Oscar Dowson
Oscar Dowson
#2826 removes the use of `constant`.
I'm going to guess not. I don't know that I've ever tested it
The assumption that all children are present and ordered has long bugged me. It shows up in a few places. Would be good to fix it properly.
This might be useful to add to a JuMP extension package for debugging and linting. I have, too often, manually traced the C API when building a MWE for Gurobi...
More changes. As evidence, it helped me get a MWE for https://github.com/jump-dev/HiGHS.jl/issues/202
This is probably related to the tolerances of the sub solvers. Have you tried experimenting with different tolerances for Hypatia and HiGHS?
Confirmed that changing the tolerance of HiGHS changes the solution: ```Julia julia> using JuMP, LinearAlgebra julia> import HiGHS, Hypatia, Pajarito julia> function main(mip_feasibility_tolerance) opt = optimizer_with_attributes(Pajarito.Optimizer, "oa_solver" => optimizer_with_attributes(HiGHS.Optimizer, MOI.Silent()...
If the MIP solver supports callbacks (like Gurobi), then Pajarito will use a single tree method: ```julia julia> using JuMP, LinearAlgebra julia> import Clarabel, Gurobi, Pajarito julia> function main() opt...
Note that tightening the `mip_feasibility_tolerance` made the solve _worse_ in this case. I don't know that I understand why.
Failures are because of a change in the `Random.MersenneTwister` stream between versions. I'll make a fix in this PR.