Random Methods for Large-Scale Optimization
Abstract
We propose to develop new class of optimization methods for problems with a large number of constraints arising in many application domains, most prominently in estimation and classification, optimal control, reinforcement learning and artificial intelligence, in general. We propose to develop and investigate a new class of algorithms, and to develop the necessarytheory that supports the proposed approach. The proposed methods use randomization in order to cope with the large number of constraints, where the constraint samples are generated on-thego.The objectives of this project are to:(1) Develop the necessary theory in support of the algorithmic development;(2) Investigate the role of symmetry structure in non-convex problems, and explore its potential advantages in the algorithmic design;(3) Design the algorithms for efficiently solving the problems with a large number of constraints, and provide their convergence and convergence rate analysis, including the performance bounds for their finite-time behavior, in terms of the problem properties and the parameters of the methods.If successful, the outcome of this research will be in the novel methods for efficiently solving a class of large-scale problems. As such, the research has a potential to impact large-scale data-driven applications such as machine learning, reinforcement learning and statistical inference. All of these applications are relevant to the Navy s missions where large collections ofdata are to be processed, such as large-area reconnaissance and monitoring, image reconstruction and language processing.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2021
- Source ID
- N000142112242
Entities
People
- Angelia Nedich
Organizations
- Arizona State University
- Office of Naval Research
- United States Navy