Professional Documents
Culture Documents
1 s2.0 S0167278920300026 Main
1 s2.0 S0167278920300026 Main
1 s2.0 S0167278920300026 Main
Physica D
journal homepage: www.elsevier.com/locate/physd
article info a b s t r a c t
Article history: Recently, the term explainable AI came into discussion as an approach to produce models from
Received 2 January 2020 artificial intelligence which allow interpretation. For a long time, symbolic regression has been used
Received in revised form 3 May 2020 to produce explainable and mathematically tractable models. In this contribution, we extend previous
Accepted 22 May 2020
work on symbolic regression methods to infer the optimal control of a dynamical system given one or
Available online 24 June 2020
several optimization criteria, or cost functions. In earlier publications, network control was achieved
Communicated by B. Hamzi
by automated machine learning control using genetic programming. Here, we focus on the subsequent
Keywords: path continuation analysis of the mathematical expressions which result from the machine learning
Explainable AI model. In particular, we use AUTO to analyze the solution properties of the controlled oscillator system
Machine learning control which served as our model. As a result, we show that there is a considerable advantage of explainable
Dynamical systems symbolic regression models over less accessible neural networks. In particular, the roadmap of future
Synchronization control works may be to integrate such analyses into the optimization loop itself to filter out robust solutions
Genetic programming
by construction.
© 2020 Elsevier B.V. All rights reserved.
https://doi.org/10.1016/j.physd.2020.132582
0167-2789/© 2020 Elsevier B.V. All rights reserved.
2 M. Quade, T. Isele and M. Abel / Physica D 412 (2020) 132582
machine learning methods as a suitable source of algorithms. mathematical model of the system, a cost function, or perfor-
Specifically, we refer to genetic programming (GP) [20] to control mance index, a specification of boundary conditions on states,
synchronization in coupled networks, including a hierarchical and additional constraints. According to the type of problem, it
network of coupled oscillators. Unlike neural networks and other is classified.
black-box artificial intelligence methods, GP allows dynamically
learning complex control laws in an interpretable symbolic form 2.1. Machine learning control
— a method that is referred as symbolic regression [21–23]. In
particular, we focus on the subsequent analysis of the optimal Here, we focus on continuous-time optimal control. If there
laws found. In contrast to previous works a full expression is are no constraints on the states or the control variables, and if
optimized, and not only parameters [11,24]. the initial and final conditions are fixed, it reads: Find the control
Based on previous results, we demonstrate the effectiveness vector u⃗ : Rnx × [ts , tf ] ↦→ Rnu that minimizes the cost function
and robustness of the control exemplarily by the analysis of ∫ tf
solutions to the system with control terms found by symbolic Γ = ϕ (⃗x(tf )) + x(t), u
L(⃗ ⃗ (⃗x, t), t)dt , (1)
regression with the aim of synchronizing or de-synchronizing two ts
coupled oscillators. Further application to the control of entire subject to
networks is technically not hard, and is to be included in the
future in the fully automatized software framework Glyph [25]. ⃗x˙ = ⃗f (⃗x, u⃗ , t), ⃗x(ts ) = ⃗xs , (2)
The setup is not only motivated by brain disorder problems in where [ts , tf ] is the time interval of interest; ⃗ x:[ts , tf ] ↦ → Rnx
the medical domain like body tremors occur when firing neurons is the state vector; ϕ : Rnx ↦ → R is a terminal cost function;
synchronize in regions of the brain [4], but may find broad appli- L:Rnx × Rnu × R ↦ → R is an intermediate cost function; and
cation in any control setup, e.g. in machinery [26]. In the case of ⃗f : Rnx × Rnu × R ↦→ Rnx is a vector field. Eq. (2) represents
brain states, if the firing of neurons is periodic, which may appear the dynamics of the system and its initial state. This problem
due to the inherent dynamics of the excitable neurons, a mutual definition is known as the Bolza problem, which is also known as
influence may give rise to synchronization [6,27]. If the coupling the Lagrange problem [29] for ϕ (x(tf )) = 0 and u ⃗ = ⃗x˙ (t). The cost
term is very large, this synchronization may extend over a whole function (or performance index) Γ is a functional, which assigns
region in the brain and thus affect many neurons. Eventually this a real value to each control function u ⃗.
