1 μ=1 j=1 Lμ ρ (j, μ; λ, ν) −
(14.3)
(j)
− λ−1
j=1 νL ρ (j, ν; λ, ν) , λ=1, N , ν = 1, I;
' 9*(
(λ) ) 8
M Lν S h (i) − M S h (i)
(λ)
ρ (λ, ν; h, i) βhν (i) =
(λ) 2
M Lν
1 ' ( ' ( ' (
= {M S λ (ν) S h (i) − M S λ (ν) M S h (i) −
Dλ (ν)
!
ν−1 !
N !
λ−1
− Dj (μ) ρ (j, μ; λ, ν) ρ (j, μ; h, i) − Dj (ν) ρ
μ=1 j=1 j=1
!
ν−1 !
N
2
!
λ−1
− Dj (μ) {ρ (j, μ; λ, ν)} − Dj (ν) {ρ (j, ν; λ, ν)}2 ,
μ=1 j=1 j=1
λ = 1, N , ν = 1, I. (14.5)
' ( !
λ−1
l1λ = s (1) − M S (1) −
λ λ
l1j ρ(j, λ; 1, 1), λ = 1, I. (14.6)
j=1
360 Reliability Control of Technical Systems based on Canonical
m(1,1)
x (1, i) = M [S(i/s(1))] = M [S(i)] + (s(1) − M [S(1)])
ρ(1, 1; 1, i), i = 1, I. (14.8)
Taking into account that the coordinate functions ρ(λ, ν, h, i), λ, h =
1, N , ν, i = 1, I are determined from the condition of the minimum of
the mean-square error of approximation in the intervals between arbitrary
values S λ (ν) and S h (i), the expression (14.8) can be generalized to the case
of forecasting S h (i), h = 1, N , i = 2, I:
m(1,1)
x (h, i) = M [S h (i/x(1))] = M [S h (i)]+
+ (s(1) − M [S(1)])ρ(1, 1; h, i), i = 1, I, (14.9)
(1,1)
where ms (h, i) is the optimal estimate of the future value S h (i), provided
that the value S(1) is used for the forecast.
(2)
Concretization of the second value l1 in (14.7) gives the canonical
expansion of the posterior sequence {S (1,2) } = S(i/s1 (1), s1 (1))2
expression (14.9):
m(1,2)
s (h, i) = M [S h (i/s(1), s(1)2 )] =
= m(1,1)
s (h, i) + [s2 (1) − m(1,1)
s (2, 1)]ρ(2, 1; h, i), i = 1, I.
(14.11)
The component S N (1) should be used as a posteriori data (for the first
section t1 in the last recurrent cycle)
m(1,N
s
)
(h, i) = M [S h (i/s(1), s(1)2 . . . s(1)N )] =
−1) −1)
= m(i,N
s (h, i) + [sN (1) − m(1,N
s ]ρ(N, 1; h, i), i = 1, I.
(14.12)
and then the forecast is specified by applying the value s(2)
m(2,1)
s (h, i) = M [S h (i/s(1), s(1)2 . . . s(1)N , s(2))] =
= m(1,N
s
)
(h, i) + [s(2) − m(1,N
x
)
(1, 2)]
ρ(1, 2; h, i), i = 1, I. (14.13)
The last recurrent cycle can be presented as the next calculation operation
(for the section t2 ),
ms(2,N ) (h, i) = M [S h (i/s(1), . . . , s(1)N , s(2), . . . , s(2)N )] =
−1) −1)
= m(2,N
s (h, i) + [sN (2) − m(2,N
s (N, 2)]
ρ(N, 2; h, i), i = 1, I.
Generalization of the obtained regularity allows writing a forecast proce-
dure for an arbitrary number of known values [34–36]:
⎧
⎪
⎪ M [S h (i)] for μ = 0;
⎪
⎪
⎪
⎪ (μ,l−1) (μ,l−1)
⎪
⎨ms (h, 1) + (sl (μ) − ms (l, μ))ρ(l, μ; h, i)
(μ,l)
ms (h, i) = for l = 1;
⎪
⎪
⎪
⎪
⎪
⎪ ms
(μ,l−1)
(h, 1) + (sl (μ) − ms
(μ−l,N )
(l, μ))ρ(l, μ; h, i)
⎪
⎩
if l = 1.
(14.14)
(μ,l)
The expression ms (h, i) = M [S h (i)/sν (j), j = l, μ − l, ν =
l, N ; sν (μ), ν = 1, l] for h = l, l = N, μ = k is the optimal estimate
14.3 Solution 363
m(k,N ) (1, i) of the future value s(i), i = k+, 1, I, provided that the values
sν (j), ν = l, N , j = 1, k are used to calculate this estimate, i.e. the results of
measurements of the sequence {S} at points tj , j = 1, k are known.
(k,N )
The expression for the desired estimate ms (1, i) can be written in the
following explicit form:
!
k !
N
(kN )
m(k,N
s
)
(1, i) = M [S(i)] + (sν (j) − M [S ν (j)])P((j−1)N +ν)
j=1 ν=1
((i − 1)N + 1), (14.15)
where
(α−1) (α−1)
(α) Pλ (ξ) − Pλ (α)γk (i), if λ ≤ α − 1
Pλ (ξ) = (14.16)
γα (ξ), for λ = α
ρ(modN (α), [α/N ] + 1; 1, [α/N ] + 1), for ξ ≤ kN ;
γα (ξ) =
ρ(modN (α), [α/N ] + 1; 1, i), if ξ = (i − 1)N + 1.
(14.17)
[.] is a rounding operation in the expression (14.17).
If the first k values are fixed, then the canonical expansion of a posterior
random sequence {S} can be written in the following way
Thus, the error of a single extrapolation has a stochastic nature that deals
with the random character of the investigated processes.
Implementation of the mathematical expectation procedure to (14.20),
taking into account the expression (14.18), give us possibility to calculate
the systematic component of the single extrapolation’s error
!
i !
N
= M [L(λ)
μ ]ρ(λ, μ, 1, i) = 0, i = k + 1, I. (14.21)
μ=k=1 λ=1
The latter Equation (14.21) illustrates that the procedures (14.14), (14.15)
provides calculation (for each predicted realization) of unbiased estimates of
future values.
This also underlines that the proposed algorithm gives possibility to
provide the unbiasedness for the average
!
i !
N
= Dλ (μ){ρ(λ, μ; 1, i)}2 , i = k + 1, I. (14.23)
μ=k+1 λ=1
Taking into account (14.22), (14.24), the expression for the mean square
of the forecast error can be written as
(k,N )
E (k,N ) (i) = [ε(k,N ) (i)]2 + DΔ (i) =
(k,N )
!
i !
N
DΔ (i) = Dλ (μ)ρ(λ, μ; 1, i), i = k + 1, I. (14.25)
μ=k+1 λ=1
Thus, the mean square of the forecast error based on the procedures
(14.14), (14.15) is equal to the variance of the posterior random sequence.
It is necessary to note that the parameters of the procedure (14.14),
calculated by the size of the statistical sample of realizations, can be specified
using new a posteriori information sL+1 (ν), ν = 1, I in real time in the
process of solving the forecasting problem:
L+1
L
sni sni L
n l=1 l=1snL+1 L M(L+1) [S n ]L + snL+1
M(L+1) [S ] = = = = ;
L+1 (L + 1)L L+1 (L + 1)
(n) 2
L+1 L
(n)
(lν,1 ) (lν,1 )2 (L − 1) (n)
l=1 l=1 (lν,1 )2
Dn(L+1) (ν) = = +
L L(L − 1) (L)
n
Dn(L+1) (ν) + (S(L+1) (ν) − m(n−1,ν) (n, ν))
;
L
L+1
(n)
lν,l (Slλ − M(L+1) [S λ (i)])
l=1
ρL+1 (n, ν; λ, i) =
LDn,(L+1) (ν)
ρL+1 (n, ν; λ, i)(L − 1) + (snL+1 (ν) − m(n−1,ν) (n, ν))
(sλL+1 (i) − M(L+1) [S λ (i)])
=
LDn,(L+1) (ν)
Thus, the results of extrapolation can be used for significant (up to
several arithmetic operations) simplification of the recalculating procedure
366 Reliability Control of Technical Systems based on Canonical
for the algorithm (14.14) parameters in the case if new statistical data of the
investigated random sequence come. It allows to use the updated parameters
for the current solution of the forecast problem.
