A quite significant bug was determined recently: if a quadratic local model is strongly inadequate the an algorithm can make a huge step with little or no reduction of the function (jump over a valley) and then terminate by ftol condition.
I added the additional condition ratio > 0.25 (which I used originally, but eliminated at some point). MINPACK, on the other hand, checks actual and predicted relative reductions and checks ratio < 2. I don't completely understand the purpose of the last check, so I need to think about it. Maybe it's better to adopt MINPACK approach (perhaps ignoring ratio < 2 if I won't determine its conceptual purpose).
A quite significant bug was determined recently: if a quadratic local model is strongly inadequate the an algorithm can make a huge step with little or no reduction of the function (jump over a valley) and then terminate by ftol condition.
I added the additional condition
ratio > 0.25
(which I used originally, but eliminated at some point). MINPACK, on the other hand, checks actual and predicted relative reductions and checksratio < 2
. I don't completely understand the purpose of the last check, so I need to think about it. Maybe it's better to adopt MINPACK approach (perhaps ignoringratio < 2
if I won't determine its conceptual purpose).