Modeling chaos from experimental data (Q4348159)

Consider a dynamical system of the form \[ x_t = f(x_{t-1}, \beta_t)+\nu_t, \] where \(x_t \in {\mathbb{R}}^d\) is the state of the system at time \(t\), \(\beta_t \in {\mathbb{R}}^k\) are time varying parameters and \(\nu_t\) represents dynamical noise. The authors consider the situation where the only information known about this system is through a scalar measurement \(y_t\) of the system: \[ y_t = c(x_t)+\xi_t. \] Here \(c\) is the read-out function and \(\xi_t\) represents the measurement noise. From this measurement they attempt to model the system as \[ y_t = F(z_t,\beta_t)+ \varepsilon_t. \] They identify a class of nonlinear models for this purpose and describe methods to select an optimal model from this class which describes the system with sufficient accuracy. They concentrate on the class of pseudo-linear models using radial basis functions. Thus \[ y_t = \sum_{i=1}^m \lambda_i f_i(z_t,\beta_t) + \varepsilon_t \] where the \(f_i\)'s are the radial basis functions. The parameters \(\lambda_i\) are estimated by minimizing the sum of squares of errors with respect to \(\lambda\). They avoid over-fitting the data by choosing the model which has the minimum ``description length'' (i.e., the model which achieves the greatest compression of data). They apply their results to the planar-forced vibrations of a stretched-uniform string. They compare the experimentally observed bifurcation diagram with their predicted one and analyse the differences seen in terms of measurement errors, etc. For the entire collection see [Zbl 0866.00065].