Confused about some eps values for ?trnlspbc_init

Confused about some eps values for ?trnlspbc_init

The documentation is not quite clear about some of the eps array containing the stop criteria for the TR solver, especially eps(5) and eps(6). Page 4164 of the MKL 11.1 documentation.

1. eps(5) : The left hand side can take negative values, thus making eps(5) potentially trigger an early termination of the solver (eps array must contain strictly positive values). I think this is actually the case, because I'm sometimes experiencing early terminations with stopping criterion equal to 6, which is caused by eps(5), and I'm wondering if there is a bug with this criterion.

2. eps(6), the trial step precision. I don't understand this one. Is it really a stopping criterion? I don't see any effect on the solver when I change this value. What does the following mean: "If eps(6) = 0, then the trial step meets the required precision (<= 1.0D-10)"? (by the way setting this value to exactly 0 raises an initialisation error).

Thank you in advance for your clarifications

6 posts / 0 nouveau(x)
Dernière contribution
Reportez-vous à notre Notice d'optimisation pour plus d'informations sur les choix et l'optimisation des performances dans les produits logiciels Intel.

Hi,

I agree with you in fact that declaration of these stopping criteria is a bit confused. I will check its implementation and return with deeper answer asap.

Thanks,

Alex

Hi Alex,

Thank you for looking into it. Any update on this subject?

Thanks

 

Simon

Hi Simon,

Sure, I have an answers:

1. eps(5) : The left hand side can take negative values, thus making eps(5) potentially trigger an early termination of the solver (eps array must contain strictly positive values). I think this is actually the case, because I'm sometimes experiencing early terminations with stopping criterion equal to 6, which is caused by eps(5), and I'm wondering if there is a bug with this criterion.

[akalinki] This is an issue in MKL branch - we use internally different parameter for eps(5) than declared one

2. eps(6), the trial step precision. I don't understand this one. Is it really a stopping criterion? I don't see any effect on the solver when I change this value. What does the following mean: "If eps(6) = 0, then the trial step meets the required precision (<= 1.0D-10)"? (by the way setting this value to exactly 0 raises an initialisation error).

[akalinki] It is not real stopping criteria, but length of step calculated by TR algorithm. If such step became lower then eps TR decided that step is almost zero and break computational - practically it is stop criteria. But in the real application such step is quite huge and other criteria execute early. 

Thanks,

Alex

 

[akalinki] This is an issue in MKL branch - we use internally different parameter for eps(5) than declared one

 

[simon] Does that mean that eps(5) passed to trnlspbc_init() is ignored? What does the associated stopping criterion #6 mean, then?

 

Thank you

 

Simon

Please, we need to know how these parameter are handled so that we correctly use them. We also need to understand if the problems we occasionaly observe come from a misinterpretation of these parameters. When can we expect an updated documentation on these stopping criteria?

Thank you!

 

Regards

Laisser un commentaire

Veuillez ouvrir une session pour ajouter un commentaire. Pas encore membre ? Rejoignez-nous dès aujourd’hui