Warm start support

[francesco-romano]

I am trying to debug the warm start support.

I attach a sample project where enabling the warm start does have some effect, passing from 8 to 1 iterations.

Important points learnt:

• get_nlp_info get called at every call to optimise. So that function is NOT the right place where we should initialise buffers (e.g. the one containing multipliers)
• there are "different" levels of warm start:
  • providing as guess the previous solution (for primal variables) already helps. In the example, it halves the number of iterations (from 8 to 4).
  • enabling warm_start_init_point makes IPOpt to ask not only for the primal variables but also for the dual variables during the get_starting_point call. Those variables can be retrieved during finalize_solution. Still, the initial guess can be modified because of how IP methods work
  • reducing how much IPOpt modifies the initial solution greatly improve the optimization: setting the various warm_start_bound_frac to 1e-6 finds the solution after 1 iteration.
  • option warm_start_entire_iterate seems to be mutually exclusive with warm_start_init_point, but I still have to understand this part.

I am trying now to "port" these discoveries to the current MPC code

hs071.zip

[francesco-romano]

Setting Options()->SetStringValue("mu_strategy", "monotone"); and

Options()->SetNumericValue("warm_start_bound_frac", 1e-10);
Options()->SetNumericValue("warm_start_bound_push", 1e-10);
Options()->SetNumericValue("warm_start_mult_bound_push", 1e-10);
Options()->SetNumericValue("warm_start_slack_bound_frac", 1e-10);
Options()->SetNumericValue("warm_start_slack_bound_push", 1e-10);

I managed to get

iter    objective    inf_pr   inf_du lg(mu)  ||d||  lg(rg) alpha_du alpha_pr  ls
   0  1.7014017e+01 3.26e-09 4.98e-09  -1.0 0.00e+00    -  0.00e+00 0.00e+00   0

Number of Iterations....: 0