This PR adds support for an ADMM solver for SLOPE. The implementation for sparse designs is currently performing badly. Also, fitting intercepts is done very crudely by adding a column of ones to X, which is fine for dense X but not really for sparse X.
I'm using the variable step size rule from Boyd's paper, but it is often not optimal.
Adaptive step size:
rho = 100
I'm sure there are other ways to handle step sizes better. Maybe someone else knows: @Klopfe, @mathurinm ?
This PR adds support for an ADMM solver for SLOPE. The implementation for sparse designs is currently performing badly. Also, fitting intercepts is done very crudely by adding a column of ones to X, which is fine for dense X but not really for sparse X.
I'm using the variable step size rule from Boyd's paper, but it is often not optimal.
Adaptive step size:
rho = 100
I'm sure there are other ways to handle step sizes better. Maybe someone else knows: @Klopfe, @mathurinm ?
Some other simulations: