Industrial engineering

Permanent URI for this collectionhttps://digital.lib.washington.edu/handle/1773/4932

Browse

Now showing 1 - 20 of 94

Using Multiple Models to Inform and Optimize Complex Systems
(2026-04-20) Morey, Danielle F.; Zabinsky, Zelda B
Modeling is crucial for decision-making in complex systems, especially when simulations or black-box models are computationally intensive. Traditional multi-fidelity optimization relies on a single most accurate model, but such a model is often unavailable in real-world scenarios with multiple, similarly uncertain models. This necessitates methods that leverage multiple models without assuming a hierarchy of accuracy. This dissertation introduces a multi-model optimization approach that utilizes the collective insights from all models by focusing on regions of consistent performance rather than a single ``best" model. The goals are to address limitations of traditional approaches, develop a methodology for models without clear accuracy ranking, and demonstrate its application in a real-world biomanufacturing system. Two applications involving complex systems are examined. The first is a military maritime communication network and the second is a variable-yield biomanufacturing process. These applications reveal the risks of assuming one model is more accurate than another. In response, the dissertation presents Set-Based Optimization with Multiple Models (S-BOMM), a framework that identifies subregions of consistently good performance across multiple models, instead of relying on a presumed hierarchy of model accuracy. Theoretical properties and empirical results are provided to further provide insight and illustration. Applying S-BOMM to the biomanufacturing application demonstrates its practical benefits, uncovering performance regions consistent across models and providing insights beyond single-model analyses. Overall, this work advances decision-making in complex systems by leveraging multiple models collectively, moving beyond traditional multi-fidelity paradigms.
Vigilance Decrement Measurement, Prediction, and Intervention Design in Safety-Critical Environments
(2026-04-20) Prendez, David; Kim, Ji-Eun; Boyle, Linda N
Vigilance is defined as the ability to remain attentive and alert over a sustained period of time, particularly toward a specific task. The deterioration of vigilance, also known as vigilance decrement, is often experienced by those completing tasks over an extended period of time. In high-risk domains, such as healthcare and driving, measuring vigilance decrement is particularly important given the resulting safety impacts. However, most research studies primarily use behavioral responses during vigilance tasks to assess vigilance levels (e.g., response time, hit rate). While these measures have proven to be effective, they are limited to post-hoc analyses only. Given recent advances in physiological sensing tools and predictive modeling strategies, it is important to also consider contextual factors and physiological indicators, allowing for more robust, continuous, and objective evaluation of vigilance decrement. For effective prevention and resolution to cases of vigilance decrement in safety-critical settings, interventions must be designed to be proactive and sustainable. This dissertation uses a combination of scoping reviews, healthcare-based studies, and a driving simulator study to answer the following research questions: 1) What contextual or contributing factors, together, affect vigilance levels in safety-critical settings?, 2) What combination of physiological measures can be used to monitor vigilance?, 3) What modeling features and strategies can be used to predict vigilance decrement?, and 4) What types of interventions can be used to maintain vigilance in a sustainable manner? Findings across the healthcare and driving studies revealed that vigilance is best predicted through a hybrid approach combining contextual factors and multi-sensor physiological measures. In the healthcare domain, sleep-related and environmental factors were significant predictors of vigilance, while heart rate, electrodermal activity, skin temperature, and eye movement-related metrics emerged as key physiological indicators. In the driving domain, driver experience, distraction susceptibility, and workload were the most influential contextual factors, while eye fixations toward a virtual meeting-based secondary task serving as the primary physiological predictor. Continuous measures of driving behavior were used as effective substitutes for peripheral physiological sensing. In the healthcare domain, probabilistic models with temporal elements, such as Dynamic Bayesian Networks, performed well for longer-duration vigilance tasks. In the driving domain, simpler linear models were most effective for single, critical event-based vigilance assessments. A scoping review of 52 studies identified 18 vigilance intervention strategies where sleep/breaks, ingestion, meditation/mindfulness, and task context changes represented the most sustainable options for real-world deployment. In the driving study, an emergency-style warning intervention produced significantly faster response times during the takeover event compared to the less urgent planned-style warning, demonstrating that salient, system-integrated alerts can serve as effective and sustainable interventions for enhancing vigilance prior to safety-critical events in semi-autonomous vehicles.
Planning and Operations with Decision-Dependent Uncertainty in Power Systems: Equity-Aware Charging and Wildfire Resilience
(2026-04-20) Zhao, Xinyi; Zhao, Chaoyue
This dissertation develops decision-making frameworks for planning and operations in power systems under decision-dependent uncertainty (DDU), motivated by two emerging challenges: equity-aware electrification of public transit and wildfire-driven reliability risks. In both settings, key uncertainties are not purely exogenous; instead, they are shaped by infrastructure investments or operational controls, creating feedback loops among decisions, system states, and future risk. On the planning side, we study the deployment of on-route fast-charging infrastructure for battery electric buses (BEBs). Unlike depot charging, on-route charging introduces stringent constraints on safe operation and local power supply capacity, and can substantially change distribution-grid operating costs. We propose an integrated planning framework that couples the bus route network with the power network and formulate the problem as a mixed-integer second-order cone program to minimize total cost, including charging equipment investment, power facility upgrades, and grid operation. To explicitly address transit equity, we introduce a fairness metric, the Regional Proportion of BEB Routes, and incorporate it into the planning model either through Jain’s index constraints or a Rawlsian objective, enabling planners to balance economic efficiency with equitable regional adoption. To capture long-term grid impacts from evolving BEB charging behaviors, we further develop a two-stage distributionally robust optimization model in which the load uncertainty set depends on siting and sizing decisions, yielding grid-aware investment plans that hedge against decision-dependent demand growth. On the operations side, we address wildfire resilience in distribution systems, where utilities must dynamically trade off service continuity against ignition risk. We model proactive switching decisions using a Markov decision process whose state transitions are governed by both exogenous wildfire conditions and endogenous operational effects. In particular, we adopt a DDU framework in which line-failure probabilities explicitly depend on power flow levels. To handle the resulting large state and action spaces, we propose an approximate dynamic programming method based on post-decision states to efficiently compute risk-aware switching policies that minimize total operational cost during wildfire events. Case studies on 54-node and 138-node systems demonstrate the scalability and effectiveness of the proposed approach. Together, these contributions provide a unified view of how DDU links long-term infrastructure planning and short-term operational control in power systems, and offer practical tools for enabling equity-aware transit electrification and wildfire-resilient grid operations.
