Organizational Unit:

H. Milton Stewart School of Industrial and Systems Engineering

Permanent Link

https://hdl.handle.net/1853/70758

Parent Organization

Organizational Unit

College of Engineering

Includes Organization(s)

Organizational Unit

William M. Keck Virtual Factory Lab (VFL)

ArchiveSpace Name Record

https://finding-aids.library.gatech.edu/agents/corporate_entities/1032

Full item page

Publication Search Results

Now showing 1 - 10 of 1822

Performance Analysis of Large-Scale Load Balancing Systems

(Georgia Institute of Technology, 2024-12-11) Zhao, Zhisheng

Advanced cloud computing platforms, such as AWS, Azure, and Google Cloud, handle millions of requests per second. Efficiently assigning tasks across servers using a load balancing algorithm is critical for the seamless functioning of these systems. This dissertation focuses on performance analysis of large-scale load balancing systems. First, we analyze the JSQ policy in the super-Halfin-Whitt scaling window. We have shown that the centered and scaled process of total number of tasks in the system converges to a certain Bessel process. Also, its stationary distribution converges to a Gamma distribution. Then, we consider a heterogeneous model with the constraint of data locality and implement JSQ(d) policy. In this work, we investigate the graph structure with which the vanilla JSQ(d) policy can achieve the throughput optimality and improve the system performance greatly compared with the random assignment. In the third project, we extend our analysis to a more general heterogeneous system where the service rate depends on both types of servers and dispatchers. We proposed a new framework for analyzing the fully heterogeneous systems and designed two simple scalable delay-optimal routing policies. In the end, we discussed an application of the load balancing concept in the EV charging networks.
Statistical Estimation, Uncertainty Quantification, and Detection for Hawkes Processes

(Georgia Institute of Technology, 2024-12-08) Wang, Haoyun

Recently Hawkes process, also known as spatio-temporal self-exciting point process, has become popular for modeling event data. Its strength lies in its ability to capture complex dependencies between events, particularly the triggering effect, which is closely related to the concept of Granger causality in time series analysis. This thesis focus on the theoretical aspects of the statistical learning problem of such effect in the linear Hawkes process model in three chapters. The topics include (parametric) uncertainty quantification in a network setting, (parametric) quickest change-point detection, and estimation of triggering effect using a general kernel function which may in practice be represented by neural networks.
Enhancing Medical Decision Support Systems for Sepsis Patients in the ICU: Real-Time Detection and Algorithmic Bias Mitigation

(Georgia Institute of Technology, 2024-12-06) Smith, Jeffrey

The research outlined in this dissertation is comprised of machine learning (ML) and statistical techniques applied to a range of healthcare challenges. Implementing machine learning methodologies in the context of patient electronic health record (EHR) data involves navigating several complex factors. These include managing the unique structure of continuous physiological data and identifying patient health conditions in the absence of standardized definitions, all while being mindful of potential clinical biases impacting model outcomes. This work aims to address these intricate problems using interpretable ML and statistical techniques. The primary objective of this research is to enhance the fairness, transparency, and efficacy of medical-AI models.
Spatio-Temporal Event Modeling Through Deep Kernel-Based Point Processes

