21090 research outputs found
Sort by
Safety and Security of Reinforcement Learning for Autonomous Driving
No full textIn the context of autonomous driving, reinforcement learning (RL) presents a powerful paradigm: agents capable of learning to drive efficiently in unseen situations through experience. However, this promise is shadowed by a fundamental concern—how can we entrust decision-making to agents that rely on trial-and-error learning in safety-critical environments where errors may carry severe consequences?
This thesis advances a step toward resolving this dilemma by integrating three foundational pillars: adversarial robustness, simulation realism, and model-based safety. We begin with a comprehensive survey of adversarial attacks and corresponding defences within the domains of deep learning (DL) and deep reinforcement learning (DRL) for autonomous vehicles. This survey reveals the porous boundary between safety and security—both natural disturbances and adversarial perturbations can destabilize learned policies. Motivated by this insight, we introduce the Optimism Induction Attack (OIA), a novel adversarial technique that manipulates an RL agent’s perception of safety, causing it to act with unwarranted confidence in hazardous situations. Evaluated in the context of an Adaptive Cruise Control (ACC) task, the OIA significantly impairs policy performance, exposing critical vulnerabilities in state-of-the-art RL algorithms.
To counter the demonstrated threats, we present a systematic defence architecture. We develop REVEAL, a high-fidelity simulation framework designed to support the training and evaluation of safe RL agents under realistic vehicle dynamics, traffic scenarios, and adversarial conditions. By narrowing the gap between abstract simulation and real-world complexity, REVEAL facilitates rigorous and nuanced testing, which is essential for safety-critical applications. To enhance learning efficiency within this environment, we employ a transfer learning (TL) strategy: policies initially trained in simplified simulators (e.g., SUMO) are adapted and fine-tuned in REVEAL, leading to faster convergence and improved safety performance during both training and deployment.
Central to our approach is the development of a Multi-Output Control Barrier Function (MO-CBF), which simultaneously supervises throttle and brake commands to enforce safety constraints in real time. Rather than relying on hard overrides, the MO-CBF operates cooperatively with the learning agent—gently adjusting unsafe actions and introducing corresponding penalties during training. This enables the agent not only to learn safe behaviour but also to internalize safety principles and anticipate potentially unsafe scenarios.
Our empirical evaluation demonstrates the effectiveness of the proposed framework across a spectrum of disturbances, adversarial inputs, and realistic high-risk maneuvers. The results consistently show improved safety and robustness, highlighting the framework’s capacity to transform RL agents from vulnerable learners into trustworthy autonomous systems.
In summary, this thesis presents a comprehensive methodology for safe and secure RL in autonomous driving. By grounding agent training in high-fidelity simulation, leveraging adversarial awareness, and embedding real-time model-based safety mechanisms, we provide a cohesive and scalable pathway toward deploying RL in the real world with confidence
The Impact of Spinoffs on the Information Environment of Peer Firms: Information Spillovers or Industry Disruption
No full textSpinoffs are a form of divestiture in which a parent firm separates a portion of its operations into a newly created and independent spinoff firm. This thesis examines how spinoffs affect the information environment of the parent firm’s peers using a sample of U.S. spinoffs from 2010 to 2018. I find that analyst forecast dispersion decreases for peers after a cross-industry spinoff. This result is more pronounced when the parent firm is operationally complex prior to the spinoff, suggesting that the separation of unrelated operations can reduce analysts’ information processing costs and facilitate information spillovers between peers. Conversely, I find that forecast accuracy decreases and forecast dispersion increases for peers after a same-industry spinoff. These results are more pronounced for peers that operate in industries that become less concentrated and peers that experience more volatile cash flows and income in the post-spinoff period. Together, these results suggest that same-industry spinoffs can be disruptive industry events that change industry compositions and outweigh the potential informational benefits of the spinoff firm’s initial financial statements, thereby complicating the forecasting process for analysts. Overall, this thesis identifies new mechanisms that helps explain the spillover effects of spinoffs on the information environment of the parent firm’s peers and demonstrates that the operational similarity between the parent and spinoff firms is an important determinant of these effects
Health and economic benefits of reducing air pollution exposure through adaptation and mitigation under a changing climate
No full textAir pollution is the world’s largest environmental health risk. Even in countries with perceived clean air, like the United States of America (U.S.) and Canada, ambient air pollution still contributes to approximately 150,000 and 17,500 annual premature mortalities, respectively.
