search

pm4py / pm4py-core

Public repository for the PM4Py (Process Mining for Python) project.
https://pm4py.fit.fraunhofer.de
GNU General Public License v3.0
722 stars 286 forks source link

Random fitness/precision result from token replay #148

Closed fmannhardt closed 4 years ago

fmannhardt commented 4 years ago

We have got a bug report over at the R pm4py binding that the token replay (which we have as a default -- which should be changed I guess) gives back random results for the same model and event log when called multiple times. You can find the original bug report here: https://github.com/bupaverse/pm4py/issues/6

Is this expected behaviour? If not I can try to make the test model and log available.

fit-alessandro-berti commented 4 years ago

Dear Felix,

We are aware of some problems if the model contains duplicate transitions (the method cannot handle duplicate transitions properly; hence in the last release we introduced a new version of the token-based replay based on a backwards state space exploration)

However, if the model does not contain duplicate transitions, then it's weird. If you can please provide us the test model and log

Sincerely Alessandro

fmannhardt commented 4 years ago

This is the model:

image

But, I just checked it directly with PM4Py and the token replay gives consistent results. It does not find that the model perfectly fits all traces (alignment-based fitness is 1.0) but it gives always the same fitness of 0.96905...

I will investigate what goes wrong in the automatic conversion from R.