Developing Interpretable Predictive Models of Driver Situational Awareness in Conditionally Automated Driving
(2026-02-05) Park, Sami; Boyle, Linda; Banerjee, Ashis
Situational awareness (SA) among human drivers is important to understand for the advancement of vehicle automation. In conditionally automated driving scenarios, where thevehicle may approach its operational limits, the driver will be required to resume control within an appropriate time frame. This would suggest that real-time information about surroundings needs to be provided to the driver in a meaningful way while ensuring they are continually aware of their road environment. The goal of this dissertation is to develop real-time predictive models of driver SA, focus- ing not only on performance but also on interpretability, providing users with insight into the driving context. The pursuit of this goal involves a structured approach, encompassing the following key steps: 1) Identification of driver-specific and environmental predictors, 2) Exploration of the relationships between three distinct levels of SA and the identified predictors, 3) Development of SA predictive models with various feature selection methods, and 4) Comparative assessment of the performance and interpretability of the developed models. To accomplish the initial two objectives and collect data for the subsequent steps, two driving simulator studies were conducted with 40 and 56 participants respectively. These studies provide insights on the predictors needed for the real-time predictive models as well as the complex relationships among different levels of SA (perception, comprehension, and projection) and the predictors. The data obtained from these studies serve as the foundational resource for objective 3: the development of real-time predictive models and Objective and 4: comparative analyses of the models in pursuit of the research goal.
Evaluating the Role of Saliency and Beliefs in the Recall of Corrections
(2026-02-05) Nandakumar, Archana; Rajivan, Prashanth
In today’s rapidly evolving online social network (OSN) landscape, health misinformation has emerged as a significant concern, especially in the aftermath of the pandemic. One effective educational strategy against health misinformation involves providing users with corrections that offer accurate, relevant facts. Although a large body of research has evaluated the effectiveness of corrections, there is still much uncertainty around the precise effects that correction messages have on reducing misinformation beliefs. Existing research acknowledges the importance of the underlying beliefs of users when it comes to the effectiveness of corrections. Several factors are found to affect the belief reinforcement and change processes such as the continued influence of misinformation, inattention, and use of intuition when evaluating facts. On the other hand, the aspects of corrections that are salient to a reader's memory are relatively understudied. In this dissertation, I investigate mechanisms grounded in the memory and decision-making literature, such as frequency and valence, that have been theorized to increase the salience of experience, and evaluate whether these mechanisms enhance the salience of corrections to misinformation in memory and how strongly they affect the recall of those corrections during misinformation judgments. Drawing inspiration from studies on human memory and learning, my initial experiments tested the hypothesis that exposing people to more frequent corrections would increase their availability and improve their ability to identify misinformation by making the corrections easier to recall during judgment. I conducted two laboratory experiments to test whether experiencing frequent corrections to misinformation improved participants' ability to discriminate between true and false news claims during extended extreme events like the COVID-19 pandemic. Results from both experiments indicate that increasing frequency of corrections may not improve the ability of participants to identify misinformation. The results also suggest that prior beliefs better determined the likelihood that individuals were likely to accept corrections. In subsequent experiments, I reoriented my study to another crucial aspect of memory salience: emotional salience. This line of inquiry sought to understand how individual preferences for content with varying degrees of emotional valence could improve the salience of corrections in memory. The goal was to design targeted corrections for health misinformation and study the effect of such corrections on recall of their content by human subjects. Finally, the results of the experiment were used to study the specific features of correction text that are likely to increase cognitive load on the user perusing the content. Through multiple experiments, this dissertation makes significant and original contributions in: experimental methodology and design, empirical findings in effectiveness of corrections, and insights into cognitive effects of textual features of corrections. Initial experiments demonstrated that merely increasing correction frequency was insufficient to increase the salience of corrections in memory. Results revealed that pre-existing individual beliefs were stronger predictors of correction effectiveness than intervention quantity. Subsequent experiments, focused on message quality and personalized design. The results showed that overall emotional targeting failed, but successful correction recall was significantly driven by individual working memory capacity. The findings from this research enables evaluation and application of correction strategies in realistic OSN environments. This research concludes by underscoring the need to move beyond statistical correlation to develop a mechanistic, individualized model based on a cognitive architecture to accurately design effective corrections.
