Bug? Backsampling with Noisy Relevance crashing

[jonnie]

Hello,
I recently discovered the possibility to activate/deactivate relevance reasoning in SMILE and promptly stumbled over some crashes.
It happened with other algorithms as well (eg EpisSampling) but one case that I have isolated is using Backsampling with noisy relevance reasoning enabled (EnableNoisyMAXRelevance). The application crashes when calling UpdateBeliefs. If I use Lauritzen, or don't enable noisy relevance, the inference runs fine.
I should say the model has a lot of noisyMAX nodes.
I also was able to reproduce the crash in GeNIe (setting DSL_NOISYMAXRELEVANCE=1 in the network's user properties). I press the lightning button and after a second Windows tells me that GeNIe 2.0 has stopped working.
Why does that happen? Is it a bug?
Regards & thanks in advance

[shooltz[BayesFusion]]

Is it a bug?

It's definitely a bug. Can you post your network here so we can reproduce it locally?

[jonnie]

I'll see if the error occurs on an obfuscated model as well, and PM you about it.
Thanks for taking care of this

[jonnie]

Update: The crash is a segfault because of a NULL pointer dereference. The bug lies in 15yo legacy code and is therefore not probable to be fixed soon.
Workaround: don't use this combination

[orzech]

Thanks for the update!

[jonnie]

Update: after conducting experiments, I found that the bug occurs:

• in ALL approximate algorithms, plus in LauritzenOld,
• when noisy relevance is enabled, and the model contains noisy nodes,
• only when evidence is set.

(Though I don't know if the latter is valid in general or only in my cases.)

[shooltz[BayesFusion]]

I can confirm the crash in four approximate algorithms and old Lauritzen implementation. However, AIS and EPIS return DSL_NO_USEFUL_SAMPLES (-5).

[jonnie]

Oh, you're correct. I tested with a slightly different model; I'll PM that to you.