What's a good stopping condition

[cuihenggang]

I'm wondering is there a good stopping condition that people usually use to decide convergence? Suppose I hope to find the function inputs that give me a function value that is allowed to differ from the optimal value by 10%. Is there a way to do that?

If that is impossible, what's the rule of thumb way of deciding how long to run Spearmint for?

Thank you very much!

[xiongzhp]

Hi do you have any updates or ideas for this question? I am also wondering that. For a model with a few hyper-parameters to optimization, it's almost impossible to arrive at a convergence as the simple demo shows. The good news is that it truly works and gets fairly good results just after a few iterations, although it could be fluctuating a lot.