Interior Point Method with Modified Augmented Lagrangian for Penalty-Barrier Nonlinear Programming
This work provides a theoretically grounded solver for general nonlinear programming problems, with convergence guarantees that address practical challenges like large penalty terms and ill-conditioned barriers.
The paper presents a numerical method for nonlinear programming that combines a modified Augmented Lagrangian technique for quadratic penalties and a primal-dual interior-point path-following technique for logarithmic barriers. The method achieves global and local quadratic convergence, and weak polynomial time-complexity for linear programs.
We present a numerical method for the local solution of nonlinear programming problems. The SUMT approach of Fiacco and McCormick results in a merit function with quadratic penalties and logarithmic barriers. Our NLP solver works by directly minimizing this merit function. In our method, we use different concepts that each shall aim at the efficient treatment of one respective special feature of this merit function. The features are: large quadratic penalty terms, and badly scaled logarithmic barriers. The quadratic penalties are treated with a modified Augmented Lagrangian technique. It enables large step sizes despite nonlinearity of the equality constraints. The logarithmic barriers we treat with a primal-dual interior-point path-following technique. We prove global convergence of the method and local quadratic convergence. We further prove weak polynomial time-complexity in the special case where the nonlinear program is a linear program. We also use a trust-funnel so to avoid that the method converges to any stationary points that are infeasible to the constraints.