Hi Yulia,
I responded to this on the Ampl disussion group this mornning.
Can you tell me where on the GAMS web-page you found this
information? This information is not correct.
KNITRO (and all NLP solvers I am aware of) assume/require that all
functions are smooth, continuously differentiable functions. KNITRO
may sometimes get lucky and work on models that have discontinuous
derivatives, but there are absolutely no guarantees that it will work
on these types of models and oftentimes it will not be able to find
the solution (since it needs the derivatives). Oftentimes, non-smooth
functions (e.g. min, max, abs) can be reformulated to make the model
smooth by adding extra variables and constraints. I don't know of a
good way to handle the "round()" function though.
Regards
-Richard Waltz