Detecting chaotic signals with nonlinear models

Hidden Markov models of chaotic signals have been used in numerical detection experiments. For broadband deterministic chaotic signals masked with noise having identical spectra at an SNR of -15 db, the experiments found flawless receiver operating characteristics. In noisy environments the performance of models trained on noise-free signals can be improved by training on signals contaminated by noise typical of the test environment. Continuous valued scalar outputs at each discrete hidden state are modeled as Gaussians with means that depend autoregressively on previous outputs.<>