Predictive Stochastic Programming: A New Class of Models and Algorithms
Abstract
This research project is intended to build fundamental mathematical concepts which arise in very large stochastic optimization, which happens to provide the foundations for developments in Artificial Intelligence and Machine learning (AI/ML). Because uncertainty is ubiquitous in several DoD applications, this project goes beyond AI/ML applications. In particular, this project focuses on designing optimization methods which employ AI/ML to improve both data science and decision-making supported by optimization. Our approach enables the use of previously observed Action/Reaction pairs (observations among adversaries) to enable optimization models to learn adversarial relationships and use such understanding to improve decision-making by incorporating a look-ahead feature within optimization models. By incorporating such recognition capabilities, our project can provide decision-support which is cognizant of potential reactions from adversaries. Predictive Stochastic Programming (PSP) is a formal mathematical approach which allows decision models to learn from prior engagements with an adversary, and to use that knowledge to recommend decisions which are more responsive, and more agile than is possible with previously known approaches such as ordinary stochastic programming (SP). While the latter paradigm (i.e., SP), provides the basis for our work, in its original form, it lacks the agility required to learn from previous experiences (i.e., observations). By adding this new predictive feature to SP, we can close-the-loop between decision cycles, thus enabling a more responsive set of recommendations.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 10, 2024
- Accession Number
- AD1228793
Entities
People
- Suvrajeet Sen
Organizations
- University of Southern California