Harnessing Parameterization For Fast And Reliable Nonconvex Optimization
Abstract
This project focused on developing novel understanding of large-scale, non-convex optimization problems by establishing robust notions of how the choice of parameterization affects the geometric and computational character of the optimization process. This understanding was used to create a methodological link between machine learning and optimal control, enabling a car to be successfully taught to drive around an unspecified track using vision-based control. Reparameterization also provided benefits for optimization of recurrent neural networks. Insights were gained into the construction of well-performing stable recurrent models for future used in machine learning.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 04, 2019
- Accession Number
- AD1090441
Entities
People
- Asuman Özdağlar
- Benjamin Recht
- Pablo Parrilo
Organizations
- University of California, Berkeley