[General] Objective value

[ocots]

@jbcaillau @PierreMartinon

In the case

$$\min f(x)$$

we return as objective value $f(x^*)$. In the case

$$\max -f(x)$$

we return also $f(x^*)$. Is it what we want?

CTDirect.jl/src/solution.jl

Lines 35 to 40 in 21c4dff

	# objective from solution
	if docp.flags.max
	objective = -docp_solution.objective
	else
	objective = docp_solution.objective
	end

[PierreMartinon]

Update as of 0.17.0

• we no longer transform max f into min -f, we use the option maximize in ADNLPModels instead
• we now recover the objective value from the NLP solution instead of recomputing it
• interestingly, Ipopt and MadNLP behave differently for maximization problem, see for instance the iterations for the Goddard problem (max altitude). While Ipopt displays and return the original maximized objective, MadNLP uses the minimized opposite. For MadNLP we manually adjust the objective sign if needed when building the OCP solution, but if you test the objective value directly from the NLP solution, you will get -f instead of f.

[jbcaillau]

@PierreMartinon regarding min / max:

• during parsing, the info is kept (= no transformation is made) https://github.com/control-toolbox/CTParser.jl/blob/489b1b65803969bb9210ae5f96aefdd3343db35c/src/onepass.jl#L185
• a good solution (i think @ocots did that in OptimalControlProblems.jl) is to formulate everything with min at @def level to avoid any mismatch depending on the solver
• later, it will be nice indeed to uniformise that (as for the other a posteriori treatments); issue somewhere?

[ocots]

When I had a bug I replaced every max by min at the@def level but some problem are truly max so I have replaced some min by max.

[jbcaillau]

When I had a bug I replaced every max by min at the@def level but some problem are truly max so I have replaced some min by max.

ah... any good reason not to have min everywhere to avoid any min / max sign issue?

[ocots]

It shouldn't be an issue :-)

[jbcaillau]

yes but it is! see above

[PierreMartinon]

I believe it is bad form to force the user to rewrite the OCP with a min.
This is better handled at the NLP modeler level.
Note that the problem Olivier encountered should be fixed now in 0.17

[jbcaillau]

agree:

• to be dealt with at modeler + nlp level to abstract this for the user
• my point was just for the benchmark (write min everywhere to keep it simple right now)

[PierreMartinon]

iirc the next update for MadNLP should fix this

[jbcaillau]

yes, expected for v0.9 MadNLP/MadNLP.jl#518

iirc the next update for MadNLP should fix this

[PierreMartinon]

bump @ocots @jbcaillau

It would seem that we still have the sign problem when maximizing with MadNLPGPU, meaning we change it whereas we should not since the sign bug only affects MadNLP and not MadNLPGPU ?

If it's problematic to distinguish the GPU case, maybe just remove the sign correction alltogether and accept that the sign for MadNLP (CPU) will be fixed by their next release ?

[ocots]

We can't maximize on GPU. It consider the problem as a min problem even if you set minimize to false.

[ocots]

Il will be fixed in the next release of MadNLP.