[Bug] Max objective for exa

[jbcaillau]

@PierreMartinon still some sign issue on the objective value retrieved from the solver (check the run from CTBenmarks below):

│  │  N : 200
│  │  ✓ | JuMP    :   3.455 s  | obj: 1.248809e+03  | iters: 321    | CPU:  72.66 MiB
│  │  ✓ | adnlp   :   1.501 s  | obj: -1.248809e+03 | iters: 263    | CPU: 286.90 MiB
│  │  ✓ | exa     :   1.011 s  | obj: -1.248809e+03 | iters: 320    | CPU:  33.51 MiB
│  │
│  │  N : 500
│  │  ✓ | JuMP    :  22.556 s  | obj: 1.247978e+03  | iters: 690    | CPU: 254.22 MiB
│  │  ✓ | adnlp   :  10.313 s  | obj: -1.247978e+03 | iters: 679    | CPU:   1.75 GiB
│  │  ✓ | exa     :   2.690 s  | obj: -1.247978e+03 | iters: 365    | CPU:  91.39 MiB
│  │
│  │  N : 1000
│  │  ✓ | JuMP    :  48.912 s  | obj: 1.247985e+03  | iters: 726    | CPU: 516.47 MiB
│  │  ✓ | adnlp   :  24.904 s  | obj: -1.247985e+03 | iters: 707    | CPU:   3.63 GiB
│  │  ✓ | exa     :  10.953 s  | obj: -1.247985e+03 | iters: 713    | CPU: 310.20 MiB
│  └─
└─

┌─ problem: insurance
│
├──┬ solver: ipopt, disc_method: trapeze
│  │
│  │  N : 200
│  │  ✓ | JuMP    : 220.169 ms | obj: 2.058581e+00  | iters: 54     | CPU: 103.47 KiB
│  │  ✓ | adnlp   : 301.016 ms | obj: 2.058581e+00  | iters: 54     | CPU:  71.60 MiB
│  │  ✓ | exa     : 179.792 ms | obj: -2.058581e+00 | iters: 54     | CPU: 220.24 KiB
│  │
│  │  N : 500
│  │  ✓ | JuMP    : 984.310 ms | obj: 2.059214e+00  | iters: 101    | CPU: 190.86 KiB
│  │  ✓ | adnlp   :   1.300 s  | obj: 2.059214e+00  | iters: 93     | CPU: 302.46 MiB
│  │  ✓ | exa     : 866.498 ms | obj: -2.059214e+00 | iters: 106    | CPU: 454.43 KiB
│  │
│  │  N : 1000
│  │  ✓ | JuMP    :   2.772 s  | obj: 2.058960e+00  | iters: 111    | CPU: 212.19 KiB
│  │  ✓ | adnlp   :   4.137 s  | obj: 2.059312e+00  | iters: 128    | CPU: 831.47 MiB
│  │  ✓ | exa     :   2.545 s  | obj: -2.059511e+00 | iters: 132    | CPU: 783.35 KiB

[ocots]

Actually the problem is not from the solve of OptimalControl.jl since we do not use it to solve in the bench. But it can come from the direct transcription or the modeller.

Problem in the bench with:

• ipopt + exa
• madnlp + (exa or adnlp)

Remark: no problem with madnlp + jump.

[ocots]

Actually madnlp returns f and not -f for a max problem: max f(x).

[jbcaillau]

@ocots @PierreMartinon looks like for both modellers adnlp and exa it is possible to do as in JuMP and indicate min or max directly inside the model. no need then to change anything (cost to -cost) before calling the solver.

@ocots do you confirm for adnlp?

see also: control-toolbox/CTParser.jl#167

[ocots]

I confirm for adnlp.

[ocots]

To update:

• pass minimize = (p.type == :min) when creating ExaCore. Need to set p.time. Then, no need to change the sign of the cost neither in Mayer nor Lagrange. Need to add some tests to check that it works: see get_minimize(nlp).
• Remove this part in CTDirect since then max problems will be handled. Here, sign will be correct. The problem was not from MadNLP but ExaModel usage. Need to add some tests to check that after a resolution of a max problem we have what we want for the 4 combinations: (ipopt / madnlp) + (adnlp / exa).
• No need to change the sign of the objective here in OptimalControlProblems. Neither here. Retest.
• Rerun the benchmarks.

-> replaced by https://github.com/orgs/control-toolbox/discussions/68#discussioncomment-14721473.