Multipredictor modelling with application to chaotic signals

The use of multipredictor models (MPMs) in the time series modeling of chaotic signals is investigated. The relation between MPMs and iterated function systems (IFSs) coupled with the ability of IFSs to generate chaotic systems motivates this approach. Emphasis is placed on two forms of MPM. The first MPM models the chaotic dynamic by way of a codebook of predictors, with both linear and nonlinear predictors discussed. It is shown how a dynamic neighborhood function can be used to improve this modeling. The second MPM can be interpreted as a predictive extension of a hidden Markov model and directly parametrized by a segmental k-means algorithm. The forms of dynamical system for which these models are best suited are considered.<>