percival

[PierreMartinon]

Percival seems to converge properly, although with much larger allocations and time.
For instance, on the double integrator (min energy) with only 25 steps

docp, nlp = direct_transcription(prob.ocp, grid_size=25)
println("ipopt")
@btime ipopt(nlp, print_level=0)
println("madnlp")
@btime madnlp(nlp, print_level=MadNLP.ERROR)
println("percival")
# @btime dsol = percival(nlp)
# same allocs/time as calling percival(nlp)
solver = PercivalSolver(nlp)
@btime solve!(solver, nlp) 

we obtain

ipopt
  1.744 ms (8308 allocations: 402.90 KiB)
madnlp
  939.964 μs (9379 allocations: 895.18 KiB)
percival
  2.552 s (24910761 allocations: 1.55 GiB)

With the standard 250 steps Percival currently fails with a timeout.
We should investigate why there is such a difference.

[github-actions]

Breakage test results
Date: 2025-03-14 16:40:10

Package name	latest	stable
OptimalControl.jl	compat: v0.13.3	compat: v0.13.3