Fix start value #138
Fix start value #138
Conversation
|
maybe I'm using it wrong but here is what I tried: function example_problem(m = Model())
@variable(m, x >= 0)
@variable(m, y <= 2)
@variable(m, P[1:2,1:2] >= 0)
@objective(m, Max, x+y + P[1,1] - P[2,2])
@constraint(m, x <= y+4)
@constraint(m, P[1,1] + P[2,2] <= 4)
@constraint(m, P[2,1] - P[1,1] <= 4)
return m
end
m = example_problem()
optimize!(m, with_optimizer(SCS.Optimizer, eps=1e-10))
solvedx = value(m[:x])
solvedy = value(m[:y])
solvedP = value.(m[:P])
m2 = example_problem()
set_start_value(m2[:x], solvedx)
set_start_value(m2[:y], solvedy)
set_start_value.(m2[:P], solvedP)
optimize!(m2, with_optimizer(SCS.Optimizer, eps=1e-10, max_iters=3, warm_start=true))
@assert termination_status(m2) == termination_status(m) # failsI can however overwrite SCS.MOISolution: primal = backend(m).optimizer.model.optimizer.sol.primal
dual = backend(m).optimizer.model.optimizer.sol.dual
slack = backend(m).optimizer.model.optimizer.sol.slack
m3 = example_problem(Model(with_optimizer(SCS.Optimizer, eps=1e-10, max_iters=3, warm_start=true)))
m3.moi_backend.optimizer.model.optimizer.sol.primal = primal
m3.moi_backend.optimizer.model.optimizer.sol.dual = dual
m3.moi_backend.optimizer.model.optimizer.sol.slack = slack
optimize!(m3)
@assert termination_status(m3) == termination_status(m) # this works |
|
What is the termination status for |
m2 = example_problem()
set_start_value(m2[:x], solvedx)
set_start_value(m2[:y], solvedy)
set_start_value.(m2[:P], solvedP)
with_SCS_3iters = with_optimizer(SCS.Optimizer, eps=1e-10, max_iters=3, warm_start=true)
optimize!(m2, with_SCS_3iters)
termination_status(m2)compare with the same problem where start values are not set: m2 = example_problem()
# set_start_value(m2[:x], solvedx)
# set_start_value(m2[:y], solvedy)
# set_start_value.(m2[:P], solvedP)
with_SCS_3iters = with_optimizer(SCS.Optimizer, eps=1e-10, max_iters=3, warm_start=true)
optimize!(m2, with_SCS_3iters)
termination_status(m2) |
|
Isn't the difference between |
|
SCS will ignore primal when dual and slack are not set: So maybe the question should be: is it possible to set dual and slack from the level of MOI? |
|
sorry for dumb questions, but the documentation is rather sparse at the moment: I tried copying solution from with_SCS = with_optimizer(SCS.Optimizer,
eps=1e-10,
# max_iters=3,
warm_start=true,
verbose=false)
with_SCS_3iters = with_optimizer(SCS.Optimizer,
eps=1e-10,
max_iters=3,
warm_start=true,
verbose=true)
m = example_problem(Model(with_SCS))
optimize!(m)
m2 = example_problem(Model(with_SCS_3iters));
MOI.set(m2, MOI.VariablePrimalStart(), m2[:x], value(m[:x]))
MOI.set(m2, MOI.VariablePrimalStart(), m2[:y], value(m[:y]))
MOI.set.(m2, MOI.VariablePrimalStart(), m2[:P], value.(m[:P]))
# is this how you iterate over constraints?
for (F,S) in MOI.get(m.moi_backend, MOI.ListOfConstraints())
cnstrs = MOI.get(m.moi_backend, MOI.ListOfConstraintIndices{F,S}())
for c in cnstrs
dual_val = MOI.get(m.moi_backend, MOI.ConstraintDual(), c)
slack_val = MOI.get(m.moi_backend, MOI.ConstraintPrimal(), c)
MOI.set(m2.moi_backend, MOI.ConstraintDualStart(), c, dual_val)
MOI.set(m2.moi_backend, MOI.ConstraintPrimalStart(), c, slack_val)
end
end
optimize!(m2)however I still get warnings: |
|
Only |
|
as noted above implementing only (I see that |
Indeed, it seems we only need to add |
|
You're missing comma after Union; unfortunately I've already tried this and I get the same warnings (and results). note: F = MOI.SingleVariable
S = MOI.EqualTo{Float64}
MOI.supports(Model(with_SCS).moi_backend, MOI.ConstraintDualStart(), MOI.ConstraintIndex{F,S})still returns the defined |
|
Good point, we need to implement starting values for bridges |
fix copy-paste typos
|
Let's first check if it works if all the constraints are natively supported, i.e. |
until jump-dev/SCS.jl#138 is merged
|
@blegat what do you mean by natively supported? reading jump-dev/MathOptInterface.jl#528 I understand that we need to wait to remove VectorOfVariables support (as this will be transformed by |
|
Indeed, I was ambiguous, I meant natively supported by the MOI wrapper (i.e. no bridge are needed). So it should work if for a model like model = Model(with_optimizer(SCS.Optimizer))
@variable(model, x[1:2])
@constraint(model, [x[1] - x[2]] in MOI.Nonnegatives(1))
@constraint(model, [1; x] in MOI.SecondOrderCone(2))
@objective(model, Max, sum(x)) |
|
I have opened an issue for start values in the bridges: jump-dev/MathOptInterface.jl#684 |
Closes #136