Test with ADNLPModels.jl

[jbcaillau]

hi @amontoison @PierreMartinon ; trying to test a hand-coded version of Goddard problem with ADNLPModels, and running into the error below (sparsity detection). both the cost and the inplace constraint evaluate properly, but when defining the model in

CTBenchmarks.jl/adnlp/goddard-ad.jl

Line 193 in 8f879ed

nlp = ADNLPModel!(f, z, lvar, uvar, con!, lcon, ucon)

the following error is encountered (apparently when trying to detect sparsisty):

any clue? the code (+ Project.toml) is here:
https://github.com/control-toolbox/CTBenchmarks.jl/blob/main/adnlp/goddard-ad.jl

[jbcaillau]

PS. @amontoison i think it is a good occasion to make your new github avatar shine 🙂

[amontoison]

@jbcaillau When you pass the functions to ADNLPModels.jl at the end of your code,
It should be the objective and the constraints.
I think you don't provide the constraints with this line:
https://github.com/control-toolbox/CTBenchmarks.jl/blob/main/adnlp/goddard-ad.jl#L193

I mean that your cons! seems to be a tuple:
https://github.com/control-toolbox/CTBenchmarks.jl/blob/main/adnlp/goddard-ad.jl#L141
I don't understand why you have nothing in the definition of the constraints.

[jbcaillau]

@amontoison thanks for the feedback. con! is inplace and returns nothing (not a tuple).

[amontoison]

If cons! is inplace, it should return the vector of constraints updated.

[amontoison]

@jbcaillau My function Lagrangian will not like that otherwise in ADNLPModels.jl:
https://github.com/JuliaSmoothOptimizers/ADNLPModels.jl/blob/main/src/sparsity_pattern.jl#L48
It will do a dot product between Nothing and Vector{Float64}.

You know what, I can call cons!(cx, x) on a different line such that this function returns anything.

[jbcaillau]

ah ok, didn't notice that (it is indeed the case in this doc example). yes, nicer not to assume that. waiting for JuliaSmoothOptimizers/ADNLPModels.jl#323 (you know i don't like to merge before tests are complete 🥲), thanks

[amontoison]

@jbcaillau

julia> include("goddard-ad.jl")
This is Ipopt version 3.14.17, running with linear solver MUMPS 5.7.3.

Number of nonzeros in equality constraint Jacobian...:     1904
Number of nonzeros in inequality constraint Jacobian.:        0
Number of nonzeros in Lagrangian Hessian.............:     1111

Total number of variables............................:      405
                     variables with only lower bounds:      102
                variables with lower and upper bounds:      202
                     variables with only upper bounds:        0
Total number of equality constraints.................:      304
Total number of inequality constraints...............:        0
        inequality constraints with only lower bounds:        0
   inequality constraints with lower and upper bounds:        0
        inequality constraints with only upper bounds:        0

