MC acceptance above one

[davidkleiven]

Hi, when I try to run the straight line example (python version) I get the following plot From the code the the lowest plot should be

levels[m]["accept"]/levels[m]["tries"]

so intuitively I would expect this to be a number between zero and one. My levels.txt files:

# log_X log_likelihood tiebreaker accepts tries exceeds visits
0.000000000000000000e+00 -1.797693134862315708e+308 0.000000000000000000e+00 6663 65021 64096 65021
-9.320907461421843454e-01 -1.749548206028707623e+08 7.795952521804472291e-02 38725 65263 27459 51512
-1.961480454579990873e+00 -7.090954872909059304e+03 3.712745756088198235e-02 20337 65240 43565 15800
-2.961163254942303702e+00 -4.855955034352015218e+02 7.277952412568250429e-01 8506 65158 56951 31507
-3.922006542673133200e+00 -3.284338445411593170e+02 2.378482294687733756e-03 604 64979 59557 44329
-4.894153694783229369e+00 -3.003078811896248226e+02 6.794595506886329517e-01 57596 65177 16272 47379
-5.767619056994849558e+00 -2.769564987772033646e+02 5.808150005686155692e-01 49014 65001 58981 3177
-6.787211705620835644e+00 -2.529338132609802301e+02 1.325988680728897862e-01 39177 65247 17309 47967
-7.657161659096917816e+00 -2.366950853781018509e+02 8.854486303168679351e-02 31586 206 51584 5065
-8.717490828306891260e+00 -2.194071224585905782e+02 6.961165282119203823e-01 24080 749 59274 42287
-9.687211881666033975e+00 -2.053333999774806671e+02 2.633039113273385468e-01 17731 642 26174 48791
-1.052014012233775730e+01 -1.944308450771352739e+02 7.086774342232374235e-02 12684 755 42448 12914
-1.164252823172497031e+01 -1.818501628319034182e+02 5.453455227779517855e-01 6229 817 36102 33485
-1.271744331596643640e+01 -1.726802094746875298e+02 1.452798424751584605e-01 65227 691 41911 26474
-1.374691369786420125e+01 -1.665984274276879660e+02 8.349636852127451103e-01 58996 614 56991 31617
-1.470776741496226592e+01 -1.632254965193020837e+02 2.302375955141304220e-01 54418 753 54584 46170
-1.570824590371078244e+01 -1.611577816882679031e+02 2.650222043896128810e-01 49454 808 61074 44102
-1.667249624742556335e+01 -1.600268133302136278e+02 5.384903420179038003e-01 44095 718 65450 53464
-1.763286425927335443e+01 -1.594315624985474926e+02 7.343026828493860458e-01 40058 440 14956 58039
-1.853183774749406965e+01 -1.591216632242262676e+02 3.138237975324571494e-01 35363 65502 6913 5736
-1.949135506592003608e+01 -1.589469510413382523e+02 3.147910860109549569e-01 31804 65400 59200 64288
-2.050191628089351781e+01 -1.588512384954648269e+02 3.213441996768060527e-01 27498 65354 44316 50236
-2.156339305503236048e+01 -1.588005126460745089e+02 5.574140749908829129e-01 23929 65340 36385 39107
-2.264740139644326078e+01 -1.587744460949023733e+02 2.224416291614284558e-01 19557 65295 59691 29866
-2.359146505214206258e+01 -1.587632270122196871e+02 7.447591844072662948e-01 16810 109 49466 53559
-2.463190509350078727e+01 -1.587567936621363174e+02 3.831551339219136376e-01 13381 251 48900 43529
-2.565668287001400216e+01 -1.587537370329235387e+02 1.187293924186407224e-01 10384 472 54660 43085
-2.665100985647396925e+01 -1.587522372612219215e+02 9.438965853890375346e-01 7391 662 18640 49057
-2.751709871697605436e+01 -1.587515238918745979e+02 2.569050943892401406e-01 5009 239 10875 12592
-2.847050189630177996e+01 -1.587510753461820059e+02 3.502633534631354184e-01 2681 65496 0 0

so one can clearly see that there are levels with more accepts than tries. Is this as it should be? In that case what is the interpretation of MH acceptance?

[eggplantbren]

Your interpretation is correct, and this shouldn't be happening. Are you using Python 2? It is not supported and I have seen strange behaviour from it before.

[eggplantbren]

I just reproduced it on Python 3. Will investigate

[eggplantbren]

I think this is beyond my pay grade. Hopefully @dfm can take a look.

[davidkleiven]

I am not sure if this helps, but when I was running with python3 and executed the dnest4.postprocess() while the code was running the code the plots looked fine until around 180 particles were saved. I included the plot I got before the strange thing happen

[dfm]

@eggplantbren: this looks like a problem with the levels.txt file (accept > tries), not the Python processing...

[eggplantbren]

When you're using a Python-based model, levels.txt appears to be created by some stuff in sampler.py and/or backends.py. It's not SamplerImpl.h - I put some diagnostic stuff in there to see what was going on, but it never got executed.

[eggplantbren]

Fixed with 7371e0f294051c45e0d18f1ca3dbd8584299a7e2. @dfm can reopen it if there's some reason my fix (using a longer unsigned integer type) is likely to cause any problems. I don't see why it would, and it seems to be working fine.

[davidkleiven]

Yes, it works fine! Thanks!

[dfm]

👍