Re: Differentiable Forth

On Wed, 17 Jul 2024 2:59:10 +0000, minforth wrote:

There is a whole website dedicated to AD, tools and applications:
https://www.autodiff.org/

Useful!

Projected to Forth, an autodiff wordset could be thought of similar
to a complex number wordset imaging the standard fp-number wordset.

That is just what I hoped to find. The Julia blurb did not mention
complex number similarities (probably I should've dug deeper).

However, care must be taken with non-differentiable regions in
e.g. fabs, fsqrt and some (inverse) trigonometric functions.
(AD libraries in C++ use operator overloading)

Exactly the problem I face when applying state-space modeling to
power electronics combined with digital control. It can be
handled relatively easily because everything already revolves
around knowing when such discontinuities happen.
-marcel