Wrong results for sigmoid function

[dendibakh]

First of all, thank you for the incredible tool!
Learned about it on HackerNews.
I tried to find a substitute for the sigmoid function:

(FPCore (x)
  :name "1/(1+pow(E,-x))"
  :pre (and (<= -1e9 x 1e9))
  (/ 1 (+ 1 (pow E (- x)))))

The tool suggests 0.5+x*0.25, saying it has 21.4x speedup and is 96.9% accurate, which I think is wrong. It is a bad approximation of sigmoid function.

​

[pavpanchekha]

Right... what's going on is that this is a good approximation, near 0, but Herbie isn't adding a branch. I'm closing this as a duplicate of #742 but please know that we're thinking about the issue. (In the next release, we hope to modify the UI to make the issue easier to avoid, and in the future we plan to explicitly track approximation steps like this one so we can correctly detect when we're using them outside the relevant range.