This is an important topic to discuss. The USUAL and most logical way would be to describe the form of the 'hidden state' in the model like "please give me information about hidden_marker X, from which i think the graph structure is left-right, using following layers....).
This would mean that a new model would be constructed on the basis of a graph structure and observations. Thus, we could fit a model and use the baum-welch algorithm to improve our simple model. The problem is, that since we do not have any actual hidden states in our interpretation of the modelling problem, drawing the transition and emission matrices directly from observed data is much more efficient and cleaner. We violate the idea of training a hidden markov model in this very special case. This should definitely be discussed in the thesis, emphasizing why exactely we violate the usual order of business concerning the training of a HMM.
This is an important topic to discuss. The USUAL and most logical way would be to describe the form of the 'hidden state' in the model like "please give me information about hidden_marker X, from which i think the graph structure is left-right, using following layers....).
This would mean that a new model would be constructed on the basis of a graph structure and observations. Thus, we could fit a model and use the baum-welch algorithm to improve our simple model. The problem is, that since we do not have any actual hidden states in our interpretation of the modelling problem, drawing the transition and emission matrices directly from observed data is much more efficient and cleaner. We violate the idea of training a hidden markov model in this very special case. This should definitely be discussed in the thesis, emphasizing why exactely we violate the usual order of business concerning the training of a HMM.