The methods of Section 2.7 
discuss ways to change
the parameters of an HMM without changing the structure of the
model (other than deleting edges or states that are not used).
In this section, I'll discuss a few techniques for modifying the model
structure, either to improve the match between the model and the seed
sequences or to decrease the size of the model.
After any operation that changes the structure of the model, the probability distributions for the states and edges involved in the change need to be recomputed, so that they continue to sum to 1.
Two operations reduce the size of the model: RemoveUseless and MergeStates. Four operations increase the size of the model: AddUseful, AddSkipEdges, UnrollSelfLoops, and SplitVagueStates.