[SciPy-dev] Another GSoC idea

[Robert Kern]

On Wed, Mar 25, 2009 at 12:21, David Warde-Farley <Dwf at cs.toronto.edu> wrote:

> Then there's PyMC, which as far as I can see has developed a *really*
> well thought out object-oriented system for specifying probabilistic
> graphical models. Of course, it's geared toward Bayesian inference via
> MCMC. In the (relatively rare) case that the posterior is analytically
> available it shouldn't be all that difficult to graft on code for
> doing that. Likewise with maximum likelihood (hyper)parameter fitting
> via EM or gradient-based optimization.

It does bring to mind an idea, though: replicate the probability
distributions from scipy.stats analytically using sympy. I think that
has reasonable scope by itself, but going through *all* of the
distributions may be more of a boring chore than I would inflict on a
student for an entire summer. However, once one gets enough
distributions to be interesting and to give assurances that one has
covered all of the use cases in the design, the student can forgo the
remaining distributions to work on a way to combine these distribution
objects into probabilistic models that can be converted to efficient
MCMC, EM, or other such numerical codes for estimation.

I haven't looked at PyMC recently, though. Maybe it has already
cornered the model construction API, and the implementation just needs
some tweaking to allow other operations on the models.

-- 
Robert Kern

"I have come to believe that the whole world is an enigma, a harmless
enigma that is made terrible by our own mad attempt to interpret it as
though it had an underlying truth."
  -- Umberto Eco