Master's DegreesMaster of EngineeringMaster of Engineering ManagementMaster of ScienceAdmissions
All Thayer Events
PhD Thesis Proposal: Navid Rashedi
3:00pm - 4:00pm ET
For info on how to attend this videoconference, please email navid.rashedi.TH@dartmouth.edu
"Data-Driven Dynamic Decision-Making Using Discrete Optimization and Supervised Machine Learning"
In the recent years, the operations research community has started to focus on systematic ways to analyze and solve combinatorial optimization problems with the help of machine learning. Data driven optimization has a great potential to expedite solving online decision-making problems in particular. This study expands on these efforts toward developing fast and high-quality approaches combining machine learning and optimization for online decision-making in transportation and healthcare industries. Specifically, this thesis is a study of two problems—one in aircraft operations recovery and the other in diagnosis of occult hemorrhage —and develops solutions to these problems via data-driven approaches by leveraging machine learning.
First, aircraft recovery process attempts to repair a disrupted aircraft schedule by minimizing the total disruption costs throughout a given airline’s network. Existing exact methods are too time-consuming – constructing and solving the resulting integer linear programming problems requires too much time to be operationally useful. In this research, supervised machine learning is employed to expedite optimization identifying near-optimal solutions by leveraging their similarity with alternative (historical) problem instances presented in the offline model-training phase. This new approach based on binary classification methods is found to be highly effective in identifying the components of near-optimal solutions efficiently, as opposed to calculating them exactly using a standard commercial solver.
For the second part of this thesis, dynamic decision-making for hemorrhage diagnosis is studied. The challenge involves identifying the patients with occult bleeding at the earliest possible time to improve prognosis via early treatment. Despite significant effort directed toward development of noninvasive diagnostics, the performance of bedside technologies is currently inadequate. In the absence of early detection, the hemorrhage could progress to a point where permanent damage to vital organs is extremely likely and the prognosis of survival is poor. So, our proposed solution involves using an advanced set of sensors and algorithms to detect which patients will need advanced surveillance and/or treatment, before it becomes too late. This problem becomes especially challenging in resource constrained settings of mass casualty events (such as a major accident or a battlefield) where number of patients coming into an emergency department of a hospital can insufficient the availability of trained medical staff for advanced surveillance/treatment. This research combines machine learning and optimization to create effective new algorithms to deal with the challenge of hemorrhage detection in resource-constrained settings. We employ approximate dynamic programming to tackle the well-known curse of dimensionality in a large-scale setting. We develop new data-driven approaches to optimizing the machine learning parameters, model hyperparameters and the optimal risk thresholds to enhance medical decision making.
- Vikrant Vaze, PhD (Chair)
- Alexandre Jacquillat, PhD
- Eugene Santos, PhD
- Jonathan T. Elliott, PhD
For more information, contact Daryl Laware at firstname.lastname@example.org.