Controlling the equilibria of nonlinear stochastic systems based on noisy data

https://doi.org/10.1016/j.jfranklin.2016.11.011Get rights and content

Abstract

For controlling an equilibrium of a nonlinear stochastic system, the problem of stabilization and synthesis with a required dispersion is studied. This problem is solved for the case where the feedback regulator uses noisy data. The new approach is based on an extension of the stochastic sensitivity synthesis method. Technically, this problem is reduced to the analysis of some quadratic matrix equations. A solution to the problem of minimizing the stochastic sensitivity is given. Details of such analysis are discussed for 2D and 3D nonlinear stochastic oscillators.

Introduction

Control problems for randomly forced systems attract a great deal of attention from science and engineering communities. The development of a stochastic control theory was initiated in [24], and continued by many investigators (see, for instance, [25], [3], [35], [15], [19] and bibliography therein).

For linear dynamic systems, the theory of control and stabilization has been well developed. Indeed, many fundamental results in linear control theory have appeared under a state-space framework. Basic advances in the Linear-Quadratic-Gaussian problem are presented in [4]. Here, the most famous is the separation principle that combines Kalman filter (linear-quadratic estimator) and a linear-quadratic regulator [34], [18]. Idea of H2 and H controllers was an important contribution to the development of the linear theory [14]. Mathematically, the most linear control problems are reduced to the analysis and solution of matrix equations and inequalities [36], [11], [21], [30], [31], [10], [32]. Computational aspects of linear control theory are discussed in [12].

Today, the control of nonlinear stochastic systems is a challenging problem in the modern engineering [33]. Besides control, there is also research about filtering on nonlinear stochastic systems [22], [23]. An interplay of nonlinearity and stochasticity can generate new unexpected dynamic regimes that are not observed in linear systems [20], [2], [17], [26]. It is important to note the essential difference in the mathematical description of stochastic dynamics of linear systems with additive Gaussian noise and nonlinear systems with state- and control-dependent noise. Solutions of the first type of systems are Gaussian processes, and its dynamics are fully described using the first two moments governed by linear deterministic equations. Analysis and control for the second type of systems are much more difficult. Indeed, a comprehensive probabilistic description of the stochastic flows in nonlinear systems is given by the Kolmogorov–Fokker–Planck equation. Analytically, this equation can be solved only in very specific cases, thereby only statistical properties are studied via numerical simulation in general. Even more unfortunately, for solving control problems, this approach is not applicable. In these circumstances, analytical approximations of probabilistic distributions become the only alternative. For this purpose, the quasipotential method is widely used [13], [16], for which a stochastic sensitivity function (SSF) technique was proposed in [28], [6], [8] and successfully applied to the stabilization of stochastic attractors and suppression of chaos [7], [9]. It was shown that the underlying reason, from an engineering point of view, for unacceptable noise-induced large-amplitude oscillations and stochastic transformations, lies in the high sensitivity of initial deterministic attractors to random disturbances.

In control systems, the stochastic sensitivity can be changed by choosing an appropriate regulator that consequently decreases the stochastic sensitivity and suppress the undesirable noise-induced oscillations. This approach was suggested for controlling the stochastic sensitivity of the equilibria of nonlinear systems [5]. In the case of having complete information, a construction of the regulator is reduced to the analysis of linear matrix equations.

Notice that, in practice, the available information on the current system state usually contains random noise or errors [29]. So, a development of the control theory for nonlinear stochastic systems with incomplete information is a very important new subject of research.

The aim of the present paper is to extend the theory of synthesis of the minimal stochastic sensitivity at equilibria to the case of incomplete information when the observations of the system states contain random errors. From a mathematical point of view, the principal difference lies in the fact that in the case of incomplete information, a synthesis of the regulator is reduced to the analysis of quadratic, but not linear, matrix equations. The main contributions of this paper are to provide a solution to the aforementioned important but challenging technical problem of minimizing the sensitivity of nonlinear stochastic systems by means of control using noisy data.

The paper is organized as follows. In Section 2, we briefly present the stochastic sensitivity function technique for the equilibria of a general nonlinear dynamic system. We devote Section 3 to the problem of the synthesis of randomly-forced system equilibria with a required dispersion. In stochastic sensitivity synthesis, we analyze an important general case with noisy observations. Main mathematical results for a general n-dimensional case are summarized in Theorem 1. In Section 4, we derive a solution of the minimization problem of the stochastic sensitivity (Theorem 2). In Section 5, for an important class of two-dimensional nonlinear oscillators, we consider details of this new control theory and give explicit formulas for the attainability sets and parameters of the stabilizing regulator. In Section 6, we apply these theoretical results to the suppression of unwanted noise-induced large-amplitude oscillations for the stochastic van der Pol system. Our general theory is finally applied to the minimization of the stochastic sensitivity in a 3D nonlinear oscillator in Section 7.

