Unweighting negative-weight events

Hi,

This is just not possible.

Cheers,

Olivier

> On Apr 11, 2016, at 16:26, Denys <email address hidden> wrote:
>
> New question #290493 on MadGraph5_aMC@NLO:
> https://answers.launchpad.net/mg5amcnlo/+question/290493
>
> Dear experts,
>
> Not all multivariate techniques (e.g. neural networks) support negative weight events in the training phase. Therefore one would like to get a sample of positive weight events from a sample which has both positive and negative ones. I wonder whether a general algorithm, possibly with low unweighting efficiency and/or non-constant positive weights, exists.
>
> Kind regards,
> Denys
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.

Hi Olivier,

Thank you for the very quick response! Is it something completely obvious? POWHEG for example has positively definite event weights, but I understand that this is related to the fact that they rearrange the calculations differently. I was wondering, if a sort of an afterburner can be devised for MC@NLO events as well.

Cheers,
Denys

Hi Denys,

> Is it something completely
> obvious?

To me it is.
Now it is possible to have procedure which remap the event to positive weight only.
But those will not correctly describe all the differential distributions and therefore I do not think that this is suitable for NN.

Cheers,

Olivier

> On Apr 11, 2016, at 17:02, Denys <email address hidden> wrote:
>
> Question #290493 on MadGraph5_aMC@NLO changed:
> https://answers.launchpad.net/mg5amcnlo/+question/290493
>
> Status: Answered => Open
>
> Denys is still having a problem:
> Hi Olivier,
>
> Thank you for the very quick response! Is it something completely
> obvious? POWHEG for example has positively definite event weights, but I
> understand that this is related to the fact that they rearrange the
> calculations differently. I was wondering, if a sort of an afterburner
> can be devised for MC@NLO events as well.
>
> Cheers,
> Denys
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.

Could you, please give a reference to remapping procedure, please?

I do not have any.

Cheers,

Olivier

> On Apr 12, 2016, at 10:27, Denys <email address hidden> wrote:
>
> Question #290493 on MadGraph5_aMC@NLO changed:
> https://answers.launchpad.net/mg5amcnlo/+question/290493
>
> Denys posted a new comment:
> Could you, please give a reference to remapping procedure, please?
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.

OK, how do you know that such a procedure exists? I am just curious to give it a try and check the effect.

Hi,

I did not say that such procedure exists just that such procedure should be technically possible depending of the property you are fine to drop.

For example, you can take all your event and make a histogram of the pt of one of your particle.
Then for each bin, you can replace the weight of each event by the mean contribution to this bin.
Then you can do normal un-weighting method if you like.

This procedure:
- only provide positive weight
- the pt of that particle will still be described correctly (up to the bin precision)
- The other variables will not be described correctly

Now this is just a crazy/bad idea of method, showing that if you drop the property that all distributions are correct then you can have a method which has only positive weight. I will not recommend to use such kind of approach.

Cheers,

Olivier

> On Apr 12, 2016, at 11:08, Denys <email address hidden> wrote:
>
> Question #290493 on MadGraph5_aMC@NLO changed:
> https://answers.launchpad.net/mg5amcnlo/+question/290493
>
> Denys posted a new comment:
> OK, how do you know that such a procedure exists? I am just curious to
> give it a try and check the effect.
>
> --
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.

Thank you for the explanations!