Advancing Time Series Forecasting: Insights from Deep Learning and Dynamic Mode Decomposition
(2026-02-05) Salazar, Christopher Alexander; Banerjee, Ashis AB
Time series forecasting presents significant challenges across engineering and scientific disciplines, particularly in handling non-stationary real-world data and providing real-time predictions from streaming sources. Deep learning approaches, including Recurrent Neural Networks (RNNs), Convolutional Neural Networks (CNNs), and Transformer-based models, have advanced the field but often fall short in interpretability, computational efficiency, and real-time adaptability. Despite their capacity for modeling complex non-linear dynamics, these models require extensive hyperparameter tuning and lack robust mechanisms for incremental updates. They also suffer from catastrophic forgetting in streaming scenarios, limiting their deployment in dynamic and resource-constrained environments. This dissertation addresses these limitations through two complementary research directions: enhancing deep learning interpretability through distance correlation analysis and developing efficient Dynamic Mode Decomposition (DMD) methods for batch and streaming forecasting. First, this work introduces a distance correlation-based framework to examine the internal mechanics of RNNs in time series forecasting. This versatile metric enables systematic analysis of information flow through RNN activation layers, revealing how these networks process temporal dependencies. Empirical analysis demonstrates that RNN activation layers effectively learn lag structures in early layers but progressively lose this temporal information in deeper layers, degrading forecast quality for series with large lag dependencies. The study further reveals fundamental limitations in RNN capabilities for modeling moving average and heteroskedastic processes. Distance correlation heatmaps provide visual comparisons across architectures and hyperparameters, demonstrating that input window size influences model behavior far more than conventional hyperparameters such as hidden units or activation functions. These findings enable practitioners to assess RNN suitability for specific time series characteristics without extensive trial-and-error experimentation. The second direction introduces novel DMD-based forecasting methods that address deep learning limitations. For batch scenarios, Incremental Kernel Dynamic Mode Decomposition (IKDMD) enhances adaptability and efficiency by integrating incremental kernel singular value decomposition and randomized linear algebra into the kernel DMD framework. Comparative analysis across real-world datasets demonstrates that IKDMD outperforms state-of-the-art deep learning methods, particularly for highly non-stationary and volatile data, while providing interpretable eigenvalue diagnostics unavailable in black-box neural networks. For streaming applications, this dissertation presents Windowed Online Random Kernel Dynamic Mode Decomposition (WORK-DMD), which integrates Random Fourier Features with online DMD to enable real-time forecasting from continuously arriving data. By employing explicit feature mappings rather than implicit kernel methods, WORK-DMD achieves fixed computational complexity per update while capturing nonlinear dynamics. Its adaptive windowing mechanism naturally handles non-stationary dynamics without catastrophic forgetting. Experimental evaluation across benchmark datasets demonstrates remarkable sample efficiency, requiring only single-pass learning while achieving competitive or superior accuracy compared to deep learning methods that demand multiple training epochs and extensive sample exposures. This efficiency translates to reduced computational costs, faster deployment, and viability for resource-constrained edge devices. Together, these contributions advance time series forecasting by providing both diagnostic tools for understanding deep learning limitations and computationally efficient alternatives that balance accuracy, interpretability, and real-time adaptability. The methods presented enable practical deployment in scenarios where traditional deep learning approaches struggle with sample efficiency, computational constraints, and evolving data dynamics.
Improving Healthcare Resource Allocation within a Geographic Area through Optimization and Machine Learning
(2025-10-02) Wang, Yinsheng; Liu, Shan
Healthcare policymakers face critical challenges in resource allocation that require: (1) practical decision-support tools, (2) robust prescriptive models that account for uncertainty, and (3) integration of multiple data sources for predictive analytics. Addressing fairness and social vulnerability is particularly crucial in healthcare resource allocation decisions. This thesis presents novel optimization and machine learning methodologies to tackle these challenges in resource-limited healthcare settings, including point-of-care testing allocation in western Kenya and mental health prevalence estimation in Washington state. The primary contributions include practical optimization models for real-world healthcare resource allocation and innovative data fusion approaches for enhanced decision-making. These methodologies provide healthcare policymakers with critical insights for informed, equitable, and effective healthcare planning decisions. This research comprises three interconnected components that comprehensively address healthcare resource allocation challenges. The first component develops decision-support tools for the strategic placement of point-of-care HIV viral load and drug resistance testing machines in Kisumu County, Kenya, optimizing resource distribution in high-need areas. The second component formulates and solves a queueing-location-allocation model using integer programming and Conditional Value at Risk (CVaR) to manage demand uncertainties in testing samples, enhancing healthcare delivery system resilience. The third component advances data fusion techniques by integrating optimization and machine learning to estimate pediatric mental health prevalence in Washington state, merging multiple national surveys with local datasets. Together this work enables healthcare policymakers to make evidence-based and strategic resource distribution decisions.
The Interplay of Optimization and Machine Learning to Solve Large-scale Black-box Noisy Functions
(2025-10-02) Maneekul, Pariyakorn; Zabinsky, Zelda B.
High-dimensional black-box optimization presents an increasingly prevalent challenge in modern science and engineering. This dissertation addresses this challenge through a novel interplay between optimization and machine learning methods, developing adaptive search algorithms that strike a balance between exploration, exploitation, and estimation. The proposed algorithms leverage machine learning techniques to construct surrogate models, thereby enhancing the efficiency of the optimization process. The dissertation proposes a multi-level Partitioning and Branch-and-Bound (PBnB) algorithm designed for level-set approximation, enhancing the original PBnB algorithm significantly. This multi-level PBnB algorithm employs importance sampling to strategically identify promising subregions of the partitioned search space. Its performance is further enhanced by integrating Gaussian processes as a surrogate model to guide local sampling exploration. During the process, the target level set is approximated by classifying subregions as either pruned (no intersection with target level set), maintained (contained within target level set), or undecided. This enhanced version of the PBnB algorithm introduces an adaptive sampling probability that strategically directs samples to the most promising regions. Since this importance sampling results in dependency amongst samples, we have applied a statistical method to construct a confidence interval on the probability of correctly classifying a subregion as pruned or maintained. The contribution to the interplay of optimization and machine learning is the local sampling within each subregion. We incorporate Gaussian processes and regularized quadratic regression, common and successful methods for prediction in machine learning for level-set approximation. The analysis of this multi-level PBnB algorithm quantifies the quality of the level set approximation by deriving probability bounds on the volume of incorrectly pruned or maintained regions, which accounts for the effects of importance sampling. To address the challenges of high dimensionality, this dissertation introduces the Branching Adaptive Surrogate Search Optimization (BASSO) framework that conceptualizes the use of branching and surrogate modeling for black-box optimization. BASSO generalizes multi-level PBnB and adapts it to optimization as opposed to level-set approximation. A finite-time analysis of BASSO proves that the expected number of BASSO function evaluations needed to first sample a point in the global optimum vicinity is linear in dimension given that two strong assumptions are satisfied. The desiredlinearity result suggests an algorithm that is scalable to high dimensions in theory. This research explores several variations to implement BASSO and partially satisfy the two assumptions. In this part of the research, methods used in machine learning are introduced to improve the chance of sampling in the improving region. One BASSO implementation incorporates Gaussian processes as a surrogate model and a second uses regularized quadratic regression as a surrogate model to predict where to sample next within a subregion. The synergy between the surrogate model and the optimization algorithm work together to balance exploration and exploitation. The local surrogate model guides sampling within a subregion, while the adaptive subregion probabilities identify promising subregions. This interplay allows the system to effectively use both local subregion information (from the surrogate model) and global information (from the adaptive probabilities) to improve its search. Numerical experiments of BASSO provide insights into the gap between theoretical ideal performance and the performance of proposed implementation with machine learning techniques to tackling high dimensional black-box problem. This dissertation also explores partitioning, clustering and decomposition as techniques for high-dimensional optimization. While the proposed multi-level PBnB algorithm and BASSO framework focus on balancing exploration and exploitation for deterministic, black-box optimization, this dissertation also considers estimation when dealing with a noisy black-box function. The dissertation extends the Single Observation Search Algorithm (SOSA) by incorporating insight from machine learning techniques. The original neighborhood averaging technique for noisy function value estimation of SOSA is replaced with a new quadratic regression, extending the concept of basis expansion. Complementing this, the search strategy is improved by incorporating optimistic sampling, a concept drawn from reinforcement learning, to more effectively guide exploration. This research contributes to the interplay of optimization and machine learning by providing quadratic regression as an estimation method within a single-observation scheme and achieving convergence results while accounting for dependency between samples. Theoretical convergence results for this SOSA extension are presented, and numerical experiments on benchmark problems demonstrate performance gains over the baseline algorithm. Finally, this dissertation identifies possible applications and future research opportunities arising from the interplay of optimization and machine learning in solving large-scale black-box noisy functions. This includes a discussion of quantum computing approaches for global optimization, considering both their theoretical promises and practical challenges.
