[Cron] Robustness of the initial guess and known local objectives #378

Activity

1. ocots

   added a parent issue on Nov 5, 2026

   • [Metadata] Computed features, reference objective, shared initial guess #306
2. github-actions

   assigned

   ocots

   on Nov 5, 2026
3. ocots commented on Nov 5, 2026

   

   ocots

   on Nov 5, 2026

   MemberAuthor

   More actions

   Input from #375 (PR #380): the first computation of the reference objectives found three disagreements between configurations, all split by solver, not by modeling or platform:

   • mountain_car: Ipopt (JuMP, OptimalControl ADNLP) 102.39757, MadNLP (OptimalControl and scalarized, Exa) 102.38100: two local solutions. This is the 1.6e-4 gap that once made the scalarized model :draft; it does not come from the scalarization.
   • jackson, rocket: MadNLP stops 1.5e-6 and 1.7e-6 above Ipopt, identically on every platform; probably the stopping test of MadNLP at tol = 1e-8 rather than another solution; to check.

   The per-solve objectives are in ext/MetaData/references/<problem>.toml (results).
4. ocots commented on Nov 6, 2026

   

   ocots

   on Nov 6, 2026

   MemberAuthor

   More actions

   Taken by an agent; a plan follows once #376 is in review.
5. ocots

   added a commit that references this issue on Nov 6, 2026

   Merge pull request #393 from control-toolbox/feat/378-local-solutions

   ...

   8b2ef8c
6. ocots

   closed this as completedin #393on Nov 6, 2026
7. ocots commented on Nov 6, 2026

   

   ocots

   on Nov 6, 2026

   MemberAuthor

   More actions

   Done in #393 (merged).

   • Known solutions of each problem in the references files (solutions), grouped within OptimalControlProblems.SAME_SOLUTION = 1e-5; solution_kind(p, objective) says :best, :local or :new; a solution column in every report; new solution flagged in cron-convergence.
   • Robustness report: test/robustness/<env>/, mode robustness, monthly cron CronRobustness.yml; Ipopt and MadNLP option sets (MadNLP's barrier passed with bypass, see MadNLP strategy: mu_init and mu_min are accepted but ignored by MadNLP 0.10; barrier is not declared CTSolvers.jl#321).
   • First robustness table (local): feat: known solutions and robustness report (#378) #393 (comment). bioreactor's reference is not the best known solution (up to 3.3e-4 better under other options); brachistochrone with LOQO reaches a worse solution.

   Next: the robustness and weekly crons are running on main (workflow_dispatch). Once they finish, the references are computed again from their artifacts (tools/references/README.md) in a pull request; bioreactor's reference then becomes the best known solution.

   🤖 Generated with Claude Code