Why return the *function* value when maxIter fails to converge to tolerance?

[fasiha]

I was minimizing a function of Dates using this fine library, with just lowerBound and upperBound and without specifying tolerance/maxIter, and was surprised to see function values instead of argument values as output, till I read in the docs that,

If tolerance is not met, returns best answer.

(This is what triggered my #1… 😅)

This is such a surprising behavior for when the tolerance is not met, I almost think that it'd be better to throw an error in this situation rather than return the function's value?

Just a suggestion. Now I've made my tolerance 1e3 (since my function operates on milliseconds since Unix epoch) and am getting perfect results, hooray!

[jaspervdg]

Having taken a quick look at the code, I suspect the intention was to return an "argument", and it's just a bug.

…

On Mon, Sep 17, 2018 at 10:02 PM Ahmed Fasih ***@***.***> wrote: I was optimizing a function of Dates, and just using this fine library with lowerBound and upperBound without specifying the tolerance/maxIter, and was surprised to see *function* outputs rather than argument outputs, till I read in the docs that, If tolerance is not met, returns best answer. (This is what triggered my #1 <#1>… 😅) This is such a surprising behavior for when the tolerance is not met, I almost think that it'd be better to throw an error in this situation rather than return the *function's value*? Just a suggestion. Now I've made my tolerance 1e3 (since my function operates on milliseconds since Unix epoch) and am getting totally sane results, hooray! — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#2>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADHOwQ_4JO2Qj6FB7mpuRsngkm4vNa8Vks5ub__VgaJpZM4WsnPE> .

[rreusser]

You're right! That makes precisely no sense at all. Gosh, it's painful to read old code.

What do you think about an optional status argument, e.g. minimize(f, options, status). After completion, it'll have, maybe:

status = {value: minimizedValue, iterations: ##, converged: true|false}`

That way you can get lazy and just use it inline (argmin = minimize(f)), and if you want to be careful about convergence, there's no overloading of outputs or try/catch (is try/catch still the optimization killer it's traditionally been?). I suppose it'd either return either the best guess or NaN (probably NaN?) if it fails to converge.

Edit: Ah, and if you return NaN, then you can still return the best guess via the status, just in case that's useful information…

[fasiha]

Fixed by #4!