Make function decrease check based on a tolerance

[btracey]

No description provided.

[btracey]

This is the code that needs to change, but I'm not sure about the comment. The current wording seems overly complicated. I'm not sure backtracking should have it. The whole point in backtracking is to find a sufficient function decrease.

[vladimir-ch]

This is not true, we compare to the value on the left side of the interval (Is it even correct? The original code compared to the lowest value found).

There is not a suitable line to put the comment on, but b.initF field and minF variable are not used in this version, is it on purpose?

[btracey]

Nope, should be minF.

[vladimir-ch]

The test

F: VariablyDimensioned
Dim: 10000
GradientThreshold: 1e-05

fails at random with LBFGS as the expansion factor is varied. It fails with ErrNoProgress on the first linesearcher step, which means that step 1 is accepted but x + grad == x. Is it fixable or is 1e-5 too stringent for this functions?

[vladimir-ch]

I would replace Wolfe with sufficient decrease or Armijo in these two lines.

[vladimir-ch]

Closing because this repository is no longer maintained. Development has moved to https://github.com/gonum/gonum.

@btracey , should we port this to the new repository or has it been done in the meantime?