PREDICTIVE STOCHASTIC PROGRAMMING: A NEW CLASS OF MODELS AND ALGORITHMS

Abstract

Stochastic Programming (SP) continues to represent some of the more applicable, yet challenging classes of formal decision models for constrained optimization under uncertainty. Its wide applicability for resource acquisition and deployment has continued to attract DoD/AFOSR attention over many years. However, there are certain handicaps which are inherent in this paradigm: a) uncertainty regarding parameters is represented under the assumption that a distribution of the random variables is available; b) even in cases where the uncertainty model is assumed to be data-driven, the data does not distinguish between predictors and responses. As a result, these models are not very responsive to information revealed via predictors; c) There is scant emphasis on model validation, so that inaccuracies in modeling can go unchecked, resulting in decisions which may not generalize to unobserved data. These issues are extremely relevant to DoD and AFOSR modeling, simulation and optimization. This proposal revolves around a new class of models where uncertain data is available in tuples representing predictors and responses. We will refer to the new class of models as Predictive Stochastic Programming (PSP). In this setting, one cannot assume that a distribution is available for all possible predictors; instead one must infer relationships among predictors and responses, and in addition, decisions must be made in a manner which reflects the relationship between predictors and responses. In this proposal, we present three new classes of models which we refer to as follows: i) coupled parametric PSP models, ii) non-parametric PSP models, iii) look-ahead PSP models. For each class of models, this proposal outlines DoD/AFOSR applications, models, and algorithmic challenges. These methods will provide decision support in situations where the data includes predictors (leading indicators) which will encourage fast-response simulations and decisionmaking, while ensuring a data-driven algorithmic and validation paradi

Document Details

Document Type
DoD Grant Award
Publication Date
Aug 12, 2021
Source ID
FA95502010006

Entities

People

  • Suvrajeet Sen

Organizations

  • Air Force Office of Scientific Research
  • United States Air Force
  • University of Southern California

Tags

Fields of Study

  • Computer science

Readers

  • Adaptive Control and Estimation with Uncertainty in Dynamic Systems.
  • Fluid Dynamics.
  • Operations Research

Technology Areas

  • AI & ML
  • AI & ML - Bayesian Inference
  • AI & ML - DoD AI Strategy
  • AI & ML - Machine Learning Algorithms