By analogy with the procedure of forming the algorithm (14.14), the
method of extrapolation of the realization of the vector random sequence {S̄},
taking into account the stochastic relations M [Sjl (μ)Shq ], μ, i = 1, I; j, h =
1, H; l, q = 1, N between the components, can be obtained as following:
⎧
⎪
⎪ M [Sh (i)], if μ = 0;
⎪
⎪
⎪
⎪
⎪
⎪
(μ,l−1) (μ,l−1) (h,q)
ms;j,h (q, i) + (slj (μ) − mx;j,h (l, μ))βj,i (μ, i),
⎪
⎪
⎪
⎪
⎪
⎨
for l > 1, j > 1;
(μ,l)
ms;j,h (q, i) = m(μ,N ) (q, i) + (sl (μ) − m(μ,N ) (l, μ))β (h,q) (μ, i),
⎪
⎪ s;j−1,h j x;j−1,j j−1,N
⎪
⎪
⎪
⎪ for l = 1, j > 1;
⎪
⎪
⎪
⎪ (μ−1,N ) (μ−1,N ) (h,s)
⎪
⎪ m (q, i)+(sl1 (μ) − ms;H,1 (l, μ))β1,1 (μ + 1, i),
⎪ s;H,h
⎩
if l = 1, j = 1;
(14.26)
(μ,l)
where ms;j,h (1, i) = M [Sh (i) /snλ (ν) , λ = 1, H, n = 1, N , ν =
1, μ − 1; snλ (μ) , λ = 1, j] is the optimal, in the mean-square sense, estimate
of the future values of the investigated random sequence, provided that a
posteriori information snλ (ν), λ = 1, H, n = 1, N is used for the forecast.
The algorithm is based on the canonical model
!
i−1 !
H !
N
(λ) (h,1)
Sh (i) = M [Sh (i)] + Lνl βlλ (ν, i)
ν=1 l=1 λ=1
!
h−1 !
N
(λ) (h,1) (1)
+ Lil βlλ (i, i) + Wih , l = 1, I; (14.27)
l=1 λ=1
(λ)
!
ν=1 !
H !
N
(l,λ)
Lνl = Slλ (ν) − M [Slλ (ν)] − L(l)
μm βmj (μ, ν)−
μ=1 m=1 j=1
!
l−1 !
N !
λ−1
(j) (l,λ)
− νm (ν, ν) −
L(j) Lνl βlj (ν, ν), ν = 1, I; (14.28)
m=1 j=1 j=1
14.3 Solution 367
' ( ' (
(λ) 2
Dl,λ (v) = M Lvl = M Sl (v) − M Sl (v)
2λ 2 λ
!
v−1 !
H !
N
2 !
l−1 !
N
2
(l,λ) (l,λ)
− Dmj (μ) βmj (μ, v) − Dmj (v) βmj (v, v)
μ=1 m=1 j=1 m=1 j=1
!
λ−1
2
(l,λ)
− Dlj (v) βlj (v, v) , v = 1, I; (14.29)
j=1
' ' ((
(λ) (λ)
M Lvl Shq (i) − M Lvl Shq (i)
(h,q)
βl,λ (v, i) =
(λ) 2
M Wvl
1 ' ( ' ( 8 9
= M Slλ (v)Shq (i) − M Slλ (v) M Shq (i)
Dl,λ (v)
!
v−1 !
H !
N
(l,λ) (h,s)
− Dmj (μ)βmj (μ, v)βmj (μ, i)
μ=1 m=1 j=1
!
l−1 !
N
(l,λ) (h,s)
− Dmj (v)βmj (v, v)βmj (v, i)
m=1 j=1
!
λ−1
(l,λ) (h,s)
− Dlj (v)βmj (v, v)βmj (v, i), λ = 1, hv = 1, i. (14.30)
j=1
The difference with (14.14) lies in that that after taking full account of the
posterior information sj (μ), s2j (μ)...sN
j (μ) at the fixed section tμ , the transi-
tion to the accumulation is carried out to predict the posterior information of
the next component Sj+1 , in the case j=N the transition to the next section
tμ+1 is carried out. Diagram in Figure 14.2 reflects the peculiarities of the
computational process when using the extrapolator (14.26).
The expression for the mean-square error of extrapolation using the
algorithm (14.26) by the known values snj (μ), μ = 1, k; j = 1, H; n = 1, N
has the form
(k,N ) 8 9
Eh (i) = M Sh2 (i) − M 2 [Sh (i)] −
!
k ! H ! N
2
(h,1)
− Djn (μ) βjn (μ, i) , i = k + 1, I. (14.31)
μ=1 j=1 n=1
368 Reliability Control of Technical Systems based on Canonical
Figure 14.2 Generative procedure for future values of a random sequence based on the
algorithm (14.26)
(K,N )
The mean-square error of extrapolation Eh (i) is equal to the variance
of the posterior random sequence
(k,N ) ) * (k,N )
Sh (i) = S i/sνl (j) , ν = 1, N , j = 1, k, l = 1, H = mH,h (1, i)
!
i−1 !
H !
N
(λ) (h,1)
!
h−1 !
N
(λ) (h,1) (1)
+ Lνl βlλ (ν, i) + Lil βlλ (i, i) + Lih ,
ν=k+1 l=1 λ=1 l=1 λ=1
i = k + 1, I. (14.32)
!
i !
H
(λ)
Sh (i) = M [Sh (i)] + vv(λ) ϕhv (i), i = 1, I, (14.34)
v=1 λ=1
where
!
v−i !
H
(j)
Vv(λ) = Sλ (v) − M [Sλ (v)] − vμ(j) ϕλμ (v)
μ=1 j=1
!
λ−1
(j)
− Vv(j) ϕλμ (v), v = 1, I; (14.35)
j=1
(λ)
(λ) M [Vv (Sh (i) − M [Sh (i)])] 1
ϕhv (i) = = (M [Sλ (v)Sh (i)]
(λ) 2 Dλ(v)
M {Vv }
!
v−1 !
H
(j) (j)
− M [Sλ (v)]M [Sh (i)] − Dj (μ)ϕλμ (v)ϕhμ (i)
μ=1 j=1
!
λ−1
(j) (j)
− Dj ϕλv (v)ϕhv (i), λ = 1, I, v = 1, i ; (14.36)
j=1
2 ' (
Dλ (v) = M Vv(λ) = M {Xλ (v)}2 − M 2 [Xλ (v)]
!
v−1 !
H
2 !
λ−1
2
(j) (j)
− Dj (μ) ϕλμ (v) − Dj (v) ϕλv (v) , v = 1, I.
μ=1 j=1 j=1
(14.37)
370 Reliability Control of Technical Systems based on Canonical
!!
μ−1 H
2 !
l−1
2
(μ,l) (j) (j)
Eh (i) = Dh (i) − Dj (v) ϕhv (i) − Dj (μ) ϕhμ (i) ,
v=1 j=1 j=1
i = μ + 1, I. (14.38)
lies within
1 1 + r(λ)(L) (v, i) zα 1 1 + r(λ)(L) (v, i) zα
ln −√ , ln +√
2 1 − r(λ)(L) (v, i) L−3 2 1 − r(λ)(L) (v, i) L−3
where zα is the value of the standard normal deviate which corresponds to the
level of confidence α. In other cases, the hypothetical value r(λ) (v, i) does not
correspond to the statistics.
Because the condition (r(λ) (v, i) = 0) is checking, the random variable
(λ)
a (v, i) takes the form [41, 42]
√ 1 + r (λ)(L) (v, i)
a(λ) (v, i) = L − 3 ln
1 − r(λ)(L) (v, i)
N = max N (v,i) .
v,i
a given distribution law and with the subsequent formation, using the
expression (14.18) or (14.32), of a set of possible continuations of the
realization of the investigated random sequence in the forecast area
[tk+1 ...tI ];
• verification of the condition of non-intersection of the boundaries of the
admissible region of change of the controlled parameters by the obtained
trajectories and determination of an estimate of the probability of failure-
free operation of a technical object;
• making a decision on the subsequent operation of the object based on
the obtained estimate of the probability of failure-free operation.
14.4 Conclusion
Thus, a method of reliability control of technical objects, the parameters
of which are of a stochastic nature, is obtained in the work. An individual
reliability estimation based on the information about the functioning of an
object during operation is a peculiarity of the proposed approach. Making
decision on suitability at future points in time is carried out in two stages: at
the first stage, the belonging of the predicted parameters to the permissible
region is checked, and at the second stage, the compliance of the proba-
bility value of no-failure operation with the required level is checked. The
mathematical model based on the canonical decomposition does not imply
any restrictions on the random sequence of change of the parameters of the
control object, which makes it possible to take into account the features of
its functioning as much as possible. Expressions for the mean-square error of
extrapolation are obtained, which allow us to estimate the accuracy of solving
the forecasting problem. The resulting method can be used to control the
reliability of arbitrary technical objects and systems [9, 10, 43–45], as well
as their components.
References
[1] A. Adarsh, R. Mangey, ‘System reliabilty: solutions and technologies’,
Taylor & Francis Group, 2019.
[2] V. Varakuta, T. Hlimancov, S. Starodubcev, S. Ptashka, ‘Determination
of risk indicators of loss of negative situations on technological danger-
ous objects’, Science and Technology of the Air Force of Ukraine, Issue
1(34), 2019, pp. 107–116.
References 375
[3] P.M. Frank, ‘Fault diagnosis in dynamic systems using analytical and
knowledge-based redundancy: A survey and some new results’, J.
Automatica 3, 459–474, 1990.
[4] R.J. Patton, P.M. Frank, R.N. Clarke., ‘Fault diagnosis in dynamic
systems: theory and application’, Prentice-Hall Inc., 1989.
