Switch to PRIMA.bobyqa for optimization?

[dmbates]

See the discussion in lme4/lme4#744 regarding PRIMA.bobyqa in linear mixed models. It may be even more useful to switch to PRIMA.bobyqa for GLMM optimization because that version of bobyqa allows for a scale argument which can be very helpful when simultaneously optimizing the fixed-effects and variance-component parameters.

If we make this change should we retain the ability to use the NLopt optimizers or allow only the PRIMA optimizers? It would be good to retain NLopt and hence compatibility with earlier versions but it could get messy with the optsum structure and with tests - not to mention the potential for some really ugly and confusing code.

[zaikunzhang]

Hi @dmbates ,

I am very glad that PRIMA turns out useful to you and your community. The Julia interface was entirely due to the efforts of @emmt and @amontoison. Ping us if you need our help.

Cheers,
Zaikun