collective firing leads to shaky movements of hands, arms or the Often, the solution to control problems cannot be found by
head, and is treated as a brain disorder. One remedy to this prob- analytical means. Then, computational methods are used to solve
lem is to implant a control device which resets the neurons and such problems. Depending on the types of cost functions, time
counteracts the collective synchronization. An evident question domain, and constraints in Eqs. (1)-(2) different methods may be
then is how to design such a controller minimizing design cost, applied [8]. The direct methods use a (temporal) discretization of
energy consumption or other medical constraints while being the control problem and solve it using nonlinear programming
as stable and reliable as possible. We analyze the found control approaches. Other methods involve the discretization of the dif-
technically by the well-known package AUTO [28]. ferential equations by defining a grid of N points covering the
To the best of our knowledge, this is the is first demonstration time interval [ts , tf ], ts = t1 < t2 < · · · < tN = tf , and solving
of machine learning control followed by an automatized bifurca- these equations using suitable numerical methods [30]. Thereby,
tion analysis. The extension and generalization is the subject of the differential equations become equality constraints of the non-
an extension of existing software. GP is used to learn a control linear programming problem. Other direct methods involve the
that brings a network of self-sustained oscillators in a desired, approximation of control and states using basis functions, such
synchronized or de-synchronized state and back. In Section 3, the as splines or Lagrange polynomials.
results of our study of GP application to networked dynamic sys- Dynamic programming is an alternative to the variational
tems are presented with focus on the path continuation analysis approach to optimal control. It was proposed by Bellman in the
which indicates the robustness of the choice of control term. 1950s and is an extension of Hamilton–Jacobi theory. A number
This publication is structured as follows: of books exist on these topics including [18,31,32]. An overview of
In Section 2, we give a brief over view of the machine learning the optimal control methods for dynamical systems can be found
methods employed. Machine learning control (Section 2.1) and in [17]. For further details, see [33].
genetic programming (Section 2.2). In Section 3, we explain the Our approach to solve a control problem uses machine learn-
application used to demonstrate the methods in the previous ing to determine the optimal control. Therefore, we adopt the
section. We explain the dynamical system used in depth in Sec- continuous-time formulation given in Eqs. (1) and (2). For multi-
tion 3.1. The following subsections focus on different control objective optimization, the reader is referred to [8]. In real-world
goals (forced synchronization in Section 3.1.1 and forced de- problems, often the derivatives are not given and one has to
synchronization in Section 3.1.2). We examine the found control reconstruct them; then it is particularly important to respect the
laws by means of bifurcation analysis in Section 3.1.3. We discuss accuracy of measured data as described in [34]. For the particular
the results an conclude in Section 4. control scheme considered here, ⃗ ⃗ are slightly reformulated,
f and u
as will be described next.
2. Methods Similar to reinforcement learning, a feedback control scheme
[35] is used here to implement the control. In Fig. 1, the architec-
In this section, we briefly recall the concept of MLC (machine ture is depicted. To follow this scheme, we use a particular form
learning control) and the method used here to solve the problem. of Eq. (2):
To control a dynamical system, one determines a manipulation of
the trajectory of the system in phase space to drive it to and keep ⃗x˙ = ⃗f (⃗x, t) + a⃗, ⃗x(ts ) = ⃗xs ,
it in a desired state. This control problem is typically formulated
such that the uncontrolled system ⃗x˙ = ⃗ x, t) is controlled by an
f (⃗
as an optimization problem with an objective function that is to
⃗, and the control function u⃗ depends on
additive actuator term a
be minimized. In the control setup, this objective is formulated
sensor measurements ⃗s ∈ Rns :
as the deviation of the state of the system from its desired one.
Consequently, an optimal control problem is formulated as a ⃗ = u⃗ (⃗s, t).
a
M. Quade, T. Isele and M. Abel / Physica D 412 (2020) 132582 3
Fig. 1. Sketch of the feedback control loop. The output of the dynamical system
⃗x is measured by sensors ⃗s which are used as input to the control function
⃗ . The control function, in turn, acts on the system via actuators a⃗ = u⃗ (⃗s, t) in
u
order to achieve a desired state. External disturbances which can be incorporated
explicitly as additional inputs to the dynamical system and the control function
are not shown here.
Fig. 4. Sketch of the machine learning loop. By its evolutionary strategy, the GP where c1,2 are the global coupling constants and κij and εij are
⃗ , called the population.
algorithm generates a set of candidate control solutions u the respective coupling matrices. This allows for several types of
The candidate solutions are then evaluated in many realizations of the control coupling such as direct, diffusive (coupling by difference), and
loop; the performance in each iteration is rated via a cost functional Γ and global coupling, or any other kind of network-like coupling. In
fed back as a cost index into the GP algorithm. The performance rating is used
to select the best solutions and to evolve them into the next generation of
the following experiments, we will use diffusive coupling in ẋi .
candidate solutions. This learning loop repeats until at least one satisfactory For GP, we use the setup described above.
control law is found (or other break conditions are met).
