Merge pull request #792 from alonsoC1s/wrap_nlpmodel

Add a conversion mechanism between NLPModel to OptimizationProblem

.github/workflows/CI.yml

@@ -36,6 +36,7 @@ jobs:
           - OptimizationQuadDIRECT
           - OptimizationSpeedMapping
           - OptimizationPolyalgorithms
+          - OptimizationNLPModels
         version:
           - '1'
     steps:

docs/Project.toml

@@ -12,6 +12,8 @@ Juniper = "2ddba703-00a4-53a7-87a5-e8b9971dde84"
 Manifolds = "1cead3c2-87b3-11e9-0ccd-23c62b72b94e"
 Manopt = "0fc0a36d-df90-57f3-8f93-d78a9fc72bb5"
 ModelingToolkit = "961ee093-0014-501f-94e3-6117800e7a78"
+NLPModels = "a4795742-8479-5a88-8948-cc11e1c8c1a6"
+NLPModelsTest = "7998695d-6960-4d3a-85c4-e1bceb8cd856"
 NLopt = "76087f3c-5699-56af-9a33-bf431cd00edd"
 Optimization = "7f7a1694-90dd-40f0-9382-eb1efda571ba"
 OptimizationBBO = "3e6eede4-6085-4f62-9a71-46d9bc1eb92b"
@@ -25,6 +27,7 @@ OptimizationManopt = "e57b7fff-7ee7-4550-b4f0-90e9476e9fb6"
 OptimizationMetaheuristics = "3aafef2f-86ae-4776-b337-85a36adf0b55"
 OptimizationMultistartOptimization = "e4316d97-8bbb-4fd3-a7d8-3851d2a72823"
 OptimizationNLopt = "4e6fcdb7-1186-4e1f-a706-475e75c168bb"
+OptimizationNLPModels = "064b21be-54cf-11ef-1646-cdfee32b588f"
 OptimizationNOMAD = "2cab0595-8222-4775-b714-9828e6a9e01b"
 OptimizationOptimJL = "36348300-93cb-4f02-beb5-3c3902f8871e"
 OptimizationOptimisers = "42dfb2eb-d2b4-4451-abcd-913932933ac1"
@@ -66,6 +69,7 @@ OptimizationManopt = "0.0.2, 0.0.3"
 OptimizationMetaheuristics = "0.1, 0.2"
 OptimizationMultistartOptimization = "0.1, 0.2"
 OptimizationNLopt = "0.1, 0.2"
+OptimizationNLPModels = "0.0.1"
 OptimizationNOMAD = "0.1, 0.2"
 OptimizationOptimJL = "0.1, 0.2, 0.3"
 OptimizationOptimisers = "0.1, 0.2"

docs/pages.jl

@@ -36,6 +36,7 @@ pages = ["index.md",
         "PRIMA.jl" => "optimization_packages/prima.md",
         "Polyalgorithms.jl" => "optimization_packages/polyopt.md",
         "QuadDIRECT.jl" => "optimization_packages/quaddirect.md",
-        "SpeedMapping.jl" => "optimization_packages/speedmapping.md"
+        "SpeedMapping.jl" => "optimization_packages/speedmapping.md",
+        "NLPModels.jl" => "optimization_packages/nlpmodels.md"
     ]
 ]

