Segmenting time series with a hybrid neural networks - Hidden Markov model

This paper describes work on a hybrid HMM/ANN system for finding patterns in a time series, where a pattern is a function that can be approximated by a recurrent neural network embedded in the state of a hidden Markov model. The most likely path of the hidden Markov model is used both for re-training the HMM/ANN model and for segmenting the time series into pattern occurrences. The number of patterns is determined from the data by first increasing the number of networks as long as the likelihood of the segmentation increases, then reducing this number to satisfy an MDL criterion. In experiments with artificial data the algorithm correctly identified the generating functions. Preliminary results with robot data show that potentially useful patterns that can be associated with low-level concepts can be induced this way.