MOI 1.0

.gitignore

@@ -5,3 +5,6 @@ Manifest.toml
 *.sublime-workspace
 *.bat
 *.sublime-project
+*.mof.json
+*.prt
+*.txt

Project.toml

@@ -1,21 +1,13 @@
 name = "QuadraticToBinary"
 uuid = "014a38d5-7acb-4e20-b6c0-4fe5c2344fd1"
 authors = ["Joaquim Garcia <joaquimgarcia@psr-inc.com>"]
-version = "0.3.0"
+version = "0.4.0"
 
 [deps]
 MathOptInterface = "b8f27783-ece8-5eb3-8dc8-9495eed66fee"
 DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
 
 [compat]
-MathOptInterface = "0.10.6"
+MathOptInterface = "1"
 DataStructures = "0.17.10, 0.18.0"
-julia = "1"
-
-[extras]
-Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
-Cbc = "9961bab8-2fa3-5c5a-9d89-47fab24efd76"
-
-[targets]
-test = ["Test", "Cbc"]
-
+julia = "1.6"

README.md

@@ -18,7 +18,7 @@ Non-convex quadratic programs are extremely hard to solve. This problem class ca
 be solved by Global Solvers such as [Couenne](https://projects.coin-or.org/Couenne).
 Another possibility is to rely on binary expasion of products terms that appear in the
 problem, in this case the problem is *approximated* and can be solved by off-the-shelf
-MIP solvers such as [Cbc](https://github.com/JuliaOpt/Cbc.jl), [CPLEX](https://github.com/JuliaOpt/CPLEX.jl), [GLPK](https://github.com/JuliaOpt/GLPK.jl), [Gurobi](https://github.com/JuliaOpt/Gurobi.jl), [Xpress](https://github.com/JuliaOpt/Xpress.jl).
+MIP solvers such as [Cbc](https://github.com/jump-dev/Cbc.jl), [CPLEX](https://github.com/jump-dev/CPLEX.jl), [GLPK](https://github.com/jump-dev/GLPK.jl), [Gurobi](https://github.com/jump-dev/Gurobi.jl), [HiGHS](https://github.com/jump-dev/HiGHS.jl), [Xpress](https://github.com/jump-dev/Xpress.jl).
 
 ## Example
 
@@ -77,15 +77,15 @@ objective_value(model) # ≈ 9.0
 @assert value(c) ≈ 4.0
 ```
 
-### MathOptInterface with Cbc solver
+### MathOptInterface with HiGHS solver
 
 ```julia
 using MathOptInterface
 using QuadraticToBinary
 const MOI = MathOptInterface
-using Cbc
+using HiGHS
 
-optimizer = MOI.instantiate(Cbc.Optimizer, with_bridge_type = Float64)
+optimizer = MOI.instantiate(HiGHS.Optimizer, with_bridge_type = Float64)
 
 model = QuadraticToBinary.Optimizer{Float64}(optimizer)
 
