Alexey Radul

What should two-argument categorical-family distributions do when the selected input changes?

3

Consider the expression `(categorical (simplex 0.5 0.5) (array (flip) (flip)))`. Suppose one coin is currently `True` and the other `False`, and the `True` one is currently selected. Suppose single-site m-h...

Two-argument categorical and dirichlet-categorical have the wrong density if their object list contains duplicates.

1

`(assess 1 categorical (simplex 0.25 0.25 0.25 0.25) (array 1 2 3 1))` should be 1/2, not 1/4. Dirichlet likewise. Plan: - [x] Fix categorical (easy) - [ ] Fix...

Either - `ripl.sample_all_with_weights`, or - `ripl.promote_particles` returning a list of (ripl, weight) pairs, with one new ripl per old particle or both. This is idea is from rather old notes...

Block enumeration on dependent principal nodes may generate wrong stationary distribution

2

Block enumeration proceeds thus: - Query each principal node's operator for the set of possible values (before detaching) - Form the Cartesian product - Evaluate the posterior at every combination...

Complete the categorical family of distributions

6

The "categorical family" are the categorical and multinomial, with their Bernoulli and binomial variants, in versions of single call, non-conjugate but tracking sufficient statistics, collapsed conjugate, and uncollapsed conjugate; and...

Plan to fix conjugate dirichlet problems

2

Our dirichlet-categorical woes, namely #407, #451, and #452, are ultimately due to representing the probabilities in terms of the indexes of the options rather than in terms of their actual...

Dirichlet categorical logDensityOfData ignores the options, provides no feedback on it.

1

Consider again the problem from #451: ``` assume x1 = flip() assume x2 = flip() force x1 = true force x2 = true assume f = make_dir_mult(array(1, 1), array(x1, x2))...

Do we want to implement a multi-variate slice sampling scheme for Venture?

1

We have access to someone very well versed in the subject matter: Madeleine B. Thompson, ["Slice Sampling with Multivariate Steps"](http://www.cs.toronto.edu/~radford/ftp/madeleine-thesis.pdf), PhD Thesis, University of Toronto, 2011.

Complete testing of GP covariance kernel functions

1

At the time the gpgradient branch was merged, the automated testing in `test/conformance/sps/test_covariance.py` was incomplete, in that it did not cover the bias, linear, and product kernels. That did not...

Preserve semantics of CRP atoms across detach+regen cycles

5

Consider what happens if the last CRP application that was assigned to a particular atom is detached and later regenerated, or its support enumerated. For example, ``` assume crp =...