Bayesian M-T clustering for reduced parameterisation of Markov chains used for nonlinear adaptive elements (Q5939333)

The paper presents one method of system identification. Here, Markov chains are considered as universal black-box models suitable for the description of nonlinear stochastic systems. The algorithm is proposed, which provides the structure estimation. This estimation is based on Bayesian estimation and a mean-tracking cluster algorithm. In doing so, a significant reduction of dimension in model parameterization is achieved. It is pointed out, to what classes of systems this algorithm may be applied, switching modes, smooth and strong nonlinearities among them. One illustrative, but complicated example is considered.