Learning to Solve Conditionally Convex Problems
Title: |
Learning to Solve Conditionally Convex Problems |
DNr: |
Berzelius-2022-222 |
Project Type: |
LiU Berzelius |
Principal Investigator: |
Paul Häusner <paul.hausner@it.uu.se> |
Affiliation: |
Uppsala universitet |
Duration: |
2022-11-01 – 2023-05-01 |
Classification: |
10201 |
Keywords: |
|
Abstract
This project aims to develop theory and algorithms for data-driven optimization schemes (learned optimization) that are specifically adapted for large-scale conditionally convex optimization problems and that preferably come with provable convergence guarantees. Large-scale conditionally convex problems pose a formidable challenge to the conventional optimization toolbox. They elude the---both theoretically and practically---attractive framework of convex optimization. On the other hand, standard approaches for non-convex problems often struggle to find a sufficiently accurate solution given a limited amount of time and compute. Thus, learning to optimize stands out as a highly promising approach, but it has proven difficult to fulfill its potential in a reliable and reproducible way.The key scientific challenge is thus to ensure that the learned optimizer generalizes.