3.1. Two coupled oscillators
where x is the state and ω, α , β > 0 are model parameters. 1 In this example, the actual/dominant frequency Ω of an oscillator can be
The parameter ω is the frequency at which the system oscil- determined numerically either by taking the frequency at the maximum of the
lates without any driving or damping force. The parameter α Fourier transform of the trajectory x, or by counting the zero-crossings of x −⟨x⟩.
M. Quade, T. Isele and M. Abel / Physica D 412 (2020) 132582 5
Fig. 5. Solutions of a single van der Pol oscillator with parameters ω = e2 , α = 3 and several values for β . Initial conditions x(0) = −0.25, ẋ(0) = 2.5.
Fig. 6. Synchronization plot of two coupled van der Pol oscillators with varying coupling strength c. The horizontal V-shaped plateau is referred to as the Arnold
tongue, it represents regions of synchronization. The parameter set and initial conditions used are stated in the left part of Table 1.
functional
Γ1 := |Ω0 − Ω1 |. (7) Table 2
Two coupled oscillators: Pareto-front solutions for forced synchronization.
It measures the difference in observed frequencies exhibited by
the two oscillators: smaller differences reduce the cost on this |Ω0 − Ω1 | Length Expression
and is defined as are then analyzed using the software AUTO [28], tracking their
⏐ ⏐ solutions and thus finding the range of parameter values where
⏐ 1 N −1 iϕ ⏐
⏐ ∑ ⏐
the control solution exists and has certain properties, like (linear)
r = ⏐⏐ e j ⏐⏐ ,
stability. This tells us how robust a certain control strategy is
⏐ N j=0
in the sense how much certain parameters may vary before a
⏐
desired solution vanishes or an undesired solution appears. As we
with ϕj being the continuous phase2 of the jth oscillator.
examine a reduced system, the stability properties of solutions
x˙ ) = −ẋ0 .
In Fig. 7, r is shown for the particular solution u(⃗ of this reduced system are not necessarily the ones of the full
The plot shows, that the controlled system completely synchro- systems. A full bifurcation and stability analysis of the examined
nizes (r ≈ const.) after passing through a short initial period of system is out of the scope of this paper and will be the subject of
de-synchronization; whereas the uncontrolled system exhibits a a subsequent publication.
permanent phase shift resulting in an oscillating graph. A more Plugging u(⃗ x˙ ) = −kẋ0 into the first oscillator equation from
detailed analysis of the solution u(⃗x˙ ) = −ẋ0 is postponed to (6) one obtains
Section 3.1.3. We will first present the results for forced de-
ẍ0 = −ω02 x0 + αẋ0 1 − β x20 + c (ẋ1 − ẋ0 ) − kẋ0
( )
synchronization.
= −ω02 x0 + (α − c − k)ẋ0 − αβ ẋ0 x20 + c ẋ1 (9)
3.1.2. Forced De-synchronization
= −ω02 x0 + a0 ẋ0 − bẋ0 x20 + c0 ẋ1 ,
The system setup for forced de-synchronization is given in
Table 1. The parameters ω1 and c are, again, chosen according with a0 := α − c − k, b := αβ , and c0 := c. Doing the same for
to Fig. 6. This time, such that the uncontrolled system follows a the second oscillator equation
synchronization regime well inside the plateau. The measure for ẍ1 = −ω12 x1 + αẋ1 1 − β x21 + c (ẋ0 − ẋ1 ) − kẋ0
( )
the degree of synchronization is now reciprocal to the previous
case = −ω12 x1 + (α − c)ẋ1 − αβ ẋ1 x21 + (c − k)ẋ0 (10)
= −ω 2
1 x1 + a1 ẋ1 − bẋ1 x21 + c1 ẋ0 ,
Γ1 := exp(−|Ω0 − Ω1 |), (8)
where a1 := α − c and c1 := c − k. One effectively obtains
and penalizes synchronization of the two oscillators. All the re- another system of two coupled van der Pol oscillators, however,
maining parameters are the same as for forced synchronization. with direct coupling in place. Such an analysis is not possible for
Results from the GP run are shown in Table 4. Two aspects of AI methods without functional output.
