Typically, you would run N number of chromatin states and then manually annotate them, and see which one best fits the data, similarly to K-means clustering. There may be a more automated way as there is for K-means, but not one I am immediately familiar with.