Github user mrkm4ntr commented on the issue:
https://github.com/apache/spark/pull/19988
> When gradient nearly zero, and line-search failed. I think this is also
possible. This should get valid model.
Yes, I agree.
> When gradient non-zero, line-search failed, will the model always be
meaning-less ?
I think the model is not always meaningless.
LineSeachFailed occurs when there is an iteration exceeding the maximum
value because the range that satisfies the strong wolfe condition is far away
from the initial value (= 1.0).
However, maxIter is 10. It is considered that the upper limit of the line
search range is 1.5 ^ 10 = 57.665, and the lower limit is 0.5 ^ 10 = 0.001
(assuming the interpolation value is at the center).
Of course, the range is not necessarily continuous, but I think whether it
can be thought that LineSearchFailed will not occur so often if the input or
intermediate calculation process is correct.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