Air pollution is expected to worsen under climate change, leading to increases in mean ozone and PM2.5 concentrations and higher extreme values. These changes could lead to more air quality alerts, which are triggered when Air Quality Index (AQI) values exceed certain thresholds. Though they are the main medium for communicating air pollution risk to the public, the effect of climate change on air quality alerts has not been previously studied.
The effectiveness of air quality alerts, and the adaptation behaviors they recommend, is also not well known. Few studies have investigated how people respond to air quality alerts, and none have looked at how behavioral responses may change in the future. Even fewer studies quantify the health benefits that adapters, or those who respond to air quality alerts, receive from their adaptation. This is critical to understand, as air pollution is a significant public health threat now and, without emission reductions, is expected to
worsen this century.
The studies included in this dissertation use modeled future data to elucidate how air quality alerts driven by PM2.5 and ozone change throughout the 21st century. They identify
who is affected by the increase in air quality alerts and model how these populations might respond. The use detailed time use, location, and building parameter data to provide
improved estimates of adaptation behaviors.
Across the three studies, we find adaptation - including limiting time outdoors, masking, and reducing infiltration - to be useful in reducing ambient air pollution exposure. However, adaptation benefits are not distributed evenly across the population. Certain populations, like seniors (aged 65+), receive much higher benefits than other groups. So too, do those who have cleaner environments in which to adapt. Reducing outdoor concentrations, through policy addressing climate change or air pollution, reduces the need to adapt and protects those who cannot adapt. However, the studies herein show that behavioral change must also be considered, as it can either offset
or amplify health improvements from ambient pollution reduction
Immersive Invaders: Privacy Threats from Deceptive Design in Virtual Reality Games and Applications
No full textVirtual Reality (VR) technologies offer immersive experiences but collect substantial user data. While deceptive design is well-studied in 2D platforms, little is known about its manifestation in VR environments and its impact on user privacy. This research investigates deceptive designs in privacy communication and interaction mechanisms of 12 top-rated VR games and applications through autoethnographic evaluation of the applications and thematic analysis of privacy policies. We found that while many deceptive designs rely on 2D interfaces, some VR-unique features, while not directly enabling deception, amplified data disclosure behaviors, and obscured actual data practices. Convoluted privacy policies and manipulative consent practices further hinder comprehension and increase privacy risks. We also observed privacy-preserving design strategies and protective considerations in VR privacy policies. We offer recommendations for ethical VR design that balance immersive experiences with strong privacy protections, guiding researchers, designers, and policymakers to improve privacy in VR environments.This project has been funded by the Office of the Privacy Commissioner of Canada (OPC); the views expressed herein are those of the author(s) and do not necessarily reflect those of the OPC, nor the University of Waterloo
Enhancing Safety and Efficiency of Underground Mining Operations Using Vision-Based Systems
No full textThis thesis investigates the design and deployment of vision-assisted monitoring and alert systems to improve safety and efficiency in underground mining operations. The research integrates advanced computer vision techniques, including object detection, pedestrian tracking, pose estimation, line detection, and Kalman filtering, for real-time operations on edge devices. Two main applications were developed and optimized: a loader monitoring system that tracks loading cycles and boom poses to provide operators with visual feedback, and a pedestrian alert system that combines detection and pose estimation to enhance safety around jumbo drills. Both systems were implemented and tested in realistic underground environments or similar settings, demonstrating their capacity to improve operational efficiency and situational awareness.
