Adaptive Estimation Techniques for Hidden Markov Models

Abstract

• ML techniques for extracting Markov signals imbedded in a mixture of white Gaussian noise and deterministic disturbances of known functional form with unknown parameters. Two such disturbances are considered: periodic disturbances and polynomial drift in the Markov states. • Adaptive on-line schemes for estimating time-varying HMMs and Hidden semi Markov models. We also propose on-line schemes for adaptively extracting Markov signals with time varying statistics imbedded in a mixture of white Gaussian noise and deterministic disturbances with time-varying parameters. In contrast to the off-line estimation techniques mentioned above, the on-line schemes can adaptively learn time-varying models. Also the memory and computational requirements are significantly reduced compared to off-line processing. The discrete-state techniques have applications which we have explored in neurobiological signal processing; in particular, for the extraction of channel currents from noisy measurements in the ionic channels of cell membranes. Extensive simulation studies have been carried out to confirm the robustness of the proposed algorithms. The HFM processing schemes and extraction schemes in the presence of deterministic interferences have also been applied to experimental data obtained from noisy measurements of channel currents in cell membranes. Show more