SNOPT FD Gradient with Jacobian Sparsity

[niwhsa9]

SNOPT is capable of estimating sparse gradients via finite difference given the sparsity structure.

I am curious if this functionality is actually exposed by PyOptSprase. When setting sens=None to use SNOPT's internal finite differencing, are the wrt keywords in the constraint groups actually passed through to SNOPT?

The language in the documentation would suggest not
Note that the wrt and jac keyword arguments are only supported when user-supplied sensitivity is used. If automatic gradients from pyOptSparse are used, the constraint Jacobian will necessarily be dense.

[ewu63]

I took a look, and the sparsity information should be available to SNOPT so in theory this should work. But I don't know in practice, maybe you can test it on a simple problem and see? Do keep us posted!

Note that in the text you provided, "automatic gradients from pyOptSparse" refers to the gradients computed by pyOptSparse (e.g. sens="FD", and not the gradients computed by the optimizer.

[niwhsa9]

sounds good, thanks. I will test it out and report back.