A comparative analysis of machine learning algorithms for hidden Markov Models

This research is a comparative analysis between the Baum-Welch and Cybenko-Crespi algorithms for machine learning hidden Markov Models (HMMs). We explore the effect of observation samples, observation length and seeding methods and their impact on the results produced by both algorithms. We show that the key component to learning an HMM is the observation sample, that the seeding method has very little impact on the results and that in most cases the Cybenko-Crespi algorithm proves to be more robust than the Baum-Welch