[5] R.J. Patton, R.N. Clark, P.M. Frank, ‘Issues of fault diagnosis for
dynamic systems’, Springer Science & Business Media, 2000.
[6] P.M. Frank, X.-C. Ding, ‘Survey of robust residual generation and
evaluation methods in observer-based fault detection systems’, Journal
of process control 7.6, 403–424, 1997.
[7] S.X. Ding, ‘A unified approach to the optimization of fault detection
systems’, International journal of adaptive control and signal processing
14.7, 725–745, 2000.
[8] N. Silva., M. Vieira, ‘Towards Making Safety-Critical Systems Safer:
Learning from Mistakes’, In: ISSREW, 2014, 2014 IEEE Interna-
tional Symposium on Software Reliability Engineering Workshops
(ISSREW), 162–167, 2014.
[9] V.M. Kuntsevich, et al (Eds), ‘Control Systems: Theory and Applica-
tions’, Series in Automation, Control and Robotics, River Publishers,
Gistrup, Delft, 2018.
[10] Y.P. Kondratenko, V.M. Kuntsevich, A.A., Chikrii V.F. Gubarev, (Eds),
‘Advanced Control Systems: Theory and Applications’, Series in
Automation, Control and Robotics, River Publishers, Gistrup, 2021.
[11] V. Varakuta, T. Hlimancov, S. Starodubcev, S. Ptashka, ‘Determination
of risk indicators of loss of negative situations on technological danger-
ous objects’, J. Science and Technology of the Air Force of Ukraine, No.
1(34), pp. 107–116. https://doi.org/10.30748/nitps.2019.34.15.
[12] A.V. Boyarchuk, etc., ‘Safety of Critical Infrastructures: Mathematical
Analysis and Engineering Methods of Analysis and Ensuring’, National
Aerospace University named after N.E.Zhukovsky “KhAI”, Kharkiv,
2011.
[13] M. Nafria, ‘Circuit reliability prediction: challenges and solutions for
the device time-dependent variability characterization roadblock’, IEEE
Latin America Electron Devices Conference (LAEDC), Mexico, 19–21
April 2021.
[14] H. Jiang, ‘Time Dependent Variability in Advanced FinFET Technology
for End-of-Lifetime Reliability’, IEEE International Reliability Physics
Symposium (IRPS) Monterey, CA, USA, 21–25 March 2021.
376 Reliability Control of Technical Systems based on Canonical
Abstract
Mathematical modeling of modern automatic control systems requires storing
and processing large amounts of data. Complex systems are described by
numerous parameters that have different significance and obey certain restric-
tions. These constraints often take the form of a system of linear inequalities.
This paper proposes a new way of describing the optimization domain with
linear constraints in the design space using Petunin ellipsoids. The statistical
properties of these ellipsoids are given, making them an effective tool for
describing data in the design space. The problem of construction of the
optimal Petunin ellipsoids in design space with linear constrains is consid-
ered in the paper. The algorithms of construction of the Petunin ellipsoids
without constraints and with linear constraints are described. The issues of
the complexity and statistical features are analyzed.
The second part is devoted to a new distribution-free approach to comput-
ing ellipsoidal conformal prediction set based on the Hill’s assumption and
the Petunin ellipsoids. The resulting prediction set is a Petunin ellipsoid with
precise confidence level depending on the number of points only. We show
that our method works for classical distributions (uniform and Gaussian) with
high performance. In addition, our method allows determining the median
of the random points set (the deepest point), outliers (outer points) and the
statistical depth for every point of the set.
379
380 Petunin Ellipsoids in Automatic Control Systems Design
15.1 Introduction
Modern computer-aided design of automatic control systems is impossible
without data analysis that describes the parameters of objects and the results
of experiments in the form of multidimensional vectors in the design space.
The complexity of the problem is due to the fact that the number of such
parameters is very large and therefore it is desirable to single out among them
those that are important from a certain point of view (for example, statistically
significant) and discard points that can be interpreted as outliers. Moreover,
each parameter, as a rule, is subject to certain restrictions. It is either limited
to the minimum and maximum values, or is related to other parameters by
linear relationships.
The problems of designing technical systems are often reduced to solving
variational problems of minimizing functionals subject to the fulfillment of
constraints in the form of a system of linear inequalities. This leads to the need
to determine the optimization area in the design space, in which the search
for the optimal solution that satisfies the given constraints is performed. A
large scatter of parameters in the design space can complicate mathematical
modeling and worsen the properties of numerical optimization algorithms,
so the optimization area should be compact and contain only statistically
significant points. In addition, many efficient optimization algorithms make
significant use of the geometric properties of the optimization domain (for
example, convexity). It is desirable that the optimization domain be convex
and described by simple equations. These requirements are met by ellipsoids
covering the given points. Ellisoids of minimum volume are often used as
such ellipsoids, ellipsoids with minimal traces, Dikin ellipsoids and other
modifications. The choice of these ellipsoids is dictated by geometric consid-
erations and does not take into account the statistical nature of the parameters
of the designed technical systems. demands without considering the data
volume.
The statistical nature of the source data dictates the choice of the appro-
priate tool. In particular, similar problems were considered in [1], where a
comparative analysis of several methods for constructing optimal descriptions
of design data was carried out, in particular, an ellipsoid of minimum volume,
an ellipsoid with a minimum trace, and the Monte Carlo method. Despite the
15.2 Petunin Ellipses and Their Statistical Properties 381
this is the same algorithm that has geometrical features dependent from the
space dimension. Case d = 2 is more descriptive and allow simple geometric
concepts to be used for illustration.
Case d = 2. Construct a convex hull of X = {(x1 , y1 ), ..., (xn , yn )} Find
a diameter of the convex hull and its ends (xk , yk ) and (xl , yl ). Construct
a segment L connecting these ends. Find points laying at the most distance
from L: (xr , yr ) and (xq , yq ).
Construct segments L1 and L2 that are parallel to L, pass through (xr , yr )
and (xq , yq ), and have the length equals to the length of L. Construct segments
L3 and L4 that are orthogonal to L and pass through (xk , yk ) and (xl , yl ).
Segments L1 , L2 , L3 and L4 form a rectangle Π . Denote a short side by a
and a long side by b (Figure 15.1).
Make translation, rotation and shrinking with a coefficient α = ab so
that Π transforms to a square Π with a center (x0 , y0 ). The random points
are transformed to points (x1 , y1 ), (x2 , y2 ), ...(xn , yn ) ∈ Π . Find distances
r1 , r2 , ..., rn from (x0 , y0 ) to (x1 , y1 ), (x2 , y2 ), ...(xn , yn ) ∈ Π . Let R =
max(r1 , r2 , ..., rn ). Construct a circle C with the center (x0 , y0 ) and radius
R. Thus, (x1 , y1 ), (x2 , y2 ), ..., (xn , yn ) ∈ C (Figure 15.2).
Performing inverse transformations of this circle we obtain an ellipse E
(Figure 15.3). The computational complexity of this algorithm is equal to
computational complexity of the construction of a convex hull O(nlgn).
15.2 Petunin Ellipses and Their Statistical Properties 383
effective and fast procedures [2, 5] that are not optimal but extremely smooth
the problem.
Table 15.1 The number of uniformly distributed points dropped out and the significance
level of the Petunin ellipse satisfying square constraints
Initial Number of Dropped Initial Significance Result Significance
Number out Points Level Level
40 19 0.976 0.955
50 7 0.980 0.977
60 9 0.984 0.981
70 19 0.986 0.981
80 24 0.988 0.982
90 50 0.989 0.976
100 66 0.990 0.971
Table 15.2 The number of Gaussian distributed points dropped out and the significance level
of the Petunin ellipse satisfying square constraints
Initial Number of Dropped Initial Significance Result Significance
Number out Points Level Level
40 7 0.976 0.971
50 14 0.980 0.973
60 12 0.984 0.983
70 13 0.986 0.985
80 8 0.988 0.986
90 8 0.989 0.988
100 9 0.990 0.989
15.2 Petunin Ellipses and Their Statistical Properties 387
As we see in the uniform case, when the sample size is greater than 40
the decreasing of the confidence level of the Petunin ellipses after peeling is
varied in the region (0.03; 0.13). When the sample size is large (90 or 100) we
388 Petunin Ellipsoids in Automatic Control Systems Design
have clear gift in the number of points (55%–66%) without significant loss of
the confidence level of the ellipse.
As we see in the Gaussian case, when the sample size is greater than 50
the decreasing of the confidence level of the Petunin ellipses after peeling is
varied in the region (0.01; 0.02). However, we had not achieved a significant
decreasing of the number of representative points. Thus, we may suppose that
this effect is connected with the kind of distribution.
Figure 15.8 illustrates changing of area of Petunin ellipses after drop-
ping one point per step. As far as the Petunin ellipses covering points are
concentric the area is monotonically decreasing. Absence of abrupt jumps
in these curves justify that the sample is homogeneous and fill the region
densely.