docs/src/optimization_packages/nlpmodels.md

@@ -0,0 +1,54 @@
+# NLPModels.jl
+
+[NLPModels](https://jso.dev/NLPModels.jl/latest/), similarly to Optimization.jl itself,
+provides a standardized modeling interface for representing Non-Linear Programs that
+facilitates using different solvers on the same problem. The Optimization.jl extension of
+NLPModels aims to provide a thin translation layer to make `NLPModel`s, the main export of
+the package, compatible with the optimizers in the Optimization.jl ecosystem.
+
+## Installation: NLPModels.jl
+
+To translate an `NLPModel`, install the OptimizationNLPModels package:
+
+```julia
+import Pkg;
+Pkg.add("OptimizationNLPModels")
+```
+
+The package NLPModels.jl itself contains no optimizers or models. Several packages
+provide optimization problem ([CUTEst.jl](https://jso.dev/CUTEst.jl/stable/),
+[NLPModelsTest.jl](https://jso.dev/NLPModelsTest.jl/dev/)) which can then be solved with
+any optimizer supported by Optimization.jl
+
+## Usage
+
+For example, solving a problem defined in `NLPModelsTest` with
+[`Ipopt.jl`](https://github.com/jump-dev/Ipopt.jl). First, install the packages like so:
+
+```julia
+import Pkg;
+Pkg.add("NLPModelsTest", "Ipopt")
+```
+
+We instantiate [problem
+10](https://jso.dev/NLPModelsTest.jl/dev/reference/#NLPModelsTest.HS10) in the
+Hock--Schittkowski optimization suite available from `NLPModelsTest` as `HS10`, then
+translate it to an `OptimizationProblem`.
+
+```@example NLPModels
+using OptimizationNLPModels, Optimization, NLPModelsTest, Ipopt
+using Optimization: OptimizationProblem
+nlpmodel = NLPModelsTest.HS10()
+prob = OptimizationProblem(nlpmodel, AutoForwardDiff())
+```
+
+which can now be solved like any other `OptimizationProblem`:
+
+```@example NLPModels
+sol = solve(prob, Ipopt.Optimizer())
+```
+
+## API
+
+Problems represented as `NLPModel`s can be used to create [`OptimizationProblem`](@ref)s and
+[`OptimizationFunction`](@ref).

lib/OptimizationNLPModels/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Vaibhav Dixit <[email protected]> and contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

lib/OptimizationNLPModels/Project.toml

@@ -0,0 +1,21 @@
+name = "OptimizationNLPModels"
+uuid = "064b21be-54cf-11ef-1646-cdfee32b588f"
+authors = ["Vaibhav Dixit <[email protected]> and contributors"]
+version = "0.0.1"
+
+[deps]
+ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"
+NLPModels = "a4795742-8479-5a88-8948-cc11e1c8c1a6"
+Optimization = "7f7a1694-90dd-40f0-9382-eb1efda571ba"
+Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
+
+[extras]
+NLPModelsTest = "7998695d-6960-4d3a-85c4-e1bceb8cd856"
+OptimizationOptimJL = "36348300-93cb-4f02-beb5-3c3902f8871e"
+Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
+Ipopt = "b6b21f68-93f8-5de0-b562-5493be1d77c9"
+OptimizationMOI = "fd9f6733-72f4-499f-8506-86b2bdd0dea1"
+
+[targets]
+test = ["Test", "NLPModelsTest", "OptimizationOptimJL", "Zygote", "Ipopt", "OptimizationMOI"]