the results indicate that de-synchronizing the pair of oscillators This system can be analyzed in the framework of synchroniza-
is a more demanding task: First, the GP algorithm comes up with tion theory which uses the method of averaging [6] under the
increasingly long expressions to achieve improvements in Γ1 ; assumption that the system is weakly nonlinear. In essence, the
second, constant optimization seems to fail in all cases where a second-order equations are rewritten as two first-order equations
constant is present (this is expressed by a value k = 1, which ẋj = y, ẏj = r.h.s. (j = 0, 1), where r.h.s. denotes the right hand
corresponds to the initial guess of the optimization procedure). side of (9) and (10) respectively. Next, the transformation
Still, the oscillating Kuramoto parameter, r, of the controlled 1
system in Fig. 8 shows, that the best solution with respect to Γ1 xj = (Aj eiωt + A∗j e−iωt ), (11)
2
performs well in de-synchronizing the oscillators. 1
The best performing solution u(⃗ x˙ ) = −ẋ0 · exp(exp(k) + yj = (iωAj eiωt − iωA∗j e−iωt ), (12)
2
cos(k)) = −k̃ẋ0 , with k̃ ≈ 26, is the same in structure as the
one found in the previous case of forced synchronization, namely is applied, where j = 0, 1, and Aj (t) = Rj (t)eiΘj (t) is the time-
dependent complex amplitude. This ansatz yields differential
the control law u(⃗ x˙ ) = −kẋ0 , with coefficient k = 1. Both are
equations for the real amplitude R and the phase Θ , which are
analyzed in more detail in Section 3.1.3.
both slowly varying. They result from the collection of terms
x˙ ) = −ẋ0 · exp(exp(k) + cos(k)) can be simplified
The fact that u(⃗
˙ with vanishing fast oscillation. The calculations are conducted
to u(⃗ x) = −26 · ẋ0 , raises the question, why the GP algorithm analogous to chapter 8 of [6], and nothing new is added here.
did not directly generate this simpler, and thus better adapted, When one writes Eqs. (9) and (10) as
solution. It is not entirely clear to the author, why that is not the
case here. One possible explanation might lie with the constant ẏ0 = −ω02 x0 + a0 y0 − by0 x20 + c0 y1 , (13)
optimization algorithm: on failure, the least squares algorithm ẏ1 = −ω 2
by1 x21 + c1 y0 ,
1 x1 + a1 y1 − (14)
returns the result of the last internal iteration. This return value
might be entirely inadequate for k, which, in turn, could lead to with the corresponding approximations (first order nonlineari-
an exploding cost-index Γ1 when integrating the dynamic system ties, slow dynamics) one arrives at new equations for A
(6), hence, disqualifying the corresponding solution. a0 b
Ȧ0 = −i∆0 A0 + A0 − |A0 |2 A0 + c0 A1 , (15)
2 8
3.1.3. Control terms and bifurcation analysis a1 b
The objective of this section is to analyze the effect of control Ȧ1 = −i∆1 A1 + |A1 |2 A1 + c1 A0 ,
A1 − (16)
8 2
exhibited by the particular results u(⃗ x˙ ) = −k · ẋ0 found on the
with ∆j = ωj − ω (j = 0, 1). For the real phases and amplitudes
synchronization of the two oscillators. The control term can be
of A(t) one then obtains a system of four real equations
arbitrarily complex, which is probable for engineering applica-
tions and other real-world examples. Here, the simplicity helps a0 b
Ṙ0 = R0 − |R0 |2 R0 + c0 R1 cos(Θ1 − Θ0 ),
to understand and evaluate our conceptual approach. For analysis, 2 8
we will use the standard nonlinear dynamics toolbox, first trans- a1 b
Ṙ1 = R1 − |R1 |2 R1 + c1 R0 cos(Θ0 − Θ1 ),
forming the two second-order differential equations to a system 2 8
of three first order ODEs using averaging. The resulting equations (17)
R1
Θ̇0 = −∆0 + c0 sin(Θ1 − Θ0 ),
R0
2 The continuous phase is computed from the analytic signal of the trajectory R0
using the Hilbert transform. For details see [44]. Θ̇1 = −∆1 + c1 sin(Θ0 − Θ1 ),
R1
M. Quade, T. Isele and M. Abel / Physica D 412 (2020) 132582 7
Fig. 7. Two coupled oscillators: Kuramoto order parameter, r, for forced synchronization. Green: the controlled, and blue: the uncontrolled system.