This work was carried out closely with industry partners, where the focus was not on setting fixed quantitative benchmarks but on delivering systems that operators and managers found useful and reliable. Instead of relying on controlled experiments or predefined metrics, the systems were shaped through an iterative cycle of design, deployment, testing, and feedback. This process often required trade-offs, such as choosing robustness and usability over purely numerical performance gains, but it ensured that the outcomes were relevant to day-to-day operations. Beyond technical development, the experience also highlighted the importance of communication with end-users, as their input directly guided adjustments to system functionality and interface design. By combining modern computer vision methods with field deployment, this thesis contributes not only practical tools for safer and more efficient mining operations, but also insights into how advanced algorithms can be adapted for adoption in real-world industrial settings. These lessons extend beyond mining, offering guidance for similar applications in other safety-critical and resource-constrained environments
Mining Time Series for Maximal Coverage with Matrix Profiles and Constraint Optimization
No full textTime Series Data (TSD) is essential in many areas of modern data analysis because they reflect how different processes change over time. Understanding these data can still be challenging. One of the main challenges is that the underlying states of a system are often hidden, making it harder to interpret patterns and draw reliable conclusions. This thesis addresses the critical task of mining recurrent patterns from systems whose internal states are neither directly observable nor controllable. It introduces a novel unsupervised approach explicitly designed to maximize coverage in TSD.
The research proposes a structured approach comprising three key steps: firstly, generating candidate patterns and their occurrences using an advanced Matrix Profile (MP) algorithm known for its efficiency and accuracy in detecting subtle recurrent patterns; secondly, translating these candidate patterns into a constraint-based model incorporating group constraints to enforce selection of all instances of the same pattern and exclusion constraints to prevent overlapping occurrences; and thirdly, selecting an optimal subset of core patterns using either a constraint solver to ensure optimal selection on shorter Time Series (TS), or scalable greedy heuristic methods that offer practical efficiency for larger or more complex datasets, thereby effectively balancing optimality with computational feasibility.
The effectiveness of the proposed approach is demonstrated through rigorous evaluations on real-world power consumption TSDs representing computer activities, alongside controlled synthetic datasets, using metrics such as coverage efficiency, computational time, and pattern compactness—striving to maximize data representation with minimal redundancy. Our experimental results show that the proposed method improves how efficiently data is covered, striking a practical balance between capturing key patterns and avoiding unnecessary repetition. This work contributes to the advancement in unsupervised pattern mining and has useful applications in areas such as forecasting, system health monitoring, anomaly detection, and policy verification
River Resilience Requires Sufficient Floodplains: Experimental Insights from a Novel Flume Study Investigating Meander Constriction
No full textGlobally, riparian zones are in poor condition. Numerous anthropogenic watershed modifications negatively affect water quality, resiliency, and habitat diversity of river systems. Research on the effects of constricting meandering rivers is limited. This has led to few methods that optimize riparian zone widths in ways that maintain adequate corridors or floodplains to support natural river processes and protect public safety. The goal of our research was to determine the effects of constraining the floodplain of a meandering river. Specifically, we studied the effects of constraining the Bow River in Calgary where the floodplain corridor is currently facing extensive development pressures. To achieve these goals a mobile bed laboratory experiment was completed. The experiment involved developing and then constricting a gravel bed meandering river from an initially straight channel. Alfalfa was grown alongside the river and within the floodplains to provide necessary bank strength. During the experiment, sediment leaving the flume was collected and aerial images were captured to allow topographic and sediment transport analysis. Results showed that as alfalfa grew, bank strength increased, limiting meander evolution. Despite the relatively fixed meanders, findings suggest that floodplain constraints significantly reduce river-floodplain connectivity, alter a channels flow regime by increasing velocity and flow depth, increase sediment transport, and narrow channel widths. The results of this study will impact river management practices, emphasize creating room for rivers as a nature-based solution, and improve laboratory methods for investigations of meandering rivers
Linking Blue Carbon Ecosystems and Water Quality in Coastal Wetlands for Viable Small-Scale Fisheries
No full textCoastal wetlands serve as critical social-ecological systems (SES), providing essential ecosystem services (ES), supporting biodiversity, and sustaining livelihoods, particularly for Small-Scale Fisheries (SSFs). However, these ecosystems are increasingly vulnerable due to the degradation of Blue Carbon Ecosystems (BCEs) such as seagrass meadows and mangrove forests, compounded by declining water quality and social-economic pressures. This doctoral research explores the interlinkages between BCEs, water quality, and SSFs in Chilika Lagoon, India, Asia's largest brackish water lagoon-through an interdisciplinary, participatory, and spatially explicit approach. Grounded in the Social-Ecological Vulnerability and Viability Nexus (SEVVN) framework, the study integrates ecological data, community knowledge, and governance analysis to understand the drivers, feedback, and outcomes influencing the resilience of coastal wetland systems.
