In this dissertation, we aim to develop efficient algorithms with theoretical guarantees for several data-driven decision making problems. Specifically, we study the data-driven deci- sion making from three different perspectives: statistical learning, nonconvex optimization, and control of stochastic system. This dissertation contains three parts. In the first part, we study the problem of learning from pairwise measurements. We consider a semiparametric statistical model and propose a contrastive estimator that is invariant to model misspecification and attains the nearly optimal statistical convergence rate. In the second part, we investigate how data outsourcing can benefit algorithm initialization in nonconvex optimization. We propose algorithms that can utilize a small amount of outsourced data to find good initial points. Both in theory and practice, we demonstrate that our algorithms perform significantly better than the random start method. Lastly, we develop a sampling-based primal-dual algorithm to solve constrained Markov decision process. We show that our algorithm converges at the nearly optimal rate. We also apply the algorithm to two classic large-scale operations management problems: multi-product inventory management and multi-class queue scheduling. Numerical experiments demonstrate that our algorithm generates controls that achieve robust and superior performance compared with state-of-art heuristics.

Creator

• Chen, Yi

DOI

• https://doi.org/10.21985/n2-h5p1-w582

Subject

• Industrial engineering

Language

• en

Alternate Identifier

• etdadmin_upload_863393
• http://dissertations.umi.com/northwestern:15869

Date created

• 2021-01-01

Resource type

• Dissertation

Rights statement

• In Copyright