Turing NUTS chains getting stuck at the parameter bounds

[fkrauer]

@dlakelan

What does the lp value look like in these different chains? I’m guessing some of these are really far from the converged lp value, they just can’t move because they’re in an island where in all directions things look worse. If that’s the case probably initial conditions are needed.

Yes, this is the main issue here. I have had this also with other models. I ended up discarding the chains that didn’t converge, but ultimately this isn’t the best approach to tackle this problem. As mentioned before, parallel tempering is probably worth a try, but this isn’t implemented in Turing yet.