Chaos, Solitons and Fractals 136 (2020) 109827

Contents lists available at ScienceDirect

Chaos, Solitons and Fractals

Nonlinear Science, and Nonequilibrium and Complex Phenomena
journal homepage: www.elsevier.com/locate/chaos

A new approach to extreme event prediction and mitigation via

Markov-model-based chaos control
Hojjat Kaveh, Hassan Salarieh∗
Department of Mechanical Engineering, Sharif University of technology, Tehran, P.O Box 11155-9567, Iran

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

Article history: Despite that the border between chaotic and stochastic systems is exactly defined, scientists, use high di-
Received 10 October 2019 mensional chaotic dynamics to model numerous stochastic models and sometimes use stochastic models
Revised 10 April 2020
to study chaotic systems. In this paper, we have investigated chaotic systems with a stochastic approach
Accepted 15 April 2020
and proposed an estimator for the chaotic system which is used to present different algorithms for chaos
control, extreme event prediction and extreme event mitigation. The stochastic estimator is constructed
Keywords: by meshing the phase space and applying the cell mapping method (with some considerations) which
Chaos control provides us with a model-free approximation of the systems. The algorithms are ideal for real-world ap-
Markov model plications where there are some noises and the model of the system is unknown. Besides, the proposed
Statistical reconstruction
methods are adopted for when the control signal is limited to a few specific values. We have tested these
Model-free
algorithms on Logistic, Henon and a physiological control system.
Extreme event prediction
Extreme event mitigation © 2020 Elsevier Ltd. All rights reserved.

1. Introduction of the system take place in numerous phenomena like the motion
of tectonic plates, oceanic waves and many famous turbulent flow
With the recently-booming data science, the investigation of examples [12–15].
the nonlinear phenomena, using novel data analysis tools, has risen Mitigation of these extreme events has also a matter of im-
much attention. With the power of deep learning and machine portance that should be distinguished here with chaos control ap-
learning many seemingly complex problems in nonlinear dynamics proaches. In extreme event mitigation, the goal is to evade from
science can be solved [1,2]. Investigation of nonlinear systems from abrupt changes in the states of the system with the least change
the statistical view, have many advantages; as a case in point, pro- of the chaotic attractor [16]; the control action should be applied
found novel theories in data science can help to dismantle many to the system in an adaptive way so as not to lose the chaotic core.
high dimensional enigmatic problems [3,4]. This problem set-up bears in mind the necessity of a prediction al-
In the last decades, many chaos control approaches have been gorithm for extreme events.
developed. Using delayed feedback method [5], conventional non- Most of the time, predicting the extreme events needs a mea-
linear control approaches [6], Takens embedding theorem [7] and sure of the probability that when increases, shows that the emer-
using the system’s entropy [8] are some of the ways for chaos gence of the extreme events is more probable [16]. So investigation
control. Among the works done in the literature for chaos con- of the systems in a stochastic way is inevitable. As an example, in
trol, few works have been done using the statistical guise of the Ref [17] authors controlled the extreme event using the probabil-
chaotic systems which might open the door for many further types ity of the emergence of the extreme events, however, the control
of research in the field of nonlinear systems jointed with data sci- action was deterministic which was multiplied by the probability
ence. Besides, in most of the previous control approaches, having of the extreme event emergence. Our idea of mitigation is to con-
a known governing equation was requisite to design a proper con- struct a stochastic representation of the system and use it for pre-
troller [9,10]. diction in tandem with control which will reduce the complexities
Unlike chaos theory, the science of extreme events in nonlin- and have some other advantages which will be discussed shortly.
ear dynamical systems is new where there are many steps to be In one part of this paper, we introduce two model-free control
taken [11]. Extreme events which are sudden changes in the states approaches which work both for chaos stabilization and extreme
event prediction and mitigation. One of these methods has a sim-
ple formulation and with easier learning procedure and the other
∗
Corresponding author. one is designed for when the control signal is restricted to a fi-
E-mail address: [email protected] (H. Salarieh). nite set of values. These two methods assume that the unstable

https://doi.org/10.1016/j.chaos.2020.109827
0960-0779/© 2020 Elsevier Ltd. All rights reserved.
2 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827

