Heuristic Techniques for Constructing Hidden Markov Models of Stochastic Processes

Three interrelated heuristic techniques for setting the parameters of hidden Markov models for implementation in pattern recognition algorithms of stochastic processes recorded in the form of sequences of observations are proposed. The techniques make it possible to obtain working models for a small number of training implementations. The first two techniques include the stage of preliminary adjustment of the initial parameters of the model using a priori data and the training stage using the Baum-Welch algorithm. At both stages, an additional procedure for adjusting the model parameters is used, which makes it possible to eliminate numerical problems when they are implemented in recognition algorithms. The third technique implements the procedure of weighted averaging of the parameters of hidden Markov models obtained by the first two techniques. The results of experimental testing of the techniques are presented, illustrating the quality of the resulting hidden Markov models used in the algorithm for pattern recognition of stochastic processes.