Figure 15.8 Area of Petunin ellipses covering 100 uniformly and Gaussian distributed points
after dropping one point per step
15.2 Petunin Ellipses and Their Statistical Properties 389
where Fξ (u) and Fη (u) are the distribution functions of ξ and η, respectively.
The probability density of ith order statistics is:
!
n !
n
fk (u) = Cni [F (u)]i [1 − F (u)] n−i
= Gi (u).
i=k i=k
390 Petunin Ellipsoids in Automatic Control Systems Design
Hence,
!
n
Fk (u) = Gi (u)
i=k
Gk (u) = Cnk [k(F (u))k−1 (1 − F (u))n−k f (u)−
− (F (u)) (n − k)(1 − F (u))n−k−1 f (u)].
k
' (
Gk+1 (u) = Cnk+1 (F (u))k+1 (1 − F (u))n−k−1
= Cnk+1 [(k + 1)(F (u))k (1 − F (u))n−k−1 f (u)−
− (F (u))k+1 (1 − F (u))n−k−2 f (u)].
Thus,
1
xi+1−1 (1 − x)n−i+1−1 dx = B (i + 1, n − i + 1)
0
Γ(i + 1)Γ(n − i + 1) Γ(i + 1)Γ(n − i + 1) i!(n − i)!
= = =
Γ(i + 1 + n + 1 − i) Γ(n + 2) (n + 1)!
j!(n − i)!
p (xn+1 < xi ) = ni−1
n−1 =
(n + 1)!
(n − 1)!j! (n − j)!
=n =
(n − 1 − j + 1)! (j − 1)! (n + 1)!
n (n − 1)!(j − 1)j! (n − j)! j
= =
(n − j)!(j − 1)!n (n + 1) (n − 1)! n+1
Therefore,
) ) ** j i j−i
p xn+1 ∈ x(i) , x(j) = p (xn+1 < xi ) = − =
n+1 n+1 n+1
Thus, if a random variable x does not depend on x1 , x2 , ..., xn and
it is drawn by sample sampling from the same population following the
distribution F (u), then
) ) ** n − 1
p x ∈ x(1) , x(n) =
n+1
) *
Remark 15.1. The confidence level of the tolerance interval J = x(1) , x(n)
n+1 , thus for n ≥ 39 the confidence level of this interval is less than 0,05.
is n−1
Figure 15.9 Ellipse covering point with uniform (left) and Gaussian (right) distribution
Figure 15.10 Areas of the ellipses covering point with uniform (left) and Gaussian (right)
distribution
Table 15.3 The results of testing confidence level of the Petunin ellipse for uniform
distribution
Observable Relative
Trial Frequency Descriptive Statistics
1 0.98 Expected probability 0.9048
2 0.95 Mean 0.9616
3 0.96 Standard Error 0.0073
4 0.92 Median 0.9700
5 0.99 Mode 0.9800
6 0.99 Standard Deviation 0.0318
7 0.94 Sample Variance 0.0010
8 0.95 Kurtosis 0.7694
9 0.98 Skewness –0.9680
10 1 Range 0.1200
11 1 Maximum 1.0000
12 0.98 Minimum 0.8800
13 0.97 Geometric Mean 0.9610
14 0.88 Harmonic Mean 0.9605
15 0.95
16 0.98
17 0.99
18 0.98
19 0.94
(Continued)
394 Petunin Ellipsoids in Automatic Control Systems Design
15.3 Conclusion
The proposed algorithm of constructing optimal Petunin ellipsoid in design
space is effective and statistically robust. It allows decreasing of the point
in design space without loss of confidence level of the ellipsoids. This algo-
rithm demonstrates robustness to outliers, high confidence level with modest
number of covering points and useful possibility to arrange the point on the
probability to cover them by the ellipsoid.
The proposed algorithm of constructing conformal prediction sets using
Petunin ellipsoid is effective and has high performance. It has theoretically
precise confidence level and allows unambiguous multivariate points arrang-
ing, computing statistical depth every points, finding median of the set, and
detecting outliers. The experiments confirmed the theoretical properties of
the Petunin ellipses.
The Petunin ellipsoid is useful tool for arrange multivariate data and
finding anomalies using statistical depth. Thus, the future research directions
within the domain of the topic are developing computationally effective
algorithms for construction of high-dimensional conformal prediction sets
using genetic and other algorithms and inventing new multivariate statistical
tests for homogeneity of multivariate samples in high-dimensional spaces.
References
[1] A. Bedrintsev, V. Chepyzhov, “Description of the design space by
extremal ellipsoids in data representation problems”, J. Commun. Tech-
nol. El., vol. 61, pp. 688–694, 2016.
[2] S. Har-Peled, “A practical approach for computing the diameter of a
point set”, In: Symposium on Computational Geometry (SOCG’2001),
pp. 177–186, 2001.
[3] S. Lyashko, D. Klyushin, V. Alexeenko, “Multivariate ranking using
elliptical peeling”, Cybern. Syst. Anal., 49, pp. 511–516, 2013.
References 395
Abstract
A rejected takeoff is the situation when it is decided to cancel the takeoff
of an airplane. When a crucial obstacle to the takeoff occurs (e.g. engine
failure), the crew compares the current velocity with the “decision-making
velocity”. Based on this velocity one can estimate whether the aircraft can
safely stop within the available distance of the interrupted flight. These values
are taken from the Aircraft Flight Manual before the takeoff. But the real
rejected takeoff distances can differ from the pre-calculated values. The goal
of our investigation is to obtain real-time estimations of the distance to full
stop. The proposed real-time algorithm of the calculation of the distance of
the interrupted flight uses the step-wise interpolation of the thrust. Using
this interpolation, the Riccati equation of the aircraft motion has explicit
solutions. Besides using known time intervals of interpolation, the time to
reach a given speed can be explicitly calculated. To check and study the
proposed algorithm for estimating the rejected takeoff distance in real time,
a computer simulation system was created. To verify the adequacy of the
algorithm, the emulation results were compared with the values obtained
from the real aircraft nomograms. The algorithm takes into account different
aircraft and environment parameters (thrust, velocity, temperature etc.) before
the moment of decision-making. As a result, the decision about the rejected
takeoff will be more reliable and safe.
399
400 On Real-Time Calculation of the Rejected Takeoff Distance
16.1 Introduction
A rejected takeoff is the situation when it is decided to cancel the takeoff of an
airplane. Analysis of accidents and serious incidents in civil aviation shows
that an aircraft exiting the runway and terminal safety strip does not happen
often when the takeoff is cancelled, but it is extremely dangerous. The reasons
for these accidents and incidents may be the following: too late actions of
the crew to stop takeoff; inconsistency of the conditions for the interrupted
takeoff with the calculated ones. In the first case, the pilot or crew reacts to the
appearance of hazard factors with a time delay greater than the allowable one.
In the second, the conditions for a rejected takeoff do not correspond to those
that were used in the preliminary calculations. In both cases, the aircraft does
not have enough time to stop within the available distance of an interrupted
takeoff (ASDA, Accelerate-Stop Distance Available), since the real distance
of an interrupted takeoff (ASD, Accelerate-Stop Distance) exceeds ASDA.
When a crucial obstacle to the takeoff occurs (e.g. engine failure), the
crew has just several seconds to analyse the situation and to make a “go–
no go” decision–to continue a takeoff, or to abort it. During this time, the
aircraft speed increases from vEF (engine failure velocity) to vEF + Δv. The
crew compares current velocity to the “decision-making velocity” V1 . Based
on V1 velocity, it is possible to estimate if the airplane still can safely stop
within the available distance of the airport. If the current speed is less than
the “decision-making velocity”, the crew makes a decision to interrupt the
takeoff; otherwise, it is necessary to continue the takeoff, avoiding overrun. It
is assumed that if the crew decides to terminate the takeoff at a speed less than
or equal to V1 , the aircraft will stop within the ASDA. The velocity V1 and
ASDA distance are calculated in advance using the “Aircraft Flight Manual”
[1]. However, the real rejected takeoff distances (the length of the braking
section until the aircraft stops completely) can differ from the pre-calculated
values used in flight manuals and nomograms due to different hazard factors
[2, 3]. Among these factors, we should consider possible differences between
the actual characteristics of the engines, the take-off weight of the aircraft and
the conditions of the airfield (especially friction during takeoff and braking)
from the characteristics used in calculating the airplane flight curves and
nomograms. Therefore, a more precise real-time calculation of the rejected
takeoff dista2nce performed by an automatic control system of the aircraft
can help the crew to make accurate and safer “go–no go” decision.
16.2 Dynamic Model of Aircraft Movement along the Runway 401
deΛl dl
= ΛeΛl = ΛeΛl v,
dt dt
Λl '
d2 eΛl de Λl dv
) *(
= Λ· v + e = Λ· Λe Λl 2
v + e Λl
−Λv 2
+ G = ΛGeΛl .
dt2 dt dt
√
1. IfΛG < 0, thenẍ = −ω 2 x, whereω = −ΛG. This is the classical
equation of small oscillations of a pendulum [15]. General solution of
this equation is
x(t) = C1 sin ω t + C2 cos ω t, (16.7)
with constants C1 , C2 that can be obtained
√ from the initial conditions.
