Efficient SMC inference in Semi-Markov models by compressing the belief state
Martin Nyolt, Thomas Kirste · 2017
Hidden semi - Markov models allow state transitions (actions) to have arbitrary duration distributions. Sequential Monte Carlo (SMC) methods provide an approximate inference for Semi-Markov models. In real-world applications with huge discrete state spaces and long-running actions with flat duration distributions, even SMC methods cannot efficiently approximate the posterior distribution. We propose a compression of the belief state to track significantly many more states with the same computational complexity. A piecewise approximation of the starting time distribution using log-linear models with constant-time online changepoint detection is used. The effect of the approach is analysed and discussed for three different datasets.