Massively Parallel Approaches for Buffered Probability Optimization and Applications

Abstract

This proposal explores a new probabilistic characteristic called buered probability of exceedance(bPOE). With bPOE, it is possible to count outcomes with value similar to a threshold value, ratherthan only outcomes exceeding the threshold value. When applying bPOE to large-scale problems,exploiting many-way parallelism and next generation computing paradigms is essential. We propose three distinct approaches to minimizing bPOE in large-scale applications that achieve massive parallelism. Our first approach uses the minimum formula for bPOE to transform the bPOE minimization problem into a nonlinear programming problem. This allows to exploit many-way parallelism through sample parallelism as well as linear algebra parallelism. Our second approach builds on the minimum formula and extends the progressive hedging algorithm to bPOE optimization. This approach decomposes the original optimization problem into numerous independent (i.e., the number of samples used to approximate bPOE) optimization problems that can be solved concurrently. Our final bPOE optimization approach exploits sensitivities of bPOE to accelerate convergence. This method will require ecient, i.e., parallel, computation of bPOE which motivates our research in parallel bPOE estimation algorithms using novel importance sampling methods.

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 28, 2018
Source ID
FA95501810391

Entities

People

  • Stan Uryasev

Organizations

  • Air Force Office of Scientific Research
  • United States Air Force
  • University of Florida

Tags

Fields of Study

  • Computer science

Readers

  • Data Mining and Knowledge Discovery.
  • Operations Research
  • Parallel and Distributed Computing.