2. If ΛG > 0, then ẍ = γ 2 x, where γ = ΛG. This is the equation of the
“inverted pendulum” [15] with general solution
x(t) = C1 t + C2 . (16.9)
time t0 . At this time, the engine control levers are moved to the “Small Gas”
position. As a result, the thrust of the engines is reduced to almost zero in 1.5–
2 seconds. Then the breaking systems are activated, including reversal thrust.
When the reverse is turned on, the negative thrust of the engines changes
from 0 to the full reverse thrust U . This process requires about 7–11 seconds.
It is assumed that the transient processes end when they reach 0.95 of the
steady-state values [16].
Complex mechanical-mathematical and thermodynamic relations describe
the transients of modern turbojet engines. However, from the point of view
of control theory, the integral effect of engine thrust on an aircraft during
the transient process is described with good accuracy by linear differential
equations [16, 17].
The equation for decreasing the takeoff thrust of the engine to zero can be
written as follows [17]
dP P
=− ,
dt T1
−t/
P (t) = P0 · e T1 , (16.10)
where the value T1 is called the engine shutdown time constant and is an
important characteristic of a particular type of turbojet engine. Function
(16.10) satisfies the relation
−t/
3T1 3T1 ) *
P (t) dt = P0 e T1 dt = P0 · T1 1 − e−3 ≈ 0, 95·
0 0
T1
P0 · T1 = 0, 95 · P0 dt.
0
We can substitute the thrust function (16.10) in the Riccati Equation (16.4)
and numerically find the distance travelled by the aircraft during the transition
process. However, given the relation (16.11), instead of the exponential thrust
function (16.10) it is possible to use the step-wise constant function
P0 , 0 ≤ t < T1 ,
P (t) =
0, T1 ≤ t ≤ 3T1 ,
Similarly, we can consider the process of the engine reversal from zero to
the nominal reverse thrust, which is described by the function
t/ −3
P (t) = U · e T2 , 0 ≤ t ≤ 3T2 , (16.12)
its derivative
deΛl(t)
= ΛeΛl(t) v(t) = C1 γchγt + C2 γshγt.
dt
We can find the constants C1 and C2 from the initial conditions l(0) = 0
and v(0) = v0 :
eΛ·0 = C1 sh(0) + C2 ch(0) = C2 ,
ΛeΛ·0 v0 = C1 γch(0) + C2 γsh(0) = C1 γ,
so we have
Λv0
C1 = , C2 = 1.
γ
We get the relation between the distance Δl1 and the speed v1 over time
Δ1 = δ + T 1
Λv0
eΛΔl1 = sh(γΔ1 ) + ch(γΔ1 ),
γ
ΛeΛΔl1 v1 = Λv0 ch(γΔ1 ) + γsh(γΔ1 ),
thus
1 Λv0
Δl1 = ln sh(γΔ1 ) + ch(γΔ1 ) , (16.14)
Λ γ
' γ (
v1 = e−ΛΔl1 v0 ch(γΔ1 ) + sh(γΔ1 ) . (16.15)
Λ
During time interval [t1 , t2 ] (with zero thrust) and [t2 , t3 ] (with reverse
thrust) we have Λ > 0, G < 0, and general solution is (16.7)
values of the parameters, we can assume that some values change at fixed
known moments of time, and at these time intervals Δi = ti+1 −ti we need to
estimate li+1 = l(ti+1 ), vi+1 = v(ti+1 ) using known ti , ti+1 , l(ti ), v(ti ).
We assume that engine failure occurs at some speedvEF , and after that, the
rejecting takeoff begins, with known time parameters δ, T1 , T2 etc. We
do not know the exact time tEF of the engine failure, this time should be
calculated, but we know that P (tEF ) = P (vEF ). As a result, during the
time interval [t∗ , tEF ] we can interpolate the thrust with constant value
1 ∗
2 [P0 + P (vEF )]. For more accuracy, we can split [t , tEF ] into two or more
parts, e.g. [t∗ , τ ], [τ, tEF ], where v(τ ) = 2 [v(t∗ ) + vEF ], and we can use
1
two constant values 12 [P0 + P (v(τ ))] and 12 [P (v(τ ) + P (vEF )]. Similarly,
we assume that we know the time when the coefficients cx , cy change their
values – it takes 1–2 second after start of the reverse thrust. The time of small
gas reverse is not known, and it must be calculated during emulation.
Taking into account the initial conditions (16.19), we can obtain explicit
dependency of time t from velocity v
1 Λω (v − v0 )
t = − arctg 2 . (16.26)
ω Λ v0 v + ω 2
We can also calculate l(t), v(t):
1 Λv0
l(t) = ln sin ω t + cos ω t , (16.27)
Λ ω
ω
v(t) = e−Λl v0 cos ω t − sin ω t . (16.28)
Λ
Thus, at each interval of interpolation of the aircraft parameters with fixed
values, we can analytically express the dependences of distance, time and
speed under given initial conditions. If time is known, we can use (16.23),
(16.24) or (16.27), (16.28), depending on the ΛG sign. If the time is unknown,
but we know the speed at the end of the time interval, then the time can found
from (16.22) or (16.26).
the temperature, slope of the runway, air density, fuel burn-up rate, number of
running engines, thrust dependence on speed, etc. are also taken into account.
You can set different initial conditions (mass of the aircraft, the speed of
decision-making, after which there is braking, etc.). You can change the time
parameters that affect the process of starting and braking (for example, the
reaction time of the crew, the time characteristics of the transition from full
thrust to reverse, time intervals of change). You can also vary the speed of
transition from full reverse thrust to small gas reverse thrust. With the help
of the system, both the entire path from start to stop (takeoff acceleration and
braking) and the process of only takeoff interruption were investigated (the
speed of decision-making is set and the distance to the moment of stop is
calculated). Different values of engine failure velocity vEF were investigated,
including the values near the decision-making velocity V1 . Actually, for real-
time calculations it is necessary to use the part of the system that emulates
only takeoff interruption. For such calculations, it is possible to use only
8 time intervals with approximation of parameters of system by the fixed
values. It took 10 time intervals to simulate both takeoff and braking, with
sufficient accuracy of emulation. To increase the accuracy of the emulation
of the acceleration part of the rejected takeoff, we can split the time interval
just before the reaching the vEF into several parts (up to 4); in his case we
will use at most 13 time intervals.
The Java programming language was chosen to develop the system. We
used a Java Development Kit version 1.8 (JDK 1.8). Apache Maven version
3.6.3 was used as a project management tool. The user interface (GUI) was
developed using Java Swing.
The main libraries used to develop the system:
Ini4J (Java API for processing configuration files in .ini format for
Windows)–configuration and initial values of system parameters.
Apache Commons Mathematics Library (library of light, independent com-
ponents of mathematics and statistics)-numerical integration of differential
equations describing the motion of the aircraft.
JChart2D (an easy-to-use component written in Java for displaying two-
dimensional tracks in a coordinate system)–visualization of changes in
parameters, speed and coordinates of the aircraft in real time.
Junit, hamcrest, powermock-api-mockito–libraries for the program testing.
A general view of the user interface is shown in Figure 16.2. Using the sliders,
you can change up to 50 parameters; their number, names, as well as the initial
410 On Real-Time Calculation of the Rejected Takeoff Distance
values of these parameters and the ranges of their changes are specified in the
configuration file config.ini. In Figure 16.2, about 30 parameters are available
for changing by sliders, and the remaining 20 are set directly in config.ini.
After changing necessary parameters, we press “Run” button, and the
system performs emulation and equation integration. The results will be
shown as plots, see Figure 16.3. Using checkboxes (left bottom), we can
select the parameters to display. The simulation results are displayed in real
time. For each parameter displayed on the screen, such as speed v, two
values are displayed simultaneously in real time. The bold line (Real) shows
the real value of the parameter obtained from the complete model (16.1)–
(16.3) by numerical integration; thin line (Emulated) represents piecewise
linear approximation of the parameter using the proposed algorithm. The
upper left corner of the screen shows the initial distance, velocity and time.
Top central part shows the distance, velocity and time of the interrupted
takeoff. REAL values are obtained by numerical integration of the complete
model, EMULATED values are obtained using the described algorithm using
two or more (up to 5) piecewise constant approximations, as selected by
16.9 Emulation Results 411
corresponding slider. The upper right corner of the screen shows the values
of the total distance and the time of the interrupted takeoff, at the moment
of full stop with v = 0. REAL values are obtained by numerical integration
of the complete model, EMULATED values are obtained using the described
algorithm using piecewise constant approximations.
line (Emulated) shows the constant emulated values for the distance, veloc-
ity, and thrust that are calculated using constant estimations of the aircraft
parameters.
Table 16.1 shows the typical values of the distance to the stop after
reaching the decision-making speed, calculated by this algorithm, in com-
parison with the results of numerical integration of the complete model.