fixed point(UFP) lies within the chaotic attractor. Then, we have have been predicted and their amplitude has been reduced. Then,
presented an algorithm, special for extreme event mitigation to re- the quality of the mitigation algorithms applied on this system are
lax the assumption of having a UFP within the chaotic attractor. To compared with each other. The stabilization of the UFP of the phys-
find the location of the UFPs of the unknown systems a remark is iological system is another part of this section. Finally, in the last
presented with the assumption of having a set of measured data section, a brief conclusion is presented.
with sufficiently large size.
One important limitation in the mitigation of extreme events is
that the utilized methods are model-based which might not work 2. Problem statement
in many different applications [17,18]. As an example, Ref [19] pro-
posed a model-based method that is applicable to specific kinds Consider a nonlinear map which represents a chaotic dynamical
of three-wave interactions. Our model-free mitigation methods can system:
open the doors for further researches in manipulating different
q(t + 1 ) = f(q(t ), u(t )). (1)
data science techniques to handle prediction and mitigation of ex-
treme events. where q(t) is the state vector in Rn , f(., .) is the nonlinear map
After the brief introduction in this section, the problem has which can be unknown here and u(t) is the control action. Assume
been clarified in Section 2 where two algorithms that work for that A is an open set around the chaotic attractor of the system.
chaos control, extreme event prediction and extreme event mit- We define m cubes denoted by Aj which cover the attractor as in
igation are presented. In Section 2, another approach special for Eq. 2. These cubes are utilized to estimate the location of the state
extreme events mitigation in tandem with some practical defi- variables stochastically and to design a Markov model for the sys-
nitions about the quantification of the quality of the prediction tem. By decreasing the size of these cubes we can increase the ac-
is presented. Finally, the numerical simulations are presented in
Section 3. We have stabilized the UFPs of the Logistic and the
Henon maps with two different algorithms. Furthermore, with the
presented algorithms, the extreme events of a physiological system 1

0.8
1
0.6
0.9
0.4
0.8

0.7 0.2
5 10 15
0.6

0.5 0.01

0.4 0.005

0.3
5 10 15 20 25 30 35 0

-0.005

0.1 -0.01
10 20 30 40

0.05
0.6

0
0.4
-0.05

0.2
-0.1

0
-0.15 5 10 15 20 25
5 10 15 20 25 30 35

Fig. 2. Algorithm II: Time series of the closed-loop Logistic map (a), control action
Fig. 1. Algorithm I: Time series of closed-loop Logistic map (a), control action (b). (b), and the Cost function (c) after applying controller with α = 1.
 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827 3

curacy of the model. ficiently long time [21]. If  is a very small positive number, and

m the state vector of the system satisfies the following condition, one
A⊂ A j. (2) can approximately estimate (qf ) using Eq. 3.
j=1
∃t : q(t + 1 ) − q(t ) <  → qf ≈ q(t ). (3)
Time series of the chaotic system is used to attain an expected
It should be noted that to find the UFP with the condition of Eq. 3,
value for the function f(., .). We have used a Markov model for
usually a large amount of data is required, which may not be pos-
this prediction by which we predict the probability of the future
sible in practice, but if we do not have the governing equations of
states of the system. Using the statistical properties of the system,
the system, the mentioned point in Remark 1 is an alternative so-
we are going to present some methods for chaos control and ex-
lution. For a system with limited available time series, the assump-
treme event prediction and mitigation.
tion of ”coming sufficiently close to the UFP”, i.e. Eq 3, should be
2.1. Chaos control satisfied by the available data of the system.

In this part, two model-free algorithms special for chaos stabi- 2.1.1. Algorithm I
lization are presented. Algorithm II is designed for when the con- The first chaos control algorithm is based on the stochastic
trol signal is limited to be a member of a finite set of values. It model of the system in Eq. 1. We propose the following control
is important to note that in algorithms I and II, we have assumed signal to stabilize UFP of the system (qf ).
that the UFP, within the chaotic attractor, is known. For when the q(t + 1 ) = f(q(t )) + u(t )
. (4)
governing equation of the system is unknown, the value of UFP (qf ) u(t ) = −ˆq(t + 1 | t ) + qf
can be approximated by using Remark 1 [20].
ˆ (t + 1 | t ) is defined as Eq. 5.
where q
Remark 1: A chaotic orbit, if it is recurrent, can enter into any
arbitrarily small neighborhood of the UFP (qf ) after passing a suf- ˆ (t + 1 | t ) = E (q(t + 1 ) | q(t ), q(t − 1 ), . . . q(t − (r − 1 ))).
q (5)

