I made some modelling and it seems making epoch length shorter just would make difficulty more bumpy. I am looking into smoothing it something like following:
- Calculate least square method based difficulty as it is done now
- Made simple diff readjustment as done in Bitcoin
- Calculate average
For most of epochs results are similar (difference with current algo is within 10%), but during turbulent times (after HF and now) this method is catching up with reality at a faster pace.
For epoch length, I guess, 128 blocks is the safest option from short ones.