search

ReactiveBayes / RxInfer.jl

Julia package for automated Bayesian inference on a factor graph with reactive message passing
MIT License
265 stars 23 forks source link

Free energy is not computed for non-terminated graphs. #179

Open albertpod opened 11 months ago

albertpod commented 11 months ago

We have added predictvars feature for the inference function. Due to some theoretical concerns, the computation of the free energy when missings are present was left aside.

Now, if one wants to compute free energy on non terminated graph, he'll get

The Bethe Free Energy computation is not compatible with the prediction functionality. Set `free_energy = false` to suppress this error.

@ismailsenoz has some good insights on this one in relation to expected free energy. IMHO, we should pick it up.

albertpod commented 10 months ago

@ismailsenoz, perhaps we just want to establish theoretical grounding. The implementation will be for another milestone.

bvdmitri commented 10 months ago

@ismailsenoz This task has been added to the milestone for tracking and prioritization.

ismailsenoz commented 10 months ago

We discussed with @albertpod that the heuristic of omitting entropies from the computation seems justified from theoretical grounds. Perhaps, we can discuss if further need be.

albertpod commented 10 months ago

There's still a small realization bottleneck. Let's discuss this onsite. Anyway, I would first address #180, and then we can code this one up.

@bvdmitri WDYT?

bvdmitri commented 10 months ago

I'm up for the discussion. However, I think #180 is orthogonal to this one and we can start those simultaneously (not necessarily you of course)

wouterwln commented 6 months ago

@albertpod @bvdmitri @ismailsenoz what is the status of this?

bvdmitri commented 6 months ago

I'm not the right person to poke about theoretical grounds of this, but again I'm open for discussions