Analysis of nonstationary time series by mixtures of self-organizing predictors

Presents a method for the analysis of time series from drifting or switching dynamics. In an extension to existing approaches that identify switches or drifts between stationary dynamical modes, the method allows one to analyze even continuously varying dynamics and can identify mixtures of more than two dynamical modes. The architecture is based on a mixture of self-organizing Nadaraya-Watson kernel estimators. The mixture model is trained by barrier optimization, a technique for constrained optimization problems. We apply the proposed method to artificially generated data and EEG recordings from the wake/sleep transition