The study explores the interlinkages between BCEs, water quality, and the viability of SSFs in Chilika Lagoon, India. It addresses four interrelated objectives using a mixed-methods approach grounded in SES thinking. First, the spatial distribution of seagrass meadows and mangrove patches is mapped and characterized through participatory mapping, remote sensing (RS), and field observations. By integrating traditional ecological knowledge (TEK) with historical and scientific data, the study reveals patterns of degradation and resilience, highlighting the role of co-produced knowledge in conservation planning. Second, key water quality parameters, salinity, turbidity, temperature, and nutrient loads, are assessed using literature, secondary datasets, and community observations. Seasonal variations and anthropogenic pressures, including plastic pollution and sedimentation, are found to affect BCE health and fishery habitats, impacting SSF productivity and livelihoods.
Third, the SEVVN framework is applied to analyze feedback loops and stressors linking BCE degradation and SSF vulnerability. Data from surveys, interviews, and focus group discussions (FGDs) show that fishers experience environmental and institutional vulnerabilities due to declining fish availability, weak governance, and limited adaptive capacity. Sectoral variations in exposure and resilience are documented across the lagoon’s Northern, Central, Southern, and Outer Channel sectors. Finally, the study identifies governance pathways for building resilience. Ecosystem-based and community-centered strategies are proposed, including TEK integration, water quality monitoring, and participatory co-management. Recommendations aim to foster inclusive decision-making and bridge scientific and local knowledge systems. Chilika Lagoon, a Ramsar site and biodiversity hotspot, provides a critical empirical context to study the trade-offs between conservation, development, and livelihood needs. This research contributes to policy and practice for coastal wetland sustainability in South Asia and beyond.
Key findings from the study underscore the urgent need for integrated social-ecological approaches in wetland governance. Participatory mapping revealed local awareness of BCE loss but limited community involvement in formal monitoring or conservation programs. Survey data showed that perceptions of water quality decline were strongly linked to observed reductions in fish catch and increased livelihood uncertainty. Participatory exercises surfaced the erosion of traditional knowledge (TK), exacerbated by policy shifts, environmental change, and generational gaps. The SEVVN analysis identified multiple entry points for intervention, including capacity-building, knowledge co-production, and inclusive policy formulation. Theoretically, the study contributes to the literature on SES, blue carbon science, and fisheries governance by developing and applying the SEVVN framework. Methodologically, it advances participatory and mixed-methods research in coastal environments. Practically, it offers actionable insights for policymakers, conservation practitioners, and local communities seeking to build adaptive capacity and enhance the viability of SSFs. By linking ecological data with social realities, and grounding analysis in community perspectives, this research contributes to sustainability science and global efforts to meet the Sustainable Development Goals (SDGs). It particularly aligns with SDG 1 (No Poverty), SDG 6 (Clean Water), SDG 13 (Climate Action), SDG 14 (Life Below Water), and SDG 17 (Partnerships for the Goals). Ultimately, the study argues for a shift from fragmented, top-down conservation models toward more inclusive, participatory, and ecosystem-based governance approaches that center the voices of those most dependent on coastal resources
Analysis of Limitations of AI Tools for Pediatric Speech Language Pathology Documentation and Mitigation Strategies
No full textSpeech Language Pathology (SLP) is a therapy discipline offered by KidsAbility, a pediatric rehabilitation clinic in Southern Ontario. Documentation is a key part of SLP and other therapy practice guidelines and can take up significant portions of a therapist’s time. AI-based clinical documentation aids have been developed to help reduce this burden, and one such tool - MutuoHealth’s AutoScribe - has been piloted by KidsAbility. Though this AI tool has been beneficial to some therapy disciplines, the SLP clinicians face unique challenges when using these tools. The model seemed unable to recognize speech therapy strategies or to parse the play-based script of pediatric appointments. This thesis seeks to explore the issues SLPs encounter with AI documentation tools and propose potential approaches to mitigate these issues.
