Erklären und Visualisieren von NLP-Lösungen

Abstract

Constrained optimization has many applications, including in robotics and design. A core problem with formulating optimization problems (i.e. specifying the costs and constraints of an NLP), is that it is hard to understand the solutions that an optimizer comes up with. The KKT conditions are differentiable, and in principle solutions should be ‘explainable’. This project explores fundamentally how optimization algorithms can not only output a solution, but also a (visual) explanation for this solution. That is, an explanation for why the found solution is a solution, or whatmake a certain configuration infeasible, where the costs come from, and how the solution would change if you modify parameters of the solution. The goal is to develop visualization techniques that allow the user to intuitively grasp solutions. In turn, this supports the user to more easily specify optimization problem.