lib/OptimizationNLPModels/src/OptimizationNLPModels.jl

@@ -0,0 +1,63 @@
+module OptimizationNLPModels
+
+using Reexport
+@reexport using NLPModels, Optimization, ADTypes
+
+"""
+    OptimizationFunction(nlpmodel::AbstractNLPModel, adtype::AbstractADType = NoAD())
+
+Returns an `OptimizationFunction` from the `NLPModel` defined in `nlpmodel` where the
+available derivates are re-used from the model, and the rest are populated with the
+Automatic Differentiation backend specified by `adtype`.
+"""
+function SciMLBase.OptimizationFunction(nlpmodel::AbstractNLPModel,
+        adtype::ADTypes.AbstractADType = SciMLBase.NoAD(); kwargs...)
+    f(x, p) = NLPModels.obj(nlpmodel, x)
+    grad(G, u, p) = NLPModels.grad!(nlpmodel, u, G)
+    hess(H, u, p) = (H .= NLPModels.hess(nlpmodel, u))
+    hv(Hv, u, v, p) = NLPModels.hprod!(nlpmodel, u, v, Hv)
+
+    if !unconstrained(nlpmodel) && !bound_constrained(nlpmodel)
+        cons(res, x, p) = NLPModels.cons!(nlpmodel, x, res)
+        cons_j(J, x, p) = (J .= NLPModels.jac(nlpmodel, x))
+        cons_jvp(Jv, v, x, p) = NLPModels.jprod!(nlpmodel, x, v, Jv)
+
+        return OptimizationFunction(
+            f, adtype; grad, hess, hv, cons, cons_j, cons_jvp, kwargs...)
+    end
+
+    return OptimizationFunction(f, adtype; grad, hess, hv, kwargs...)
+end
+
+"""
+    OptimizationProblem(nlpmodel::AbstractNLPModel, adtype::AbstractADType = NoAD())
+
+Returns an `OptimizationProblem` with the bounds and constraints defined in `nlpmodel`.
+The optimization function and its derivatives are re-used from `nlpmodel` when available
+or populated wit the Automatic Differentiation backend specified by `adtype`.
+"""
+function SciMLBase.OptimizationProblem(nlpmodel::AbstractNLPModel,
+        adtype::ADTypes.AbstractADType = SciMLBase.NoAD(); kwargs...)
+    f = OptimizationFunction(nlpmodel, adtype; kwargs...)
+    u0 = nlpmodel.meta.x0
+    lb, ub = if has_bounds(nlpmodel)
+        (nlpmodel.meta.lvar, nlpmodel.meta.uvar)
+    else
+        (nothing, nothing)
+    end
+
+    lcons, ucons = if has_inequalities(nlpmodel) || has_equalities(nlpmodel)
+        (nlpmodel.meta.lcon, nlpmodel.meta.ucon)
+    else
+        (nothing, nothing)
+    end
+    sense = nlpmodel.meta.minimize ? Optimization.MinSense : Optimization.MaxSense
+
+    # The number of variables, geometry of u0, etc.. are valid and were checked when the
+    # nlpmodel was created.
+
+    return Optimization.OptimizationProblem(
+        f, u0; lb = lb, ub = ub, lcons = lcons, ucons = ucons, sense = sense, kwargs...)
+end
+
+end

