Robust parameter estimation of chaotic systems (Q2282743)

The authors provide a robust method for a parameter estimation and an uncertainty quantification that requires neither the knowledge of initial values for the given dynamical system nor good guesses for the unknown parameters provided that long enough time series data are available. The method uses a new distance concept to characterize the variability of chaotic dynamical systems. For \(R > 0\), the modified correlation sum is defined as \[ C(R, N, \mathbf{\theta}, \mathrm{x}, \tilde{\mathbf{\theta}}, \tilde{\mathrm{x}})=\frac{1}{N^{2}}\sum_{i,j\leq N}\# (\|s_{i}-\tilde{s}_{j}\|<R \] where \(\# (\|s_{i}-\tilde{s}_{j}\|<R)\) denotes the number of points of the different trajectories lying within a sphere of radius \(R\). The authors define a likelihood for a given dynamical system with a fixed parameter value \( \mathbf{\theta}\) as follows. Fix a set of radii \(R_{k}, k=1, \dots, M\), and consider a vector \(y\) of dimension \(M\) with \(\mathbf{\theta}=\tilde{\mathbf{\theta}}\) \[ y_{k}=C(R_{k}, N, \mathbf{\theta}, \mathrm{x}, \tilde{\mathrm{x}}). \] This expression provides a summary statistics that maps samples from a chaotic attractor into a Gaussian feature vector. As the Gaussian likelihood is constructed, the authors search for its maximum likelihood point and perform the subsequent Markov chain Monte Carlo (MCMC) runs for estimating the parameter posterior distributions. After the convergence of the Differential Evolution (DE) optimization, the authors study the parameter distributions by MCMC. The authors consider the following extended system: \(\dot{x}=f(x)\), \(\dot{W}=-K \cdot (W-X)\), where \(W\) is a fast scalar variable and \(K>0\) is a large constant associated with the characteristic time of adjustment of \(W(t)\) to \(X(t)\). For the original system of equations containing both fast and slow variables, the authors are interested to estimate the coefficient for both, the slow and the fast, parts of the dynamical system as a part of one procedure combining the classical deterministic and the chaotic system parameter identification approaches. The aim of this paper is to present a robust solution to the chaotic system estimation problem by using a distance concept. The authors provide: (i) the estimation of the chaotic system parameters from noisy data, including cases where the initial state values are unknown and the initial model parameters lie far away from those producing the attracting chaotic manifold; (ii) extensions of the distance concept by improving many results as: more accurate confidence intervals for the estimated parameter values, estimating the dynamics, including the cases where the system involves vastly different time scales, or where the parameter estimation needs to be performed simultaneously for coupled chaotic and deterministic parts of a system. The authors also demonstrate the application of the approach for estimating the parameters from sparse data, with poor initial guesses for parameter values and unknown initial values of the state vector. Further, in numerical experiments, they apply the method to the analysis of numerous well-known chaotic systems and they show how the proposed extension greatly improves the parameter identification, especially, in more complex cases. With several examples of chaotic systems, the authors demonstrate how likelihood construction allows for a robust parameter estimation, together with subsequent MCMC sampling of the parameter posteriors. The authors illustrate the results with a large number of chaotic test cases, and extend the method, thus increasing the accuracy of the estimation results. In the Appendix, the authors discuss stochastic optimizers, practical implementations of the described algorithms and provide 14 attractors for which they performed the investigation.