Sequential Monte Carlo estimation of transition probabilities in mixture filtering problems

Physical systems switching between various working regimes are often encountered in practical applications. However, transition probabilities, according to which a system switches from the current regime to another one, are commonly designed as a priori known parameters, and their misspecification can degrade the performance of the algorithms filtering (or estimating) latent variables of the system. To overcome the misspecification, the present paper proposes a novel Sequential Monte Carlo procedure for estimating the transition probabilities. More specifically, it extends the concept of Rao-Blackwellization to the Dirichlet distribution, which represents the model of these probabilities. The experiments show that the proposed technique outperforms some of the classical methods in terms of the estimation precision and also the precision stability.