Efficient and Effective Learning of HMMs Based on Identification of Hidden States

Figure 6

Identification performance for a 10-state discrete persistent HMM. (a) and (b) show a comparison of log-likelihoods during iterative trainings with different models: the 20 repetitive BW models with an unknown or known number of states , the SCT models, the selected best BW models without or with a given , the ground truth HMM, the multinomial model, the best one-state simpler model , and the best two-state simpler model . (c) shows a comparison of the model parameter heat maps for the ground truth HMM, the best BW models with an unknown or a given , and the SCT model.

(a) Log-likelihoods during iterative training, BW with unknown

(b) Log-likelihoods during iterative training, BW given correct

(c) Heatmap of HMM model parameters