The NLopt package provides both gradient-based (where you have to supply the analytical gradient) and derivative-free (where you only supply the objective function) optimizers.
It is not really practical to do optimization of "very" high dimensional problems without knowing the gradient analytically. (Normally you shouldn't need to provide a Hessian, however.) But I think of "very" as being 1000s of dimensions; if you only have tens of dimensions, that is fine for derivative-free optimizers. On Thursday, January 16, 2014 7:03:27 PM UTC-5, jbeginner wrote: > I am trying to use Julia's Ipopt interface for an optimization problem. I > have two questions. Firstly, is it possible to only provide the objective > function and starting values and not bother about the gradient, hessian, > etc, or alternatively would providing the objective function and gradient > suffice? I know that this greatly reduces the performance of the solver but > is it possible? My function is very high dimensional and it would be very > cumbersome to compute those manually. >