Fig. 8. Two coupled oscillators: Kuramoto order parameter, r, for forced synchronization. Green: the controlled, and blue: the uncontrolled system. The horizontal
axis is scaled to a limited time window in order to make the oscillations visible.
and for the phase difference ∆Θ = Θ0 − Θ1 an asymmetric Adler- system, this corresponds to a Hopf-bifurcation because the full
type equation, which may or may not show synchronization for system shows oscillations.
the given parameters: To this end, we use the path-following and bifurcation analysis
R1 R0 package AUTO-07p [28]. In the scope of this work, this analysis
∆Θ̇ = −∆ω − (c0 + c1 ) sin(∆Θ ). (18) has been done manually, it is however a sensible, possible and
R0 R1 planned extension of the MLC software to do this in an automated
with ∆ω := ω0 − ω1 . If the stationary state for the phase differ- fashion once a control term has been found.
ence ∆Θ̇ = 0 has a solution, one can find synchronization, else The paths of stationary solutions, as observed in the following,
not. A special solution where also the amplitudes are stationary are interpreted using dynamical systems ideas and terminology,
(Ṙ0 = Ṙ1 = 0) has to be determined numerically. cf. e.g. [45]. Stable solutions are displayed by solid lines, unstable
Thus, the full coupled system for stationary solutions of (17) ones by dotted lines. First, we consider the uncoupled case (c =
reads: 0) for varying k. Next, solutions for R0 , R1 , ∆Θ are tracked against
a0 b0 varying k(or c) for several values of c(or k) Fig. 11 (Fig. 10 resp.).
0= R0 − |R0 |2 R0 + c0 R1 cos(∆Θ ),
2 8 The uncoupled case (c = 0). of (19) is shown in Fig. 9. Since the
a1 b1 radius equations are cubic, one obtains three solutions until the
0= |R1 | R1 + c1 R0 cos(∆Θ ),
R1 − 2
(19)
2 8 damping (introduced by k) becomes stronger than energy input
R1 R0 and no nontrivial solution is possible. The control term introduces
0 = ∆ω + (c0 + c1 ) sin(∆Θ ), a coupling ‘‘through the backdoor’’ into the second equation of
R0 R1
the system via the term c1 R0 cos(∆Θ ). So, when varying k, the
with parameters a0 = α − c − k, a1 = α − c, b0 = b1 = α · β , bifurcation diagram of R0 shows a plain pitchfork bifurcation, the
c0 = c, c1 = c − k and ∆Θ = Θ0 − Θ1 . one for R1 shows a distorted version of the pitchfork bifurcation
First, one can observe that the term kẋi (i = 0, 1) introduces due to the ‘‘quasi-coupling’’. Solutions attain and lose stability
an asymmetry in Eqs. (17), such that one or the other oscillator through a Hopf bifurcation.
might be ‘‘favored’’ by the dynamics, since it may have a different
damping depending on the parameter settings. In an uncoupled The coupled case (c ̸ = 0). is visualized for varying parameter c in
system (c = 0) this is of course relevant if one needs to study a Fig. 10 and for varying k in Fig. 11.
real application. For a better understanding of the implications of In Fig. 10, one can see that the value range of c, where solu-
the control term kẋ1 , we note that in the uncoupled case, the first tions are stable is enlarged by non-vanishing control k. The tran-
equation of (19) reduces to sition from stable to unstable occurs through either Hopf- or limit
point bifurcation. The exact bifurcation structure is complicated
a b
0= R0 − |R0 |2 R0 , (20) and out of the scope of this paper.
2 8 A coarse understanding comes from the following argument:
with rescaled parameters a = α − k, b = α · β . This is the normal In a three-dimensional system with cubic terms one can obtain, in
form of a pitchfork bifurcation. In the original, non-averaged principle, more than three solutions. Since we have no additional
8 M. Quade, T. Isele and M. Abel / Physica D 412 (2020) 132582
Fig. 9. Stationary solutions for the system (19) without coupling (c = 0). (a)
R0 , (b) R1 , (c) ∆Θ vs. k are plotted. Both stable (green, solid) as unstable
solutions (red, dashed) are shown. Both trivial (R0 = R1 = ∆Θ = 0) and non-
trivial solution are possible. Bifurcation points are marked BP (Branch points),
LP (limit points) and HB (Hopf bifurcation). Beyond k = α no nontrivial solution
is possible. Physically this reflects the fact that the damping is so large that no
oscillation is possible.
