Augmented Markov Models
Dani Goldberg, Maja J. Matarić · 1999
This technical report presents augmented Markov models (AMMs), and provides detailed descriptions of their structure and one model construction algorithm. Augmented Markov models are essentially probabilistic transition networks similar to hidden Markov models (HMMs), except that the hidden state assumption is removed. Additional statistics (augmentations) are maintained in the links and nodes of AMMs and may be employed in model construction and utilization. The model construction algorithm we present is designed to have relatively low computational and space overheads, and provide useful models on-line and in real-time. 1 Introduction An augmented Markov model (AMM) is essentially a finite state automaton with probabilities associated with transitions from state to state, similar to a Markov chain. AMMs may also be viewed as a degenerate form of hidden Markov model (HMM) (Rabiner 1989) in which the observation symbol probability in each state is 1:0 for a particular symbol and 0:0 f...