[Coin-ipopt] warm start from AMPL

Andreas Waechter andreasw at watson.ibm.com
Tue Jan 8 09:12:21 EST 2008


Hi Adela,

Great, I'm glad to hear it is working fine for you.

One more thing you could try if you wanted to:

Instead of choosing the initial mu yourself in the restart, you can try if 
it gets faster if you use the adaptive barrier parameter option, which 
particularly for a warmstart might work better.  In general, this option 
often leads to fewer iterations, but the cost per iteration can be higher.

The relevant options are

mu_strategy adapitve
mu_oracle (where you might want to try quality-function, loqo, probing)

Cheers,

Andreas

On Tue, 8 Jan 2008, [iso-8859-1] Adela Pagès Bernaus wrote:

> Hi Andreas,
>
> Thanks for your comments.
>
>> 3. Are you saying that you would like to intercept Ipopt in each > iteration? Usually, this is what the intermediate_callback method of the > TNLP is for, if you are using C++ to formulate your problem; this can be > used to decide to possibly terminate the optimization early, or to do some > output. However, from AMPL, this is not accessible. You can do warm > starts from AMPL, but if you would only do one iteration in Ipopt, and > then do a warmstart from AMPL, the internal state in Ipopt will not be the > same as when you left the previous solve (things like the barrier > parameter value, filter status etc will be different). So, in case this > is what you meant, the answer is probably, that you can't do that. (Why > do you want to do that anyway...?)>
>
> The problem I presented was only to check that I store all the information I need for doing some sort of warm starting from AMPL. Doing this exercise I realized that the information about mu is also a key point, but I didn't know about the filter status etc information.
>
> Anyway, with the current information in some problems I've halved the computation time, which is great news!
>
>>> I'm trying to figure out this hypothesis with a file_print_level of 8 > > but the information issued is not clear to me. Can anyone help me? Is > > there any option which forces Ipopt to take the point given by the user?> > The Victor's AMPL warmstart option, maybe that is what you are looking > for...
>
> Yes, thank you.
>
> Regards,
> Adela
>
> _________________________________________________________________
> Express yourself instantly with MSN Messenger! Download today it's FREE!
> http://messenger.msn.click-url.com/go/onm00200471ave/direct/01/


More information about the Coin-ipopt mailing list