Nonparametric identification of a cascade nonlinear time series system (Q2277418): Difference between revisions

This paper is generally a well written paper. It describes a nonparametric identification of a Hammerstein system, which is composed of a memoryless, odd nonlinearity, driven by a stationary random process and a linear asymptotically stable system, cascaded to the nonlinearity. An algorithm is described to identify the linear system. A nonrelated algorithm is derived to identify the nonlinear part, both for the stationary and nonstationary cases. This algorithm is based on using Legendre orthogonal polynomials as bases for the nonlinear estimate, as opposed to the authors previous approach of using kernel estimates. Convergence rate of \(O(n^{-2/3})\) is established, slightly worse than the convergence rate for parametric algorithms (generally requiring the inversion of a matrix) of \(O(n^{-1})\). Faster convergence rates can be attained by posing stricter continuity condition on the nonlinearity. The convergence rate of the authors kernel algorithm of \(O(n^{-4/5})\) can thus be reached and surpassed. Even stricter continuity conditions on the nonlinearity allow the convergence rate to approach that of parametric algorithms. Results derived are supported by various simulation results. One problem left for future study is how to choose the optimum number of Legendre polynomials, the cost function has shown strong oscillatory behavior in some cases.