Simulation-Based Decision Making with Streaming Data
Author(s)
Wang, Yuhao
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This dissertation studies simulation-based decision making in environments where data arrive sequentially over time. Many real-world decision problems involve complex stochastic systems whose performance cannot be evaluated analytically and must instead be estimated through simulation. At the same time, the stochastic inputs driving these systems are often unknown and must be inferred from data, which may be collected gradually over time. These challenges motivate the development of data-driven methodologies that integrate simulation, optimization, and learning under uncertainty.
This work develops frameworks and algorithms for incorporating streaming data into both simulation-based optimization and reinforcement learning. In the context of ranking and selection, we propose a data-driven approach that simultaneously allocates resources for simulation and input data collection, enabling efficient reduction of both simulation uncertainty and input uncertainty. For simulation-based bi-level optimization problems, we develop sequential pruning--optimization procedures that improve computational efficiency by eliminating suboptimal alternatives early while refining promising solutions, supported by statistical guarantees.
In the context of reinforcement learning, we study model uncertainty using a Bayesian risk framework. We propose algorithms that incorporate streaming observations to update posterior distributions of model parameters and learn policies that balance robustness and performance. The Bayesian risk formulation provides a flexible mechanism to account for uncertainty while avoiding excessive conservativeness. We further extend this framework to online learning settings, where data are collected adaptively through interaction with the environment, and develop algorithms with provable performance guarantees.
Overall, this dissertation contributes to the development of adaptive, data-driven decision-making methodologies that integrate streaming data, uncertainty quantification, and simulation-based optimization. The proposed approaches provide a principled foundation for solving complex stochastic decision problems in modern data-rich environments.
Sponsor
Date
2026-05
Extent
Resource Type
Text
Resource Subtype
Dissertation (PhD)