Hey Conrad, this is great. The only suggestion I'd make is that often times you want a probability distribution which is updatable over time. This happens in online learning as well as Gibbs sampling where you always only implicitly store the posterior b/c it's changing constantly. So in my research code, I typically seperate the sufficient statistics of a distribution (raw counts in the case of a multinomial as well as a total count to avoid wasteful re- computations) from the logic of yielding a probability (in the case of multinomials just return count/total-count). This way it's easy to update distributions and make changes on the fly such as altering your smoothing parameters.
On Jan 8, 1:54 am, Konrad Hinsen <konrad.hin...@laposte.net> wrote: > I have just added a new module for handling (finite) probability > distributions to clojure-contrib. Some examples are included as well, > but for those who want to see the examples without downloading the > clojure-contrib source code, I also uploaded them to the files > section of this group: > > http://groups.google.com/group/clojure/web/ > probability_distributions.clj > > This module was initially inspired by the PFP library for Haskell > (seehttp://web.engr.oregonstate.edu/~erwig/pfp/, and in particular > the first paper under "further information"), but ended up being > quite different: > > 1) PFP uses lists of (value probability) pairs to represent > distributions. Clojure's maps are a more natural fit, so I use maps > from values to probabilities. > > 2) At the moment, I have only implemented deterministic > transformations of finite probability distributions. PFP also offers > random sampling, which I plan to add in the future. > > 3) I added a second monad, cond-dist, which allows a straightforward > implementation of Bayesian filters. > > 4) A significant part of the PFP code does nothing else but hide the > implementation details behind a few abstract data types. This makes > the code much more difficult to understand. In Clojure we don't have > to deal with types, so this overhead disappears. > > Any feedback is welcome, as always! > > Konrad. --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "Clojure" group. To post to this group, send email to clojure@googlegroups.com To unsubscribe from this group, send email to clojure+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/clojure?hl=en -~----------~----~----~----~------~----~------~--~---