Fig. 3. Meshed phase portrait of the Henon map with finer mesh around the fixed point.
 4 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827

Fig. 4. Probability distribution for q(t + 1 ) when q(t), q(t − 1 ) and q(t − 2 ) are known and u(t ) = 0.35 .

where r is the order of the Markov model and is recommended to ˆ can be calculated using
After completing these steps matrix P
be chosen a value more than but close to the order of the system. Eq. 6.
When the governing equation of the system is unknown, one can
C
use the method introduced in [22] to find r. Note that E(.) is the Pˆls = m ls . (6)
expected value operator and q ˆ (t + 1 | t ) is the expected value of z=1 Clz
the open-loop system at time t + 1. By using this simple strategy,
where Pˆls is the lth row and sth column in Pˆ and Cls is the lth row
it can be easily shown that the expected value of the system would ˆ will converge
and sth element in the Matrix C. The definition of P
converge to the UFP of the system. To find the expected value of
to the probability (P(.)) of any future event in the attractor based
the system, we use statistical properties of the attractor by which
on its past r values. We estimate the future state of the system
we propose a Markov chain to predict the future states. For this
with the Eq. 7.
purpose, the following steps should be taken.

m
ˆ (t + 1 | t ) 
q cj .P (q(t + 1 ) ∈ A j | q(t ), q(t − 1 ), . . . ,
j=1
(i) Mesh the phase portrait of the system with ‘m’ cubes, with the
q(t − (r − 1 ))). (7)
extra fine meshes around the UFP.
(ii) Construct a Matrix (Matrix C) with mr rows and m columns where cj denotes center of the jth grid (Aj ) which is in the and Rn
with its elements initially set to zero. Note that m is the num- can be found by averaging of all of the points in Aj and P(.) is the
ber of cubes and r is the order of Markov model. This matrix probability function.
is 2-dimensional whose rows show the location of q(t ), q(t −
1 ), . . . , q(t − (r − 1 )) within the meshed phase space, and its
2.1.2. Algorithm II
columns denote the location of q(t + 1 ).
This algorithm is devised for when the control actions are lim-
(iii) For any q(t + 1 ) find its corresponding cube within the meshed
ited to a finite set of k discrete values. Assume that the u(t) in
space and add number one to its corresponding element.
Eq. 1 is restricted to be a member in U in Eq. 8.
(iv) Do the last procedures for sufficient steps to have a fine proba-
bility distribution for any incident. U = {u1 , u2 , . . . , uk }. (8)
 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827 5

of q(t ), q(t − 1 ), . . . , q(t − (r − 1 )) and also the value of con-

1.5 trol signal u(t), and its columns denote the location of q(t + 1 ).
Since we mesh the phase space with m cubes and the control
signal has k discrete values, the number of rows is mr k. Because
1 q(t + 1 ) may set outside the meshed phase space, we set m + 1
columns for our Markov model (Matrix C), the last one for any
0.5 q(t + 1 ) which is not inside any mesh.
(iii) For each t in Eq. 1, randomly choose a control signal from the
available ones (between k discrete values) in Eq. 8.
0
(iv) Do the step (iii) in Algorithm I
(v) Do the step (iv) in Algorithm I
-0.5
After completing these steps Matrix P̄ can be calculated using
Eq. 11.
-1
Cls
P̄ls = m+1 . (11)
-1.5 z=1 Clz
120 130 140 150
where P̄ is similar to Pˆ in Algorithm I, unless the rows of the P̄
not only denotes the past events but also a corresponding control
signal.

0.1
2.2. Extreme events prediction and mitigation

In this part, after reviewing some definitions from Ref

[16] about extreme events, the applications of Algorithm I and Al-
0.05 gorithm II in prediction and mitigation are presented. Then, an-
other extreme event mitigation algorithm is introduced to relax
the assumptions of the first two algorithms. It is important to note
that we do not assume that the extreme events are deterministic,
since we use a stochastic representation of the dynamical system
0 and in case of having a good indicator for the stochastic extreme
events, these algorithms work.

