Skip to main content

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [List Home]
Re: [recommenders-dev] Jayes - Likelihood Evidence

Hi Marcel,

Regarding Performance: i think so. Regarding release: i am not in a hurry, and if Christian is, he could get the code from gerrit once i put it there, right? So no need for a branch on my side.

Regards, Michael

Am 25.07.2014 10:48 schrieb "Marcel Bruch" <marcel.bruch@xxxxxxxxxxxxxx>:
Thanks Michael. 

Regarding release: Yes, we are in a quiet week right now until Wednesday (announced by Johannes a couple of days ago). Then we release 2.1.2. The next release 2.1.3 will probably be in 2 1/2 weeks from now. Maybe a branch is a good solution for this ATM?

Regarding performance: If you still have the code in place… :-)

Marcel


Am 25.07.2014 um 10:06 schrieb Michael Kutschke <kutschke.michael@xxxxxxxxxxxxxx>:

Hi Marcel,

no, the only change that is necessary is a support for virtual evidence, which has no impact unless you want to use it, and is not that big of a change (implemented it yesterday evening). I implemented it as an addition, no changes to the way "normal" hard evidence is used. The algorithm for soft evidence is going to be a wrapper that uses the virtual evidence support in an iterative algorithm, so no changes will be required on top of the virtual evidence support. For you guys, there should not be any measurable change, and the API you use won't change either.

I can measure the performance tonight or tomorrow, just for you ;-)

Anyways, is it safe to push right now, any releases soon?

Regards, Michael

Am 25.07.2014 09:26 schrieb "Marcel Bruch" <marcel.bruch@xxxxxxxxxxxxxx>:
Hi Michael,


I wonder, which implications such a change would have on the performance of the current system. Would this affect the performance of the existing recommenders in any way?

Thanks,
Marcel


Am 25.07.2014 um 07:45 schrieb Michael Kutschke <kutschke.michael@xxxxxxxxxxxxxx>:

Hi Christian,

I found the information i needed. Looks more complicated than i initially thought, but less complicated than i thought yesterday.  I will push a change tomorrow.

Regards, Michael

Am 24.07.2014 15:17 schrieb "Christian Lemke" <lemke@xxxxxxxxxxxx>:
Hi Michael,

given the distinction between soft and virtual evidence, a handling of soft evidence is what is needed for my use case, because the given information/evidence the system recives (from an external source sharing a subset of random variables of the BN in use) has actually to deal with concrete probability distributions which should be treated the same way as hard evidence would be. In this sense (proper) hard evidence can be interpreted as a special case of soft evidence represented by probability distributions containing only zero and one values.

Greetings,
 christian


Am 24.07.14 09:18 schrieb "Michael Kutschke" unter <kutschke.michael@xxxxxxxxxxxxxx>:

Hi,

what makes a difference is this: Are we talking about soft evidence, which means that you observe an actual distribution, or virtual evidence, where you have a distribution P (x is observed| x is the actual state). I assume most people only need the latter, and that can be directly encoded in your model, no changes to the algorithm required. You would just need to add an extra node for the observation.  The point here being that the real value is also modeled as a hidden variable,  distinct from the observation itself.

Regards, Michael
Am 24.07.2014 09:01 schrieb "Michael Kutschke" <kutschke.michael@xxxxxxxxxxxxxx>:

Hi Christian,

i would look into JuntionTreeAlgorithm, inside the method updateBeliefs, there should somewhere be a call to the method that sets the evidence. That could be a starting point. Until i understand the reqirement better and read up  a bit on soft evidence, i have to admit though that anything i say about what it takes to make it work like you expect is entirely guts feeling.

Regards, Michael
Hi Michael, thanks for your immediate response to my question!

I consider to use Jayes as an inference component in a system which recives evident information with certain degrees of uncertainty. So it's possible that the requirements of my intended use case are different from those Jayes is designed for.

What are the crutial parts in the code to be changed, to enablbe a proper handling of soft evidence in the Inferer classes?

Greetings,
 christian


-----Ursprüngliche Nachricht-----
Von: recommenders-dev-bounces@xxxxxxxxxxx im Auftrag von Michael Kutschke
Gesendet: Mi 23.07.2014 15:06
An: Recommenders developer discussions
Betreff: Re: [recommenders-dev] Jayes - Likelihood Evidence

Hi Christian,

Only hard evidence is supported. You could simulate soft evidence by changing the probability distribution of your variable, but that would be a lot less efficient because the internal structures would have to be rebuilt. Supporting the use case in an efficient manner would not be too hard to implement though, i think. It just has not appeared as requirement so far. What is your goal?

Regards, Michael


Am 23.07.2014 14:57 schrieb "Christian Lemke" <lemke@xxxxxxxxxxxx>:


        Hello,
       
        I have a question about the evidence concepts supported by the Jayes implementation of a Bayesian network.
        Does Jayes offer the possibility to set an arbitrary (but valid) probability distribution (e.g. [0.75,0.25]) as likelihood evidence for a node in the BN or is only the usage of hard evidence ([1.0,0.0]) supported?
       
        Thanks,
         christian

        _______________________________________________
        recommenders-dev mailing list
        recommenders-dev@xxxxxxxxxxx
        To change your delivery options, retrieve your password, or unsubscribe from this list, visit
        https://dev.eclipse.org/mailman/listinfo/recommenders-dev
       



_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev


_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev

_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev
_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev


_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev
_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev


_______________________________________________
recommenders-dev mailing list
recommenders-dev@xxxxxxxxxxx
To change your delivery options, retrieve your password, or unsubscribe from this list, visit
https://dev.eclipse.org/mailman/listinfo/recommenders-dev

Back to the top