Implement support for second order numerical derivatives (forward over adjoint)

[casadibot]

Created by jaeandersson at: 2010-10-22T18:55:07
Last updated at: 2010-11-26T16:26:23

---

Comment 1 by jaeandersson at 2010-11-26T16:26:23

This is not really necessary. Instead the same can be achieved by a combination of symbolic AD (to generate an analytic function for the first order derivatives) followed by a numerical AD (if a directional derivative is enough) or symbolic AD (if the whole hessian, with sparsity, is necessary).

Most things already exist for this and it works already today, (see CSTR example).

To make this work more smoothly, and generally:

1. Implement AD by SCT for the MX class, couple this to the "jacobian"-function in the MXFunction class (involved!) Only AD forward at a first stage.
2. Write a Jacobian function for the integrator class(es). This function should create a new integrator for the forward sensitivity equations (easy)