From Table 16.1 we can see that the calculation of the interrupted take-
off distance using the proposed algorithm corresponds to the results of
the integration of the motion Equations (16.1)–(16.3). Thus, the piecewise
16.11 Conclusion
The proposed algorithm for calculating the rejected takeoff distance uses step
interpolation of aircraft thrust and other parameters. This interpolation allows
to obtain explicit solutions of the Riccati equation that describes the aircraft
motion. The calculations by explicit formulae on a small number of time
intervals can be performed very quickly, in real time. In any moment of the
takeoff, we can obtain an estimation of the distance to the full stop. Algorithm
takes into account aircraft parameter just in the moment of emergency. With
414 On Real-Time Calculation of the Rejected Takeoff Distance
References
[1] Airplane AN-124-100. “Manual for flight” [in Russian], Book 1, 1993.
[2] NLR Air Transport Safety Institute, Report no. NLR-TP-2010-177,
April 2010, https://reports.nlr.nl/xmlui/bitstream/handle/10921/15
8/TP-2010-177.pdf.
[3] “Reducing the risk of runway excursion”. RUNWAY SAFETY INITIA-
TIVE. Flight Safety Foundation, May 2009, https://flightsafety.org/files
/RERR/fsf-runway-excursions-report.pdf.
[4] Kondratenko, Y.P., Kuntsevich, V.M., Chikrii, A.A., Gubarev, V.F.
(Eds). Advanced Control Systems: Theory and Applications. Series in
Automation, Control and Robotics, River Publishers, Gistrup, 2021.
[5] Kondratenko, Y.P., Chikrii, A.A., Gubarev, V.F., Kacprzyk, J. (Eds).
Advanced Control Techniques in Complex Engineering Systems: The-
ory and Applications. Dedicated to Professor Vsevolod M. Kuntsevich.
Studies in Systems, Decision and Control, Vol. 203. Cham: Springer
Nature Switzerland AG, 2019.
[6] A.A. Chikrii. “Control of Moving Objects in Condition of Conflict”.
In: Kuntsevich, V.M. et al. (Eds). Control Systems: Theory and Appli-
cations. Series in Automation, Control and Robotics, River Publishers,
Gistrup, Delft, 2018, pp. 17–42.
[7] A.A. Chikrii, et al. Method of Resolving Functions in the Theory of
Conflict-Controlled Processes. In: Kondratenko, Y.P., Chikrii, A.A.,
Gubarev, V.F., Kacprzyk, J. (Eds) Advanced Control Techniques in
Complex Engineering Systems: Theory and Applications. Dedicated to
Prof. V.M.Kuntsevich. Studies in Systems, Decision and Control, Vol.
203. Cham: Springer Nature Switzerland AG, 2019, pp. 3–33. DOI:
https://doi.org/10.1007/978-3-030-21927-7_1
[8] A.A. Belousov, V.V. Kuleshyn, “Game approach to control of running
start of aircraft on its take off”, J. of Automation and Information
Sciences, 2012, Volume 44, Issue 8, pp. 78–84.
References 415
Abstract
In this chapter we develop a cognitive map (CM) of the use of cryptocur-
rency in the financial market and describe dynamic model of CM impulse
processes as a difference equation system (Roberts equations) based on
that CM. We chose an external control vector for the CM impulse process
provided by means of CM nodes varying. A closed-loop CM impulse process
control system was implemented. This system includes multidimensional
discrete controller, based on automated control theory methods, that generates
selected control vector and directly affects respective CM nodes by varying
their coordinates. We solved three problems of a discrete controller design for
automated dynamic processes control applied to cryptocurrency in financial
markets. The first problem is unstable cryptocurrency rate stabilization based
on modal CM impulse process control. The second problem is constrained
external and internal disturbances suppression during CM impulse processes
control based on invariant ellipsoid method. The third problem is minimizing
generalized variance of CM nodes coordinates and controls for stabilizing
coordinates at given levels. In this chapter, we developed a system for
identifying CM weighting coefficients based on a recurrent least-squares
417
418 Automated Control Problem for Dynamic Processes Applied to Cryptocurrency
method. We did some performance research for each of the designed discrete
controllers.
17.1 Introduction
To study the dynamic processes applied to cryptocurrency the chapter uses
cognitive modelling, which is one of the most relevant areas of scientific
and practical research of complex systems of different nature now [1–4].
It is based on the notion of a cognitive map (CM), which is a weighted
directed graph [5]. The nodes (nodes) are coordinates (factors) of the com-
plex system and weighted edges (arcs) of the graph describe causal links
between CM nodes. When disturbances affect CM nodes, we can observe
impulse transitional process. Equation (17.1) describes the dynamics of that
process [6]:
!n
Δyi (k + 1) = aij Δyj (k), (17.1)
j=1
options of modal control with one and with several controls and carried out
the simulation of a real modal control system of the CM impulse process in
the commercial bank.
its development. Exchanges form the price of bitcoin. During formation of the
initial price of bitcoin (during mining), it’s worth in the range from 2000 to
3000 dollars, but expectations of bitcoin becoming a conventional payment
system allow it to maintain its price at 9000-10000 dollars. But any financial
news might drastically decrease the price.
There are various perturbations to the processes of using cryptocurrency,
aimed at reducing the level of confidence in the use of cryptocurrency.
Therefore, there are the following risks that can arise in crypto operations.
1. Risk of losing users, which leads to a decrease in the price of bitcoin;
2. Risk associated with inappropriate general expectations of many users
at the same time, which traders create by manipulating the trades on the
exchanges or breaking crypto-related news;
3. The risk of a sharp change in the price of cryptocurrency as a result of
fraud on exchanges like high-frequency trading that usually lies in the
high speed of buying monetary assets before the large operation of an
investor and then selling it to the investor, before the information about
the purchase reaches all exchanges;
4. The risk associated with the lack of guarantee of the safety of capital
invested in the purchase of cryptocurrency, which leads to a certain
hysteria of users during trading on exchanges;
Figure 17.1 shows the CM, which is developed on the basis of causal
relationships.
We introduce the following notations of the CM nodes:
1. rate of cryptocurrency (bitcoin value),
2. trading volume of cryptocurrency,
3. capitalization volume,
4. number of cryptocurrency users,
5. investment volume (interest in bitcoin from institutional investors),
6. volume of speculation in cryptocurrency,
7. supply of cryptocurrency,
8. variance of cryptocurrency rate,
9. demand for cryptocurrency,
10. indirect profit,
11. level of confidence in cryptocurrency,
12. integrated level of risk when using cryptocurrency.
We can measure or compute nodes 1-8. However, we cannot measure nodes
9-12 or it’s hard to measure them.
17.2 Cognitive Mapping of Cryptocurrency Usage in Financial Markets 421
Here the size m of the control vector is smaller than the size n of the vector of
CM nodes coordinates Ȳ (k). This version of the modal control can be found
in [10]. If the pair (A, B) in (17.3) is controllable, the following algorithm
determines the feedback matrix K.
1. We set the desired spectrum of modes λ1 , . . . , λn of the closed-loop
system ΔY (k + 1) = (A − BK)ΔȲ (k), where all λj , j = 1, . . . , n are
different and modulo less than one. Moreover, there are no eigenvalues
of matrix A among λj .
2. We introduce Rj , j = 1, . . . , n – eigenvectors of the state matrix of the
closed system (A-BK), for which this relation is satisfied:
Consider the case when control actions Δu1 (k), Δu2 (k), Δu3 (k),
Δu4 (k), and Δu5 (k) are applied to the following nodes of CM respectively:
2 - trading volume, 3 - capitalization volume, 5 - investment volume, 6 -
speculation volume, 7 - supply of cryptocurrency. In this case, the matrix B
has the following form:
⎛ ⎞
0 1 0 0 0 0 0 0 0 0 0 0
⎜ 0 0 1 0 0 0 0 0 0 0 0 0 ⎟
⎜ ⎟
B =⎜
T ⎟
⎜ 0 0 0 0 1 0 0 0 0 0 0 0 ⎟.
⎝ 0 0 0 0 0 1 0 0 0 0 0 0 ⎠
0 0 0 0 0 0 1 0 0 0 0 0
We form the matrix P for the design of the modal controller as follows:
⎛ ⎞
1 1 1 1 1 1 1 1 1 1 1 1
⎜ 1
⎜ 0 0 1 1 0 0 1 0 1 1 0 ⎟ ⎟
P =⎜ ⎜ 0 0 1 0 0 1 0 0 0 0 1 1 ⎟ ⎟.
⎝ 0 1 1 1 1 1 0 1 0 1 1 1 ⎠
1 1 0 0 0 0 0 1 1 1 0 0
Based on methods (17.6), (17.7), (17.8), we define the feedback matrix of the
modal controller (17.9).
Figure 17.2 shows the results of modelling the CM system of equations
under five control actions, given that initial negative impulse “−1” affects
node 11 “confidence level”. The graphs show transients of the CM node
coordinates yi , i = 1, . . . , n, coordinate increments Δyi (k), and controls
increments Δu1 (k), Δu2 (k), Δu3 (k), Δu4 (k), and Δu5 (k).
