DNM: trigger Gurobi get obj bound issue

[chriscoey]

@odow

the issue is reproduced if you run include("test/myruntests.jl") (not the usual "runtests.jl")

what I see is:

no conic constraints need outer approximation
Set parameter FeasibilityTol to value 1e-09
Set parameter MIPGap to value 1e-09
Set parameter IntFeasTol to value 1e-09
Set parameter TimeLimit to value 9.9999999973702431e+05
JuMP.primal_status(opt.oa_model) = MathOptInterface.FEASIBLE_POINT
_soc1: Error During Test at /Users/coey/.julia/dev/MOIPajarito/test/JuMP_tests.jl:34
  Got exception outside of a @test
  Gurobi Error 10005: Unable to retrieve attribute 'ObjBound'
  Stacktrace:
    [1] _check_ret
      @ ~/.julia/packages/Gurobi/A9YqC/src/MOI_wrapper/MOI_wrapper.jl:325 [inlined]
    [2] get(model::Gurobi.Optimizer, attr::MathOptInterface.ObjectiveBound)
      @ Gurobi ~/.julia/packages/Gurobi/A9YqC/src/MOI_wrapper/MOI_wrapper.jl:3141
    [3] get (repeats 2 times)
      @ ~/.julia/packages/MathOptInterface/eoIu0/src/Bridges/bridge_optimizer.jl:815 [inlined]
    [4] get(model::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{Gurobi.Optimizer}}, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}}, attr::MathOptInterface.ObjectiveBound)
      @ MathOptInterface.Utilities ~/.julia/packages/MathOptInterface/eoIu0/src/Utilities/cachingoptimizer.jl:803
    [5] _moi_get_result(model::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{Gurobi.Optimizer}}, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}}, args::MathOptInterface.ObjectiveBound)
      @ JuMP ~/.julia/packages/JuMP/2IF9U/src/JuMP.jl:1226
    [6] get(model::JuMP.Model, attr::MathOptInterface.ObjectiveBound)
      @ JuMP ~/.julia/packages/JuMP/2IF9U/src/JuMP.jl:1251
    [7] objective_bound(model::JuMP.Model)
      @ JuMP ~/.julia/packages/JuMP/2IF9U/src/objective.jl:30
    [8] setup_models(opt::MOIPajarito.Optimizer)
      @ MOIPajarito ~/.julia/dev/MOIPajarito/src/optimize.jl:499
    [9] optimize(opt::MOIPajarito.Optimizer)
      @ MOIPajarito ~/.julia/dev/MOIPajarito/src/optimize.jl:95
   [10] optimize!
      @ ~/.julia/dev/MOIPajarito/src/MOI_wrapper.jl:212 [inlined]
   [11] optimize!
      @ ~/.julia/packages/MathOptInterface/eoIu0/src/MathOptInterface.jl:81 [inlined]
   [12] optimize!(m::MathOptInterface.Utilities.CachingOptimizer{MOIPajarito.Optimizer, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}})
      @ MathOptInterface.Utilities ~/.julia/packages/MathOptInterface/eoIu0/src/Utilities/cachingoptimizer.jl:285
   [13] optimize!
      @ ~/.julia/packages/MathOptInterface/eoIu0/src/Bridges/bridge_optimizer.jl:348 [inlined]
   [14] optimize!(m::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{MathOptInterface.Utilities.CachingOptimizer{MOIPajarito.Optimizer, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}}}, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}})
      @ MathOptInterface.Utilities ~/.julia/packages/MathOptInterface/eoIu0/src/Utilities/cachingoptimizer.jl:294
   [15] optimize!(model::JuMP.Model, optimizer_factory::Nothing; bridge_constraints::Bool, ignore_optimize_hook::Bool, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
      @ JuMP ~/.julia/packages/JuMP/2IF9U/src/optimizer_interface.jl:195
   [16] optimize! (repeats 2 times)
      @ ~/.julia/packages/JuMP/2IF9U/src/optimizer_interface.jl:167 [inlined]
   [17] _soc1(opt::MathOptInterface.OptimizerWithAttributes)
      @ Main.TestJuMP ~/.julia/dev/MOIPajarito/test/JuMP_tests.jl:53

here is the line where it occurs: https://github.com/chriscoey/MOIPajarito.jl/blob/e161bc8f0a1559f970b66eb8b3b2ca141277f063/src/optimize.jl#L498

it is solving a MISOCP with Gurobi just a few lines above that (MOIPajarito does not use outer approximation because Gurobi already supports MISOCP and this instance is MISOCP).

[odow]

Do you have the Gurobi log? I don't have a good solution for you. That's an internal Gurobi error, not from us.

What happens if you use Gurobi directly? There might be other parameters that you need to set.

[chriscoey]

OK I think I figured it out. the problem being solved is continuous, not MISOCP. so we can't call objective_bound:

using Test
import JuMP
import Gurobi
const MOI = JuMP.MOI
TOL = 1e-4

m = JuMP.Model(Gurobi.Optimizer)
JuMP.@variable(m, x)
JuMP.@objective(m, Min, -x)
JuMP.@constraint(m, [3.5, x] in JuMP.SecondOrderCone())
JuMP.optimize!(m)
@test JuMP.termination_status(m) == MOI.OPTIMAL
@test JuMP.primal_status(m) == MOI.FEASIBLE_POINT
@test isapprox(JuMP.objective_value(m), -3.5, atol = TOL)
@test isapprox(JuMP.value(x), 3.5, atol = TOL)
@test isapprox(JuMP.dual_objective_value(m), -3.5, atol = TOL) # EDIT: also ERROR
@test isapprox(JuMP.objective_bound(m), -3.5, atol = TOL)   # ERROR

So I can fix this in MOIPajarito.

In general I don't know whether it is intended that objective_bound only work for MIPs and not for continuous problems. But I guess it was Gurobi's choice in this case.

[chriscoey]

actually, JuMP.dual_objective_value(m) errors too (same error). so I can't get dual obj or obj bound for this SOCP. that's a bit surprising

[odow]

Yeah that's a deliberate Gurobi choice. The perils of solver-independence...

You should probably use a try-catch.

[chriscoey]

eh i'll just use the primal objective value

[chriscoey]

changed my mind. does this look OK? I'm not 100% sure if there's a problem with using return inside try.

# try to get an objective bound or dual objective value, else use the primal objective value
function get_objective_bound(model::JuMP.Model)
    try
        return JuMP.objective_bound(model)
    catch
    end
    try
        return JuMP.dual_objective_value(model)
    catch
    end
    return JuMP.objective_value(model)
end

[odow]

Using return inside try is fine.

Is this for MILP models? What even is the dual objective value?

[chriscoey]

the OA solver isn't necessarily a MIP solver. at some point with a separation-cuts-only algorithm, we can give cuts to a conic solver that doesn't support all the cones we want