Multistage Stochastic Programs with Dynamic Learning
Abstract
Many decision-making problems arising in various domains, such as logistics and transportation, defense applications, and energy systems management, are stochastic and dynamic, where the values of random parameters are revealed over time, and decisions are made sequentially with the available information. Despite advances in stochastic dynamic decision-making, most research assumes that random parameters are known a priori or their distributions belong to a set. In particular, they assume that the distributions for random variables are non-adaptive. While these assumptions are in part due to facilitating mathematical and computational analysis, they rarely hold in practice. This proposal introduces a novel paradigm for stochastic dynamic decision-making in the presence of uncertain parameters, where their distributions are adaptive and impacted by decisions. The proposed paradigm is amenable to mathematical analysis and allows sequential learning of random parameters from the observed impacts of past decisions. Novel computational solution methodologies, for which the available techniques do not apply, will be developed. The core of the proposed methodologies will be creating surrogate problems and providing computationally efficient and scalable algorithms. If successful, the proposed framework and solution methodologies can potentially address several problems the U.S. Department of Defense (DoD) troops face during military deployments and humanitarian relief efforts.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 06, 2025
- Source ID
- FA95502410241
Entities
People
- Hamed Rahimian
Organizations
- Air Force Office of Scientific Research
- Clemson University
- United States Air Force