<div dir="ltr">Hi Ce Qin,<div><br></div><div>Typically the MT line search accepts the first trial point, it is mostly used to throw out a few terrible guesses and to help prevent stalling. You should get a general idea of how much time is spent line searching just be checking the number of TAO iterations versus the number of function evaluations in -tao_view, and there are simple ways to access this</div>
<div>information directly if you need to.</div><div><br></div><div>There are some parameters for the line search that make it more or less selective, but saving a few function evaluations here in the line search will probably cost more evaluations in the broader view of the optimization in general.</div>
<div><br></div><div>If you haven't done so yet, try the lmvm algorithm as a substitute for cg, it works on the same information and usually performs a little better.</div><div><br></div><div>Jason</div></div><div class="gmail_extra">
<br><br><div class="gmail_quote">On Sun, May 25, 2014 at 8:12 AM, Ce Qin <span dir="ltr"><<a href="mailto:qince168@gmail.com" target="_blank">qince168@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<div class="HOEnZb"><div class="h5">
<div>
Thanks, Jason. That's what I need.<br>
<br>
One more question, How many function and gradient evaluations one MT line search need?
<div>I'm doing geophysical inversion, function and gradient evaluation is very expensive, so I want to minimize the function calls. Do you have any suggestions?</div>
</div>
</div></div></blockquote></div><br></div>