repeat calls to gradient

[Nick Larson]

Is it possible to speed up problem.gradient() if I want to test a list of perturbations to a parameter? Right now I am calling .gradient() for every perturbation and it takes a long time.

What I am hoping is I can get the gradient of the variable with respect to each parameter as a matrix with shape of num(variables) by num(parameters). Is this possible? This way for any new vector of deltas I can compute the expected values as a first order approximation without having to reapply differentiation.

Thanks for the help!

Nick

[Steven Diamond]

In theory it should be possible to do something here, but it will be a while before any developer had time to look into this.