152. Decision rules for sequential decision-making under uncertainty
Invited abstract in session FC-5: Plenary IV, stream Plenaries.
Friday, 12:15 - 13:15Room: E III
Authors (first author is the speaker)
| 1. | Merve Bodur
|
| The University of Edinburgh |
Abstract
Sequential decision-making emerges in a broad range of fields and is often impacted by uncertainty. Multistage stochastic programming (MSP) and multistage adjustable robust optimization (MSARO) are suitable modelling frameworks for sequential decision-making under uncertainty, among others. Those problems are theoretically and computationally challenging and, as such usually solved by means of approximations. In that regard, a commonly employed approach is decision rules (DRs), which restrict the policies to follow a certain functional form of the observed random outcomes. In this talk, we will review traditional as well as recently proposed primal and dual DRs for general MSP and MSARO problems. Our review will include two-stage DRs and Lagrangian dual DRs, which make solution algorithms designed for two-stage problems amenable to generating high-quality policies for multistage problems. In particular, for MSPs with potentially mixed-integer recourse, our review will also feature a Markov-chain-based variant of two-stage DRs to leverage the underlying stochastic process and a certain class of mixed-integer DRs for nonlinear MSPs with a large number of stages. The latter, by design, leads to smooth restricted problems (making stochastic gradient decent methods amenable) and highly interpretable decision policies. We will present numerical results on applications from various domains to illustrate the presented ideas.
Keywords
- Optimization under uncertainty and applications
Status: accepted
Back to the list of papers