Fig. 10. Stationary solutions for the system (19). Solution variables (a) R0 , (b)
R1 and (c) ∆Θ are plotted against varying coupling c and a set of fixed control
objective in the GP run, it is clear that the particular choice of parameters k. Stability is shown as stable (solid lines), unstable (dashed lines).
k in such a simple control term as kẋi is just a representative of The trivial solution (R0 = R1 = ∆Θ = 0) has been omitted in the plot for better
overview.
a larger class of control laws. To fix it to a specific value, or to
enforce a symmetric situation, one has to add the corresponding
terms in the cost function. Since the control term is asymmetric, oscillators. The value can be read off of the figures of Fig. 11 for
with increasing k, one changes the bifurcation scenario from per- a given coupling strength.
fect to imperfect. This qualitative behavior is found in all graphs The GP-algorithm yields a value of k = 1, clearly this is
shown hereafter. larger than the critical value of approximately 0.05 (for c =
The behavior of R0 and R1 with increasing coupling and fixed 0.022, see Table 3), read off Fig. 11. Now, one may ask why this
control can also be understood with a close look at Fig. 10. At value is chosen and not another one k > 0.05. The reason is
highest values of c where stationary solutions exist, the values of simple, but it is hidden in the problem formulation: The objective
R0 and R1 both vanish. As the coupling strength diminishes, the function only requires that synchronization be destroyed. This is
two radii take increasingly different values as for low values of possible for many functions and in particular for many simple
c, the coupling is dominated by the asymmetric terms with k. As functions with complexity 2. Among them, the control function
the control strength k is increased, stationary, stable solutions at kẋ is particularly appealing due to its simplicity. The algorithm
higher coupling strengths c exist. is now free to choose any k > 0.05 and so it does. The value 1
These observations explain also, why the simple control term is probably appearing because it is the first guess for a constant
kẋi can push dynamics from synchronized to de-synchronized: it in the constant-optimization step of the algorithm and because it
adds a damping of one oscillator i that desynchronizes the two satisfies the objective.
M. Quade, T. Isele and M. Abel / Physica D 412 (2020) 132582 9
Table 4
Two coupled oscillators: Pareto-front solutions for forced de-synchronization.
exp(−|Ω0 − Ω1 |) Length Expression Constant
0.248 9 −ẋ0 · exp(exp(k) + cos(k)) k=1
0.258 7 cos(exp(ẋ1 + cos(cos(ẋ0 ))))
0.875 4 cos(exp(exp(ẋ0 )))
0.912 3 sin(exp(ẋ0 ))
0.999 2 exp(k) k=1
1.000 1 ẋ1
1.000 1 ẋ0
1.000 1 k k=1
is measured using the Kuramoto parameter. In both cases, syn- [9] Impulsive stabilization for control and synchronization of chaotic systems:
chronization and de-synchronization, the found control laws are Theory and application to secure communication, IEEE Trans. Circuits Syst.
I 44 (10) (1997) 976–988, http://dx.doi.org/10.1109/81.633887.
tested for stability.
[10] Z. Li, X. Cao, N. Ding, Adaptive fuzzy control for synchronization of nonlin-
...... ear teleoperators with stochastic time-varying communication delays, IEEE
The results in all setups highlight the importance of designing Trans. Fuzzy Syst. 19 (4) (2011) 745–757, http://dx.doi.org/10.1109/tfuzz.
the objective functions appropriately. In the current setup, we 2011.2143417.
simply relied on learning the control laws by minimizing the [11] H. Shokri-Ghaleh, A. Alfi, Optimal synchronization of teleoperation sys-
tems via cuckoo optimization algorithm, Nonlinear Dyn. 78 (4) (2014)
error with the desired output. We did not specify any symmetry 2359–2376, http://dx.doi.org/10.1007/s11071-014-1589-5.
or energy function to be minimized, nor did we restrict the [12] C. Hammond, H. Bergman, P. Brown, Pathological synchronization in
number of oscillators to be controlled. These details appear to parkinson’s disease: Networks, models and treatments, Trends Neurosci.
be important for using our methods in a real-world application. 30 (7) (2007) 357–364, http://dx.doi.org/10.1016/j.tins.2007.05.004.
The difficulty in the very general approach is recognized by the [13] F. Dorfler, M. Chertkov, F. Bullo, Synchronization in complex oscillator
networks and smart grids, Proc. Natl. Acad. Sci. 110 (6) (2013) 2005–2010,
asymmetry of the control laws, which is a disadvantage in our http://dx.doi.org/10.1073/pnas.1212134110.
opinion. One thus has to carefully analyze the objectives and may [14] D.E. Kirk, Optimal Control Theory: An Introduction, Courier Corporation,
upgrade them step by step if unwanted solutions occur. 2012.
