Unintuitive dakota behavior?

[Guest Mr. Benoît Bidaine]

I am using Sandia's dakota optimization tool (Design Analysis Kit for Optimization and Terascale Applications) and I understood that some of you also do so.

I am wondering about 2 "unintuitive" behaviors of several SCOLIB algorithms.
- The DIRECT algorithm "DIviding RECTangles" can receive a maximum number of evaluations as input. However this maximum is not strictly enforced. Could you explain why?
- The COBYLA (Nelder-Mead extension) and Pattern Search algorithms start to probe the design space "at the bottom left", i.e., for the lower bound of all optimization parameters. Is there some reason for this? Would it be possible to start from another point, e.g., the middle or a user-defined location?

Thanks a lot in advance for your answers!

[Jörg 29]

Hi,

 

Thanks for this post. Yes, we have an interface for this toolkit in CAESES, and many users work with it.

 

However, I guess it is better to directly ask the makers of these algorithms for detailed information about the behavior. There is a mail list for these kinds of questions, and usually they reply very fast.

 

I hope this helps,

Joerg

[Guest Mr. Benoît Bidaine]

Hi Joerg,

 

Thanks for your quick reply!

I sent my question to the mailing list you referred me to.

 

Best regards,

Benoît