Re: [rules-users] Planning Under Uncertainty
drools-planner + drools-chance = planning under uncertainty
I am looking forward to the first alpha release of drools-chance,
so I can experiment with "planning under uncertainty" examples,
such as "investment portfolio optimization" etc.
If you want to do it today, you got 2 options as I see it:
1) Use drools-chance in your score DRL (and contribute to and fix
drools-chance)
https://github.com/droolsjbpm/drools-chance
2) Write the uncertainty calculations yourself in your score DRL
Both ways require you to fully understand the math behind "uncertainty
calculations",
which in your case are "probability calculations" I think.
Here's just the beginning of the begin (Bayes' Rule etc):
http://www.khanacademy.org/video/probability--part-6
http://www.khanacademy.org/video/probability--part-7
http://www.khanacademy.org/video/probability--part-8
From a drools-planner perspective, there is no uncertainty:
The score DRL's need to produce a Score for a solution which is
comparable the Score of another solution of the same problem.
Even if the Score instance contains probability parts (which is
allowed), Planner doesn't care:
you still need to define which of the 2 Score instances is better (by
compareTo), something which is domain specific
(and might even depend on input data like how much risk you're willing
to take).
--
With kind regards,
Geoffrey De Smet
Op 24-09-11 00:55, Chris Spencer schreef:
On Fri, Sep 23, 2011 at 6:39 PM, Davide
Sottara<dsotty(a)gmail.com> wrote:
> I'm adding support native for uncertainty to drools: the code is in a
> sub-project repository called "drools-chance". I'll probably commit an
> update next week, supporting the declaration of beans with distributions as
> fields.
That's great to hear. I'm glad I'm not the only one interested in this
functionality.
> Full support in the rules, instead, will take some more time. I would like
> to understand better what you mean by "uncertain planning": are you
> considering actions with uncertain preconditions / effects or both? or are
> you referring to randomized search algorithms?
> Of course, you can always manage uncertainty explicitly, using additional
> facts and calculating the probabilities, but I would not recommend that
> unless you need a working prototype very quickly.
I'm mainly interested in planning with uncertain effects, similar to
what I've found in most probabilistic graphplan implementations (e.g.
http://www.cs.cmu.edu/~avrim/pgp.html).
> Best,
> Davide
>
> p.s.
> The paper you found is old and obsolete, definitely not worthy looking at (
> shame on the main author ;) )
Heh, I'm glad you're still involved with the community after all these
years. Thanks for not dropping the research. I've seen a lot of
interesting graduate research just die as soon as the project/thesis
is over.
Regards,
Chris
_______________________________________________
rules-users mailing list
rules-users(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/rules-users