Further work will extend current methods in several ways: [15] Numerical Optimization, Springer-Verlag, 1999, http://dx.doi.org/10.1007/
b98874.
a subsequent automatic stability analysis would be performed
[16] S. Sra, S. Nowozin, S.J. Wright, Optimization for Machine Learning, MIT
for the numerical experiments. This way one can immediately Press, Cambridge, USA, 2011.
distinguish stable and useful control dynamics from unstable [17] V. Becerra, Optimal control, Scholarpedia 3 (1) (2008) 5354, http://dx.doi.
ones. Second, we aim to look into the design of better objective org/10.4249/scholarpedia.5354.
functions, taking into consideration prior domain knowledge, e.g. [18] Optimal Control Theory, Springer-Verlag, 2000, http://dx.doi.org/10.1007/
0-387-29903-3.
in the form of additive symmetry terms. Finally, we plan to
[19] Handbook of Chaos Control, Wiley-VCH Verlag GmbH Co. KGaA, 2007,
integrate methods in our framework that would search for the http://dx.doi.org/10.1002/9783527622313.
optimal sensor (for measurement) and pressure (for actuation) [20] J.R. Koza, Genetic Programming: On the Programming of Computers by
points in the network for a better control. Means of Natural Selection, MIT Press, 1992.
[21] M. Schmidt, H. Lipson, Distilling free-form natural laws from experimental
data, Science 324 (5923) (2009) 81–85, http://dx.doi.org/10.1126/science.
CRediT authorship contribution statement 1165893.
[22] E. Vladislavleva, G. Smits, D. den Hertog, Order of nonlinearity as a
Markus Quade: Investigation, Methodology, Software, Vali- complexity measure for models generated by symbolic regression via
dation, Visualization. Thomas Isele: Investigation, Software, Vi- pareto genetic programming, IEEE Trans. Evol. Comput. 13 (2) (2009)
333–349, http://dx.doi.org/10.1109/tevc.2008.926486.
sualization, Writing. Markus Abel: Conceptualization, Funding
[23] M. Quade, M. Abel, K. Shafi, R.K. Niven, B.R. Noack, Prediction of dynamical
acquisition, Methodology, Supervision, Writing - review & editing. systems by symbolic regression, Phys. Rev. E 94 (1) (2016) http://dx.doi.
org/10.1103/physreve.94.012214.
Declaration of competing interest [24] H. Shokri-Ghaleh, A. Alfi, A comparison between optimization algorithms
applied to synchronization of bilateral teleoperation systems against time
delay and modeling uncertainties, Appl. Soft Comput. 24 (2014) 447–456,
The authors declare that they have no known competing finan- http://dx.doi.org/10.1016/j.asoc.2014.07.020.
cial interests or personal relationships that could have appeared [25] M. Quade, J. Gout, M. Abel, Ambrosys/glyph: V0.3.2, 2017, http://dx.doi.
to influence the work reported in this paper. org/10.5281/zenodo.801819.
[26] Y.E.S.M.P. Oswald, S. Sattler, P. Kumar, R. Radespiel, C. Behr, M. Sinapius,
Acknowledgments J. Petersen, P. Wierach, M. Quade, M. Abel, B.R. Noack, R. Semaan, Open-
and closed-loop control investigations of unsteady coanda actuation on
a high-lift configuration, 2018, http://dx.doi.org/10.2514/6.2018-3684, URL
We thank J. Gout for computer wisdom, and R. Niven and L. https://arc.aiaa.org/doi/abs/10.2514/6.2018-3684 arXiv:https://arc.aiaa.org/
Cordier for helpful discussion and the unknown reviewers for doi/pdf/10.2514/6.2018-3684.
their comments. This work was supported economically by 4Cast [27] D.A. Wiley, S.H. Strogatz, M. Girvan, The size of the sync basin, Chaos 16
(1) (2006) 015103, http://dx.doi.org/10.1063/1.2165594.
GmbH & Co. KG and Ambrosys GmbH, Germany.
[28] E.J. Doedel, T.F. Fairgrieve, B. Sandstede, A.R. Champneys, Y.A. Kuznetsov,
X. Wang, Auto-07p: Continuation and Bifurcation Software for Ordinary
References Differential Equations, Tech. Rep., 2007.
[29] H.H. Goldstine, A History of the Calculus of Variations from the 17th
[1] R. Goebel, A. Chander, K. Holzinger, F. Lecue, Z. Akata, S. Stumpf, P. through the 19th Century, Springer New York, 1980, http://dx.doi.org/10.
Kieseberg, A. Holzinger, Explainable ai: The new 42?, in: A. Holzinger, P. 1007/978-1-4613-8106-8.
Kieseberg, A.M. Tjoa, E. Weippl (Eds.), Machine Learning and Knowledge [30] W.H. Press, Numerical Recipes 3rd Edition: The Art of Scientific Computing,
Extraction, Springer International Publishing, Cham, 2018, pp. 295–303. Cambridge university press, 2007.