2.2.1. Definitions
-0.05 Definition 1 For the system of Eq. 1 the observer h: A → R
120 130 140 150 is a measure by which we define a threshold for the system (he )
which signifies the extreme events [16] (A is an open set around
the chaotic attractor). As an example, for a heat transfer system,
Fig. 5. Algorithm I: Time series of the closed-loop Henon map (a), control action the function h, can be the average temperature of the system and
(b).
(he ) can be a value close to the ultimate allowed temperature.
The extreme event set Eh (he ) of a dynamical system is defined
To find the best control signal at each step we define a cost func- based on h: A → R, corresponding to the threshold he :
tion (L) which we are going to minimize (Eq. 9). Eh ( he ) = {q(t ) ∈ A : h(q(t ) ) > he } (12)
L(t, u(t )) = E (| q(t + 1 ) − qf | +α u(t )2 ). (9)
For the purpose of this paper, the observer h: A → R is defined
where α is a positive real number. Eq. 9 can be rewritten (by some as one of the states of the system, named by qi (t), whose value
simplifications) as Eq. 10. plays the role of extreme event measure.
Definition 2 For a dynamical system the indicator g(.) is a mea-

m
L(t, u(t ))  (|cj − qf |.P (q(t + 1 ) ∈ A j | q(t ), q(t − 1 ), . . . , sure by which we signalize an upcoming extreme event. The indi-
j=1 cator g(.), with a threshold ge is ideal when if g(.) > ge it implies
that h(.) > he [16]. In this paper we define the function g: Ar → [0,
q(t − (r − 1 )), u(t ))) + α u(t )2 . (10)
1] as Eq. 13:
To estimate the probability function and finding the proper 
Markov model, the following steps should be made which are akin g(q(t ), q(t − 1 ), . . . , q(t − (r − 1 ))) = ˆ
s∈S Pls. (13)
to the Algorithm I: S = {s | h ( cs ) > he }

(i) Mesh the phase portrait of the system as the same as the In Eq. 13, Pˆ is the probability matrix which was defined in Algo-
meshes in Algorithm I, in tandem with setting k equal to the rithm I, l is the corresponding row which signifies the meshes of
number of available control signals. the past events (q(t ), q(t − 1 ), . . . , q(t − (r − 1 ))) and cs is the cen-
(ii) Construct a Matrix (Matrix C) with mr k rows and m + 1 ter of the sth mesh. g(.) gets past r events of the system and pro-
columns with its elements initially set to zero. Note that m poses the probability of an extreme event emergence.
is the number of cubes, r is the order of the Markov model Definition 3 Pee is the probability of the extreme event emer-
and k is the number of available values for the control sig- gence which measures the probability of q(t + 1 ) ∈ Eh (he ) for a
nal. This matrix is 2-dimensional whose rows show the location given g(q(t ), q(t − 1 ), . . . , q(t − (r − 1 ))) [16].
6 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827

1.5

0.5

0
1.5

-0.5

-1 1
5 10 15 20 25 30 35

0.5
0.2

0
0 5 10 15 20 25 30 35

-0.2

-0.4
10 20 30 40

Fig. 6. Algorithm II: Time series of the closed-loop Henon map (a), control action (b), and the Cost function (c) after applying controller with α = 0.05.

2.2.2. Application of Algorithm I and II in extreme event mitigation 16

The mitigation laws based on the Algorithm I and Algorithm
II are proposed as Eqs. 14 and 15. These two algorithms assume 14
that the UFP lies within the attractor and has a value less than he .
For a system with unknown governing equation one should refer
12
to Remark 1 to find the value of the UFP. Algorithm II is designed
for when the control signal is limited to a few specific values. The
10
mitigation strategy in these two algorithms is stabilizing the UFP
of the system as soon as the indicator predicts the extreme events.
The statistical representation of the system helps us to predict ex- 8
treme events (g(.)) and to apply a proper control signal by provid-
ing an estimation of the system. 6
⎧
⎨if: ge ≤ g(q(t ), q(t − 1 ), . . . , q(t − (r − 1 )))
⎪ 4
u(t ) = −ˆq(t + 1 | t ) + qf ,
. (14)
⎪if: ge > g(q(t ), q(t − 1 ), . . . , q(t − (r − 1 )))
⎩ 2
u(t ) = 0
⎧
⎨if: ge ≤ g(q(t ), q(t − 1 ), . . . , q(t − (r − 1 )))
⎪ 0
u(t ) = −argmin(L(t , u(t ))) + qf 120 125 130 135 140 145
. (15)
⎪if: ge > g(q(t ), q(t − 1 ), . . . , q(t − (r − 1 )))
⎩
u(t ) = 0 Fig. 7. Time series of the physiological control system .

