| In this dissertation, we derive inference and learning procedures for the segmental semi-Markov model, and propose novel learning algorithms and applications within this framework. The application of the model to real-world supervised and unsupervised learning problems is investigated for both real-valued sequences (for time series patterns) and discrete symbol sequences. A variety of experiments are reported on real-world data such as semiconductor manufacturing time series data, ECG biomedical time series, Unix command sequences, and English text, as well as on simulated data sources. The experimental results confirm that the model can outperform competing non-Markov methods across a variety of selected pattern recognition tasks. |
