Optimizing First-Order Method Parameters via Differentiation of the Performance Estimation Problem
This thesis treats the problem of finding optimal parameters for first-order optimization methods. In part, we use the Performance Estimation Problem (PEP), a framework for convergence analysis of first-order optimization methods. The fundamental idea of the PEP is to formulate the problem of finding the worst-case convergence rate of a first-order optimization algorithm, as an optimization proble