2.2.3. Algorithm III

In algorithm III the assumption of having an UFP within the
chaotic attractor which has a value less than he has been relaxed. t < t  < t + w − 1. Our goal here is to push the system to the de-
In this part, we define qd(w ) (t + 1 ) as the desired future state at sired state (qd(w ) (t + 1 )) with a proper control signal (u(t)) that
(t + 1 ) such that it guarantees the future w states of the sys- eliminates the necessity of the control action at t < t  < t + w − 1.
tem to have the value less than he without any input signal for The following equations give a proper estimation of qd(w ) (t + 1 ) for
 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827 7

Fig. 8. Meshed phase space of the physiological control system.

our purpose: As an example, qd(2 ) (t + 1 ) for when the Markov order is two, is
any q ∈ A which has the following condition:

h(qd(2) (t + 1 )) < he
(w )
h(qd (t + 1 )) < he g(qd(2) (t + 1 ), q(t )) < ge (17)
if: w = 1
E (q(t + r )), . . . , E (q(t + 2 )), qd(w ) (t + 1 ) ∈ A E (q(t + 2 ) | qd(2) (t + 1 ), q(t )) ∈ A
⎧
⎪ h(qd(w ) (t + 1 )) < he
⎪
⎪ It is important to note that qd(w ) (t + 1 ) might not have a unique
⎪
⎪ g(E (q(t + 1 )), q(t ), . . . , q(t − (r − 2 ))) < ge
⎪
⎪ g(E (q(t + 2 )), E (q(t + 1 )), . . . , answer in A. Using only one of the answers of the qd(w ) (t + 1 ) in
⎪
⎪
⎨ q(t − (r − 3 ))) < ge the control action might push the system into an unstable periodic
if: w > 1 .. orbit which is undesired for the purpose of mitigation where sav-
⎪
⎪ . ing the chaotic core is required. To avoid stabilization of unwanted
⎪
⎪
⎪
⎪ g(E (q(t + w − 1 ), E (q(t + w − 2 ), . . . , periodic orbits, one should choose different values of qd(w ) (t + 1 )
⎪
⎪ E (q(t + w − r )) < ge
⎪
⎩ in the control procedure. By using Eq. 18 as the control signal, the
E (q(t + r )), . . . , E (q(t + 2 )), qd(w ) (t + 1 ) ∈ A expected value of the system at time t + 1 would be qd(w ) (t + 1 )
which will save future w states of the system from extreme be-
(16)
haviour.
u(t ) = qd(w ) (t + 1 ) − q
ˆ (t + 1 | t ). (18)
It is recommended that to have E (q(t + 1 )), E (q(t + where qˆ (t + 1 | t ) can be calculated using Pˆ in Algorithm I. It is
2 )), . . . , E (q(t + r )) to be in the attractor, small values for the worth mentioning that Algorithm I acts like a special case of Al-
control signal (Eq. 18 which will be discussed shortly) should be gorithm III, since qf is the limiting case of qd(w ) (t + 1 ) for when w
used. goes to infinity.
8 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827

25 (20)

Although matrix C seems to be large, most of its rows are equal

20 to zero (almost 86%), stemming from the relatively high order of
the Markov model and a fine mesh which has made the probability
of many incidents to be equal to zero.
15 The closed-loop controlled system based on Algorithm I and its
corresponding control signal is depicted in Fig. 1. Time series of
10 the Logistic map after applying the control signal based on the Al-
gorithm II which has been found by minimizing the cost function
of Eq. 10 is depicted in Fig. 2 (a) and its control signal and the cost
5 function with α = 1 is plotted in Fig. 2 (b,c).

