PhD Oral Exam - Pedram Farghadani-Chaharsooghi, Industrial Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Casualty Response Planning (CRP) is a critical component of disaster management, where timely and coordinated medical intervention can greatly reduce mortality and improve outcomes for affected populations. Large-scale disasters often cause widespread damage to infrastructure, disrupt hospital operations, and generate uncertain and fluctuating demands across diverse patient groups. In these uncertain environments, effective planning requires advanced optimization techniques that can support both immediate and evolving decisions. In this thesis, we present a sequence of stochastic programming models and solution methods that address key operational challenges in casualty response, including the management of different classes of patients, blood inventory logistics, and hospital evacuation, with a focus on optimizing system-wide performance under uncertainty.

First, we present a two-stage stochastic mixed-integer programming model for CRP that jointly considers uncertain patient demand and hospital treatment capacity. In the first stage, we determine the location of Alternative Care Facilities (ACFs) and allocate resources such as rescue vehicles, physicians, and medical equipment. In the second stage, we assign patients with varying injury types to hospitals or ACFs and allow for transfers between facilities if needed. We also introduce an equivalent but more compact model that improves computational performance. To solve these models efficiently, we develop enhanced decomposition algorithms, including an accelerated branch-and-Benders-cut method enriched with Lagrangian decomposition, multi-cut strategies, and Pareto-optimal cuts. We demonstrate the effectiveness of our approach through extensive computational experiments and a case study based on the 2011 Van earthquake in Turkey.

Second, we extend the CRP framework to a multi-stage stochastic setting that reflects the dynamic and evolving nature of post-disaster operations. We develop a model that incorporates both apheresis and whole blood collection methods, deploys portable apheresis machines, and enables platelet transshipments among medical facilities. We treat patient demands, blood donor availability, and treatment capacity as uncertain parameters that unfold over time. Our objective is to optimize timely surgeries and compatible platelet transfusions, improving patient outcomes. To solve this problem, we implement a stochastic dual dynamic programming algorithm and enhance it using randomized quasi-Monte Carlo sampling, K-means++ clustering for scenario reduction, and strong valid inequalities. We evaluate our method through computational experiments and demonstrate its real-world potential using the Van earthquake case study.

Third, we develop a two-stage stochastic model for hospital evacuation, an often overlooked aspect of disaster response. In the first stage, we make strategic decisions regarding the location of ACFs, allocation of land and aerial rescue vehicles, and designation of backup hospitals. In the second stage, after uncertainty is revealed, we schedule patient evacuations and transfers over time, aiming to minimize evacuation risk, transportation risk, and operational cost. To solve this complex and large-scale problem, we propose a deep-reinforcement-learning-based matheuristic that integrates deep reinforcement learning with adaptive large neighborhood search. We preserve scenario diversity using decision-based clustering and improve solution quality using quasi-Monte Carlo sampling. Our algorithm delivers near-optimal solutions for small instances and scales efficiently to larger ones, outperforming benchmark methods such as Progressive Hedging.