search

jump-dev / MatrixOptInterface.jl

An interface to pass matrix form problems
MIT License
11 stars 4 forks source link

Support quadratic objective in conic form #19

Open blegat opened 3 years ago

blegat commented 3 years ago

See https://github.com/oxfordcontrol/COSMO.jl/issues/126 This might be breaking as it changes the type

matbesancon commented 3 years ago

So we would represent the quadratic part with an abstract matrix right?

blegat commented 3 years ago

The objective would be an MOI.AbstractScalarFunction. The plan is to have matrix forms be parametrization of MOIU.GenericModel (see https://github.com/jump-dev/MathOptInterface.jl/issues/1261). It would be parametrized by 1) the objective type 2) the container for variable constraints 3) the container for other constraints Currently, in MOIU.GenericModel, the objective type is hardcoded as Union{MOI.SingleVariable, MOI.ScalarAffineFunction{T}, MOI.ScalarQuadraticFunction{T} however, this type should be a parameter of MOIU.GenericModel instead. Then, similarly to how https://github.com/jump-dev/MathOptInterface.jl/pull/1238 got ScalarAffineFunction generalized with the AbstactVector of terms, we could generalize ScalarQuadraticFunction with an AbstractMatrix of terms. One option could be to store the coefficients in a sparse matrix.

matbesancon commented 3 years ago

Given that in MatOI we take and store the elements of the optimization problem directly as matrices and vectors, I would be for the second option of having an abstract matrix