Fast convergence identification of Hidden Markov models using risk-sensitive filters

In this paper we derive recursive risk-sensitive filters which may be used for both on-line and off-line identification of hidden Markov models (HMMs). The identification is achieved by first taking risk-sensitive conditional mean estimates of the number of state transitions (jumps) and occupation times. These values are then used to estimate the parameters of the system. This risk-sensitive scheme has the advantage of offering the possibility of more rapidly convergent parameter estimates. In turn, there can be more rapid optimization of systems. To demonstrate the potential of the risk-sensitive filters over existing identification schemes, we present a number of simulation studies.