(Georgia Institute of Technology, 2024-12-05) Dong, Zheng

As the data volume and complexity in modern applications continue to grow, there is an increasing need in parallel for advanced point process models that can effectively capture intricate event dependencies and dynamics. This thesis focuses on advancing point process modeling by developing deep influence kernels for spatio-temporal event data. Combining statistical modeling principles with the expressive power of deep learning, the proposed methods effectively capture complex event dependencies, improve model estimation efficiency, and enhance interpretability. The thesis also demonstrates the practicality of deep kernel-based point processes in various real-world applications, such as in modeling COVID-19 transmission dynamics and urban crime events\footnote{Implementations are open-sourced at \url{https://github.com/McDaniel7}}. I hope that the contributions presented here will not only extend to a broader range of methodological and real-world applications but also inspire future research in the rapidly evolving area of spatio-temporal event modeling and neural point processes. For instance, when looking from a methodological standpoint, using neural networks as a flexible tool offers an opportunity to investigate more complex, higher-order statistics of point process models. On the application side, the use of neural networks allows the integration of comprehensive external data sources within the statistical frameworks, such as demographic or mobility data, to enhance the realism of real-world implementations. The neural point processes also have the potential to be adopted in controlled experiments for the study of variable effects, providing researchers with diverse options. These models could assist domain experts by suggesting new hypotheses derived from robust statistical perspectives, motivating interdisciplinary collaboration.
Incorporating Travel Behaviors into Transit Network Designs: Methods, Applications, and Extensions

(Georgia Institute of Technology, 2024-12-02) Guan, Hongzhao

Over the past few decades, urban areas have witnessed consistent population growth, leading to a significant rise in privately owned vehicles. Consequently, there's a growing need for new transit systems to address these challenges effectively. One such promising solution is On-Demand Multimodal Transit Systems (ODMTS), which integrate on-demand shuttles with fixed transit services to provide cost-effective and convenient transportation options. Chapter 2 investigates ODMTS from two crucial perspectives: network design and demand modeling. Advanced machine learning models are commonly employed for modeling demand, while optimization frameworks with fixed demand are typically used for network design problems. The chapter also discusses real-world ODMTS deployments, such as MARTA Reach in 2022 and CAT Smart in 2024. Chapter 3 studies on the ODMTS Design with Adoptions (ODMTS-DA) problem, aiming to incorporate choice models into optimization frameworks to handle latent demand while designing ODMTS. It proposes a path-based optimization model called P-Path to address computational difficulties, achieving significant computational improvements compared to existing approaches. Similarly, Chapter 3 extends the concept of ODMTS-DA to Transit Network Design with Adoptions (TN-DA) and designs heuristic algorithms to solve the problem efficiently. The chapter also provides guideline metrics for transit agencies and conducts extensive large-scale case studies on different transit systems. Lastly, Chapter 5 applies the concepts introduced in earlier chapters to a different domain—public school redistricting. It proposed a Contextual Stochastic Optimization framework and applies it to study the impact of redrawing elementary school attendance boundaries on socioeconomic segregation. Computational results reveal the effectiveness of the framework in predicting school choice and its potential to reduce segregation in schools, offering valuable insights for policymakers and academics.
Exploring Strength in Weakness: Multifaceted Weakly Supervised Learning Approaches for Telehealth and Industrial Quality Predictions

(Georgia Institute of Technology, 2024-12-02) Alenezi, Dhari F.

The rapid advancement in data collection technologies has significantly impacted the fields of telehealth and industrial manufacturing. However, the challenge of obtaining comprehensive and precise labels for this data poses significant difficulties for traditional machine learning models. This thesis addresses these challenges by proposing three novel weakly supervised learning (WSL) approaches, specifically targeting the domains of telehealth and manufacturing quality prediction. The first study introduces a Ranking-Based Weakly Supervised Learning (RWSL) model for assessing disease severity in telemonitoring, with a focus on Parkinson’s disease. Using data from the mPower app, this model integrates both labeled and ranked samples to improve predictive accuracy, overcoming the challenge of limited labeled data. By leveraging weak supervision, the RWSL model provides a more accurate and timely assessment of disease progression. The second study presents a Physics-Informed Weakly Supervised Learning (PWL) framework designed for quality prediction in industrial manufacturing processes. This approach integrates the physics-based understanding of manufacturing processes with machine learning models, improving prediction accuracy despite the scarcity of labeled samples. The PWL model bridges the gap between theoretical physics-based models and practical machine learning applications, enabling more effective quality control in manufacturing. The third study proposes a Multi-source Multi-task Weakly Supervised Transfer Learning (M2WeST) approach for telehealth, which addresses the heterogeneity of disease manifestations across patients. This model combines data from multiple patients, incorporating both strong and weak labels to provide personalized disease severity predictions. The M2WeST framework significantly improves the robustness and accuracy of predictions, even in scenarios with limited labeled data. These three studies contribute to the advancement of weakly supervised learning in both telehealth and industrial applications, offering innovative solutions for extracting meaningful insights from weakly labeled data.
Statistical Theory for Neural Network-Based Learning