Evaluating text representations in cognitive models of human decision making: The case of phishing attacks
(2025-10-02) Mehrabi, Elaheh; Rajivan, Prashanth PR
Phishing attacks use deceptive emails to manipulate individuals into revealing sensitive information or performing harmful actions. They continue to remain a widespread threat because they exploit human cognitive vulnerabilities rather than technological weaknesses. Unlike hardware or software weaknesses that can often be patched, human susceptibility to phishing cannot be intervened with a one-size-fits-all anti-phishing training solution. People differ in the types of phishing threats to which they are susceptible, which require personalized and adaptive anti-phishing training solutions capable of tailoring learning experiences to individual needs and context. Cognitive models could enable analyses of cognitive processes predictive of individuals susceptibilities to phishing threats. They could enable anti-phishing training solutions to identify gaps in individuals' knowledge about phishing threats, assess people's ability to recall training when faced with real threats, and estimate how well individuals can generalize training received to recognize new types of attacks. However, a key challenge in the developing such cognitive models lies with instance representation. In the context of models of phishing decision making, it would be the information communicated through email text. It remains unclear how people encode and recall past email conversations when making decisions about phishing attacks. This dissertation addresses this critical gap by investigating how different text representations affect cognitive models of human decision-making in phishing contexts. Specifically, it explores how people encode and recall textual information from emails when deciding whether to engage with or ignore them, and how to generate representations aligned with human memory processes for use in computational cognitive models. My research began by integrating existing cognitive models into a simulated environment emulating user interactions with phishing emails. Early experiments revealed that these models struggled to capture the processes by which humans recall and interpret complex email content when making response decisions, highlighting a fundamental limitation in current modeling approaches. As a first step toward addressing this gap, I developed a privileged learning framework that was trained on a set of high-level cues derived from human-labeled data. Once trained, the model was able to generate similar representation cues for new, unseen emails that lacked these annotations. This approach demonstrated the potential to bridge the gap between raw email text and more abstract decision-making features, while also highlighting the need for deeper analysis of how different representation methods impact cognitive modeling. Motivated by these findings, I systematically evaluated two broad categories of text representations within the instance-based learning (IBL) framework. The first category included lower-order, text-based representations, such as embeddings of full sentences, embeddings of email summaries from generative AI, and recall-based keywords. The second category encompassed higher-order, intent-driven representations capturing cues like tone, sentiment, and inferred communicative intent—features inspired by speech act theory hypothesized to better reflect how humans remember the intent behind what was communicated through an email.Through extensive evaluation within the instance-based learning theory framework, my findings demonstrate that higher-level, abstract representations—such as inferred intent, sentiment, and tone—more accurately model human decision-making than full-text embeddings or even embeddings based on email summaries. These representations not only improved prediction performance but also enhanced robustness when simulating users whose behaviors were less consistent or dissimilar from the majority, suggesting greater generalizability across diverse individuals. In a subsequent exploratory effort, I implemented IF-THEN production rules based on intent representations within the ACT-R cognitive architecture to examine how rule-based cognitive modeling could enhance model alignment with human responses. While preliminary, this effort highlighted the potential of explicit rule-guided frameworks to support future human decision-making modeling directions and offer interpretable decision mechanisms. Together, these contributions advance our understanding of how cognitively grounded representations of email content influence the modeling and analysis of phishing decisions. They underscore the value of aligning computational models with human memory processes to support the development of personalized, adaptive, and cognitively-informed anti-phishing interventions.
Toward Enhancing Multitasking Performance: Modeling, Prediction, and Intervention Design
(2025-08-01) Li, Jiaxin; Kim, Ji-Eun JE; Boyle, Linda LN
Multitasking involves performing more than one task in parallel or in a serial manner. Improving multitasking performance is particularly crucial in safety-critical environments, where performance decrement can lead to serious or even life-threatening consequences. For example, engaging in secondary tasks while driving, commonly known as distracted driving, is one of the main contributors to traffic accidents. In the field of aviation, several fatal aircraft crashes have also been linked to pilots' errors during multitasking. Despite the crucial need to enhance multitasking performance, existing studies continue to rely on retrospective behavioral measurements, which are insufficient for continuously tracking and predicting individuals’ multitasking performance. Furthermore, despite the widespread use of automation in human-system interactions, there is a lack of research on how to design automation in multitasking environments. This dissertation aims to answer four research questions: 1) Which factors impact multitasking performance in practical scenarios? 2) Which neurophysiological responses indicate changes in multitasking performance? 3) How can multitasking performance be estimated over time using probabilistic modeling? 4) How multitasking performance can be enhanced by automation? To answer these research questions, a mix of survey, behavioral, and neurophysiological data recorded from both controlled experiments and field studies was used. The findings of this dissertation can be applied in safety-critical settings to reduce operators’ multitasking errors, enable timely and effective interventions, and ultimately enhance multitasking performance and mitigate safety risks.
