Merge pull request #3140 from AayushSabharwal/as/optsys-fixes

fix: respect bounds in `modelingtoolkitize`, make bounded variable irreducible

Author: ChrisRackauckas

src/systems/optimization/modelingtoolkitize.jl

@@ -33,6 +33,15 @@ function modelingtoolkitize(prob::DiffEqBase.OptimizationProblem;
     end
     _vars = reshape(_vars, size(prob.u0))
     vars = ArrayInterface.restructure(prob.u0, _vars)
+    if prob.ub !== nothing # lb is also !== nothing
+        vars = map(vars, prob.lb, prob.ub) do sym, lb, ub
+            if iszero(lb) && iszero(ub) || isinf(lb) && lb < 0 && isinf(ub) && ub > 0
+                sym
+            else
+                Symbolics.setmetadata(sym, VariableBounds, (lb, ub))
+            end
+        end
+    end
     params = if has_p
         if p_names === nothing && SciMLBase.has_sys(prob.f)
             p_names = Dict(parameter_index(prob.f.sys, sym) => sym

src/systems/optimization/optimizationsystem.jl

@@ -103,6 +103,17 @@ function OptimizationSystem(op, unknowns, ps;
     ps′ = value.(ps)
     op′ = value(scalarize(op))
 
+    irreducible_subs = Dict()
+    for i in eachindex(unknowns′)
+        var = unknowns′[i]
+        if hasbounds(var)
+            irreducible_subs[var] = irrvar = setirreducible(var, true)
+            unknowns′[i] = irrvar
+        end
+    end
+    op′ = substitute(op′, irreducible_subs)
+    constraints = substitute.(constraints, (irreducible_subs,))
+
     if !(isempty(default_u0) && isempty(default_p))
         Base.depwarn(
             "`default_u0` and `default_p` are deprecated. Use `defaults` instead.",
@@ -113,7 +124,8 @@ function OptimizationSystem(op, unknowns, ps;
         throw(ArgumentError("System names must be unique."))
     end
     defaults = todict(defaults)
-    defaults = Dict(value(k) => value(v)
+    defaults = Dict(substitute(value(k), irreducible_subs) => substitute(
+                        value(v), irreducible_subs)
     for (k, v) in pairs(defaults) if value(v) !== nothing)
 
     var_to_name = Dict()

src/variables.jl

@@ -106,6 +106,7 @@ isoutput(x) = isvarkind(VariableOutput, x)
 # Before the solvability check, we already have handled IO variables, so
 # irreducibility is independent from IO.
 isirreducible(x) = isvarkind(VariableIrreducible, x)
+setirreducible(x, v) = setmetadata(x, VariableIrreducible, v)
 state_priority(x) = convert(Float64, getmetadata(x, VariableStatePriority, 0.0))::Float64
 
 function default_toterm(x)

test/modelingtoolkitize.jl

@@ -67,6 +67,16 @@ sol = solve(prob, BFGS())
 sol = solve(prob, Newton())
 @test sol.objective < 1e-8
 
+prob = OptimizationProblem(ones(3); lb = [-Inf, 0.0, 1.0], ub = [Inf, 0.0, 2.0]) do u, p
+    sum(abs2, u)
+end
+
+sys = complete(modelingtoolkitize(prob))
+@test !ModelingToolkit.hasbounds(unknowns(sys)[1])
+@test !ModelingToolkit.hasbounds(unknowns(sys)[2])
+@test ModelingToolkit.hasbounds(unknowns(sys)[3])
+@test ModelingToolkit.getbounds(unknowns(sys)[3]) == (1.0, 2.0)
+
 ## SIR System Regression Test
 
 β = 0.01# infection rate

test/optimizationsystem.jl

@@ -1,5 +1,5 @@
 using ModelingToolkit, SparseArrays, Test, Optimization, OptimizationOptimJL,
-      OptimizationMOI, Ipopt, AmplNLWriter, Ipopt_jll
+      OptimizationMOI, Ipopt, AmplNLWriter, Ipopt_jll, SymbolicIndexingInterface
 using ModelingToolkit: get_metadata
 
 @testset "basic" begin
@@ -347,3 +347,15 @@ end
     prob = @test_nowarn OptimizationProblem(sys, nothing)
     @test_nowarn solve(prob, NelderMead())
 end
+
+@testset "Bounded unknowns are irreducible" begin
+    @variables x
+    @variables y [bounds = (-Inf, Inf)]
+    @variables z [bounds = (1.0, 2.0)]
+    obj = x^2 + y^2 + z^2
+    cons = [y ~ 2x
+            z ~ 2y]
+    @mtkbuild sys = OptimizationSystem(obj, [x, y, z], []; constraints = cons)
+    @test is_variable(sys, z)
+    @test !is_variable(sys, y)
+end