Alexey Radul
Alexey Radul
Student-T with few degrees of freedom. In any case, that's a non-issue for Venture. Nothing interests itself in means or higher moments -- just the sampler and the density.
Except for the assessor.
Also part of #570.
Generic workaround: force a fake random choice: ``` venture[script] > observe normal(exactly(0), 1) = 10 venture[script] > global_log_likelihood [-50.918938533204674] ```
This is a mirror of the problem I had with Stan doing the same thing. If we are interested in Venture-RIPL models being usable as components of other probabilistic systems,...
The commentary on #455 provides a hint to this problem. When the uncollapsed kernel understands the local posterior well enough, it should be able to report its size (as a...
Apparently it also has derivatives implemented, and might possibly even come with an automatic differentiation system that we could use to get gradient methods to work in Puma
Paper: http://arxiv.org/abs/1509.07164
By code inspection, I assume that `mapv` currently has this problem, but it is masked by its output PSP not defining a gradient method at all.
Assuming there are no underwater stones (which assumption may be false, as @riastradh-probcomp tends to be pretty thorough), finishing this seems like a good exercise for @Schaechtle , who is...