Based on the analysis of the graphs, we can state the following patterns
of changes in the CM node coordinates and their increments during the
transition process:
• The coordinates: y7 (supply of cryptocurrency) and y10 (indirect profit)
increased;
• The following coordinates decreased: y1 (cryptocurrency rate), y2 (trad-
ing volume) and y4 (number of users), y8 (rate variance), y12 (risk
level).
The node coordinate increments Δyi (k) , i = 1, . . . , 12 during the sta-
bilization transient converge to zero, i.e. we suppress all perturbations.
The increments of all control actions Δui (k) converge to zero after the
coordinates of nodes yi are stabilized.
426 Automated Control Problem for Dynamic Processes Applied to Cryptocurrency
y (k ) y (k ) y (k ) y (k )
1 2 3 4
5 6 5 5
4 4 4
3 2 3
2 0 -5 2
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
y (k ) y (k ) y (k ) y (k )
5 6 7 8
8 8 12 6
10
7 7 4
6 6 2
6
5 5 4 0
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
y (k ) y (k ) y (k ) y (k )
9 10 11 12
10 10 5 5
8 8 4
4.5
6 6 3
4 4 4 2
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
y (k ) y (k ) y (k ) y (k )
1 2 3 4
0 5 0 1
-1 0.5
-0.5
0 -2 0
-1
-3 -0.5
-1.5 -5 -4 -1
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
y (k ) y (k ) y (k ) y (k )
5 6 7 8
1.5 4 6 2
1 4 0
2
0.5 2 -2
0
0 0 -4
-0.5 -2 -2 -6
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
y (k ) y (k ) y (k ) y (k )
9 10 11 12
2 2 1 1
0.5
1 1 0
0 0 -1
-0.5
-1 -1 -1 -2
0 5 10 15 20 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
0
0
-5
-10 -5
0 5 10 15 20 0 5 10 15 20
u (k ) u (k )
3 4
2 5
1
0
0
-1 -5
0 5 10 15 20 0 5 10 15 20
u (k )
5
5
-5
0 5 10 15 20
C, Ψ are compiled from the coefficients of the matrix of the initial model
(17.2) of the CM impulse process. The matrices D, Ψ show the relationships
between the first (17.10) and the second (17.11) parts of the initial CM (17.2).
The values of non-measurable coordinates ΔZ(k) are taken into account as
external constrained perturbations with unknown probability characteristics
in the first system of equations (17.10) of the CM model compiled for impulse
processes with measurable coordinates X.
We synthesized a control vector to suppress constrained perturbations
ΔZ(k) by implementing static state feedback:
which acts directly on the measured coordinates of the nodes X of the first
control system according to the state equation:
in which the vectors and matrices have the following dimensions: dim ΔX =
n; dim ΔZ = p;dim Δw = n;A1 (n × n); D(n × p), I is a unit matrix of
dimension n × n. We assume that the internal and external perturbations are
17.4 The Problem of Constrained Internal and External Disturbances 429
if from ΔX(0) ∈ εΔX the condition ΔX(k) ∈ εΔX follows for all discrete
moments of time k = 1, 2, 3, . . . . Then the matrix P is called the matrix of
the ellipsoid εΔX .
Let us prove the invariance condition of the ellipsoid (17.17) under
perturbations (17.16). For this purpose, using the method [17], we introduce
the quadratic Lyapunov function V (X(k)) = ΔX̄ T QΔX̄ at Q > 0, con-
structed on the solutions of the system (17.15). To ensure that the trajectories
ΔX(k) 6 of the system (17.15) do not go 7beyond the boundary of the ellipsoid
εΔX = ΔX(k) ∈ Rn : V (X(k)) ≤ 1 , it is required that V (ΔX(k)) ≤ 1
at V (X(k)) ≤ 1, so we have:
ΔX̄ T (k + 1)QΔX̄(k + 1) =
8 9 I 1
= ΔX̄ T A1 T + Δw̄T (k) ΔZ̄ T (k) Q
DT
) * Δw(k)
A1 ΔX̄(k) + I1 D ≤1
ΔZ(k)
By multiplication, we get:
ΔX̄ T (k + 1)QΔX̄(k + 1) =
8 8 9 9
= ΔX̄ T Δw̄T (k) ΔZ̄ T (k)
⎡ ⎤
A1 T QA1 A1 T Q (I1 D )
T ⎡ ⎤
⎢ ⎥ ΔX(k)
⎢ ⎥⎣ Δw(k) ⎦
⎣ I1 I1 T ⎦
QA1 Q (I1 D ) ΔZ(k)
DT DT
430 Automated Control Problem for Dynamic Processes Applied to Cryptocurrency
⎡ ⎤
ΔX(k)
Let us apply the S-procedures [17, 19]. Let S = ⎣ Δw(k) ⎦. Then the
ΔZ(k)
quadratic forms we can represent as follows:
T
f0 (S) = S M0 S =
8 8 9 9
= ΔX̄ T Δw̄T (k) ΔZ̄ T (k)
⎡ T ⎤⎡ ⎤ ;
A1 T QA TQ
A1 1(I D ) ΔX(k)
⎣ I1 1
I1 T ⎦⎣ Δw(k) ⎦ ≤ 1
QA1 Q (I1 D )
DT DT ΔZ(k)
8 8 9 9
f1 (S) = S̄ T M1 S = ΔX̄ T Δw̄T (k) ΔZ̄ T (k)
⎡ ⎤
ΔX(k)
Q 0 ⎣
Δw(k) ⎦ ≤ 1;
0 0
ΔZ(k)
8 8 9 9
f2 (S) = S̄ T M2 S = ΔX̄ T Δw̄T (k) ΔZ̄ T (k)
⎡ ⎤
ΔX(k)
Q 0 ⎣
Δw(k) ⎦ ≤ 1,
0 I2
ΔZ(k)
If τ2 = 1 − τ1 then we get:
τ1 Q ≥ A1 T
−1
I1 ) * I1
Q + Q(I1 QT ) (1 − τ1 )I2 − Q I1 D Q A1
DT DT
(17.19)
According to the lemma of matrix inversion [20], we will have:
) * I1 ) * −1 I1
Q + Q I1 Q T (1 − τ1 )I2 − Q I1 D Q
DT DT
−1
) * I1
== Q−1 − (1 − τ1 )−1 I1 D
DT
Then expression (17.19) we can write as follows:
−1
−1 −1
) * I1
τ1 Q = A1 Q − (1 − τ1 )
T
I1 D A1
DT
Multiply from the left by A1 , and then from the right by A1 T , and
renominate τ1 = α. Then the linear matrix inequality will take the final form:
1 I1 + DDT
A1 P A1 T − P + ≤0 (17.20)
α (1 − α)
432 Automated Control Problem for Dynamic Processes Applied to Cryptocurrency
1 I1 + DDT
(A1 − BKp )P (A1 − BKp )T − P + ≤0 (17.23)
α (1 − α)
We consider the minimization of the trace of the ellipsoid matrix (17.17)
as the optimality criterion for the synthesis of the regulator (17.12) in this
chapter:
trP (α) → min, α∗ ≤ α < 1, (17.24)
This ensures minimization of the size ofthe invariant
ellipsoid (17.17) with
Δw(k)
the largest suppression of perturbations , which are constrained
ΔZ(k)
only by the maximum range (17.16). After multiplying the factors in the
inequality (17.23), we obtain:
1
(A1 P A1 T − BKp P A1 T − A1 P Kp T B T + BKp P Kp T B T )
α
I1 + DDT
−P + ≤0 (17.25)
(1 − α)
Inequality (17.25) is nonlinear with respect to P and Kp , which need to be
optimized. In [17] a linearization by replacement L = Kp P and introduction
of an additional constraint is proposed:
R L
≥ 0, (17.26)
LT P
17.4 The Problem of Constrained Internal and External Disturbances 433
with the estimated values of α̂, P̂ , L̂, R̂, providing minimization of criterion
(17.24) under constraints (17.26), (17.27).
⎛ ⎞
0.3 0.1 0.15 0.2 0.2 0.3 0 0
⎜
⎜ 0 0 0.1 0.1 0.1 0.35 0.35 0 ⎟ ⎟
⎜
⎜ 0.25 0 0.5 0 0.1 0 0 0 ⎟ ⎟
⎜ 0.4 0 0 0.3 0 0 0 0 ⎟ ⎟,
A1 = ⎜
⎜
⎜ 0 0.4 0 0 0.4 0 0 0 ⎟ ⎟
⎜
⎜ 0 0.5 0 0.2 0 0 0.6 −0.2 ⎟
⎟
⎝ 0 0 0 0.35 0.3 0 0 0 ⎠
−0.3 0 0 0 0 0 0 0
434 Automated Control Problem for Dynamic Processes Applied to Cryptocurrency
⎛ ⎞
0.2 0 0.3 −0.3
⎜ −0.3 0 0.3 −0.2 ⎟
⎜ ⎟
⎜ 0 0 0.5 0 ⎟
⎜ ⎟
⎜ 0 0 0.2 −0.2 ⎟
D=⎜ ⎟.