(Georgia Institute of Technology, 2024-12-02) Ko, Hyunouk Andy

In Chapter 1, we introduce the general problem of binary classification and the significance of studying statistical properties of neural network-based classifiers. In addition, we include a high-level overview of the main results in this thesis along with a brief review of relevant literature. In Chapter 2, we provide the necessary technical preparations for the statement and proofs of the results in the rest of the thesis. Specifically, we define the classification problem and the space of neural networks, give a short introduction to concepts from rate distortion theory used in Chapter 4, and define the Barron approximation space used in Chapter 5. Furthermore, we conclude with a discussion of the relationship between regression and classification and related results from the literature. In Chapter 3, we show that random classifiers based on finitely wide and deep neural networks are consistent for a very general class of distributions. Consistency is a highly desirable property for a sequence of classifiers that guarantees that the classification risk converges to the smallest possible risk. This result improves the classical result of Farago and Lugosi (1993) by extending the consistency property for shallow, underparametrized neural networks with sigmoid activations to wide and deep ReLU neural networks without complexity constraints. In Chapter 4, we give several convergence rate guarantees of the excess classification risk for a semiparametric model of distributions indexed by Borel probability measures on [0, 1]d and regression functions belonging to L2 class of functions with finite Kolmogorov-Donoho optimal exponents. Furthermore, we give explicit characterizations of distributional regimes in which neural network classifiers are minimax optimal. In Chapter 5, we show that for a semiparametric model of distributions defined by regular marginal distributions and regression functions that locally belong to the Barron approximation space, neural network classifiers achieve a rate of $n^{(1+\alpha)/(3(2+\alpha))}$. We also show that this rate is minimax optimal up to a logarithmic factor.
Dynamic 3D Shape Modeling and Control for 3D and 4D Printing Processes

