Closed riastradh-probcomp closed 9 years ago
There is a candidate fix in the 20150924-riastradh-drawconstraint branch, to force a particular value and not just a particular cluster from which to draw in simple_predictive_sample_unobserved
, with an OK from @axch. What remains:
simple_predictive_sample_observed
.I suspect it doesn't matter for observed rows, because we never ask PREDICT x GIVEN x = 0
expecting it to give 0; in fact bayeslite PREDICT
explicitly omits any current value for the requested column in the constraints. But I'm not sure there is no application of simple_predictive_sample_observed
for which you might pass a constraint that you're asking about.
20150924-riastradh-drawconstraint now addresses simple_predictive_sample_observed
with an essay arguing why this is not sensible and inviting arguments to the contrary before we reject such queries.
Fixed in v0.1.29, abc491c56be9ea639210c2b0c2aeec0aef4de372.
If you constrain column 3 to be 42, and sample column 3, Crosscat draws randomly instead of returning 42 as one might expect. This results in strange results in bayesdb like: