Sampling Controlled Numerical Optimization
Abstract
The rapid rise of big data and machine learning has led to numerous problems from classical statistics receiving renewed attention due to their occurrence within modern application areas such as natural language processing,anomaly detection, text-mining, online sensor networks, and epidemic modeling. The heart of many of theseformulations is the data-driven simulation optimization (SO) problem ~ essentially an optimization problem where the objective and constraint function(s) are unknown but can be estimated either by drawing data from a large database, by numerical quadrature, or from repeated executions of a stochastic simulation.We propose to develop theory, methods, and implementations for the following three major algorithm classes to solve (virtually all) important flavors of data-driven SO problems.T.1 The Adaptive Sampling Line Search Method;T.2 The Adaptive Sampling Derivative-Free Trust-Region Method; andT.3 The Adaptive Sampling Interior Point Method.Each of T.1, T.2 and T.3 is rooted in an important algorithm class in the deterministic context, and results from replacing needed objects, e.g., derivatives, appearing within the class with their stochastic counterparts. The stochastic counterparts are in turn constructed using a crucial idea called adaptive sampling that is designed to automatically endow optimal convergence properties in an algorithm, by linking sampling decisions to algorithm trajectory. Adaptive sampling in T.1, T.2 and T.3 is accomplished through a simply stated guiding principle.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 03, 2017
- Source ID
- N000141712295
Entities
People
- Raghu Pasupathy
Organizations
- Office of Naval Research
- United States Navy
- University of Virginia