Adaptive Time Series Analysis Using Predictive Inference and Entropy

摘要

Research is reported on adaptive time series methods for detecting and trackingboth abrupt and slow changes in both structure and parameters of dynamic systems. The methods are based on a unified statistical framework which is motivated by statistical inferences and entropy arguments. The method yields estimates of multivariate input/output dynamics and noise statistics. It also gives estimate of system order that is optimal in the sense of an information theoretic criterion. The integrated approach is known as CVA-AIC. Many theoretical issues have been explored under the scope of this project. The relationship between this technique and another powerful framework for estimation known as E-M algorithmic approach has been established. It the CVA-AIC technique is embedded properly in an E-M framework, it leads to maximum likelihood estimates and recursive algorithms for system identification. (jhd)