0 3.2. Henon map

50 100 150 200
The governing equation on the Henon map is as Eq. 21 and
is chaotic when γ = 1.4, β = 0.3 [23]. Without using Eq. 21,
with 20,0 0 0 data points the UFP of this attractor is found to be
qf ≈ [0.631, 0.188].

0 q1 (t + 1 ) = q2 (t ) + 1 − γ q1 (t )2 + u(t )
q2 (t + 1 ) = β q1 (t ). (21)
-2 where u(t) is the control signal which is assumed to be applied to
the first state. After meshing the phase space (Fig. 3) with m = 35,
-4 C matrix for Algorithm I and Algorithm II is constructed respec-
tively from which P ˆ and P̄ are derived. The order of the Markov
-6 model is considered to be 3 and the control signal for Algorithm II
has the following restriction:
-8
u(t ) ∈ {−0.35, 0.35, −0.1, 0.1, −0.05, 0.05, −0.02, 0.02, −0.015, 0.015, 0}.

-10 (22)

Fig. 4 shows the probability distribution for the q(t + 1 ) when q(t),
50 100 150 200
q(t − 1 ) and q(t − 2 ) are located in the positions specified in the
figure and the control signal is assumed to be 0.35 (Algorithm II).
Fig. 9. Stabilization of UFP of the physiological control system (a), Control action The closed-loop system based on the control law of Algorithm
(b). I is plotted in Fig. 5. The time series of the closedloop system, the
control action and the cost function for the Algorithm II are plot-
3. Simulations and results ted in Fig. 6. It is evident from the control action that the expected
value of u(t) is zero but it has a variance around its mean. To re-
In this section, two famous chaotic maps, the Logistic and the duce this variance, more applicable control signals (Eq. 22) and a
Henon maps, and a physiological control system are tested with finer mesh should be used.
the proposed Algorithms. The Algorithms I and II are applied to
the Logistic and the Henon system. Then, the physiological con- 3.3. Physiological control system
trol system is stabilized using Algorithm I. Moreover, the extreme
events of this system are predicted and their amplitude is reduced The governing equation of the discrete-time hematopoiesis
using Algorithm I, II and III. Then, the quality of the prediction and model with mature circulating cells has the following form [24]:
the mitigation methods are evaluated.

b c
q1 (t + 1 ) = aq1 (t ) + + + u(t )
3.1. Logistic map 1 + q2 (t )M 1 + q1 (t )M . (23)
q2 (t + 1 ) = q1 (t )
The governing equation on the Logistic map with control input This system has a chaotic response for when M = 15, a = 0.3, b =
u(t) is as Eq. 19 where μ is a constant and the system is chaotic 10 and c = 12 and its extreme events are evident in its time
when μ = 4. We have not used Eq. 19 in our simulations and only series which is depicted in Fig. 7. The UFP of this system is
the time series of the Logistic map is used here. Using remark 1 qf ≈ 1.2373499 and if Eq. 23 is unknown, it can be approximated
with 20,0 0 0 data points, the UFP is estimated to be qf ≈ 0.749. by Remark 1 to be qf ≈ 1.239 with 20,0 0 0 data points.
q(t + 1 ) = μq(t )(1 − q(t )) + u(t ). (19) We are going to stabilize the UFP of the system. Then, using
the different algorithms we are going to predict and mitigate its
After meshing the phase space with the considerations ex-
extreme events. To do so, we should mesh the phase space to con-
plained in the previous section, the matrix C is constructed with
struct a statistical representation of the system. We have used 75
the number of meshes (m) equal to 37 and the Markov order of 3.
meshes (m) and Markov order 2 (r). The meshed phase space is
For the simulations based on algorithm II, available control signals
depicted in Fig. 8. To control this system, we have used Algorithm
are assumed to have the limitation of Eq. 20. It is worth mention-
I by which the stabilized UFP of the system and the control signal
ing that the least non-zero magnitude of the available control sig-
are shown in Fig. 9.
nal should be small enough to handle small perturbations around
In the next step, we are going to predict the extreme events
the UFP.
using function (g(.)). In this example, we have used the Markov or-
u(t ) ∈ {−0.1, 0.1, −0.05, 0.05, −0.02, 0.02, −0.01, 0.01, −0.001, 0.001, 0}. der r = 2 for the function (g(.)) to present the probability of the
 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827 9

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 0 0.5 1

Fig. 10. The quality of the extreme events prediction.

