zl376 / admm_newton_con

Newton ADMM method with linear inequality constraints

Newton ADMM method with linear inequality constraints

Note:

• Use penalty indicator for inequality constraint
• Split (canonical) loss and penalty with equality constraint
• Solve equality constrained problem using ADMM (alternating direction method of multipliers)
• Primal sub-problem is solved using Newton method (one step)
• Newton inversion can be done using direct inversion or CG (conjugate gradient)

Require:

• (Canonical) loss has explicit gradient and hessian.

An example for logistic regression