@@ -118,7 +118,8 @@ MOI.optimize!(model)
 @assert MOI.get(model, MOI.ConstraintPrimal(), c) ≈ 4.0
 ```
 
-Note that duals are not available because the problem was approximated as a MIP.
+Note:
+that duals are not available because the problem was approximated as a MIP.
 
 ## QuadraticToBinary.Optimizer Attributes
 
@@ -140,6 +141,16 @@ MOI.set(model, QuadraticToBinary.VariablePrecision(), vi, val)
 The precision for each varible will be `val * (UB - LB)`. Where `UB` and `LB` are,
 respectively, the upper and lower bound of the variable.
 
+Note:
+binary expansion problem can be numerically challenging for high precision. You
+might need to modify solver options accordingly. In the case of HiGHS:
+
+```julia
+tol = 1e-9
+MOI.set(highs_high_tol, MOI.RawOptimizerAttribute("mip_feasibility_tolerance"), tol)
+MOI.set(highs_high_tol, MOI.RawOptimizerAttribute("primal_feasibility_tolerance"), tol)
+```
+
 ### Bounds
 
 For the sake of simplicity, the following two attributes are made available:

src/moi.jl

@@ -229,7 +229,7 @@ end
 
 # MOI.supports(model::Optimizer, args...) = MOI.supports(model.optimizer, args...)
 
-# TODO, call this on inner solver
+# Q2B is always incremental because it uses add_constraint to cacehc onversions
 MOI.supports_incremental_interface(::Optimizer) = true
 
 function MOI.supports(model::Optimizer, attr::MOI.VariableName, tp::Type{MOI.VariableIndex})
@@ -270,6 +270,7 @@ function MOI.set(model::Optimizer, attr::Union{
     MOI.Silent,
     MOI.NumberOfThreads,
     MOI.TimeLimitSec,
+    MOI.RawOptimizerAttribute,
 }, val
 )
     return MOI.set(model.optimizer, attr, val)
@@ -307,8 +308,8 @@ quadratic_type(::Type{MOI.VectorAffineFunction{T}}) where T = MOI.VectorQuadrati
 
 
 function MOI.get(model::Optimizer, ::MOI.SolverName)
-    return "Binary reformulation of quadratic model with solver " *
-        MOI.get(model.optimizer, MOI.SolverName()) * " attached"
+    return "QuadraticToBinary [" *
+        MOI.get(model.optimizer, MOI.SolverName()) * "]"
 end
 
 function MOI.supports_add_constrained_variables(
@@ -340,20 +341,8 @@ function MOI.supports_add_constrained_variables(
     return MOI.supports_add_constrained_variables(model.optimizer, S)
 end
 
-# function MOI.set(model::Optimizer, param::MOI.RawOptimizerAttribute, value)
-#     # if in a subset of the q2b save it
-#     # otherwise pass it forward
-# end
-
-# function MOI.get(model::Optimizer, param::MOI.RawOptimizerAttribute)
-# end
-
-function MOI.Utilities.supports_default_copy_to(model::Optimizer, val::Bool)
-    return true # val # MOI.Utilities.supports_default_copy_to(model.optimizer, val)
-end
-
-function MOI.copy_to(dest::Optimizer, src::MOI.ModelLike; kwargs...)
-    return MOI.Utilities.automatic_copy_to(dest, src; kwargs...)
+function MOI.copy_to(dest::Optimizer, src::MOI.ModelLike)
+    return MOI.Utilities.default_copy_to(dest, src)
 end
 
 
@@ -989,25 +978,14 @@ end
 
 function delete_additional_constraints(model)
     ch = model.index_cache
-    if true
-        MOI.delete(model.optimizer, ch.sa_eq)
-        MOI.delete(model.optimizer, ch.sa_lt)
-        MOI.delete(model.optimizer, ch.sa_gt)
-        MOI.delete(model.optimizer, ch.vi)
-    else
-        for ci in ch.sa_eq
-            MOI.delete(model.optimizer, ci)
-        end
-        for ci in ch.sa_lt
-            MOI.delete(model.optimizer, ci)
-        end
-        for ci in ch.sa_gt
-            MOI.delete(model.optimizer, ci)
-        end
-        for vi in ch.vi
-            MOI.delete(model.optimizer, vi)
-        end
-    end
+    MOI.delete(model.optimizer, ch.sa_eq)
+    MOI.delete(model.optimizer, ch.sa_lt)
+    MOI.delete(model.optimizer, ch.sa_gt)
+    MOI.delete(model.optimizer, ch.sv_zo)
+    MOI.delete(model.optimizer, ch.sv_lt)
+    MOI.delete(model.optimizer, ch.sv_gt)
+    MOI.delete(model.optimizer, ch.vi)
+    return nothing
 end
 
 function lower(model, info)
@@ -1481,7 +1459,8 @@ function MOI.set(
     ::GlobalVariablePrecision,
     value::Union{Nothing, Float64}
 )
-    val = value === nothing ? NaN : value
+    # default global precision is 1e-4
+    val = value === nothing ? 1e-4 : value
     @assert 0 < val <= 1
     model.global_initial_precision = val
     return

test/Project.toml

@@ -0,0 +1,7 @@
+[deps]
+HiGHS = "87dc4568-4c63-4d18-b0c0-bb2238e4078b"
+MathOptInterface = "b8f27783-ece8-5eb3-8dc8-9495eed66fee"
+Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+
+[compat]
+HiGHS = "1.1.0"