COMPRESSION AND RANDOMIZATION FOR EXTREME-SCALE TRAINING AND OPTIMIZATION (CREST OPT)
Abstract
Many applications of interest to the Department of Defense (DoD) can be formulated as simulation-aided decision problems that are commonly fraught with uncertainty. For example, autonomous aircraft in the Department of the Air Force Vanguard program Skyborg must detect potential air and ground threats, determine threat proximity, analyze imminent danger, and identify suitable options for striking or evading enemy aircraft . Beyond Skyborg, simulation-constrained optimization arises in most areas of science and engineering, including the control of ventilation systems to mitigate the spread of contagions such as COVID-19 and the design of electromagnetic materials and devices for use in surveillance and warfighter applications. A common thread between the above-mentioned applications is the need to efficiently optimize a dynamical system. Such optimization problems face significant challenges. Two primary solution approaches for dynamic problems are (i) maintain the dynamical system as an explicit constraint and the trajectories as optimization variables, and (ii) eliminate the dynamical system by including the trajectory variables as implicit functions of the controls or designs. The first approach often reduces the computational complexity by relaxing the dynamical system constraint at the cost of memory limitations since the entire trajectory must be stored in memory. In contrast, the second approach only maintains the control or design variables in memory, but requires the solution to a large dynamical system. This project will develop novel, rapidly converging algorithms for nonsmooth optimization problems including constrained, bilevel, and risk-averse problems. To ensure that the methods are computationally feasible for dynamic problems, the investigators will augment their algorithms with randomized preconditioning and compression techniques. If successful, the proposed work will enable scalable automated design, data analysis and optimization of highly nonlinear dynamical systems with uncertainty. It will develop a first-of-a-kind scalable technology for solving risk-averse optimization problems and optimization problems with dynamical system constraints. By taking a dynamical system view of neural networks, the algorithms will also enable scalable training of the network using massive data sets that are currently required for effective scientific machine learning. In total, the technologies developed in this project will enable numerical optimization for use in some of the most challenging problems across the DoD research portfolio.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 07, 2023
- Source ID
- FA95502210248
Entities
People
- Harbir Antil
Organizations
- Air Force Office of Scientific Research
- George Mason University
- United States Air Force