Section snippets

Stochastic sensitivity of equilibrium for nonlinear system

Consider a nonlinear stochastic system of the formẋ=f(x)+εσ(x)ξ(t),where f(x)Rn is a continuously differentiable vector-function, ξ(t)Rm is a δ-correlated white Gaussian noise vector satisfying Eξ(t)=0,Eξ(t)ξ(τ)=δ(tτ)I, and I is the identity matrix. Here, σ(x) is an (n×m) -matrix-function characterizing the dependence of disturbances on the states, and ε is a scalar parameter of the noise intensity.

It is supposed that the corresponding deterministic system (1) (with ε=0 therein) has an

Control of stochastic sensitivity by regulator with noisy data

Consider a nonlinear controlled stochastic system,ẋ=f(x)+g(x)u+εσ(x)ξ(t),with x,fRn,uRl,gRn×l, and the control input u is synthesized by the feedback regulatoru=K(xx¯)with a constant (l×n) -matrix K.

It is assumed that the corresponding deterministic system (6) (with ε=0 and u=0 therein) has an equilibrium x¯ whose stability is undetermined. The synthesis of a required stochastic sensitivity function for the equilibrium x¯ of system (6), (7) with full information was studied in [5].

For the

Minimization of stochastic sensitivity

In many practically important engineering tasks, it is required to build a controller which substantially reduces the variation of random states around the equilibria, corresponding to the basic operating mode. For the fixed level of noise, the decrease of this variation means a decrease of the stochastic sensitivity (see (5)). Therefore, we study the following control problem connected with the minimization of the stochastic sensitivity.

Consider a criterion functionJ(W)=W,Q.Here, the scalar

Stabilizing the equilibrium of a 2D nonlinear stochastic oscillator

Consider a nonlinear stochastic oscillator with control input in the form ofx¨=f(x,ẋ)+g(x,ẋ)u+εσ(x,ẋ)ξ,where f,g and σ are scalar functions, ξ is white Gaussian noise satisfying Eξ(t)=0,Eξ(t)ξ(τ)=δ(tτ), and ε is the noise intensity.

Rewrite Eq. (29) as a system:x1̇=x2,x2̇=f(x1,x2)+g(x1,x2)u+εσ(x1,x2)ξ.Assume that the uncontrolled deterministic system (30) (with u=0,ε=0 therein) has an equilibrium (x¯1,0). Now consider a feedback,u=k1(y1x¯1)+k2y2,y1=x1+εφ1(x1,x2)η1,y2=x2+εφ2(x1,x2)η2,where k1

Synthesis of the minimal stochastic sensitivity for the stochastic van der Pol oscillator

Consider the following stochastic system:x1̇=x2,x2̇=x1+(δ+μx12x14)x2+u+εσξ,where ξ is standard white Gaussian noise. The deterministic uncontrolled system (35) (with ε=0,u=0 therein) is the well-known van der Pol oscillator [1], [27] with saddle-node bifurcation at δ=μ2/8. For parameters μ=1,δ=0.1, this system exhibits (see Fig. 1a) a stable equilibrium (0,0) (black circle) and a stable limit cycle (solid line) separated by an unstable limit cycle (red dashed line).

Under stochastic

Minimization of stochastic sensitivity for a 3D nonlinear stochastic oscillator

Consider a nonlinear stochastic oscillator with control input in the form ofd3xdt3=f(x)+u+εσ(x)ξ(t),where f and σ are scalar functions, ξ(t) is white Gaussian noise satisfying Eξ(t)=0,Eξ(t)ξ(τ)=δ(tτ), and ε is the noise intensity. Assume that the deterministic uncontrolled oscillator (37) (with u=0,ε=0) has an equilibrium x¯:f(x¯)=0.

Rewrite Eq. (37) asx1̇=x2,x2̇=x3,x3̇=f(x1)+u+εσ(x1)ξ(t).Consider a feedback with noisy observations,u=k1(x1x¯+εφ1η1)+k2(x2+εφ2η2)+k3(x3+εφ3η3),where k1,k2,k3 are

8. Conclusion

In this paper, a stabilization problem for nonlinear stochastic system with incomplete information was considered. A new constructive method for stochastic sensitivity synthesis was suggested. Mathematically, this problem of synthesis of feedback regulator was reduced to the solution of the corresponding quadratic matrix equation for regulator׳s parameters. An analysis of the attainability of the assigned stochastic sensitivity was discussed. Explicit formulas of the regulator are given in

Acknowledgement

This research was supported by the Hong Kong Research Grant Council under the GRF Grant CityU11201414, the Government of the Russian Federation (Act 211, Contract no. 02.A03.21.0006) and RFBR (16-08-00388).

