Separate function for direct transcription (ocp to nlp)

[PierreMartinon]

TODO:

• add an export of the NLP problem alone i.e the member docp.nlp (note that we then lose the link to the original OCP, in terms of initial guess and solution)

docp = DirectTranscription(ocp1, grid_size=100, init=init_constant)
nlp = getNLP(docp)
print(docp.nlp.meta.x0)

• add possibility to pass initial guess to solveDOCP (currently passed to DirectTrancription since initial guess is part of NLPModel). Using for instance a previous solution to generate the initial guess

init_sol = OptimalControlInit(sol)

one can then set the initial guess in the DOCP

setDOCPInit(docp, init_sol)
sol = solveDOCP(docp, print_level=5, tol=1e-12)

or pass it to the solver call (will bypass the initial guess stored in docp)

sol = solveDOCP(docp, init=init_sol, print_level=5, tol=1e-12)

Note: some indicators currently defined in DOCP could be migrated up to the OCP, such as has_free_final_time, has_variable etc, similarly to the existing is_min(). Using these higher level functions is probably cleaner than looking at the internal members of OCP.

[codecov]

Codecov Report

Attention: Patch coverage is 98.78788% with 4 lines in your changes are missing coverage. Please review.

Project coverage is 97.85%. Comparing base (e63a595) to head (15c12e8).
Report is 2 commits behind head on main.

Files	Patch %	Lines
src/solve.jl	82.60%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #73      +/-   ##
==========================================
- Coverage   99.15%   97.85%   -1.30%     
==========================================
  Files           6        6              
  Lines         471      466       -5     
==========================================
- Hits          467      456      -11     
- Misses          4       10       +6     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[jbcaillau]

@PierreMartinon @tmigot linked to this issue

[tmigot]

@PierreMartinon I am just reading the comments, so I am not sure what was blocking, but it is possible to broadcast an initial guess to ADNLPModel to change the initial guess, i.e., something like

nlp = ADNLPModel(...)
nlp.meta.x0 .= ones(n)

or the alternative is to pass an initial guess directly to the solver (it takes nlp.meta.x0 only by default).

[PierreMartinon]

Thanks @tmigot for the clarification ! I guess people usually prefer to pass the initial guess to the solver directly. Both options added.

[jbcaillau]

thanks @tmigot ; @PierreMartinon please send a message when PR is ready to merge 🤞🏾

[PierreMartinon]

Ok for me, @joseph-gergaud do you have more changes before the merge ?