10/07/19 11:57:12
続きです
Semi-Markov models are a generalization of Markov models that allow the time spent in a state
to be represented by an arbitrary probability distribution and the state transition probabilities to be
dependent on the time that has been spent in that state (Puterman 1994).
In a Markov model, time spent in a state can be modeled only as a self-transition, which has an exponential distribution.
One way of avoiding this limitation is to transform the semi-Markov model into an approximately equivalent fully Markov model.
For instance, Younes finds approximate solutions for generalized semi- Markov decision process models
by replacing each semi- Markov state with a phase type distribution, a series of fully Markov states that
approximate a continuous distribution, and then solving the resulting continuous-time Markov decision process (Younes & Simmons 2004).
This work has only been applied to fully observable planning problems.
However, phase type distributions have been used by Duong et.al. to approximate hidden semi-Markov
models for action recognition of everyday household activities (Duong et al. 2005).