iter    objective    inf_pr   inf_du lg(mu)  ||d||  lg(rg) alpha_du alpha_pr  ls
   0 -1.0100000e+00 9.00e-01 2.00e+00   0.0 0.00e+00    -  0.00e+00 0.00e+00   0
   1 -1.0090722e+00 8.99e-01 6.78e+01   1.3 1.67e+02    -  3.70e-03 5.93e-04f  1
   2 -1.0000907e+00 8.73e-01 1.93e+02   1.0 6.31e+00    -  3.37e-02 2.97e-02h  1
   3 -1.0023839e+00 8.36e-01 1.06e+04   1.0 6.92e+00    -  1.91e-01 4.27e-02f  1
   4 -1.0025122e+00 7.66e-01 8.42e+03   1.4 4.31e+00    -  1.00e+00 8.33e-02f  1
   5 -1.0033877e+00 7.13e-01 2.32e+05   2.3 3.74e+00    -  5.60e-01 6.86e-02f  1
   6 -1.0137050e+00 8.96e-03 9.94e+04   2.3 7.13e-01    -  3.84e-01 9.90e-01h  1
   7 -1.0099666e+00 3.16e-03 3.40e+05   1.8 4.32e-01    -  5.44e-01 9.90e-01h  1
   8 -1.0068998e+00 2.81e-04 1.55e+06   0.6 2.17e-01    -  7.02e-01 9.91e-01h  1
   9 -1.0068334e+00 4.51e-06 2.15e+07   0.1 1.12e-01    -  7.18e-01 1.00e+00h  1
iter    objective    inf_pr   inf_du lg(mu)  ||d||  lg(rg) alpha_du alpha_pr  ls
  10 -1.0068338e+00 1.72e-10 6.68e+05  -5.1 3.74e-04    -  9.89e-01 1.00e+00h  1
  11 -1.0068350e+00 4.32e-10 7.58e+03  -7.0 4.82e-04    -  9.89e-01 1.00e+00h  1
  12 -1.0079594e+00 6.12e-04 8.99e+03  -3.0 4.61e-01    -  6.49e-01 7.54e-01f  1
  13 -1.0081967e+00 3.11e-06 7.67e+03  -9.0 1.15e-02    -  8.76e-01 1.00e+00h  1
  14 -1.0092440e+00 1.72e-04 4.11e+02  -4.5 2.26e-01    -  1.00e+00 8.65e-01h  1
  15 -1.0106706e+00 2.19e-04 1.52e+02  -4.7 2.25e-01    -  1.00e+00 7.05e-01h  1
  16 -1.0118589e+00 9.30e-05 1.65e-03  -5.3 9.93e-02    -  1.00e+00 1.00e+00h  1
  17 -1.0123711e+00 5.50e-05 1.92e+01  -5.6 9.70e-02    -  1.00e+00 9.16e-01h  1
  18 -1.0125010e+00 1.98e-05 2.19e-04  -6.1 9.58e-02    -  1.00e+00 1.00e+00h  1
  19 -1.0125534e+00 1.02e-05 1.90e+00  -6.7 9.04e-02    -  1.00e+00 9.68e-01h  1
iter    objective    inf_pr   inf_du lg(mu)  ||d||  lg(rg) alpha_du alpha_pr  ls
  20 -1.0125690e+00 4.93e-06 1.19e-01  -7.5 9.66e-02    -  1.00e+00 9.94e-01h  1
  21 -1.0125708e+00 1.93e-06 4.96e-06  -8.0 1.46e-01    -  1.00e+00 1.00e+00h  1
  22 -1.0125715e+00 7.00e-07 2.30e-03  -8.8 1.09e-01    -  1.00e+00 9.97e-01h  1
  23 -1.0125716e+00 2.60e-08 3.61e-04  -9.4 1.92e-02    -  1.00e+00 9.97e-01h  1
  24 -1.0125716e+00 8.67e-11 4.72e-10 -11.0 1.43e-03    -  1.00e+00 1.00e+00h  1

Number of Iterations....: 24

                                   (scaled)                 (unscaled)
Objective...............:  -1.0125716188852516e+00   -1.0125716188852516e+00
Dual infeasibility......:   4.7236740752417433e-10    4.7236740752417433e-10
Constraint violation....:   1.8882118091312350e-11    8.6655016495740256e-11
Variable bound violation:   5.1043528986882691e-09    5.1043528986882691e-09
Complementarity.........:   1.2638315525660684e-11    1.2638315525660684e-11
Overall NLP error.......:   1.8882118091312350e-11    4.7236740752417433e-10


Number of objective function evaluations             = 25
Number of objective gradient evaluations             = 25
Number of equality constraint evaluations            = 25
Number of inequality constraint evaluations          = 0
Number of equality constraint Jacobian evaluations   = 25
Number of inequality constraint Jacobian evaluations = 0
Number of Lagrangian Hessian evaluations             = 24
Total seconds in IPOPT                               = 10.350

EXIT: Optimal Solution Found.
lcon    : true
ucon    : true
dynamics: 1.0859130379933743e-10
This is Ipopt version 3.14.17, running with linear solver MUMPS 5.7.3.

Number of nonzeros in equality constraint Jacobian...:     1904
Number of nonzeros in inequality constraint Jacobian.:      304
Number of nonzeros in Lagrangian Hessian.............:     1111

Total number of variables............................:      405
                     variables with only lower bounds:        0
                variables with lower and upper bounds:        0
                     variables with only upper bounds:        0
Total number of equality constraints.................:      304
Total number of inequality constraints...............:      304
        inequality constraints with only lower bounds:      102
   inequality constraints with lower and upper bounds:      202
        inequality constraints with only upper bounds:        0

