Fast and Scalable Stochastic Derivative-free Optimization
Abstract
Many engineering systems with naval capabilities are too complex to be optimized directly through analytical means. Instead, computer simulations are built and validated to replicate the behavior of these systems. One primary purpose of using simulations is optimal decision-making. Still, the challenge for solving such problems is that the underlying problem being optimized is in the form of a black box, and extracting direct derivative information from it to facilitate the optimization is either impossible or costly. Derivative-free stochastic optimization is a field that attempts to support decision-making in such a setting, particularly in continuous search spaces. Such optimization settings are relevant to situational awareness, electromagnetic maneuver warfare (EMW) technology calibrations, and unmanned vehicle mobility problems. The current derivative-free optimization tools are slow in higher dimensionsbecause of the additional effort they expend to draw information about the local behavior of the system, which is often uncertain. This proposal offers innovative tools from applied probability, statistics, and geometry to evolve the computational ability of derivative-free stochastic optimization tools for computationally efficient high-dimensional search. The research objective is to develop a sampling scheme integrated with a trust-region optimization engine that maximizes the information gain and balance between exploration and exploitation of the decision space to improve the algorithm s efficiency. To accommodate high-dimensional decision spaces, a novel coupling of random-subspace and sieve methods will be developed to reduce the computation by focusing on the most active subspace. For rigorous analysis and additional insight for implementation, we also establish a new analysis framework for the proposed algorithms using strong approximation theory. Coupled with the strategies to efficiently estimate the function values and approximate the best next steps, we will also address special nonsmooth SDFO cases for these novel ideas broader applicability. The successof this project provides new scientific knowledge for computing and large-scale data management that leads to enhanced real-time capabilities beneficial to the Office of Naval Research.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 13, 2024
- Source ID
- N000142412398
Entities
People
- Sara Shashaani
Organizations
- North Carolina State University
- Office of Naval Research
- United States Navy