Thank you so much for your valuable inputs, Alejandro, Paolo, and Olivier.
Based on your feedback, i have decided to drop the idea of RL-learning for
sklearn.
Instead, here is an alternate proposal: Supervised learning with Bayesian
Networks.
Bayesian networks provide highly interpretable models, and have been used
with notable success in medical diagnosis. If you know the graph topology
(structure) of the bayes-net, inference operations reduce to
maximum-likelihood/expectation maximization. However in practice, the
graph-structure is often unknown.
I propose implementing a bayes-net learner that can learn both the
structure as well as parameters of a bayesian network from a given dataset,
and demonstrate it's uses for prediction / inference tasks.
So, what do you guys think about my new proposal? :)
regards
shankar.
On Wed, Mar 14, 2012 at 10:22 AM, Olivier Grisel
<[email protected]>wrote:
> Le 13 mars 2012 07:53, Alejandro Weinstein
> <[email protected]> a écrit :
> > On Tue, Mar 13, 2012 at 6:37 AM, Shankar Satish <[email protected]>
> wrote:
> >> Do you think my proposal about implementing reinforcement-learning
> >> algorithms (subject line: "GSOC project idea: online learning
> algorithms")
> >> is something that is well suited for integration into scikit-learn? Do
> you
> >> think it makes more sense to start a new scikit focussed on
> reinforcement
> >> learning?
> >
> > Just a couple of comments. There are some RL Python implementations,
> > e.g. PyBrain (http://pybrain.org/) and RL-Glue/RL-Library
> > (http://glue.rl-community.org/wiki/Main_Page). It seems that none of
> > these are being actively developed.
> >
> > The nature of RL problems implies that the architecture of the code is
> > different than the "single script" approach used in scikit-learn. For
> > instance, in RL-Glue/RL-Library you run three independent programs
> > (the agent, environment and experiment programs) plus the RL-Glue
> > process. This approach is natural because it mimics the actual RL
> > problem, where the agent and the environment are two different
> > entities. Also, in the case of RL-Glue, you can combine environments
> > and agents written in different languages. I wonder how this different
> > architecture of RL would match with the scikit-learn ecosystem.
>
> I globally agree with that view: RL does not really fit in the current
> sklearn API. Modeling agents / environment interactions currently
> looks out of the scope of the project. PyBrain is probably a better
> project for this kind of models. Maybe they will take part in this
> year GSoC too.
>
> I must admit I haven't thought through the problem too much though as
> I don't know the RL literature enough to make an informed judgment.
>
> --
> Olivier
> http://twitter.com/ogrisel - http://github.com/ogrisel
>
>
> ------------------------------------------------------------------------------
> Virtualization & Cloud Management Using Capacity Planning
> Cloud computing makes use of virtualization - but cloud computing
> also focuses on allowing computing to be delivered as a service.
> http://www.accelacomm.com/jaw/sfnl/114/51521223/
> _______________________________________________
> Scikit-learn-general mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>
------------------------------------------------------------------------------
Virtualization & Cloud Management Using Capacity Planning
Cloud computing makes use of virtualization - but cloud computing
also focuses on allowing computing to be delivered as a service.
http://www.accelacomm.com/jaw/sfnl/114/51521223/
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
