Using MOI to Warm Start Duals Values on Nonlinear Constraints

[dsigler1234]

Hello,

I am trying to warm start a non-linear model that I am solving using Ipopt. I can warm start the primal variables just fine and the dual variables for all constraints except the ones made using the non-linear interface. Here is a simple example:

using JuMP
import Ipopt
using MathOptInterface
const MOI = MathOptInterface

model = JuMP.Model(Ipopt.Optimizer)
@variable(model, x)
@constraint(model, con, x >= 0)
@NLconstraint(model, nl_con, cos(x) >= 1.0 )
@objective(model, Min, x)

JuMP.optimize!(model)

con_dual_val = JuMP.dual(con)
nl_dual_val = JuMP.dual(nl_con)

MOI.set(model, MOI.ConstraintDualStart(), con, con_dual_val) # works just fine
MOI.set(model, MOI.ConstraintDualStart(), nl_con, nl_dual_val) # causes the error below

ERROR: LoadError: MethodError: no method matching throw_set_error_fallback(::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{Ipopt.Optimizer}, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}}, ::MathOptInterface.ConstraintDualStart, ::NonlinearConstraintIndex, ::Float64)
Closest candidates are:
throw_set_error_fallback(::MathOptInterface.ModelLike, ::Union{MathOptInterface.AbstractConstraintAttribute, MathOptInterface.AbstractVariableAttribute}, ::Union{MathOptInterface.VariableIndex, MathOptInterface.ConstraintIndex}, ::Any; error_if_supported) at /Users/dsigler/.julia/packages/MathOptInterface/IIN1o/src/attributes.jl:499
throw_set_error_fallback(::MathOptInterface.ModelLike, ::Union{MathOptInterface.AbstractModelAttribute, MathOptInterface.AbstractOptimizerAttribute}, ::Any; error_if_supported) at /Users/dsigler/.julia/packages/MathOptInterface/IIN1o/src/attributes.jl:487
throw_set_error_fallback(::MathOptInterface.ModelLike, ::MathOptInterface.ConstraintFunction, ::MathOptInterface.ConstraintIndex{F, S}, ::F; error_if_supported) where {F<:MathOptInterface.AbstractFunction, S} at /Users/dsigler/.julia/packages/MathOptInterface/IIN1o/src/attributes.jl:1571
...
Stacktrace:
[1] set(::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.Bridges.LazyBridgeOptimizer{Ipopt.Optimizer}, MathOptInterface.Utilities.UniversalFallback{MathOptInterface.Utilities.Model{Float64}}}, ::MathOptInterface.ConstraintDualStart, ::NonlinearConstraintIndex, ::Float64)
@ MathOptInterface ~/.julia/packages/MathOptInterface/IIN1o/src/attributes.jl:481
[2] set(model::Model, attr::MathOptInterface.ConstraintDualStart, cr::NonlinearConstraintRef{ScalarShape}, value::Float64)
@ JuMP ~/.julia/packages/JuMP/lnUbA/src/JuMP.jl:1321
[3] top-level scope
@ ~/PowerModelsDistribution.jl/examples/issue.jl:18
[4] include(fname::String)
@ Base.MainInclude ./client.jl:444
[5] top-level scope
@ REPL[7]:1
in expression starting at /Users/dsigler/PowerModelsDistribution.jl/examples/issue.jl:18

I was wondering if there is a way to do this or if this is not supported?

Best,

Devon

[mlubin]

See https://jump.dev/MathOptInterface.jl/stable/reference/nonlinear/#MathOptInterface.NLPBlockDualStart.

[dsigler1234]

Thanks for the quick response. So should I replace

MOI.set(model, MOI.ConstraintDualStart(), nl_con, nl_dual_val) # causes the error below

with

MOI.set(model, MOI.NLPBlockDualStart(), nl_con, nl_dual_val)

because that still throws this error

ERROR: LoadError: MethodError: no method matching set(::Model, ::MathOptInterface.NLPBlockDualStart, ::NonlinearConstraintRef{ScalarShape}, ::Float64)
Closest candidates are:
set(::Model, ::MathOptInterface.AbstractConstraintAttribute, ::ConstraintRef, ::Any) at /Users/dsigler/.julia/packages/JuMP/lnUbA/src/JuMP.jl:1313
set(::Model, ::MathOptInterface.AbstractModelAttribute, ::Any) at /Users/dsigler/.julia/packages/JuMP/lnUbA/src/JuMP.jl:1297
set(::MathOptInterface.ModelLike, ::Union{MathOptInterface.AbstractConstraintAttribute, MathOptInterface.AbstractModelAttribute, MathOptInterface.AbstractOptimizerAttribute, MathOptInterface.AbstractVariableAttribute}, ::Any...) at /Users/dsigler/.julia/packages/MathOptInterface/IIN1o/src/attributes.jl:480

sorry for the naive question.

Best,

Devon

[mlubin]

NLPBlockDualStart is an attribute of the model. You pass a single vector with the warm starts for all nonlinear constraints. JuMP-related "how do I do X" aspects are better discussed on Discourse: https://discourse.julialang.org/c/domain/opt/13.

[dsigler1234]

Ok thanks for your help.

[mlubin]

@dsigler1234 You may be interested in jump-dev/JuMP.jl#2879. Until that's released, you can replicate the functionality manually.