(426g) On the Implementation of a Structured-Exploiting Interior-Point Solver for Structured Nonlinear Programs

We describe the implementation of a fillter line-search interior point solver capable of exploiting multiple embedded structures at the linear algebra level (PIPS-NLP). We place special emphasis on the issue of inertia detection because we argue that this prevents the use of modular linear algebra implementations and thus hinders scalability. We prove that a simple test for curvature along the computed direction is sucient to guarantee global convergence. In addition, using CUTEr and energy optimization problems, we demonstrate that the strategy is as eective as inertia detection obtained through symmetric indenite factorizations in converging to second order points. We provide scalability results for our linear algebra implementation using stochastic optimal control problems of natural gas networks in which we exploit stochastic and reduced space structure.