Maxtime and compilation

[BioTurboNick]

I'm setting maxtime to avoid an occasional infinite cycle on some inputs, but it triggers during the first execution due to compilation of the objective function.

[odow]

What would you expect to happen? I don't think NLopt can tell whether the time is because of a long-running function, or because of compilation. It can only terminate once maxtime has passed. If it's a problem, you could evaluate the objective once before calling NLopt?

[BioTurboNick]

That's fair; if Julia had the ability to track non-compilation elapsed time, NLopt.jl could detect this condition and try again.

[stevengj]

No, the maxtime is actually based on a feature from the NLopt C library — it is just wall-clock time.

Mainly, I find it useful for global optimization (where there aren't good stopping criteria), e.g. "give me the best possible result in 3 hours". But usually in this situation the maxtime is huge compared to any conceivable compilation time.

[BioTurboNick]

I understand it's used by NLopt C library, but the Julia library could handle the exceptional condition specific to Julia.

I've used maxtime just recently when most optimizations take less than a second and occasionally runs for minutes at a time on certain inputs to keep the total time to a manageable level.

[odow]

If it's a problem, you could evaluate the objective once before calling NLopt?

Did you consider doing this? Or run NLopt twice? Once with a time limit of 0.0 and once with your larger time limit?

[BioTurboNick]

Sure I could. That's just annoying 😆

[stevengj]

You could also just use maxeval instead.