When dwell time histogram is over-binned, ML-DPH will find an optimum model in Erlang transition scheme with ridiculously populated states and such an overestimation of state degeneracy make the transition rate matrix calculation infinitely long.
To prevent infinite calculations, an additional test is run iteratively on the best ML-DPH fits ordered by goodness of fit (BIC):
ML-DPH is run on the non-binned histogram
if diagonal DPH transition probabilities are too low, the model is discarded and the next best fit is tested
the first model yielding proper diagonal DPH transition probabilities is selected as best fit
This way, over-binning may preferably underestimate state degeneracies, which still yield the wrong result, but is less problematic for the inference time of transition rates.
When dwell time histogram is over-binned, ML-DPH will find an optimum model in Erlang transition scheme with ridiculously populated states and such an overestimation of state degeneracy make the transition rate matrix calculation infinitely long. To prevent infinite calculations, an additional test is run iteratively on the best ML-DPH fits ordered by goodness of fit (BIC):
This way, over-binning may preferably underestimate state degeneracies, which still yield the wrong result, but is less problematic for the inference time of transition rates.
fixes #113