15

10

15
5

10
0
0 100 200 300 400 500
5

15
0
0 100 200 300 400 500
10

0
0 100 200 300 400 500

Fig. 11. Extreme event amplitude reduction of the chaotic physiological system, Algorithm I (a), Algorithm II (b), Algorithm III (c).
10 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827

0.3

0.25

0.2

0.15

0.1

0.05

0
2 4 6 8 10 12

Fig. 12. PDF of the time-series of the system before and after applying different mitigation algorithms.

extreme event emergence (Eq. 13). The quality of the prediction is Moreover, the stochastic representation of the systems is used
depicted in Fig. 10. It is evident that when the indicator (g(.)) esca- to predict extreme events and to design some methods for their
lates, the observer of the system (h(.)) becomes higher and higher mitigation. The extreme events of a physiological system are pre-
until it reaches extreme behaviour (Fig. 10(a)). Moreover, there is dicted using the concepts proposed in this paper. To mitigate ex-
a monotone function between Pee and g which shows that if the treme events, three methods are presented. Two of the algorithms
value of the function g increases, the extreme behavior of the sys- assume that the UFP of the system, within the chaotic attractor, is
tem is more probable. available, or (in case of having enough data set) is previously es-
To reduce the amplitude of the extreme events, Algorithm I, timated by some numerical analysis. It is also assumed that the
II and III are applied to the system. For when the algorithm II mentioned UFP is in a region in which we do not have extreme
is applied to the system, the control signal has the restriction of events. The third algorithm is proposed to relax these assumptions.
Eq. 24 and when the algorithm III is applied to the system, w is Using PDF before and after the mitigation, the quality of the three
set to 1. algorithms in the sense of preserving the primary chaotic attractor,
has been evaluated. It is inferred that stabilization of UFP of the
u(t ) ∈ {−10, −8, −6, −1, −0.3, −0.2, −0.15, 0.15, −0.1, 0.1
system for a short period (after predicting the extreme event emer-
, −0.03, 0.03, −0.01, 0.01, −0.005, 0.005, 0}. (24) gence) can be regarded as a serious solution for extreme event
It is important to note that Algorithms I and II do not necessar- mitigation beyond the examples of this paper.
ily guarantee the system to have values less than he , since these In the future works, the application of the concepts of this work
algorithms try to stabilize UFP for when g > ge , and this stabiliza- on higher dimensional systems will be evaluated. Moreover, we
tion might lead to values more than he in transient time. To avoid are going to reduce the computational costs by meshing only the
this in Algorithm I and II, small control signals should be used and phase space around the UFP in the chaotic region and relax the
the number of control actions after sensing a high probability of condition of the availability of all states using Takens embedding
extreme event emergence, should be found using trial and error theorem.
procedure. The amplitude reduction of the extreme event of this
system based on three different algorithms has been depicted in Declaration of Competing Interest
Fig. 11. To judge the quality of the extreme event amplitude re-
duction in the sense of preserving the chaotic core, the Probability The authors declare that they have no known competing finan-
Distribution Function (PDF) of the time-series of the system before cial interests or personal relationships that could have appeared to
and after applying Algorithm I, II and III are depicted in Fig. 12. influence the work reported in this paper.
Based on Fig. 12 Algorithm I has the closest PDF to the open-loop
system. CRediT authorship contribution statement

4. Discussion and conclusion Hojjat Kaveh: Methodology, Software, Formal analysis, Writing
- original draft. Hassan Salarieh: Conceptualization, Methodology,
We have used the Markov model in tandem with the cell map- Writing - review & editing, Supervision.
ping method, to have a statistical model-free representation of
chaotic systems. The statistical representation of the system is used References
in two different areas: 1- chaos control 2- extreme event predic-
[1] Anderson PW. The economy as an evolving complex system. CRC Press; 2018.
tion and mitigation.
[2] Wang B, Mao Z, Huang K. Detecting outliers for complex nonlinear systems
For chaos control, we have assumed that the UFP of the system with dynamic ensemble learning. Chaos Solitons Fractals 2019;121:98–107.
lies within the chaotic attractor. Two algorithms are presented in [3] Yamins DLK, DiCarlo JJ. Using goal-driven deep learning models to understand
this part, one of which can control chaos with limitation on the sensory cortex. Nat Neurosci 2016;19(3):356.
[4] Jokar M, Salarieh H, Alasty A. On the existence of proper stochastic markov
control signal. The algorithms are tested on three different systems. models for statistical reconstruction and prediction of chaotic time series.
The results show the stabilization of the UFPs of these systems. Chaos Solitons Fractals 2019;123:373–82.
 H. Kaveh and H. Salarieh / Chaos, Solitons and Fractals 136 (2020) 109827 11

