Brandon T. Willard

We should introduce a few basic macros and interface changes so that running `miniKanren` from [Hissp](https://github.com/gilch/hissp) and [Hy](https://github.com/hylang/hy) will behave just like it does in standard Lisp implementations (e.g. a...

This PR updates the `pymc3-hmm` codebase so that it works with PyMC3 `v4`. - [x] Finish converting `DiscreteMarkovChain` - [x] Convert `FFBSStep` - [ ] Convert `TransMatConjugateStep` This requires the...

Create and test a DP mixture example

1

We can use the truncated DP mixture approach given [in the PyMC3 examples](https://docs.pymc.io/notebooks/dp_mix.html#Dirichlet-process-mixtures) to estimate the number of mixture components. Let's see how well that works with our HMM models...

Speed up posterior predictive sampling

5

Posterior predictive sampling takes considerably longer than model estimation itself, and that's completely unreasonable! Last time I tried, the new [`fast_sample_posterior_predictive`](https://github.com/pymc-devs/pymc3/blob/master/pymc3/distributions/posterior_predictive.py#L124) function wouldn't work with our custom distribution classes. If...

Try the new JAX functionality in Theano

1

We should see if our implementations can be converted to JAX, and whether or not that offers any performance advantages. We start by attempting conversion (e.g. using [`jax_funcify`](https://github.com/pymc-devs/Theano-PyMC/blob/master/theano/sandbox/jaxify.py#L156)) of our...

FFBS performance improvements

1

There are a few straightforward things we can do to improve sampling performance: - Re-use allocated log-likelihood arrays - These arrays can be fairly large (i.e. length of the series...

Basic model assumptions and sampler tests

3

We need to create some toy/example series and build tests around them. See the examples [here](https://github.com/brandonwillard/amimodels/blob/master/tests/test_step_methods.py#L30) for reference.

In cases for which we are only estimating Poisson rate parameters, it would be nice to have a simple conjugate sampler when we set gamma priors (on the rate parameters)....

Consider implementing the Conway-Maxwell-Poisson distribution

1

The [Conway-Maxwell-Poisson distribution](https://en.wikipedia.org/wiki/Conway%E2%80%93Maxwell%E2%80%93Poisson_distribution) offers another way means of capturing over- and under-dispersion in a Poisson-like framework. It has also been [studied under HMM models](https://www.tandfonline.com/doi/abs/10.1080/00031305.2018.1505656?journalCode=utas20) and shown to "cope with the...

We need some simple unit tests that confirm proper estimation of the initial state probabilities (e.g. `gamma_0` with a Dirichlet prior in most of our models).