Data-driven approaches to disaster preparedness: Integrating natural language processing and machine learning for enhanced infrastructure system resilience
(2025-08-01) Lensing, Julia Carier; Choe, John Y
In an era marked by intensifying disasters caused by natural hazards, improving resilience and preparedness has become increasingly important for safeguarding communities and the critical infrastructures on which they depend. Earthquakes, in particular, pose severe threats to human well-being and the integrity of critical infrastructure systems, such as bridges, which serve as key conduits for transportation, emergency response, and economic continuity during and after crises. The rapid acceleration of data generation presents opportunities and challenges; if effectively leveraged, diverse and complex data sources can support more robust, data-driven decision-making for disaster preparedness. This dissertation presents three analytically rigorous and adaptable frameworks to address key challenges using high-dimensional, real-world data to improve disaster preparedness and infrastructure assessment. This dissertation first presents a text mining framework and an accompanying open-source code designed to extract insights from a novel corpus of practical disaster reports. This chapter highlights the ability to synthesize lessons from past disasters, providing observations that inform future preparedness initiatives. Second, this work introduces a machine learning-based framework for predicting key bridge characteristics related to seismic vulnerability. This framework reduces the need for manual data collection while increasing the availability of network-level characteristics, thus allowing for a more comprehensive understanding of infrastructure resilience. Third, this work proposes an integrated framework that combines natural language processing, feature engineering, and uncertainty quantification to improve the reliability and interpretability of seismic vulnerability assessments based on small, information-rich datasets. Together, these contributions lay the groundwork for future research and practice in disaster resilience, particularly in contexts characterized by limited data availability and high data complexity. By demonstrating how natural language processing and machine learning can be used to harness textual, structured, and high-dimensional data, we aim to advance intelligent, data-efficient methods that support more informed and effective decision-making in disaster preparedness and infrastructure system management.
Improving Effectiveness and Equity of Healthcare Delivery through Systems Optimization
(2024-10-16) Sun, Xiaonan; Liu, Shan
Effective and equitable healthcare delivery is crucial for advancing health outcomes, reducing resource waste, alleviating healthcare disparities, and improving overall individual well-being and community welfare. With increasing costs, limited resources, growing demand for patient-centered services, and advancements in remote technology, resource allocation has gained significant attention as a key strategy to optimize care delivery. The objective of this dissertation is to improve the effectiveness and equity of healthcare services through developing decision-analytic, machine learning, and optimization models using patient-level data, with a focus on both remote and in-person healthcare settings. In remote care settings, we explored how technologies could enhance healthcare resource utilization for chronic disease management. Remote monitoring has emerged as a promising option with high personalization and adaptability. However, the cost-effectiveness of these technologies remained uncertain. We used chronic depression as a case study and evaluated the cost-effectiveness of remote monitoring strategies compared to rule-based follow-up and fixed-frequency follow-up strategies. We developed a decision-analytic Markov-cohort model to simulate disease progression for patients with different risks, incorporating optimal treatment switching. Results showed that remote monitoring technology can be cost-effective and identified requirements for it to work more effectively. It provided a novel assessment framework that can guide the development of emerging technologies and highlighted the bright future of improving care delivery through remote monitoring. In in-person care settings, we aimed to optimize trauma care delivery, given its critical role in emergency healthcare. We began by investigating the variability in care delivery within statewide trauma systems. Hospitals are designated as trauma centers (TCs) with level I-V, or non-trauma centers (non-TCs), based on their medical and research resources. To explore trauma care delivery patterns and their association with trauma designation levels, we performed three sets of unsupervised clustering analyses on statewide TCs and non-TCs based on hospital features with a focus on surgical care. We found that the resulting clusters only partially aligned with the TC designations, implying not all hospitals with the same TC level provide equivalent care. The results highlight the performance variability and help us better understand trauma system functioning, guiding the subsequent study to optimize the trauma system at the hospital level. To optimize statewide trauma systems, we developed a systematic framework for improving care quality while addressing population equity. This objective is achieved by establishing and assigning hospital profiles representing performance targets which can be used to guide resource allocation and operational adjustment decisions. While many studies have focused on optimizing emergency transport services, care quality and equity have often been overlooked. Using state data, we established a set of comprehensive trauma care quality metrics for distinct population groups formed by sociodemographic factors and Injury Severity Score (ISS). We then created a quality index to represent trauma care quality accounting for hospital variations using a Principal Component Analysis (PCA) analysis of the quality metrics. Next, we created hospital profiles using a quality index of each population group, which were estimated from data and imputed using a linear mixed-effects model. We formulated a mixed-integer linear program (MILP) to maximize the quality index of targeted population groups under various equity objectives. The model identified optimal hospital profile assignments as proxies for performance targets for the hospitals. These results help identify necessary resources for performance enhancement, guiding hospitals in making targeted improvements to better serve diverse patient populations. Overall, this dissertation advances healthcare effectiveness and equity by evaluating remote care technologies, uncovering variability in trauma systems, and establishing optimal performance targets for hospital trauma care delivery. Our findings offer actionable guidelines to enhance chronic disease management in remote settings and improve the quality and equity of statewide acute trauma care systems.
