Fail to reach target error -- adjusting phase-space integration?

Asked by Radha Mastandrea on 2021-02-08

I am trying a process with a custom Lorentz model (by default, I ask for 10,000 events) that can be controlled with a 2-dimensional coupling. I am finding that I sometimes get a "Fall to reach target" error (only for large values of the coupling -- the process works fine for intermediate values so I don't think the issue is inherently in the model).

I have tried changing the value of the hard_survey parameter, but this doesn't seem to work:

-- when hard_survey = 0, I get 1,814 events
-- when hard_survey = 1, I get 1,096 events
-- when hard_survey = 2, I get 22 events
-- when hard_survey = 3, I get 25 events

I am wondering if there are any other parameters I can try to increase the number of events (such as second_refine_threshold)? If it helps, I am looking at the process u d > g u d, and I know exactly which of the 5 corresponding Feynman diagrams is causing the issue, as the corresponding subprocess is associated with a very low luminosity (2 orders of magnitude lower than the other 4 diagrams).

Question information

Language:
English Edit question
Status:
Answered
For:
MadGraph5_aMC@NLO Edit question
Assignee:
No assignee Edit question
Last query:
2021-02-09
Last reply:
2021-02-09

Hi,

Did you check 2102.00773 <https://arxiv.org/abs/2102.00773> ?
This contains a lot of information about 2.9.x new version which contains (as explained in the paper)
a lot of tricks to improve the situation.

I would in particular check what happen if I change the SDE_strategy for that process.

Cheers,

Olivier

> On 8 Feb 2021, at 19:15, Radha Mastandrea <email address hidden> wrote:
>
> New question #695430 on MadGraph5_aMC@NLO:
> https://answers.launchpad.net/mg5amcnlo/+question/695430
>
> I am trying a process with a custom Lorentz model (by default, I ask for 10,000 events) that can be controlled with a 2-dimensional coupling. I am finding that I sometimes get a "Fall to reach target" error (only for large values of the coupling -- the process works fine for intermediate values so I don't think the issue is inherently in the model).
>
> I have tried changing the value of the hard_survey parameter, but this doesn't seem to work:
>
> -- when hard_survey = 0, I get 1,814 events
> -- when hard_survey = 1, I get 1,096 events
> -- when hard_survey = 2, I get 22 events
> -- when hard_survey = 3, I get 25 events
>
> I am wondering if there are any other parameters I can try to increase the number of events (such as second_refine_threshold)? If it helps, I am looking at the process u d > g u d, and I know exactly which of the 5 corresponding Feynman diagrams is causing the issue, as the corresponding subprocess is associated with a very low luminosity (2 orders of magnitude lower than the other 4 diagrams).
>
> You received this question notification because you are an answer
> contact for MadGraph5_aMC@NLO.

Radha Mastandrea (rrm50) said : #2

I've switched to the new MG version (it seems to be a lot faster which is great)! However, I'm running into another problem with the integration (which may not be specific to the version as I didn't make these checks for version 2.8). Namely, the "fail to reach target" seems to be highly variable with the random seed, which is not behavior I would expect.

As an example:
-- random seed 5: makes all 10,000 events
-- random seed 6: fail to reach target with 1,993 events
-- random seed 7: fail to reach target with 449 events
-- random seed 8: makes all 10,000 events

I am wondering how MadGraph decides to discard events, and why the failure is so variable?

Thank you!

Hi,

Not sure I have the full picture myself.
This is a bit of a chaotic system with machine learning algorithm and all the issue of false minima, over-fitting,...
and all the condition linked to threshold effect in a lot of place inside the code.
So it is very difficult to have a good understanding of the impact of any change (This is why I'm so proud of the above paper)

Cheers,

Olivier

Can you help with this problem?

Provide an answer of your own, or ask Radha Mastandrea for more information if necessary.

To post a message you must log in.