Fast change point detection in switching dynamics using a hidden Markov model of prediction experts

We present a framework for modeling switching dynamics from a time series that allows for a fast online detection of dynamical mode changes. The method is based on a hidden Markov model (HMM) of prediction experts. The predictors are trained by expectation maximization (EM) and by using an annealing schedule for the HMM state probabilities. This leads to a segmentation of the time series into different dynamical modes and a simultaneous specialization of the prediction experts on the segments. In a second step, an input-density estimator is generated for each expert. It can simply be computed from the data subset assigned to the respective expert. In conjunction with the HMM state probabilities, this allows for a very fast online detection of mode changes: change points are detected as soon as the incoming input data stream contains sufficient information to indicate a change in the dynamics