(Georgia Institute of Technology, 2024-11-14) Biehler, Michael

Author: Michael Biehler Advisor: Dr. Jianjun Shi Length of PhD Thesis: 150 pages Abstract: The world around us is comprised of dynamically evolving 3D shapes. For instance, think of 3D printing, where products are manufactured one layer at a time. The shape of each part depends on the previous layer’s 3D shape and additional operations in the current layer. Or think of landslides on mountains, which evolve based on the 3D topography and changing weather conditions. Advances in sensing technologies have made contactless scanning of these 3D shapes possible through acquisition devices such as laser and LiDAR scanners, resulting in unstructured 3D point clouds containing millions of data points. However, modeling the spatio-temporal 3D evolution of such phenomena, whether in 3D or 4D printing processes, poses significant challenges due to the large volume, permutation invariance, and unstructured nature of these 3D point clouds. To tackle these challenges, this doctoral thesis presents a series of methodologies for process modeling, control, and optimization, all grounded in the analysis of dynamically evolving 3D point clouds and heterogeneous inputs. The proposed methods have been implemented and validated in real-world systems. Specifically, this thesis explores three key topics to address the aforementioned challenges: (1) Nonlinear Dynamic Evolution Modeling of Time-Dependent 3D Point Cloud Profiles: Modeling the evolution of a 3D profile over time as a function of heterogeneous data and previous time steps’ 3D shapes presents a challenging yet fundamental problem in many applications. To address this, a novel methodology for the nonlinear modeling of dynamically evolving 3D shape profiles has been developed. This model integrates heterogeneous, multimodal inputs that influence the evolution of 3D shape profiles. Both forward and backward temporal dynamics are utilized to preserve the underlying physical structures over time. The approach leverages the theoretical Koopman framework to create a deep learning-based model for nonlinear, dynamic 3D modeling with consistent temporal dynamics. (2) Real-Time Control of Time-Dependent 3D Point Cloud Profiles: In modern manufacturing processes, ensuring the precision of 3D profiles is critical. However, achieving this accuracy is challenging due to the complex interactions between process inputs and the data structure of 3D shape profiles. To overcome this, a control framework for 3D profiles has been developed, which actively adapts and controls the manufacturing process to improve the accuracy of 3D shapes. Since 3D profile scans serve as the ultimate measure of part quality, using them as system feedback for control purposes provides the most direct and eAective approach. The eAectiveness of this framework is demonstrated in a case study on wire arc additive manufacturing. (3) Analysis and Optimization of Process Parameters in 4D Printing for Dynamic 3D Shape Morphing Accuracy: Additive manufacturing (AM), commonly known as 3D printing, has made significant advancements, particularly in the area of stimuli-responsive, 3D-printable, and programmable materials. This progress has given rise to 4D printing, a fabrication technique that combines AM with intelligent materials, adding dynamic functionality as the fourth dimension. Among these materials, shape memory polymers have gained prominence, especially for critical applications in stress-absorbing components. However, the accuracy of 3D shape morphing in 4D printed products is influenced by both the 3D printing conditions and the stimuli activation, making precise control challenging. To model and optimize the dynamic 3D evolution of 4D printed parts, a novel machine-learning approach that extends the concept of normalizing flows has been developed. This method not only optimizes the dynamic 3D profile evolution by refining the process conditions during both 3D printing and stimuli activation but also provides interpretability of the intermediate shape morphing process.
Strategic resource coordination for detecting illegal activity

(Georgia Institute of Technology, 2024-09-06) Bahamondes Pizarro, Bastian Matias