[2] E. Ott, C. Grebogi, J.A. Yorke, Controlling chaos, Phys. Rev. Lett. 64 (11) [31] F.L. Lewis, D.L. Vrabie, V.L. Syrmos, Optimal Control, John Wiley & Sons,
(1990) 1196–1199, http://dx.doi.org/10.1103/physrevlett.64.1196. Inc, 2012, http://dx.doi.org/10.1002/9781118122631.
[3] Chaos Control, Springer Berlin Heidelberg, 2003, http://dx.doi.org/10.1007/ [32] A.E. Bryson (Jr.), Y. Ho, Applied Optimal Control, Halsted Press, 1975.
b79666. [33] M. Athans, P.L. Falb, Optimal Control: an Introduction to the Theory and
[4] H. Haken, Brain Dynamics: Synchronization and Activity Patterns in its Applications, Dover Publications, 2006.
Pulse-Coupled Neural Nets with Delays and Noise, in: Springer Series in [34] K. Ahnert, M. Abel, Numerical differentiation of experimental data: local
Synergetics, Springer Berlin Heidelberg, 2006, URL https://books.google.de/ versus global methods, Comput. Phys. Comm. 177 (10) (2007) 764–774,
books?id=8elDAAAAQBAJ. http://dx.doi.org/10.1016/j.cpc.2007.03.009, URL http://www.sciencedirect.
[5] J.M. Schwalb, C. Hamani, The history and future of deep brain stimulation, com/science/article/pii/S0010465507003116.
Neurotherapeutics 5 (1) (2008) 3–13, http://dx.doi.org/10.1016/j.nurt.2007. [35] G.F. Franklin, J.D. Powell, A. Emami-Naeini, Feedback Control of Dynamic
11.003. Systems, seventh ed., Pearson, 2014.
[6] A. Pikovsky, M. Rosenblum, J. Kurths, Synchronization, Cambridge [36] H.U. Voss, P. Kolodner, M. Abel, J. Kurths, Amplitude equations from
University Press, 2001, http://dx.doi.org/10.1017/cbo9780511755743. spatiotemporal binary-fluid convection data, Phys. Rev. Lett. 83 (1999)
[7] S.H. Strogatz, Sync: How order emerges from chaos in the universe, nature, 3422–3425, http://dx.doi.org/10.1103/PhysRevLett.83.3422, URL https://
and daily life, Hyperion (2003). link.aps.org/doi/10.1103/PhysRevLett.83.3422.
[8] J. Gout, M. Quade, K. Shafi, R.K. Niven, M. Abel, Synchronization control [37] H. Voss, M.J. Bünner, M. Abel, Identification of continuous spatiotemporal
of oscillator networks using symbolic regression, Nonlinear Dynam. 91 (2) systems, Phys. Rev. E 57 (1998) 2820–2823, http://dx.doi.org/10.1103/
(2018) 1001–1021, http://dx.doi.org/10.1007/s11071-017-3925-z. PhysRevE.57.2820, URL https://link.aps.org/doi/10.1103/PhysRevE.57.2820.
M. Quade, T. Isele and M. Abel / Physica D 412 (2020) 132582 11
[38] N.L. Cramer, A representation for the adaptive generation of simple [42] Y. Kuramoto, in: H. Araki (Ed.), International Symposium on Mathematical
sequential programs, in: J.J. Grefenstette (Ed.), Proceedings of an Interna- Problems in Theoretical Physics, in: Lecture Notes in Physics, vol. 39,
tional Conference on Genetic Algorithms and the Applications, Psychology Springer, 1975, p. 420.
Press, 1985, pp. 183–187. [43] Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence, Springer Berlin
[39] Genetic Programming, Springer Berlin Heidelberg, 1999, http://dx.doi.org/ Heidelberg, 1984, http://dx.doi.org/10.1007/978-3-642-69689-3.
10.1007/3-540-48885-5.
[44] L. Cohen, Time Frequency Analysis: Theory and Applications, first ed.,
[40] X.-S. Yang, Metaheuristic optimization, Scholarpedia 6 (8) (2011) 11472,
Prentice Hall, 1994.
http://dx.doi.org/10.4249/scholarpedia.11472.
[45] S.H. Strogatz, Nonlinear Dynamics and Chaos, second ed., Westview Press,
[41] M.A.M. Quade, J. Gout, Glyph: symbolic regression tools, J. Open Res. Softw.
1 (7) (2019). 2015.