Empirical Data in Disaster Recovery: A Data Pipeline to Investigate a Pandemic's Impact on Community Mobility
(2024-09-09) Martell, Matthew; Choe, Youngjun
Disaster events are becoming an increasingly common part of life for people across the world. Climate and weather-based disasters are on the rise, in addition to higher probabilities of future disease emergence. One of the main ways we understand past and current events and work to mitigate the effects of future events is through the use of data. Data, when used appropriately, can help us understand the possible impact of disaster events and our mitigation efforts. Data’s impact is maximized when data sets and methods are shared. Unfortunately, there is a dearth of understanding of what data are available in the disaster research field, leading to slow progress in community resilience efforts. This dissertation provides a greater understanding of the data being used in the disaster recovery space, and generates a new data set on community recovery from COVID-19 using a novel method. First, we conduct a review of data use in lifeline infrastructure restoration modeling. The review covers publications from the 1980s to 2019. It describes both the types of data used and their features, in addition to breakdowns of how different modeling approaches obtained and utilized data differently. The review concludes with some discussion of future research directions for the field and data management best practices, most notably data publication and documentation for fully reproducible methods. Next we create methods for utilizing street-view images to study community mobility during disaster events. The framework is an open-source data pipeline for street-view images. It demonstrates the data management best practices from the review while showing it is possible to gather insights into the effects of community events and public policy on foot traffic using the street-view images. The framework is usable to detect any type of object where an appropriate computer vision algorithm is readily available, and has applications in disaster recovery, urban planning, and even demographics estimation. Our implementation uses the images to count pedestrians and uses the count as a metric for community mobility. Additionally, the entire data set and a sample code base are available for reproducibility, or use in other research efforts. The final chapter of this dissertation implements a distance estimation algorithm on the street-view image data set to understand how outdoor social distancing practices changed over the course of the pandemic. The analysis includes vaccine availability, weather, and time of the week as predictors in addition to socioeconomic factors at the census tract level. In addition to looking at the city of Seattle as a whole, this analysis also subsets the data into various places of interest such as schools or hospitals to understand the distancing trends at these locations.
Measuring and Predicting Driver Situation Awareness
(2024-09-09) Xing, Yilun; Boyle, Linda Ng
Situation awareness (SA) encompasses the perception, comprehension, and projection of elements within a given situation, aligning with the three levels of SA. It plays a critical role in driver safety, as drivers must continuously maintain awareness of dynamic road conditions, significantly influencing traffic crash rates associated with human error. Advanced driver assistance systems (ADAS) are designed to enhance driving performance. However, overly complex or ambiguous information may overwhelm drivers' limited SA capacity. To address this challenge, ADAS models should consider the operator's perception and understanding of the environment to tailor assistance effectively. This dissertation proposes methods to measure and predict driver SA, focusing on objects of interest within the scene. Initially, an experimental approach utilizing real-world driving videos and a web-based touch recorder was introduced to capture driver SA, which demonstrates efficacy in capturing various levels of driver SA. Furthermore, it was observed that drivers often rely on the memory of element trajectories to understand or predict the location of OOIs. Subsequent analysis identified the impacts of environmental features, object characteristics, and driver demographics on driver SA. Incorporating these feature groups into predictive models proved reasonable, with environmental factors such as the number of objects in the scene, scene visual complexity and roadway type, object features such as object size and type, and driver demographics such as gender showing significant impacts on driver SA. Next, gaze-point-based and visual sensory-dependent features were processed from the eye-tracking data. Predictive models incorporating different feature groups including environmental, object, and driver features, and those extracted from the eye-tracking data were fitted and compared. Two phases of SA were distinguished: object localization and recognition. Binary classification models were developed and rigorously evaluated for each phase. Recognizing the impracticality of drivers wearing eye-trackers during daily driving, an alternative for eye-tracking data, which utilizes computer vision to estimate visual attention from forward-view driving videos was proposed. Gaze-related features extracted from these videos demonstrated comparable performance to those from eye-tracking data, suggesting their viability for predicting driver SA. This insight could inform the design of ADAS systems, enabling low-cost selective assistance to drivers and ultimately enhancing driver safety.
Modeling Healthcare Policy: From Calibration to Optimization
(2024-09-09) Lee, Serin; Liu, Shan SL; Zabinsky, Zelda ZZ
Setting effective healthcare policy is complex, particularly due to the dynamic and heterogeneous nature of individual behaviors. While modeling studies offer valuable insights, their computational complexity and calibration requirements can limit practical applications. An important question when identifying optimal healthcare policies is how to consider heterogeneous individual health behavior dynamics while at the same time find ways to consume less time and computational resources. This dissertation addresses three main objectives. The first objective aims to optimize public health interventions to minimize the disease burden of the COVID-19 pandemic. An agent-based model (ABM) is developed to evaluate non-pharmaceutical interventions (NPIs), and vaccination policies under continuous virus mutation. This ABM simulates a heterogeneous population and offers flexibility for addressing diverse policy questions. By addressing parameter uncertainty through calibration and simulating multiple scenarios, the model identifies robust strategies, such as periodic vaccination and adaptive social distancing, for effective disease control. The second objective is to design optimal vaccination promotion campaigns that increase vaccination uptake and improve public health. This approach integrates coupled dynamics, social contagion, and evolutionary game theory to model how vaccination behavior shifts within an ongoing epidemic and with word-of-mouth vaccination campaigns. This model overcomes the limitations of prior studies that assumed static vaccination willingness in vaccination allocation problem. The study offers population-level insights into how resources and messaging should be targeted across demographic groups while considering the societal contexts surrounding vaccination within communities. The final objective is to enhance calibration practices for simulation models by introducing a representative calibration framework. This approach is particularly valuable for complex models where uncertainties exist, and evaluation for policy analysis is computationally expensive. This framework identifies minimal sets of parameter values to limit computational expense while capturing diverse model behaviors. By focusing on representativeness in calibration, this research yields reliable implications for real-world decision-making, filling a gap where other methods emphasize precision at the risk of not accounting for data and model uncertainty. Overall, this dissertation advances healthcare policy modeling by addressing complex and heterogeneous individual behaviors, as well as addressing uncertainties arising from data limitations and simulation complexity through calibration. This research will equip healthcare policymakers to derive informed, data-driven insights from modeling studies, despite model uncertainties and complexities.