In an increasingly complex and interconnected world, ensuring security and resilience requires effective allocation of inspection resources to detect illegal activities. The evolving nature of threats, coupled with resourceful adversaries and limited inspection resources, makes it imperative to develop strategic inspection operations. Challenges include coordinating multiple resources, accounting for imperfect detection capabilities and asymmetric valuations of targets. New opportunities, such as advances in sensing technologies and data analytics, offer potential solutions to enhance the effectiveness of inspection operations. This thesis leverages game theory for the strategic coordination of inspection resources, focusing on Nash Equilibria (NE) as the main solution concept. It aims to provide valuable insights and efficient algorithms for inspection operations across various security domains. Chapter 2 examines a variant of the hide-and-seek game, motivated by the challenge of detecting smuggled commodities hidden by criminal organizations. In this game, a seeker inspects multiple hiding locations to find multiple items hidden by a hider. Each hiding location has a maximum hiding capacity and a probability of detecting its hidden items upon inspection. The seeker (resp. hider) aims to minimize (resp. maximize) the expected number of undetected items. We develop a two-step solution approach to compute NE for this zero-sum game. First, we solve a lower-dimensional continuous game to derive closed-form expressions for the equilibrium marginal distributions. Second, we design a combinatorial algorithm to compute mixed strategies that satisfy these marginal distributions. Our approach reveals novel equilibrium behaviors influenced by the complex interplay of game parameters and computes NE in quadratic time with linear support. Chapter 3 explores a nonzero-sum variant of the hide-and-seek game, driven by the asymmetric valuations that security agencies and criminal organizations place on the outcomes of their interactions. Here, a seeker inspects multiple locations with unit hiding capacities to find items hidden by a hider. Each location is associated with different utility values for the seeker and hider. The seeker (resp. hider) aims to maximize the utility from inspected (resp. uninspected) locations containing hidden items. We extend the previous two-step approach to obtain NE by deriving closed-form expressions for the equilibrium marginal distributions and computing compatible mixed strategies, resulting in a quadratic-time algorithm for solving this nonzero-sum game. Our analysis not only reveals complex equilibrium behaviors influenced by the players' asymmetric and heterogeneous valuations, but also addresses strategic interactions in various contexts beyond security domains, such as animal behavior and political campaigns. By offering both an intuitive analysis and an efficient solution method, this work bridges a gap in the study of equilibrium behavior in nonzero-sum games of strategic mismatch. Chapter 4 addresses strategic inspection problems in critical infrastructure resilience through a network inspection game, where a defender positions detectors on a network to detect multiple attacks on its components caused by an attacker. Each detector location has a probability of detecting attacks within its monitored components. The defender (resp. attacker) aims to minimize (resp. maximize) the expected number of undetected attacks. This model extends the hide-and-seek game of Chapter 2 by allowing for detection from multiple locations. To compute NE for this large-scale zero-sum game, we formulate a linear program with a small number of constraints and solve it using Column Generation. We provide an exact mixed-integer program for the pricing problem, which entails computing a defender's pure best response, and leverage its supermodular structure to derive two efficient approaches for obtaining approximate NE with theoretical guarantees: a Column Generation and a Multiplicative Weights Update (MWU) algorithm with approximate best responses. Each iteration of our MWU algorithm requires computing a projection under the unnormalized relative entropy, for which we provide a closed-form solution and a linear-time algorithm. Our computational results in real-world gas distribution networks demonstrate the performance and scalability of our solution approaches.
Design and Throughput Capacity Evaluation of Pickup and Dropoff (PUDO) Facilities

