Theses (Computer Science)

Permanent URI for this collection

https://dspace.library.uvic.ca/handle/1828/76

Browse

Now showing 1 - 20 of 755

Thresholded linear bandits
(2025) Nguyen, Trang Thu; Mehta, Nishant
Thresholded linear bandits is a novel bandit problem that lies in the intersection of several important multiarmed bandit (MAB) variants, including active learning, structured bandits, and learning halfspaces. To achieve sublinear regret in the presence of exponentially many arms, one method is to exploit the structure of the reward function. However, the presence of an unknown threshold component makes previously known algorithms for structured bandits unsuitable. Moreover, the threshold introduces a discontinuity to the reward function, making the problem significantly more difficult. In this thesis, we study the union of axis-parallel halfspace variant of the thresholded linear bandits problem. We suggest an algorithm that achieves sublinear regret and provide theoretical guarantees on the performance of the algorithm
AI-driven security in software-defined networks: A unified framework for intrusion detection and mitigation
(2025) El Gadal, Walid; Ganti, Sudhakar
Over the past decade, data networks have evolved from static resource deployment to a more dynamic and adaptive paradigm. Software-Defined Networking (SDN) is one of the most creative network technologies where network control is separated from forwarding. It is directly programmable and has been proposed as a way to programmatically control networks, facilitating the deployment of new applications and services, as well as tuning network policies and performance. However, various challenges have hindered achieving strong cybersecurity within the dynamic network configurations of Software-Defined Networking. Traditional cybersecurity measures, especially in programmable and dynamic network infrastructures like SDNs, are not sufficient to mitigate cyber threats. This dissertation explores the capabilities of SDN and examines how AI-driven methods can enhance intrusion detection and mitigation. The study begins by providing a comprehensive introduction to SDN, outlining its fundamental capabilities and comparative advantages over traditional network architectures. In addition, it explores SDN vulnerabilities and addresses complex security challenges. The objective of this thesis work is to improve the detection and mitigation of threats in SDN environments. For this, we first present a dynamic defense framework that includes Machine Learning and Deep Learning techniques for attack detection and mitigation. Furthermore, a novel hybrid Coot-Lyrebird optimization algorithm is developed to specifically choose the most impactful features in the network. The selected features are given to the proposed hybrid network that combines Convolutional Neural Network (CNN), SE-ResNeXt, and Long Short-Term Memory (LSTM) networks. Finally, the proposed Deep Q-Network (DQN) model performs attack mitigation measures. The results indicate that the proposed dynamic defense has an accuracy of 0.999571%. In addition, we extended our study to include more complex environments. Software-Defined Internet of Things (SD-IoT) networks enabled intelligent network management through their dynamic features, but expose centralized infrastructure to complex cyberattacks that put the system in great danger. In order to address this, a novel federated secure intelligent intrusion detection and mitigation framework with automated attack reporting for SD-IoT network is presented.
MountainScape semantic segmentation of historical and repeat images
(2025) Mahindrakar, Aniket; Tzanetakis, George; Higgs, Eric
Semantic segmentation of ultra-high resolution images is challenging due to high memory and computation requirements. Current approaches to this problem involve cropping the ultra-high resolution image into small patches for individual processing in order to provide local context, or under-sampling the images to provide global context, or following a combination of both which gives rise to global-local refinement pipelines. In this thesis, we present the MountainScape Segmentation Dataset (MS2D) which comprises high-resolution historic (grayscale) manually segmented images of Canadian mountain landscapes captured from 1861 to 1958 and their corresponding modern (colour) repeat images. Additionally, we analyze the characteristics of the dataset, define evaluation criteria, and provide a baseline to serve as a reference benchmark for automated land cover classification using the Python Landscape Classification Tool (PyLC), an existing software tool. The main contribution of this thesis is the experimental exploration of various deep learning architectures to address the tiling artifacts and spatial context loss faced by PyLC in its tile-based processing of ultra-high-resolution images, alongside a comprehensive investigation using a larger dataset than that employed in the original PyLC study to solve this tiling problem.
Two views of cryptography and the gap in-between
(2025) Wu, Zehou; Kapron, Bruce; Lu, Yun