[5] Pyragas K. Continuous control of chaos by self-controlling feedback. Phys Lett [15] Sinha A, Horvath PA, Beason T, Roos KR. Simulation of a financial mar-
A 1992;170(6):421–8. ket: the possibility of catastrophic disequilibrium. Chaos Solitons Fractals
[6] Shirkavand M, Pourgholi M. Robust fixed-time synchronization of fractional or- 2019;125:13–16.
der chaotic using free chattering nonsingular adaptive fractional sliding mode [16] Farazmand M, Sapsis TP. Extreme events: mechanisms and prediction. Appl
controller design. Chaos Solitons Fractals 2018;113:135–47. Mech Rev 2019;71(5).
[7] Kaveh H, Salarieh H. Control of continuous time chaotic systems with un- [17] Farazmand M, Sapsis TP. Closed-loop adaptive control of extreme events in a
known dynamics and limitation on state measurement. J Comput Nonlinear turbulent flow. Phys Rev E 2019;100:033110. doi:10.1103/PhysRevE.100.033110.
Dyn 2019;14(1):11007. [18] Saha A, Feudel U. Extreme events in fitzhugh-nagumo oscillators coupled with
[8] Januário C, Grácio C, Mendes DA, Duarte J. Measuring and controlling the two time delays. Phys Rev E 2017;95(6):62219.
chaotic motion of profits. Int J Bifurc Chaos 2009;19(11):3593–604. [19] Santos MS, Szezech JD, Batista AM, Iarosz KC, Caldas IL, Viana RL. Dragon-kings
[9] Lee HJ, Park JB, Chen G. Robust fuzzy control of nonlinear systems with para- death in nonlinear wave interactions. Physica A 2019;534:122296. doi:10.1016/
metric uncertainties. IEEE Trans Fuzzy Syst 2001;9(2):369–79. j.physa.2019.122296.
[10] Salarieh H, Kashani SMM, Vossoughi G, Alasty A. Stabilizing unstable fixed [20] Kaveh H, Salarieh H, Hajiloo R. On the control of unknown continuous time
points of discrete chaotic systems via quasi-sliding mode method. Commun chaotic systems by applying takens embedding theory. Chaos Solitons Fractals
Nonlinear Sci Numer Simul 2009;14(3):839–49. 2018;109:53–7.
[11] Galuzio PP, Viana RL, Lopes SR. Control of extreme events in the bubbling on- [21] Gulick D. Encounters with chaos. McGraw-Hill; 1992.
set of wave turbulence. Phys Rev E 2014;89(4):40901. [22] Kennel MB, Brown R, Abarbanel HDI. Determining embedding dimension for
[12] Farazmand M, Sapsis TP. Dynamical indicators for the prediction of bursting phase-space reconstruction using a geometrical construction. Physical review
phenomena in high-dimensional systems. Phys Rev E 2016;94(3):32212. A 1992;45(6):3403.
[13] Mattheakis M, Pitsios IJ, Tsironis GP, Tzortzakis S. Extreme events in complex [23] Cao Y, Mao JM. The non-wandering set of some hénon maps. Chaos Solitons
linear and nonlinear photonic media. Chaos Solitons Fractals 2016;84:73–80. Fractals 20 0 0;11(13):2045–53.
[14] Gualandi A, Michel S, Avouac J-P. Slow slip events in cascadia: evidences of [24] Li L. Bifurcation and chaos in a discrete physiological control system. Appl
chaotic behavior from geodetic position time series. In: EGU General Assembly Math Comput 2015;252:397–404.
Conference Abstracts, vol. 20; 2018. p. 11162.