(Georgia Institute of Technology, 2024-07-27) Liu, Xinyu

Pickup and dropoff (PUDO) facilities have long been important to facilitate passenger transportation and goods delivery. Typical examples of passenger PUDO facilities are expected at inter-mobility stations or terminals where passengers switch between different transportation modes, outside the stadiums where large-scale sports events or concerts are hosted, and at schools where students are dropped off in the morning and picked up in the afternoon. Meanwhile, the last mile of the urban freight delivery system has gauged unprecedented and continuing attention since the booming of e-commerce and time-sensitive delivery services, leading to frequent pickups and dropoffs of goods close to their end consumers. The recent increase in the adoption of on-demand transportation provided by transportation network companies, and the anticipated continuation of this trend with the introduction of self-driving vehicles, imply that PUDO facilities will become more widespread and substantial. The current practice of mostly using curbsides for PUDO operations will not be sustainable as demand increases. This dissertation presents stochastic modeling and computational tools to evaluate the throughput capacity of PUDO facilities in various contexts. This dissertation also studies the optimal design decisions and characterizes near-optimal approximate policies to inform practical and implementable solutions in the real world. Chapter 1 presents a modeling and computational framework to evaluate layout designs and operational policies for passenger PUDO facilities, with realistic representations of the conflicting movements among vehicles and the resulting trade-off between space utilization and vehicle throughput rate. The results show the throughput capacities of PUDO facilities exhibit decreasing returns to scale, which are not captured by existing analytical approaches such as linear curb length models and queueing approximations. Several stochastic models are proposed to compute optimal and near-optimal throughput capacities. Renewal-reward process calculations are presented for single-lane facilities often encountered at schools and taxi ranks. For more general facility layouts, we propose a continuous-time Markov decision process to optimize the operational policies and to evaluate the performance of approximate policies. A microscopic trajectory-based simulation is developed to validate the stochastic models. Findings and computational insights are presented, including comparisons of facility layouts and operational policies, as well as insights into the behavior of congested systems. In Chapter 2, we study the throughput capacities of airport PUDO facilities. An airport serves as an interface between ground and air transportation, and therefore the efficient processing of ground transportation arrivals and departures is an important part of airport operations. At many airports, the current PUDO locations for taxis and other passenger cars are along the terminal curb or in existing parking facilities, and many of these PUDO facilities suffer from excessive congestion. The increased adoption of ride-hailing services has contributed to the growing use of on-demand ground transportation to and from airports, which aggravates congestion at terminal curbs. Since most airports are severely space-constrained, there is a need to consider PUDO facilities that are more efficient than terminal curbs, in terms of vehicle throughput per unit area. We consider the effect of the facility layout and operational rules on conflicts between the movements of different vehicles, the resulting delays in the movements of vehicles, as well as the spatial requirements of different layouts. We demonstrate the impact of mean service times, variability in service times and vehicle movement times, and operational rules on the relative throughput capacities of different facility layouts. Chapter 3 discusses computational models to evaluate the throughput capacity of accessible PUDO facilities with a mix of wheelchair-accessible and regular spots, making provisions for passengers with reduced mobility. Improving transportation accessibility and usability for passengers with reduced mobility has long been a major objective of transportation service providers and facility designers. The design of accessible transportation facilities is guided and enforced by legislation and regulations, such as the 2010 ADA Standards. With the growing use of mobility services involving pickups and dropoffs of passengers, there is a pressing need for the design and operations of PUDO facilities that are both efficient and accessible to mobility-challenged passengers. We compare the throughput capacities of different facility layouts and operational policies, as a function of facility sizes. The major decision variables are (i) the number of regular and accessible spots in the facility; and (ii) the operational policies to control the use of both types of spots by vehicles with and without handicap registration. We propose a continuous-time Markov decision process and show the optimality of a family of threshold policies analytically, such that accessible spots can be assigned to vehicles with and without handicap registration as long as a sufficient number of them are available. The optimal thresholds and accessible spot counts can be computed using a high-fidelity microscopic vehicle trajectory simulation and an efficient search procedure. Unlike parking facilities, our numerical results show the throughput capacity per unit area of accessible PUDO facilities can be increased by constructing a relatively large number of accessible spots and allowing vehicles without handicap registration to use them subject to the threshold policies. In other words, it is optimal or near-optimal to make nearly all spots accessible and flexible when vehicle service times are short. In Chapter 4, we consider the operational policy design problem for loading and unloading zones, also known as delivery bays in the urban context. The current and continuing surge in e-commerce and on-demand delivery services has contributed to a growing need for goods to be delivered directly to their end consumers. Furthermore, an increased demand for time-insensitive parcel deliveries as well as time-sensitive grocery and meal deliveries was stimulated during the COVID-19 pandemic when the mobility and movement of many people were limited or affected. One may expect the continuation of this trend due to the convenience and flexibility brought to consumers. This last step of goods delivery is often fulfilled at the curbsides or designated freight bays where delivery vehicles can be parked close to the final destinations. However, the sustainability and feasibility of exploiting the curbside to fulfill the delivery services are questionable, due to the various functions of the curb that compete over limited curb space. It is thus essential to envision designated spaces that serve as facilities for goods to be picked up or dropped off efficiently, with minimized effects on the local traffic. This work aims to provide computational models and methods that evaluate and characterize the optimal operational policies of loading and unloading facilities for goods delivery, passenger transportation, or a mix of both. The important operational decision to make is whether to admit or delay an arriving vehicle and where to assign an admitted vehicle. We specify the problem using a continuous-time Markov Decision Process and propose multiple low-fidelity models that are easier to solve and are upper bounds on the system performance. We motivate and propose near-optimal approximate policies that are computationally efficient. The approximate policies are validated using a microscopic trajectory-based simulation and compared against a benchmark policy. Lastly, Chapter 5 presents an empirical study to understand how facility layout design decisions impact the average vehicle enter and exit maneuver times. Real vehicle maneuvers are recorded at selected parking and PUDO facilities. Various factors including spot dimensions and angles, sizes of vehicles, and the occupancy of neighboring spots during the maneuvers are also collected. A regression analysis is conducted to delineate the impacts of these factors on the averages of different vehicle maneuvers. The results and insights from this empirical work may inform the optimization of PUDO facility layout design decisions.