There are two popular views of cryptography. One is formal (symbolic), which uses expressions to model the ideal functionality of encryption functions and is easy to verify. The other is computational, which is what cryptographic assumptions rely on and is used for most security definitions. The challenge of reconciling these two views of cryptography lies in security under the presence of encryption cycles. In this thesis, we provide a proof of completeness for Abadi-Rogaway symbolic logic with respect to KDM security, a strong form of circular security. Further, we provide a larger set of expressions for which Micciancio's symbolic logic is complete with respect to CPA security, extending Micciancio's completeness, which holds only for the set of acyclic expressions. We also give an alternate characterization of Micciancio's logic. On the computational side, we give a proof that circular insecurity is maintained as cycle length decreases, which is not a previously shown result.
Dynamic and cost-efficient deployment of large language models using uplift modeling and multi armed bandits
(2025) Tongay, Ninad; Chester, Sean
The rapid advancement of large language models (LLMs) has brought about a new class of challenges in balancing performance, cost, and scalability. As organizations seek to deploy these models in production environments, a key question arises: how can we maintain the quality of responses delivered by advanced LLMs while reducing the significant computational and financial costs associated with them? Relying entirely on high-end models like GPT-4 can ensure quality but often proves economically unsustainable, while defaulting to smaller, cheaper models may sacrifice performance and user satisfaction. This tension calls for more intelligent decision-making strategies ones that dynamically allocate queries to the most appropriate model depending on the task’s complexity and expected value. To address this, we propose a hybrid decision-making framework that brings together causal uplift modeling and multi-armed bandits to drive cost-aware, adaptive model selection. Uplift modeling enables the system to reason causally about the benefit of using a stronger model for a specific query, thereby offering interpretable, feature-informed decisions from the outset. These predictions serve as a strong offline prior. The bandit component builds on this by adapting the policy in real time learning from feedback, correcting for model mispredictions, and responding to shifts in query distribution or underlying model performance. This fusion of causal inference and online learning results in a system that is not only efficient and scalable, but also interpretable and responsive to real-world variability. We validate the approach through controlled simulations that mimic real deployment conditions, including concept drift, shifts in user query types, and the emergence of unseen domains. Across these scenarios, the hybrid consistently achieves a more favorable balance between quality and cost than baseline strategies. Furthermore, the system is designed to expose its decision-making logic, offering transparency through uplift scores and feature-based justifications a critical requirement for high-stakes AI deployments. By combining performance, cost awareness, and explainability, this work contributes a practical solution to the growing need for intelligent model orchestration in the multi-LLM landscape.
Morphology agnostic multi-agent character control
(2025) Zhang, Rui; Haworth, Brandon
Crowd simulation plays a crucial role in various applications, from urban planning to virtual reality, by modeling realistic pedestrian behavior and interactions. Traditional approaches typically utilize simplified agent representation such as particles, whereas recent advancements have introduced fully physical character models in crowds, which relies on morphology-specific motion control, limiting their applicability to heterogeneous agents with diverse body structures and movement capabilities. This thesis introduces a morphology-agnostic multi-agent character control framework that integrates physics-based locomotion with hierarchical reinforcement learning. A low-level locomotion controller utilizes generalized goal conditioning to enable robust and adaptable movement across agents with different morphologies through parameter sharing, eliminating the need for predefined gait cycles or morphology-specific trajectory planning. A high-level navigation controller processes morphology-agnostic state observations and integrates visual attention sampling to improve decision-making. The navigation controller provides goal conditioning to the locomotion controller, guiding agents toward their target positions in dynamic environments. The proposed system improves generalizability in multi-agent settings by decoupling locomotion control from agent-specific kinematics while maintaining stability and responsiveness.
Fast trips: A scalable insertion operator approach for ridesharing over time-dependent road networks
(2025) Mukherjee, Aaditya; Chester, Sean; Nascimento, Mario A.
