forum.alglib.net

Hi,

I have been experimenting with the minnlc optimizer (the SQP algorithm, set with minnlcsetalgosqp).

The documentation states that the scaling coefficients set with minnlcsetscale are used to test stopping conditions, and when using numerical differentiation. My problem has an analytical gradient, so as per the documentation, the scaling coefficients shall only affect stopping conditions. I presume these stopping conditions will come into effect when the optimizer will be doing some kind of line search internally.

Is the minnlcsetscale documentation accurate? Are the scaling coefficients really used only to test stopping conditions, or does the SQP solver do any kind of preconditioning based on these scaling coefficients?

Thanks,
Matevž

Hi!

Yes, you are right and the documentation is not exactly accurate here. The SQP solver also uses scales for preconditioning. In particular, internal quasi-Newton SR1 update uses inverted squared variable scales as an "initial" diagonal matrix for the update.

I have fixed the documentation, the upcoming 4.02 release (will be in several days, a week at most) will feature it.

Sergey.

Thanks!