You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Nov 23, 2018. It is now read-only.
Default Printer reports NaNs as objective function values when GradientDescent is used and every iteration printer. The reason is that Backtracking requires only FunctionEval, so when the step is finished, GradientEval is requested. evaluate() for this evalType puts math.NaN() into location.F unconditionally (original location.F is then restored back by Backtracking).
One option would be to test if xNext is different from location.X and put NaN into location.F only if it is. However, this implies extra work when Dim is large. Any other options?
The text was updated successfully, but these errors were encountered:
local.go:evaluate() puts unconditionally a NaN in location.F if
GradientEval is requested, or into location.Gradient if FunctionEval is
requested. This may cause some issues especially in Recorders.
Therefore, use NaNs only if location.X and xNext are different.
Fixes#11
Default Printer reports NaNs as objective function values when GradientDescent is used and every iteration printer. The reason is that Backtracking requires only FunctionEval, so when the step is finished, GradientEval is requested. evaluate() for this evalType puts math.NaN() into location.F unconditionally (original location.F is then restored back by Backtracking).
One option would be to test if xNext is different from location.X and put NaN into location.F only if it is. However, this implies extra work when Dim is large. Any other options?
The text was updated successfully, but these errors were encountered: