"Oxberry, Geoffrey Malcolm" <oxber...@llnl.gov> writes: > I realize this point was brought up earlier, but doesn’t this > discussion still assume that evaluating g at zero is defined and makes > sense? Though I see the appeal in this design, I’m not sure it will > necessarily work in practice. For instance, we definitely have > formulations that involve terms like sqrt(x), which isn’t > differentiable at zero, and would break this interface. (We would > bound the feasible set away from zero, so the optimization algorithm > should still work.)
Isn't the context of this particular comment that the user claims their model is quadratic? Of course evaluating at 0 has no special meaning for a general nonlinear model. >> On Aug 18, 2016, at 11:02 AM, Munson, Todd <tmun...@mcs.anl.gov> wrote: >> >>> >>> People are free to use MatShell to create a "matrix" that is >>> actually a nonlinear operator. Solvers won't work properly if it's >>> not, but that's their problem. >> >> The quadratic programming solvers in our case will happily go and >> tell you it solved the problem...however, the problem it solved uses >> >> c = grad[g(0)] H = hess[g(0)] >> >> I'd be happier if the solver barfed and told the user to select a >> method appropriate for their real problem -- it it truly is nonlinear >> -- rather than going off and solving the wrong problem, which is what >> is done today. >> >> Todd. >>
signature.asc
Description: PGP signature