lib/OptimizationNLPModels/test/runtests.jl

@@ -0,0 +1,117 @@
+using OptimizationNLPModels, Optimization, NLPModelsTest, Ipopt, OptimizationMOI, Zygote,
+      OptimizationOptimJL
+using Test
+
+@testset "NLPModels" begin
+    # First problem: Problem 5 in the Hock-Schittkowski suite
+    # https://jso.dev/NLPModelsTest.jl/dev/reference/#NLPModelsTest.HS5
+    # Problem with box bounds
+    hs5f(u, p) = sin(u[1] + u[2]) + (u[1] - u[2])^2 - (3 / 2) * u[1] + (5 / 2)u[2] + 1
+    f = Optimization.OptimizationFunction(hs5f, Optimization.AutoZygote())
+    lb = [-1.5; -3]
+    ub = [4.0; 3.0]
+    u0 = [0.0; 0.0]
+    oprob = Optimization.OptimizationProblem(
+        f, u0, lb = lb, ub = ub, sense = Optimization.MinSense)
+
+    nlpmo = NLPModelsTest.HS5()
+    converted = OptimizationNLPModels.OptimizationProblem(nlpmo, Optimization.AutoZygote())
+
+    sol_native = solve(oprob, Optimization.LBFGS(), maxiters = 1000)
+    sol_converted = solve(converted, Optimization.LBFGS(), maxiters = 1000)
+
+    @test sol_converted.retcode == sol_native.retcode
+    @test sol_converted.u ≈ sol_native.u
+
+    # Second problem: Brown and Dennis function
+    # https://jso.dev/NLPModelsTest.jl/dev/reference/#NLPModelsTest.BROWNDEN
+    # Problem without bounds
+    function brown_dennis(u, p)
+        return sum([((u[1] + (i / 5) * u[2] - exp(i / 5))^2 +
+                     (u[3] + sin(i / 5) * u[4] - cos(i / 5))^2)^2 for i in 1:20])
+    end
+    f = Optimization.OptimizationFunction(brown_dennis, Optimization.AutoZygote())
+    u0 = [25.0; 5.0; -5.0; -1.0]
+    oprob = Optimization.OptimizationProblem(f, u0, sense = Optimization.MinSense)
+
+    nlpmo = NLPModelsTest.BROWNDEN()
+    converted = OptimizationNLPModels.OptimizationProblem(nlpmo, Optimization.AutoZygote())
+
+    sol_native = solve(oprob, BFGS())
+    sol_converted = solve(converted, BFGS())
+
+    @test sol_converted.retcode == sol_native.retcode
+    @test sol_converted.u ≈ sol_native.u
+
+    # Third problem: Problem 10 in the Hock-Schittkowski suite
+    # https://jso.dev/NLPModelsTest.jl/dev/reference/#NLPModelsTest.HS10
+    # Problem with inequality bounds
+    hs10(u, p) = u[1] - u[2]
+    hs10_cons(res, u, p) = (res .= -3.0 * u[1]^2 + 2.0 * u[1] * u[2] - u[2]^2 + 1.0)
+    lcons = [0.0]
+    ucons = [Inf]
+    u0 = [-10.0; 10.0]
+    f = Optimization.OptimizationFunction(
+        hs10, Optimization.AutoForwardDiff(); cons = hs10_cons)
+    oprob = Optimization.OptimizationProblem(
+        f, u0, lcons = lcons, ucons = ucons, sense = Optimization.MinSense)
+
+    nlpmo = NLPModelsTest.HS10()
+    converted = OptimizationNLPModels.OptimizationProblem(
+        nlpmo, Optimization.AutoForwardDiff())
+
+    sol_native = solve(oprob, Ipopt.Optimizer())
+    sol_converted = solve(converted, Ipopt.Optimizer())
+
+    @test sol_converted.retcode == sol_native.retcode
+    @test sol_converted.u ≈ sol_native.u
+
+    # Fourth problem: Problem 13 in the Hock-Schittkowski suite
+    # https://jso.dev/NLPModelsTest.jl/dev/reference/#NLPModelsTest.HS13
+    # Problem with box & inequality bounds
+    hs13(u, p) = (u[1] - 2.0)^2 + u[2]^2
+    hs13_cons(res, u, p) = (res .= (1.0 - u[1])^3 - u[2])
+    lcons = [0.0]
+    ucons = [Inf]
+    lb = [0.0; 0.0]
+    ub = [Inf; Inf]
+    u0 = [-2.0; -2.0]
+    f = Optimization.OptimizationFunction(
+        hs13, Optimization.AutoForwardDiff(); cons = hs13_cons)
+    oprob = Optimization.OptimizationProblem(f, u0, lb = lb, ub = ub, lcons = lcons,
+        ucons = ucons, sense = Optimization.MinSense)
+
+    nlpmo = NLPModelsTest.HS13()
+    converted = OptimizationNLPModels.OptimizationProblem(
+        nlpmo, Optimization.AutoForwardDiff())
+
+    sol_native = solve(oprob, Ipopt.Optimizer())
+    sol_converted = solve(converted, Ipopt.Optimizer())
+
+    @test sol_converted.retcode == sol_native.retcode
+    @test sol_converted.u ≈ sol_native.u
+
+    # Fifth problem: Problem 14 in the Hock-Schittkowski suite
+    # https://jso.dev/NLPModelsTest.jl/dev/reference/#NLPModelsTest.HS14
+    # Problem with mixed equality & inequality constraints
+    hs14(u, p) = (u[1] - 2.0)^2 + (u[2] - 1.0)^2
+    hs14_cons(res, u, p) = (res .= [u[1] - 2.0 * u[2];
+                                    -0.25 * u[1]^2 - u[2]^2 + 1.0])
+    lcons = [-1.0; 0.0]
+    ucons = [-1.0; Inf]
+    u0 = [2.0; 2.0]
+    f = Optimization.OptimizationFunction(
+        hs14, Optimization.AutoForwardDiff(); cons = hs14_cons)
+    oprob = Optimization.OptimizationProblem(
+        f, u0, lcons = lcons, ucons = ucons, sense = Optimization.MinSense)
+
+    nlpmo = NLPModelsTest.HS14()
+    converted = OptimizationNLPModels.OptimizationProblem(
+        nlpmo, Optimization.AutoForwardDiff())
+
+    sol_native = solve(oprob, Ipopt.Optimizer())
+    sol_converted = solve(converted, Ipopt.Optimizer())
+
+    @test sol_converted.retcode == sol_native.retcode
+    @test sol_converted.u ≈ sol_native.u
+end