Designing a Wearable Hemodialysis System: a Human Factors Engineering Approach
(2024-04-26) Jónsdóttir, Auður Anna; Kim, Ji-Eun
End-stage renal disease (ESRD) is a medical condition of permanent kidney failure requiring a patient to either receive a kidney transplant or adhere to long-term dialysis treatments. Dialysis is a life-sustaining method of cleaning and filtering a patient's blood using a dialysis device. The majority of ESRD patients receive hemodialysis treatments where the patient's blood is cleaned and filtered outside of the patient's body using a hemodialysis system. Current hemodialysis systems have not been significantly updated since first pioneered in 1943. Patients still carry a high symptom burden with loss of mobility and independent living while undergoing treatment. However, recent technological advances in hemodialysis treatments allow for the development of a new generation of wearable hemodialysis systems bringing hope to improve ESRD patients' quality of life. Despite the potential to transform the lives of dialysis patients, many obstacles have impeded the development and use of a wearable hemodialysis system. One major hindrance to its development is the lack of adopting a human factors engineering design process that incorporates users' perspectives throughout the entire design process. To this day, no studies exist that gather and characterize the perspectives of distinct user groups at the beginning of the design cycle for a wearable dialysis system. This dissertation aims to fill that gap. The overall goal of this dissertation is to identify users' perspectives on a wearable hemodialysis system. In particular, by gathering responses from distinct user groups and employing both qualitative and quantitative methods, this dissertation aims to answer four research questions: (R1) what are patients' and care partners' needs and perspectives of a wearable hemodialysis system? (R2) What are patients' and care partners' needs and expectations for monitoring and training procedures for a wearable hemodialysis system? (R3) What are clinicians' perspectives on a wearable hemodialysis device? and (R4) What is the relationship among user characteristics, human factors design principles, and proposed design concepts of a wearable hemodialysis device? The results from research questions R1-R3 aim to help developers of a wearable hemodialysis system set design and usability goals to help ensure an optimal design process with a resulting system that meets and supports users' needs. Additionally, the results from research question R4 aim to help developers better understand the relationship among users' demographic characteristics, human factors design principles for wearable medical devices, and designs of a wearable hemodialysis device. Understanding this relationship may help identify important factors that contribute to users' adoption behaviors. As a result, researchers and designers may refine their objectives to help ensure a wearable hemodialysis system that is designed in accordance with users' needs.
Robust Optimization Methods for Improving Virtual Power Plant Reliability and Classification Fairness
(2024-04-26) Yang, William; Zhao, Chaoyue
In this dissertation, we present three different robust optimization approaches: distributionally robust optimization (DRO), target-oriented robust optimization (TORO), and Rawlsian fairness. We apply DRO and TORO frameworks to address virtual power plant reliability, and Rawlsian fairness to address the fair binary classification problem. A virtual power plant (VPP) is an entity that aggregates smaller solar and wind farms with other heterogeneous distributed energy resources (DERs) to increase their visibility to Independent System Operators (ISOs) and allows them to participate in the energy market. Typically, smaller solar and wind farms are unable to participate in the wholesale energy market, so a VPP's ability to integrate them into the energy market is a crucial step for reducing the global carbon footprint and combating climate change. Our proposed multi-stage DRO framework allows us to schedule VPP operations in the presence of intermittent renewable energy output by dynamically coordinating the heterogeneous DERs in a reliable and cost-effective manner. Our proposed TORO method helps us address another challenge that arises from increased renewable energy penetration, which is the assessment of the VPP’s flexibility. The uncertainty brought by renewable energy makes it harder to balance energy supply and demand, and failing to do so can result in expensive renewable energy curtailment or blackouts. Our TORO method provides a flexibility assessment framework that identifies the maximum amount of net load deviation the system can tolerate. Traditional binary classification algorithms are prone to producing unfair results that favor certain demographic groups over others. This inequity is often exacerbated in unbalanced datasets where the number of entries from a majority group significantly outweighs the entries from a minority group. We use a MIP framework to formulate our Rawlsian fairness to address these inequities. Our methodology prioritizes the performance of the worst-off demographic group, and our specific formulation can produce interpretable solutions by directly optimizing sparsity. Additionally, it provides flexibility for users to achieve interpretable solutions in multiple ways.
Distributionally Robust Optimization for Reinforcement Learning
(2024-04-26) Song, Jun; Zhao, Chaoyue
Reinforcement learning (RL) has received remarkable success in many domains, including video games, board games, robotics and continuous control tasks. Despite the success and attention that RL has received during the past decades, it struggles with several issues that degrade its performance and lead to suboptimality. In model-based RL, the uncertainty in environment dynamics can significantly deteriorate the learnt agent’s ability to recommend good actions. While in model-free RL, learnt agent's performance can be greatly affected by the restrictive parametric assumption on policy distribution. In this dissertation, our goal is to utilize distributionally robust optimization (DRO) to overcome the above-mentioned limitations of RL, and to develop novel and practical RL algorithms with improved robustness and performance. To achieve the goal, we follow two main objectives. The first objective is to adopt DRO to add robustness to the uncertainty in the environment dynamics of the model-based RL. We propose a new Distributionally Robust Markov Decision Process (DRMDP) framework where the distribution of environment dynamics does not have predetermined parametric values, and we consider the worst-case probability distribution of these transition probabilities within a decision-dependent ambiguity set. The second objective is to utilize optimistic DRO to develop nonparametric policy optimization methods for the model-free RL. Since the policy learnt is not confined to the scope of parametric functions, this opens up the possibility of converging to a better optimality. Following this objective, we propose three different nonparametric policy optimization frameworks, with Kullback–Leibler, Wasserstein and Sinkhorn constraints respectively to control the size of policy update. For each framework, we derive the closed-form policy update solution to the corresponding optimistic DRO problem using Lagrangian duality, and propose practical RL algorithms to perform the policy updates. We further improve the sample efficiency of the proposed nonparametric policy optimization frameworks, by incorporating human guidance through imitation learning techniques.
