MODELING DOUBLE SCROLL TIME-SERIES

The ubiquity of strange attractors in nature suggests that nonlinear modeling techniques can improve performance in some signal processing applications. We introduce mixed state Markov models (MSMM's), a refinement of hidden filter HMM's, and apply both to a synthetic double scroll time series. Forecasts by HFHMM's diverge after a few steps. Using ad hoc procedures, forecasts by MSMM's, even models generated by crude methods without iterative optimization, can be made more stable.