zl376 / admm_newton_con

Newton ADMM method with linear inequality constraints

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

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

alt text

About

Newton ADMM method with linear inequality constraints


Languages

Language:Jupyter Notebook 84.7%Language:Python 15.3%