Implement local_optimizer for MOI interface (#173)

README.md

@@ -134,6 +134,22 @@ got 0.5443310477213124 at [0.3333333342139688,0.29629628951338166]
 Note that the MathOptInterface interface sets slightly different convergence tolerances by default (these default values are given by the `NLopt.DEFAULT_OPTIONS` dictionary),
 so the outputs from the two problems are not identical.
 
+Some algorithms need a local optimizer. These are set with `local_optimizer`, e.g.,
+```julia
+model = Model(NLopt.Optimizer)
+set_optimizer_attribute(model, "algorithm", :AUGLAG)
+set_optimizer_attribute(model, "local_optimizer", :LD_LBFGS)
+```
+To parametrize the local optimizer, pass use the `NLopt.Opt` interface, e.g.,
+```julia
+model = Model(NLopt.Optimizer)
+set_optimizer_attribute(model, "algorithm", :AUGLAG)
+local_optimizer = NLopt.Opt(:LD_LBFGS, num_variables)
+local_optimizer.xtol_rel = 1e-4
+set_optimizer_attribute(model, "local_optimizer", local_optimizer)
+```
+where `num_variables` is the number of variables of the optimization problem.
+
 ## Reference
 
 The main purpose of this section is to document the syntax and unique

src/MOI_wrapper.jl

@@ -84,6 +84,7 @@ const DEFAULT_OPTIONS = Dict{String, Any}(
     "population" => 0,
     "seed" => nothing,
     "vector_storage" => 0,
+    "local_optimizer" => nothing,
 )
 
 function Optimizer()
@@ -776,6 +777,15 @@ function MOI.optimize!(model::Optimizer)
     # TODO: Reuse model.inner for incremental solves if possible.
     num_variables = length(model.variable_info)
     model.inner = Opt(model.options["algorithm"], num_variables)
+    local_optimizer = model.options["local_optimizer"]
+    if local_optimizer !== nothing
+        if local_optimizer isa Symbol
+            local_optimizer = Opt(local_optimizer, num_variables)
+        else
+            local_optimizer = Opt(local_optimizer.algorithm, num_variables)
+        end
+        local_optimizer!(model.inner, local_optimizer)
+    end
 
     # load parameters
     stopval!(model.inner, model.options["stopval"])

test/MOI_wrapper.jl

@@ -23,9 +23,28 @@ end
     @test !MOIU.supports_default_copy_to(optimizer, true)
 end
 
-@testset "Non-Linear tests" begin
+function test_nlp(solver)
+    optimizer = MOI.instantiate(solver)
     MOIT.nlptest(optimizer, config)
 end
+@testset "Non-Linear tests" begin
+    test_nlp(solver)
+    test_nlp(MOI.OptimizerWithAttributes(
+        NLopt.Optimizer,
+        "algorithm" => :AUGLAG,
+        "local_optimizer" => :LD_LBFGS,
+    ))
+    # NLP tests have different number of variables so we
+    # cannot run through them all with the same `local_optimizer`.
+    # Let's just do hs071.
+    local_optimizer = Opt(:LD_LBFGS, 4)
+    opt.xtol_rel = 1e-6
+    MOIT.hs071_test(MOI.instantiate(MOI.OptimizerWithAttributes(
+        NLopt.Optimizer,
+        "algorithm" => :AUGLAG,
+        "local_optimizer" => local_optimizer,
+    )), config)
+end
 
 @testset "Testing getters" begin
     MOI.Test.copytest(MOI.instantiate(solver, with_bridge_type=Float64), MOIU.Model{Float64}())
@@ -98,7 +117,16 @@ end
         # NumberOfThreads not supported
         "number_threads",
         # Infeasibility and unboundedness not detected by NLopt
-        "solve_unbounded_model"
+        "solve_unbounded_model",
+        "solve_farkas_interval_lower",
+        "solve_farkas_lessthan",
+        "solve_farkas_equalto_lower",
+        "solve_farkas_equalto_upper",
+        "solve_farkas_variable_lessthan",
+        "solve_farkas_variable_lessthan_max",
+        "solve_farkas_greaterthan",
+        "solve_farkas_interval_upper",
+        "solve_farkas_lessthan",
     ]
     MOIT.unittest(bridged, config, exclude)
 end

test/runtests.jl

@@ -1,3 +1,4 @@
 include("tutorial.jl")
 include("fix133.jl")
 include("MPB_wrapper.jl")
+include("MOI_wrapper.jl")