Open yuki-koyama opened 5 years ago
This is probably not a bug but a correct behavior.
This document, https://dl.acm.org/citation.cfm?id=279236, said
If the line search is unable to find a point with a sufficiently lower value of the objective after 20 evaluations of the objective function, [...]
which indicates that it is possible to fail the line search to find a solution that satisfies the necessary conditions. This document also said
In this case all correction vectors [...] are discarded and the iteration is restarted along the steepest descent direction.
So it may be better to follow this restarting strategy. Anyway, this issue needs to be examined still.
I've just found that the demo of GPR sometimes fails to optimize hyperparameters using BFGS or L-BFGS in this library; it throws run-time exceptions during line search. This needs to be examined carefully.