Skip to content

consistently treat machine-precision sized intervals in Learner1D

Joseph Weston requested to merge bugfix/termination into master

If an interval reaches machine precision in size, but it's loss does not reach tolerance, then the loss should be set to zero. The choice of minimal distance is set by machine precision and the largest floating point numbers that may occur in the domain of the learned function.

Closes #54 (closed).

Merge request reports