References (36)

  • I. Bashkirtseva et al.

    Stochastic equilibria control and chaos suppression for 3D systems via stochastic sensitivity synthesis

    Commun. Nonlinear Sci. Numer. Simul.

    (2012)
  • G.N. Mil׳shtein et al.

    A first approximation of the quasipotential in problems of the stability of systems with random non-degenerate perturbations

    J. Appl. Math. Mech.

    (1995)
  • A.A. Andronov et al.

    Theory of Oscillators

    (1966)
  • V.S. Anishchenko, V.V. Astakhov, A.B. Neiman, T.E. Vadivasova, L. Schimansky-Geier, Nonlinear Dynamics of Chaotic and...
  • K.J. Astrom

    Introduction to the Stochastic Control Theory

    (1970)
  • M. Athans (Editor), Special issue on the Linear-Quadratic-Gaussian problem, IEEE Trans. Autom. Control 16(6)...
  • I. Bashkirtseva et al.

    On control of stochastic sensitivity

    Autom. Remote Control

    (2008)
  • I. Bashkirtseva et al.

    Sensitivity analysis of stochastic attractors and noise-induced transitions for population model with Allee effect

    Chaos

    (2011)
  • I. Bashkirtseva et al.

    Stochastic sensitivity analysis of the noise-induced excitability in a model of a hair bundle

    Phys. Rev. E

    (2013)
  • I. Bashkirtseva et al.

    Stabilizing stochastically-forced oscillation generators with hard excitement: a confidence-domain control approach

    Eur. Phys. J. B

    (2013)
  • P. Benner, M. Bollhfer, D. Kressner, C. Mehl, T. Stykel (Eds), Numerical Algebra, Matrix Theory, Differential-Algebraic...
  • S. Boyd, L. El Ghaoui, E. Feron, V. Balakrishnan, Linear matrix inequalities in system and control theory, Volume 15 of...
  • C. Brezinski

    Computational Aspects of Linear Control (Numerical Methods and Algorithms)

    (2002)
  • M. Dembo et al.

    Large Deviations Techniques and Applications

    (1995)
  • J. Doyle et al.

    State-space solutions to standard H2 and H control problems

    IEEE Trans. Autom. Control

    (1989)
  • W.H. Fleming et al.

    Deterministic and Stochastic Optimal Control

    (1975)
  • M.I. Freidlin et al.

    Random Perturbations of Dynamical Systems

    (1984)
  • L. Gammaitoni et al.

    Stochastic resonance

    Rev. Mod. Phys.

    (1998)
  • Cited by (19)

    • Stochastic sensitivity synthesis in nonlinear systems with incomplete information

      2020, Journal of the Franklin Institute
      Citation Excerpt :

      In the case with complete information, the theory of synthesizing the assigned stochastic sensitivity was elaborated for both continuous [24] and discrete [27] systems. A problem in the synthesis of the stochastic sensitivity for systems with incomplete information by static feedback regulators was studied in [28,29]. The present paper is devoted to developing a theory of the stochastic sensitivity control for nonlinear systems with incomplete information for a wider class of dynamic regulators.

    • Recurrence quantification analysis of business cycles

      2018, Chaos, Solitons and Fractals
      Citation Excerpt :

      Therefore to it applies the research on stochastic dynamic systems, e.g. Kushner [35], Guo and Wang [29], Fleming and Rishel [25], as well as that on controlling chaotic dynamical systems, e.g. Romieras et al. [55], Grebogi and Laib [28], Calvo and Cartwright [13], Pettini [51]. More recently random noise, coupled with incomplete information, has been approached in terms of static Bashkirtseva et al. [6], Bashkirtseva [5] or dynamic Bashkirtseva [4] feedback regulators via the synthesis of stochastic sensitivities. However, the way in which this can be empirically applied to steer the economy has not been resolved yet.

    • Robust finite-time state estimation for uncertain discrete-time Markovian jump neural networks with two delay components

      2018, Neurocomputing
      Citation Excerpt :

      At present, there are nearly 40 kinds of neural networks and have been widely used in fields such as pattern recognition, optimization design, static and dynamic image processing and so on. Therefore, the study of neural networks has been one hot topic in recent decades, and it has attracted a large number of experts and scholars, see [1–6]. Many studies on neural networks need to know the state of the neural network, such as [7,8].

    View all citing articles on Scopus
    View full text