⎜ 0 0.4 0 −0.3 ⎟
⎜ ⎟
⎜ 0 0 0 −0.4 ⎟
⎜ ⎟
⎝ 0.5 0.2 0 0 ⎠
0 0 0 0
The control matrix B has the form given in the example of paragraph 3 (cut
to 8 rows).
We consider the step impacts with the amplitude of 1, acting at the initial
moment of time on the two unmeasured nodes 10 and 11 in the direction
of reducing the indirect profit (node 10) and confidence level (node 11) for
modelling the dynamics of the closed-loop control system of the impulse
process of this CM cryptocurrency using the proposed method, as external
disturbances. The internal perturbations in the simulation are formed as fol-
lows: the value of non-zero matrix coefficients vary at each sampling period
by the formula A1var (k) = A1 ξ(k), where ξ(k) is a normally distributed
random variable (Gaussian white noise). Only the values of A1 are used for
control, while A1var remains unknown. The initial levels of all coordinates of
the CM nodes are assumed zero for convenience.
The charts of the transients of the CM node coordinates, their increments
and controls are shown in Figure 17.3.
5 4 4
0 4.5 3 3
4 2 2
-5 3.5 1 1
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
y y y y
5 6 7 8
5 7 5 8
4.5 4
6 7
4 3
5 6
3.5 2
3 4 1 5
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
y y y y
9 10 11 12
5 5 5 9
4 4 8
3 3 0 7
2 2 6
1 1 -5 5
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
y y y y
1 2 3 4
0.5 0.5 0.5 0.5
0 0 0
-1 -0.5 -1 -1
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
y y y y
5 6 7 8
0.2 0.5 0.5 0.4
0
0 0.2
-0.2 0
-0.5 0
-0.4
y y y y
9 10 11 12
0.2 0.5 0 0.4
0
0 0.2
-0.2 -0.5
-0.5 0
-0.4
-0.6 -1 -1 -0.2
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
0.2 0.5
0 0
-0.2 -0.5
0 10 20 30 40 50 0 10 20 30 40 50
u u
3 4
0.4 1
0.2 0.5
0 0
-0.2 -0.5
0 10 20 30 40 50 0 10 20 30 40 50
u
5
0.4
0.2
-0.2
0 10 20 30 40 50
R1 0
in expression (17.32) is diagonal and positively determined,
0 R2
since on the basis of the rule of formation of matrices B and L(k)
it turns out that B T L(k) = 0, LT (k)B = 0. As a result, it is
shown
in [21] that the eigenvalues of the diagonal matrix (17.32) I −
B(B T B + R1 )−1 B T 0
will be mod-
0 L(k)(LT (k)L(k) + R2 )−1 LT (k)
ulo less than unity. As a result, the closed-loop system (17.31) will be
stable.
0 0 0 0 0 0 1 0
y (k ) y (k ) y (k ) y (k )
1 2 3 4
8 6 8 8
7 5.5 7 7
6 5 6 6
5 4.5 5 5
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
y (k ) y (k ) y (k ) y (k )
5 6 7 8
5.5 5.5 6 5
5.5
5 5 4.5
5
y (k ) y (k ) y (k ) y (k )
9 10 11 12
7 6.5 8 5
6 6 7 4.8
5 5.5 6 4.6
4 5 5 4.4
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
1
Ki (k) = Pi (k − 1)X i (k)X̄iT (k)Pi (k − 1)
1+ X̄iT (k)Pi (k − 1)X i (k)
(17.36)
1
Pi (k) = Pi (k − 1) −
1 + X̄iT (k)Pi (k − 1)X i (k)
Pi (k − 1)X i (k)X̄iT (k)Pi (k − 1).
The recurrent procedure (17.36) should be performed for each node of CM
Δyi (k), i = 1, 2, . . . , n at each sampling step. We use the obtained estimates
ˆ (k) as the values of the adjacency matrix A coefficients at the current step
Θ̄ i
of the control algorithm. For parametric identification of the adjacency matrix
A, we can also apply non-recurrent identification methods, outlined in [8, 26].
442 Automated Control Problem for Dynamic Processes Applied to Cryptocurrency
17.7 Conclusion
This chapter considers the cryptocurrency CM and solves three problems of
its application in the financial markets. The first problem covers the case of
instability when the cryptocurrency rate is unstable and decision-makers want
to stabilize it. For this case control actions are designed based on the modal
controller, which help to stabilize the system and assign desired values to the
closed-loop system poles. The second problem considers the cryptocurrency
system affected by unmeasurable internal and external disturbances, and the
only thing we know about these disturbances is that they are constrained.
For this case controls are designed based on the modified invariant ellipsoids
method which ensures that the system’s “trajectory” is limited inside of the
minimal size ellipsoid. The third problem discussed in the chapter relates to
the case when it is required to set some of the CM nodes at the new levels.
This problem is solved based on minimization of the generalized variance
criterion. Additional feature of this approach is that we may not only change
some resources associated with CM nodes but also change how one node
affects another one, i.e. to vary CM edges weights. As a result of solving
all three problems decision-makers obtain a sequence of control actions
which they can apply to the real cryptocurrency market. All the results were
mathematically simulated and the effectiveness of the suggested approaches
was proven.
References
[1] E. Papageorgiou, J. Salmeron, “A review of fuzzy cognitive maps
research during the last decade”, IEEE transactions on fuzzy systems,
vol. 21, no. 1, 2013, pp. 66–79.
[2] G. Gorelova, N. Pankratova (Eds.). Innovational development of socio-
economic systems based on methodologies of foresight and cognitive
modeling, Kyiv, Naukova dumka, 2015, 464 p. (In Russian)
[3] C. Stylios, P.P. Groumpos, “Modeling complex systems using fuzzy
cognitive maps”, IEEE Transactions on Systems, Man, and Cybernetics-
Part A: Systems and Humans, 34((1), 155–162, 2004.
[4] J. Aguilar, “A survey about fuzzy cognitive maps papers”, International
Journal of Computational Cognition, 3, N 2, P. 27–33, 2005.
[5] R. Axelrod, “The structure of decision: Cognitive maps of political
elites”, Princeton University Press, 1976.
References 443
B E
Bounded errors 167 Equilibrium problem 145, 148, 159,
Boundedness 63, 64, 67, 73, 76, 84, 163
91 Estimation 62, 88, 90, 154, 264, 370,
Braking distance 400 373, 407, 413
Estimation algorithm 83, 168
Euler–Lagrange and Hamiltonian
C inclusions 31
Complex discrete system 167 Evasion 4, 16, 104, 142
Conflict-controlled processes 3, 114, Extremal aiming rule 3
401
Constrained optimization 32, 380
Cryptocurrency 417, 418, 420, 421, F
423, 433 Feedback control system 65, 66, 89
Cyber-physical system 328, 329, Fuzzy controller 216, 218, 222, 264,
332, 333, 336 281
Fuzzy system 200, 202, 203, 224,
284
D
Data mining 236, 237, 380
Design space 265, 379, 380, 384, 394 G
Design and structural optimization Game theory 28, 128, 146
197, 204, 214, 228, 229 Games of approach 131, 138, 401,
Differential game 104, 113, 132, 142 415
445
446 Index
I Q
Identification 77, 81, 99, 167, 169, Quadrotor UAV 289, 293, 294, 311,
257 312, 313
Information technology 197, 199, Quadrotor unmanned aerial vehicle
204, 207, 209, 210, 212, 229 197, 199, 214, 229
Integral inequality 100, 101, 105
Invariant ellipsoid method 417
R
Real-time algorithm 399, 415
L Regularization 145, 149, 167, 175,
Linear regression 167, 171 177
Linear matrix inequalities 292, 427 Regularized dimension 167
Linguistic terms 197, 198, 201, 204, Resolving function 7, 118, 121, 125,
210, 212, 219, 222, 229 135
Logistics 198, 328, 335, 346, 348 Rejected takeoff 399, 400, 404, 409,
412, 413
Riccati equation 341, 344, 399, 401,
M 413
Marine moving object 291, 292 Robust adaptive controller 62
Minimum trace ellipsoid 380 Robust-optimal control 289, 294,
Minimum volume ellipsoid 380 308, 309
Modal control 291, 418, 419, 423 Route planning 327
O
Optimal control 31–36, 42, 147, 289, S
294, 303, 309 Scheme of iterative regularization
Optimality conditions 31, 35, 43, 46, 145, 146, 159
294, 317, Set-valued mapping and its
Outlier detection 380 selections 3
Index 447
State constraints 4, 17, 19, 20, Two-level problem 146, 147, 148,
48 116 159
State controller 423, 432 Two-stage proximal method 145
Stroboscopic strategy 6, 117
Strong convergence 41, 145, 160 U
Successive pursuit 3, 4, 6, 22, 25, 116 Uncertainty 4, 61, 66, 89, 114, 265,
SVD 174, 176, 177, 180 305, 422
Sweeping processes 31, 33, 38, 48,
59
V
Variational analysis 31, 33, 36, 43,
T 48, 55
Trajectory 4, 8, 25, 104, 182, 218, Variational inequality problem 145,
242, 301, 303 146, 158, 159
About the Editors
449
450 About the Editors