[工程科技]Observableoperatormodelsfordiscretestochastictimeseries.pdfVIP

Observable operator models for discrete stochastic time series1 Herbert Jaeger German National Research Center for Information Technology (GMD) SET.KI Schloss Birlinghoven, D-53754 Sankt Augustin phone +49 2241 14 2253, fax +49 2241 14 2384 email: herbert.jaeger@gmd.de May 10, 1999 1This article has been accepted for publication in Neural Computation. Abstract A widely used class of models for stochastic systems is Hidden Markov mod- els. Systems which can be modeled by hidden Markov models are a proper subclass of linearly dependent processes, a class of stochastic systems known from mathematical investigations carried out over the last four decades. This article provides a novel, simple characterization of linearly dependent pro- cesses, called observable operator models. The mathematical properties of observable operator models lead to a constructive learning algorithm for the identification of linearly dependent processes. The core of the algorithm has 3 a time complexity of O(N + nm ), where N is the size of training data, n is the number of distinguishable outcomes of observations, and m is model state space dimension. 1 Introduction Hidden Markov models (HMMs) (Bengio, 1996) of stochastic processes have been investigated mathematically long before they became a popular tool in speech processing (Rabiner, 1990) and engineering (Elliott, Aggoun, Moore, 1995). A basic mathematical question was to decide when two HMMs are equivalent, i.e. describe the same distribution (of a stochastic process) (Gilbert, 1959). This problem was tackled by framing HMMs within a more general class of stochastic processes, now termed linearly dependent processes (LDPs). Dec