The AI documentation process was divided into the transcription pipeline, where an audio file input produced a corresponding transcript output, and the generation pipeline, where an input transcript produced a draft SOAP note. The SLPs who had participated in the AutoScribe pilot test were interviewed about their experiences with the tool and its integration into their workflows. The issues reported by the therapists were sorted into those more closely related to the transcript and those more closely related to the drafted SOAP note. A set of sample SLP appointments from KidsAbility were gathered from an extended AutoScribe pilot, with 10 selected as examples of appointment data (audio, transcripts, drafted and final SOAP notes) to test the transcription and generation pipelines. An augmented automatic speech recognition (ASR) pipeline based on a Whisper model was used to test improvements to the transcript. However, the generated transcripts were not significantly improved from the pilot test. Instead, ground truth transcriptions were manually created from the audio files to use for testing the generation pipeline. For SOAP note generation, the addition of discipline-specific context tailored to appointment type was tested. This context was curated in collaboration with SLPs from KidsAbility to include SOAP templates, definitions of key concepts, and information about speech data. A Llama 3.3 70B model was used for SOAP note generation with ground truth transcriptions and SLP specific RAG-adjacent information as context. The input context was optimized over several iterations based on clinicians’ evaluations of generated SOAP note quality.
KidsAbility’s SLPs had flagged sessions targeting speech practice as having particular difficulties with AutoScribe. The model seemed unable to make inferences about the child’s speech quality from the transcript alone. Methods of quantitatively assessing speech based on session audio were explored as ways to provide additional context on speech quality to the SOAP generation model. A sample appointment was selected for testing, and child speech samples of the targeted sound were sliced from the audio and assigned quality categories. These samples were then compared against correct productions using the cosine distance between their mel-spectrograms. The samples were also passed through a phoneme-based ASR model to get the layer activations. The cosine distances and layer outputs were then tested as predictive measures of articulation accuracy, with layer outputs yielding the best results. The resulting speech accuracy scores were then passed into the generation model as additional context, with the output containing correct statements about the nature of the child’s articulations.
Though clinicians’ availability limited the extensiveness of generated SOAP note evaluations, the SOAP notes generated with SLP-specific context showed improvement compared to the basic model generation. The model also tended to repeat information from previous SOAP notes if examples were provided. It was found that quantitative speech analysis does seem possible using phoneme model layer activations and cosine distances between the mel-spectrograms of correct articulations. Based on these findings, further optimizations to the generation pipeline and work on making effective AI tools for KidsAbility’s EY SLPs will continue
Perceiving Change in Uncertain Times: Public Accuracy and its Individual Differences in Estimating Past Societal Shifts During COVID-19
No full textUnderstanding where society is headed requires a clear grasp of where it has been, to make sense of what factors shape direction of societal change. Yet during periods of heightened uncertainty, such as the COVID-19 pandemic, most people struggle to accurately perceive the direction of societal shifts. In this thesis, I used the first year of the COVID-19 pandemic as a naturalistic context of uncertainty to investigate how the public estimated societal change during such volatile periods, what domains may have been more accurately perceived, and what individual factors may have shaped this accuracy. U.S. participants (N = 644) estimated societal change across thirteen domains over six-month (April-October 2020) and one-year periods (April 2020-April 2021), either via free-text or on a -50% to +50% slider, providing their confidence for their estimates per domain. Next, they completed measures assessing general knowledge, confidence in their knowledge judgments, and metacognitive engagement (e.g., reflecting on limits of one’s knowledge when discussing social issues). I further assessed deliberation-related engagement by tracking time spent on each estimation. Results showed that Americans held a largely pessimistic view of societal changes over the pandemic, especially when considering actual change in the domains of depression rates, mortality, violent crimes, unemployment, and charitable donations. Out of all domains, the majority of the sample correctly estimated the direction of societal change for depression rates, life satisfaction, explicit prejudice, charitable giving, and religiosity. However, for other domains participants were either at chance or got the direction of change wrong. Participants were also more accurate in estimation of change for shorter (vs. longer) time frames, and when using open-ended response options (vs. percentage-based slider). Moreover, individuals showing greater task deliberation, metacognitive engagement, and confidence in their estimates, but not greater general knowledge or calibration in confidence and accuracy of their knowledge, were more accurate. Additionally, effects of knowledge calibration and deliberation varied as a function of domain, either improving estimates or increasing error or bias. These findings suggest that how people engage with information matters more than simply what they know. The thesis concludes by discussing implications for understanding public perceptions of pandemic-related societal changes and identifies factors that may help align these perceptions with actual social trends