get gradient value out of IPOPT solver

[RFTV]

Dear,

I'm new to CasADi but I think it is great! I made a NMPC controller using IPOPT interface in CasADi (MATLAB) and it works great! Only now I was wondering if I could get the gradient value out of the IPOPT solver. The standard output of the solver is:

- sol.f  --> objective function value

- sol.g (gradient???- or constraint value?)

- sol.lam_g (??, maybe Lagrange multipliers?)

- sol.lam_p (??, maybe Lagrange multipliers?)

- sol.lam_x (??, maybe Lagrange multipliers?)

- sol.x --> output of decision variables

Could somebody maybe clear this up please? I could not find it in the manual. 

Also, if sol.g is not the gradient, is there a way to extract this from IPOPT?

Thanks in advance,

RFTV

[Bruno M L]

Hi!

I believe sol.g is the value of the nonlinear constraints.

[Joris Gillis]

Hi there!

This is documented in `help(nlpsol)` or https://web.casadi.org/python-api/#nlp

NLPSOL_X	x	Decision variables at the optimal solution (nx x 1)
NLPSOL_F	f	Cost function value at the optimal solution (1 x 1)
NLPSOL_G	g	Constraints function at the optimal solution (ng x 1)
NLPSOL_LAM_X	lam_x	Lagrange multipliers for bounds on X at the solution (nx x 1)
NLPSOL_LAM_G	lam_g	Lagrange multipliers for bounds on G at the solution (ng x 1)
NLPSOL_LAM_P	lam_p	Lagrange multipliers for bounds on P at the solution (np x 1)

Ipopt requires several callbacks for numerical evaluation of objective, gradient of objective, constraints etc.

You can introspect these functions with sol.get_function.

To get a gradient at the solution you would do:

nlp_grad_f = sol.get_function('nlp_grad_f');

[~,grad] = nlp_grad_f(sol.x, p) % p = fixed parameters of the problem; might be empty ( [] ) in your case

Best regards,

   Joris

[RFTV]

Okay, thanks a lot!

I found the help information, but it was not clear to me how to actually acquire it. 

"nlp_grad_f = sol.get_function('nlp_grad_f');

[~,grad] = nlp_grad_f(sol.x, p) % p = fixed parameters of the problem; might be empty ( [] ) in your case"

Okay, thanks! No, p is used a reference state/control and also as estimated states of different subsystems. So I guess I should use that too. 

Thanks a lot for the quick replies and easy solution.

This forum is great!

Best,

RFTV