On Feb 28, 2012, at 11:53 AM, Sascha Merz wrote:
Thanks a lot for providing this great optimization library. One
question that I have is if there is a certain reason that gradients
are computed in the MMA inner iterations.
Two reasons:
First, if it is the last inner iteration, then I would have to compute
the gradients anyway for the next outer iteration, so I'd have to call
the objective function again. So, there is a tradeoff here, but if
computing the gradient is cheap compared to computing the objective
function (as is often the case) then it seemed worth it to compute it
in each inner iteration just in case that inner iteration is the last.
Second, I use a slight modification of the Svanberg algorithm, in that
if an inner iteration finds a better feasible point but doesn't
satisfy the termination conditions for the inner iteration, then I
continue the inner iterations with a new approximant based on the new
feasible point.
However, it would be straightforward to modify the code to change the
above two behaviors, going back to the original Svanberg algorithm and
not computing the gradient during the inner iterations (but
recomputing the objective after the inner iterations complete).
_______________________________________________
NLopt-discuss mailing list
[email protected]
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss
