delete and add anonymous constraints in direct model seems to perform weird

[mzy2240]

I have a dict container of VariableRef and it looks like this

model[:a] = Dict{Int64,Vector{ConstraintRef}}()

and it is populated as below

model[:a][k] = @inbounds @constraint(model, [i = 1:num_of_blocks], x[i] <= mws[i] * (1 - u[i]))

where x is a vector of continuous vars and u is a vector of binary vars. When I tried to fix u JUST in this constraint container (instead of fix globally) by removing the constraints using delete(model, model[:a][k]) and then adding the "fixed" constraints again like

model[:a][k] = @inbounds @constraint(model, [i = 1:num_of_blocks], x[i] <= mws[i] * (1 - uv[i])) # uv is the value.(u)

I would get weird dual value (0.0) from an equality constraint in the relaxed model (the context is optimal power flow so you need relaxed model in order to calculate the system lambda). The correct dual value should be 30-ish and I can get the correct result if 1) I added the above "fixed" constraints as additional cut instead of deleting the original constraints, or 2) I dont delete and add any constraints. I know its very hard to reproduce based on the description, but I am just wondering if this delete&add anonymous constraints in direct model has some potential issues.

[odow]

I'm not aware of any issue related to this. I don't think we can debug anything without a reproducible example.

Does your model have a single unique dual solution? Do you have evidence that 0.0 is wrong?

[odow]

Is this the type of example you are talking about? I don't think this is a bug. If c is deleted then a change to the right-hand side of d does not change the objective.

julia> using JuMP, CPLEX

julia> begin
           model = direct_model(CPLEX.Optimizer())
           set_silent(model)
           @variable(model, x[1:3])
           @variable(model, 0 <= u[1:3] <= 1, Bin)
           @objective(model, Max, sum(i * x[i] for i in 1:3))
           d = @constraint(model, sum(u) == 2)
           c = @constraint(model, [i in 1:3], x[i] <= (1 - u[i]))
           optimize!(model)
       end

julia> uv = value.(u)
3-element Vector{Float64}:
 1.0
 1.0
 0.0

julia> unset_binary.(u);

julia> c2 = @constraint(model, [i in 1:3], x[i] <= (1 - uv[i]));

julia> optimize!(model)

julia> dual(d)
3.0

julia> delete.(model, c);

julia> optimize!(model)

julia> dual(d)
-0.0

[odow]

I think I'm going to close this issue as not reproducible.

The most likely scenario is that the dual is correct, and that there's just a misunderstanding about what the expected result is from replacing some variables in a model with constants. That's not the same model as fixing variable bounds.

But, if you have a reproducible example that demonstrates a bug, please comment below and I will gladly reopen this issue and investigate further 😄

[mzy2240]

Is this the type of example you are talking about? I don't think this is a bug. If c is deleted then a change to the right-hand side of d does not change the objective.

julia> using JuMP, CPLEX

julia> begin
model = direct_model(CPLEX.Optimizer())
set_silent(model)
@variable(model, x[1:3])
@variable(model, 0 <= u[1:3] <= 1, Bin)
@objective(model, Max, sum(i * x[i] for i in 1:3))
d = @constraint(model, sum(u) == 2)
c = @constraint(model, [i in 1:3], x[i] <= (1 - u[i]))
optimize!(model)
end

julia> uv = value.(u)
3-element Vector{Float64}:
1.0
1.0
0.0

julia> unset_binary.(u);

julia> c2 = @constraint(model, [i in 1:3], x[i] <= (1 - uv[i]));

julia> optimize!(model)

julia> dual(d)
3.0

julia> delete.(model, c);

julia> optimize!(model)

julia> dual(d)
-0.0

Hey @odow, thank you for the reply! Yeah this example is exactly what I encountered in the production, you are amazing! I just dont understand why we have c2 (a fixed c) now but dual(d) is 0.0 instead of 3.0. In my opinion, c2 alone should have the same result as c + c2 (when c and c2 co-exist in the model).

Edit: nvm I understand now. Thank you for the help!