I293 nuts better dual averaging

[martinjrobins]

Fixes #293

This is an update of the previous nuts pr, with better dual averaging to more closely match STAN

also removes slice sampling option, as was too difficult to keep in step with the monomial (sp?) sampling

[MichaelClerx]

@martinjrobins I've done the simpler coverage issues, but there's still a few to tick off in the actual nuts/dual-averaging code. (You can see them on the "Diff" tab, not sure why the other tabs in codecov don't work for this)

[martinjrobins]

@MichaelClerx, all changes done, can you check if this looks ok then we can merge? I've tested everything except for one branch of an if statement, which I don't think is worth testing explicitly.

[MichaelClerx]

Are you sure that whatever causes the inf we test for here, doesn't cause problems somewhere else?

[martinjrobins]

you mean the nan coming from the logistic model? It might do, do we have a pints policy on models returning nans? we need to test nuts more thoroughly, and in conjunction with transformations since that is what is done in Stan etc. So I'd like to merge this in (and the transform and interfaces branch) before doing more testing.

[MichaelClerx]

you mean the nan coming from the logistic model? It might do, do we have a pints policy on models returning nans? we need to test nuts more thoroughly, and in conjunction with transformations since that is what is done in Stan etc. So I'd like to merge this in (and the transform and interfaces branch) before doing more testing.

Is that where the NaN is coming from, from a NaN returned by the logistic model? We've got an open issue about a "policy" for this at #794
But my concern is that this NaN could easily cause trouble elsewhere in the method, so I'm not sure hiding this with a pragma is a good idea?

I'll do a quick check and then we can merge, maybe leave open an issue to check its behaviour if a NaN gets returned?

[MichaelClerx]

🥳