Warm start initial guess

User Questions
1

Hi,

I am using the python interface with acados and would like to know if my process of warm starting the initial guess with a naive trajectory is appropriate.

With the naive trajectory which I have computed, I use ocp_solver.set_flat('x', inital_guess) before the first solve is called. I do not have a guess of the inputs and hence have left that out. After doing this the controller is run in closed loop with a simulator. Both with and without, the warm start, the trajectory is very similar. I was just wondering if I have

  1. Warm started in the correct way?
  2. Do I need to warm up the computational graphs with some initial solves first?
  3. or is there some flags/settings which I am missing.

Thank you
Nico

2

Hi Nico,

the way you initialize the state trajectory is correct.
If you know the initial parametrization of the problem or something close, you can of course try to get closer by doing solves before starting the actual simulation.
You could monitor the residuals over the closed loop simulation to see how close they are, instead of just looking at the closed-loop trajectory.

Best,
Jonathan