Practical Multi-Objective Optimization Approaches for Decision Making in Health Care Considering Infectious Disease Dynamics and Uncertainties
(2024-02-12) Greene, Chelsea Amanda; Zabinsky, Zelda B
Operations research methods have been commonly used to inform decisions in health care related to inventory management, policy implementation, and resource allocation. However, the current research does not address many of the unique challenges and objectives faced by decision makers and stakeholders involved in managing infectious diseases. For example, many optimization approaches have a single objective, and do not balance overall health outcomes with health equity metrics. The goal of my dissertation research is to develop practical, effective, and equitable approaches that address challenges in the management of infectious diseases. To achieve this goal, my research combines state-of-the-art mathematical modeling methodologies including optimization, dynamic transmission compartmental modeling, and statistics with the perspectives of several organizations and researchers across multiple disciplines, including epidemiologists, economists, and public health officials. The challenges and objectives addressed in this research include but are not limited to, dynamics and uncertainties in demand, supply, and health outcomes, multiple objectives, and vulnerabilities of populations. This research contributes to three realistic applications, including:• inventory and order management for multiple healthcare commodities and populations during an infectious disease outbreak, and • policy analysis of tuberculosis and HIV health care program interventions in KwaZulu-Natal, South Africa, and • budget allocation to regional tuberculosis and HIV health care programs across the nine provinces of South Africa. The three applications and methodologies provide practical approaches that address these challenges and pave the way for new, more practical, effective, and equitable approaches to decision making in health care for the management of infectious disease.
Robust Markov decision processes with data-driven, distance-based ambiguity sets
(2023-09-27) Ramani, Sivaramakrishnan; Ghate, Archis
We consider finite- and infinite-horizon Markov decision processes (MDPs), with unknown state-transition probabilities. These transition probabilities are assumed to belong to certain ambiguity sets, and the goal is to maximize the worst-case expected total discounted reward over all probabilities from these sets. Specifically, the ambiguity set is a ball — it includes all probability distributions within a certain distance from the empirical distribution constructed using historical, independent observations of state transitions. We therefore call these problems robust MDPs (RMDPs) with data-driven, distance-based ambiguity sets. The literature on data-driven robust optimization mentions (i) robust value convergence with respect to sample size, (ii) out-of-sample value convergence with respect to sample size, (iii) probabilistic performance guarantees on out-of-sample values, and (iv) a probabilistic convergence rate, as desirable properties. The research objective of this dissertation is to establish essentially a minimal set of conditions under which RMDPs with data-driven, distance-based ambiguity sets exhibit these four properties. We first achieve this for the (s, a)-rectangular RMDPs ((s, a)-RMDPs) studied in Chapter 2. There, the ambiguity set for the whole MDP equals a Cartesian product of ambiguity sets for individual state-action pairs. We establish robust and out-of-sample value convergence under a generalized Pinsker’s inequality, if the radii of the ambiguity balls approach zero as the sample-size diverges to infinity. This inequality links convergence of probability distributions with respect to the distance function, to their topological convergence in a Euclidean space. We also establish that, for finite sample-sizes, the optimal value of the RMDP provides a lower bound on the value of the robust optimal policy in the true MDP, with a high probability. This probabilistic performance guarantee relies on a certain concentration inequality. Under these generalized Pinsker and concentration inequalities, we also derive a probabilistic rate of convergence for the robust and out-of-sample values to the true optimal value. These two inequalities hold for several well-known distance functions including total variation, Burg, Hellinger, and Wasserstein. We present computational results on a generic MDP and a machine repair example using total variation, Burg, and Wasserstein distances. These results illustrate that the generality of our framework provides broad choices when attempting to trade-off conservativeness of robust optimal policies against their out-of-sample performance by tuning ambiguity ball radii. In Chapter 3, we extend results from Chapter 2 to a so-called s-rectangular framework. In this more general context, the ambiguity set for the MDP is a Cartesian product of ambiguity sets for individual states. In that chapter, we introduce a family of distance-based s-rectangular RMDPs (s-RMDPs) indexed with ρ ∈ [1, ∞]. In each state, the ambiguity set of transition probability mass functions (pmfs) across different actions equals a sublevel set of the ` ρ -norm of a vector of distances from reference pmfs. Setting ρ = ∞ in this family recovers (s, a)-RMDPs. For any s-RMDP from this family, there is an (s, a)-RMDP whose robust optimal value is at least as good; and vice versa. This occurs because the s- and (s, a)-RMDPs can employ different ambiguity set radii, casting a nuanced doubt on the anecdotal belief that (s, a)-RMDPs are more conservative than s-RMDPs. More strongly, if the distance function is lower semicontinuous and convex, then, for any s-RMDP, there exists an (s, a)-RMDP with an identical robust optimal value. This suggests that appropriate caution should be exercised before interpreting too broadly any anecdotal claims that (s, a)-RMDPs are more conservative than s-rectangular ones. We also study data-driven versions of our s-RMDPs,where the reference pmf equals the empirical pmf constructed from state transition samples. We prove that the robust optimal values, and the out-of-sample values of robust optimal policies both converge to the true optimal, asymptotically with sample sizes, if the distance function satisfies the generalized Pinsker’s inequality introduced in Chapter 2. The robust optimal value also provides a probabilistic lower bound on the out-of-sample value of a robust optimal policy, when the distance function satisfies the concentration inequality. This finite-sample guarantee admits a surprising conclusion — (s, a)-RMDPs are the least conservative among all s-RMDPs within our family. Like in Chapter 2, under these generalized Pinsker and concentration inequalities, we also derive a probabilistic rate of convergence for the robust and out-of-sample values to the true optimal value. Though similar asymptotic and finite-sample results were developed for (s, a)-RMDPs in Chapter 2, the proof techniques in this chapter are different and more sophisticated. These more involved proofs are needed because the structure of s-RMDPs introduces new analytical hurdles in our attempt toestablish the sufficiency of generalized Pinsker and concentration inequalities. For example, it is no longer adequate to limit attention to deterministic policies — randomization may be needed for optimality. We also present computational experiments on a machine repair example using the total variation distance and ρ = 1. The results of those experiments validate the claims established in that chapter. Finally, in Chapter 4, we develop a data-driven, distance-based RMDP framework on separable complete metric spaces. We extend our asymptotic and finite-sample results to that setup. Unlike our first two studies, this more general endeavor relies on measure-theoretic concepts from minimax control.

Browse

Recent Submissions