Adjoint (dynamics multipliers) parsing

[jbcaillau]

@PierreMartinon As discussed: probable shift in parse_DOCP_solution_dual when retrieving the multipliers associated with the dynamics (= adjoints). Bug seem to be activated for problems with nonlinear constraints (= not mere boxes on slices of state / control). See for instance this variant of Goddard that, contrary to this one, has a such a nonlinear constraint: control-toolbox/BVP-DAE#1 (comment)

Similar behaviour for these other projects that also have nonlinear (control or mixed) constraints:
@alesiagr https://github.com/mctao-inria/sail-therm
@PhysicDev https://github.com/PhysicDev/McTao_lunar_solar_sail

[jbcaillau]

@PierreMartinon in the current code below, compared to the previous tags, I don't see where you would skip the multipliers associated with constraints other than the dynamics (e.g. state constraints) in order to retrieve only the dynamics:

"""
$(TYPEDSIGNATURES)

Recover OCP costate from DOCP multipliers
"""
# +++todo: parse multipliers for the path and boundary constraints (use tuples)
function parse_DOCP_solution_dual(docp, multipliers)

    # constraints, costate and constraints multipliers
    N = docp.dim_NLP_steps
    P = zeros(N, docp.dim_NLP_x)
    
    if !isnothing(multipliers)
        index = 1
        for i in 1:N
            # state equation multiplier for costate
            P[i,:] = multipliers[index:index+docp.dim_NLP_x-1]
            index = index + docp.dim_NLP_x
        end
    end

    return P
end

[PierreMartinon]

@jbcaillau
Should be fixed in 0.9.6, can you try updating ?

[alesiagr]

J'arrive pas à update la version 0.9.6, j'ai tjrs la 0.9.4. Sinon je pourrais tester de suite

[jbcaillau]

👍🏽 @PierreMartinon vu pour la ligne ci-dessous, je teste

CTDirect.jl/src/solution.jl

Line 46 in d2321e2

index = index + dim_path_constraints(docp.ocp)

[jbcaillau]

@PierreMartinon CTDirect 0.9.6 now available. Latency? Anyway, seems much better 🙂 @ocots @alesiagr

[alesiagr]

BINGO! @jbcaillau @PierreMartinon @ocots Merci 🙏🏼
Qu'ils sont beaux ces adjoints quand ils oscillent pas 🥹

[PierreMartinon]

Il faudra que je complete les tests pour le parsing de solution, verifier qu'on recupere bien ce qu'il faut. Actuellement on ne teste que l'objectif.

[ocots]

Top si on peut corriger ce bug. Il faudra relancer les simulations avec Paul.

[ocots]

Je ne sais pas si ça peut aider mais il y des choses sur l'adjoint dans le cas de contraintes d'états dans ce papier : https://inria.hal.science/hal-01367918

Il y a cette proposition page 15

et cette figure juste après

[alesiagr]

Je sais pas si ça s'adressait à moi ton message mais merci @ocots :D

[ocots]

Plutôt à @PierreMartinon et @jbcaillau mais tu peux regarder 😀

[jbcaillau]

Top si on peut corriger ce bug. Il faudra relancer les simulations avec Paul.

@PierreMartinon has corrected. Leaving the issue open so that we remember to add the relevant unitary test in CTDirect.

Check control-toolbox/BVP-DAE#1 (comment)

[jbcaillau]

@PierreMartinon still missing unit tests before closing, right?

[PierreMartinon]

@PierreMartinon still missing unit tests before closing, right?

Yes. Ideally we would test the closeness of the interpolated trajectory and costate to a known functional solution (evaluate both on a given time grid then use isapprox) . Does anyone have a basic problem with simple expressions for x,u,p (preferably non-constant) ?

[jbcaillau]

@PierreMartinon is our basic example OK? (One of the adjoint is not constant.)

[ocots]

@PierreMartinon is our basic example OK? (One of the adjoint is not constant.)

There is a problem here : https://control-toolbox.org/OptimalControl.jl/stable/tutorial-double-integrator.html

[PierreMartinon]

@PierreMartinon is our basic example OK? (One of the adjoint is not constant.)

There is a problem here : control-toolbox.org/OptimalControl.jl/stable/tutorial-double-integrator.html

Ok, I'll take the double integrator !
@ocots, the second one does not converge on the page :D

[ocots]

@PierreMartinon is our basic example OK? (One of the adjoint is not constant.)

There is a problem here : control-toolbox.org/OptimalControl.jl/stable/tutorial-double-integrator.html

Ok, I'll take the double integrator ! @ocots, the second one does not converge on the page :D

Do you understand why?

[PierreMartinon]

I'll have a look as soon as I can get CTBase 0.12 (there is usually some latency)

[ocots]

I'll have a look as soon as I can get CTBase 0.12 (there is usually some latency)

It is ready. You can get it whenever you want.

[PierreMartinon]

When retesting locally the problem converges.

Let's try to rebuild the doc and check ?

[PierreMartinon]

On the other hand, tests for state, control and costate added and OK for the double integrator

[ocots]

It converged with 100 points bu not with 200. Here how many do you have?

[PierreMartinon]

New default in CTDirect is 250, so 250 probably

[jbcaillau]

New default in CTDirect is 250, so 250 probably

@PierreMartinon 👍🏽

NB. probably safer to use double integrator with $L^2$ cost (no issue of convergence, whatever the grid size)