[Coin-ipopt] ensuring decrease of objective

[James Lu]

On Thursday 03 May 2007 00:22, Andreas Waechter wrote:
> Hi James
>
> What Graham describes is the intended way to give users the opportunity to
> stop the algorithm prematurely, or to collect some intermediate
> information.  So, if that is what you are interested in doing, that's the
> way to go.
>
> But I'm not sure I understand you original question:  When you describe
> that the objective function increases at the end of the optimization, does
> that happen because you interrupt the optimization early?  (I assume here
> that you have no constraints except for bound constraints for the
> variables...)  In that case you might want to try if the optimization is
> better when you either use a small initial value of the barrier parameter
>
> mu_init 1e-4   [or some other small value]
>
> or if you use the automatic choice of the barrier parameter
>
> mu_strategy adaptive
>
>
> Regards,
>
> Andreas

hi Andreas:

Thanks for the hints! 

Yes, we stop the optimization early: for computational efficiency, my 
colleague tried max_iter = 5, 10 for each call to local solve (within scatter 
search). Our problem has ~30 unknowns, and only bound constraints for each. 
In your experience, is this too few to ensure descent? 

One of the problem might be that sometimes (due to the scatter search 
strategy) the initial point is right at the bound. We have tried setting 
mu_init small; will see if the mu_strategy = adaptive helps.

best regards,

James