iter    objective    inf_pr   inf_du lg(mu)  ||d||  lg(rg) alpha_du alpha_pr  ls
   0 -1.0125716e+00 8.67e-11 2.81e-02  -1.0 0.00e+00    -  0.00e+00 0.00e+00   0
   1 -1.0111610e+00 8.31e-05 1.02e+01  -1.7 1.12e-01    -  5.78e-01 9.90e-01f  1
   2 -1.0101393e+00 1.56e-04 9.07e+02  -1.7 1.76e-01    -  4.77e-01 9.30e-01f  1
   3 -1.0091937e+00 5.94e-05 2.35e+02  -1.7 1.02e-01    -  1.00e+00 9.91e-01f  1
   4 -1.0083592e+00 5.20e-05 4.81e-01  -1.7 3.74e-01    -  1.00e+00 1.00e+00f  1
   5 -1.0083388e+00 1.19e-08 1.87e-02  -3.8 1.42e-03    -  1.00e+00 1.00e+00h  1
   6 -1.0091035e+00 1.69e-05 5.29e-03  -3.8 8.04e-02    -  1.00e+00 1.00e+00h  1
   7 -1.0091568e+00 5.46e-08 6.42e-06  -3.8 8.26e-03    -  1.00e+00 1.00e+00h  1
   8 -1.0097679e+00 7.71e-06 1.20e+03  -5.7 4.81e-02    -  9.19e-01 1.00e+00h  1
   9 -1.0118339e+00 1.53e-04 1.17e+02  -5.7 1.53e-01    -  9.03e-01 8.03e-01h  1
iter    objective    inf_pr   inf_du lg(mu)  ||d||  lg(rg) alpha_du alpha_pr  ls
  10 -1.0123562e+00 4.80e-05 7.58e-02  -5.7 9.11e-02    -  1.00e+00 8.37e-01h  1
  11 -1.0124172e+00 1.16e-05 1.53e-04  -5.7 8.18e-02    -  1.00e+00 1.00e+00f  1
  12 -1.0124174e+00 2.43e-07 1.63e-06  -5.7 1.54e-02    -  1.00e+00 1.00e+00h  1
  13 -1.0125387e+00 6.16e-06 6.55e+00  -8.6 5.90e-02    -  8.31e-01 7.95e-01h  1
  14 -1.0125666e+00 5.76e-06 9.91e-02  -8.6 8.90e-02    -  8.09e-01 8.45e-01h  1
  15 -1.0125708e+00 2.33e-06 1.01e-01  -8.6 8.91e-02    -  8.58e-01 8.73e-01h  1
  16 -1.0125714e+00 1.56e-06 5.52e-02  -8.6 1.69e-01    -  9.15e-01 9.78e-01f  1
  17 -1.0125714e+00 1.26e-07 8.14e-08  -8.6 3.89e-02    -  1.00e+00 1.00e+00f  1
  18 -1.0125714e+00 1.89e-10 1.11e-10  -8.6 2.05e-03    -  1.00e+00 1.00e+00h  1

Number of Iterations....: 18

                                   (scaled)                 (unscaled)
Objective...............:  -1.0125714090188309e+00   -1.0125714090188309e+00
Dual infeasibility......:   1.1143212901759571e-10    1.1143212901759571e-10
Constraint violation....:   1.8903026296628088e-10    1.8903026296628088e-10
Variable bound violation:   0.0000000000000000e+00    0.0000000000000000e+00
Complementarity.........:   2.5126078866705649e-09    2.5126078866705649e-09
Overall NLP error.......:   2.5126078866705649e-09    2.5126078866705649e-09


Number of objective function evaluations             = 19
Number of objective gradient evaluations             = 19
Number of equality constraint evaluations            = 19
Number of inequality constraint evaluations          = 19
Number of equality constraint Jacobian evaluations   = 19
Number of inequality constraint Jacobian evaluations = 19
Number of Lagrangian Hessian evaluations             = 18
Total seconds in IPOPT                               = 6.471

EXIT: Optimal Solution Found.
"Execution stats: first-order stationary"

[amontoison]

I do a new release of ADNLPModels.jl this morning after JuliaSmoothOptimizers/ADNLPModels.jl#323.

Update: JuliaRegistries/General#124103

[jbcaillau]

👍🏽 @amontoison same result replacing

con!(c, z) = (build(c, z, N :: Int); nothing)

by

con!(c, z) = (build(c, z, N :: Int); c)

[jbcaillau]

nicer not to assume that the value is changed inplace and returned

[PierreMartinon]

Hi there. Yes @jbcaillau , if you check the CTDirect code in docp.jl you'll see the constraints vector is explicitely returned. I added a comment some time ago to remember why :D

[jbcaillau]

@PierreMartinon this code makes a comparison of OptimalControl.jl with a hand-coded ADNLPModels.jl Goddard:

• first solve with OC to find a solution that will be used as a common initial guess for both solves
• results indicate more allocations for OC, to be compared with the upcoming OC / CTDirect implementation
• run below for N = 200
• NB : regarding run times, the computation doesn't seem fair for ADNLP as no bounds on the unknowns are given (generic constraints are used, instead)