Effective Route planning for shared mobility (RPSM) is crucial for optimizing the goals of transportation services such as ridesharing, logistics, and food delivery. Route planning requires online integration of new transportation requests into existing routes of transportation workers while accounting for real-world conditions such as traffic congestion, variable travel speeds, and changing demand patterns. A core component of route-planning systems is the insertion operator, a state-of-the-art method that integrates new transportation requests into existing worker routes with minimal additional travel time. Although effective and fast for route-planning simulations on static road networks, route-planning simulations experience significant performance degradation when applied to real-world, time-dependent road networks (TDRNs), where travel times between roads fluctuate due to varying traffic conditions. This thesis addresses this scalability challenge by introducing an informed approach to partitioning the data used by the insertion operator into separate, disjoint batches. I propose a partioning method utilizing K-means clustering complemented by an opportunistic allocation of workers to clusters. This method reduces the large number of shortest path query invocations inherent to time-dependent insertions, significantly decreasing RPSM simulation speeds without sacrificing, and in some cases even improving the quality of the solutions. Through extensive experimental evaluations using large-scale, real-world datasets from major Chinese cities, the proposed method is compared against the sequential time-dependent insertion operator. The results indicate a minimum of 7X acceleration in RPSM simulation times and maximum speedups up to 24X, while consistently matching or surpassing the original insertion operator in terms of solution quality. Furthermore, the flexibility of the clustering approach allows for customizable trade-offs between simulation speed and service quality, ensuring adaptability to diverse operational goals of transportation services. Ultimately, this thesis offers a scalable, adaptable, and computationally efficient insertion operator framework capable of handling realistic scenarios in dynamic shared mobility environments, providing valuable tools for transportation and logistics companies seeking operational optimization.
Fast database join on ray-tracing core equipped GPU
(2025) Wu, Yijie; Chester, Sean
With the increase in GPU memory and computing power, GPU databases have become more popular, driving extensive research on GPU-based indexing. One study introduced a novel approach called RTX(Ray-tracing Index), which utilizes ray-tracing cores(RT cores) to accelerate GPU indexing. However, RTX suffers from a large build size and slow range queries. A follow-up work called cgRX(Coarse-granular Indexing), optimized the construction and range query algorithms, improving throughput by 1.5x–3x in relation to memory footprint, the range query time by 2x, and 5.5x faster updatability compared to RTX. However, the experimental results of cgRX may be inaccurate because RTX was not properly optimized as a baseline in cgRX, at least for the range query. To optimize the RTX, this thesis explores multiple OptiX(Nvidia's Ray-tracing Software API) optimization strategies for RTX, including a revised range query algorithm, BVH partitioning, reverse mapping, and spatially closed query mapping. Additionally, the best configurations are applied to other baselines, including cgRX. All these improvements together are used to reproduce the experiments in cgRX. The evaluation is first based on the impact of each optimization technique on RTX. These optimizations reduce RTX's memory usage during construction and improve range query performance. Then, cgRX, optimized RTX, and other baselines are compared using the same experimental setup as cgRX, all using their best configurations. The re-evaluated results differ significantly from those in cgRX. In summary, this thesis contributes to RTX optimization by exploring the effects of multiple optimization techniques. The optimized RTX and baselines configured with optimized settings collectively aim to develop a high-performance GPU database index.
A power-aware IoT-fog-cloud architecture for telehealth applications
(2025) Guo, Yunyong; Ganti, Sudhakar
This dissertation presents an energy-efficient model for integrating Internet of Things (IoT) devices with fog and cloud computing platforms, specifically designed for telehealth applications. As the deployment of telehealth IoT devices continues to grow, the demand for efficient, real-time data processing and energy conservation becomes increasingly critical. This research addresses these challenges by proposing a hybrid architecture that combines the low-latency benefits of fog computing with the scalable resources of cloud computing. The model reduces energy consumption by processing data locally through fog nodes, minimizing the need for constant communication with cloud servers. This not only decreases latency but also optimizes the use of computational resources, making the system more adaptable to the dynamic demands of telehealth services. The model is further enhanced by an adaptive resource scaling algorithm, which dynamically adjusts processing capacity based on workload, ensuring both efficiency and reliability in critical healthcare applications. Simulations studies demonstrate the effectiveness of the model in reducing energy consumption and improving system performance for real-time telehealth monitoring. The results show significant improvements in data processing speed, energy efficiency, and resource utilization compared to traditional cloud-only architectures. This work contributes to the ongoing development of sustainable telehealth solutions by providing a robust framework for IoT-fog-cloud integration that meets the stringent demands of modern healthcare systems.
CNN-based models for pitch estimation, modification, and auto-tuning
(2024) Jiang, Jiazhuo; Tzanetakis, George
Pitch estimation and pitch modification are fundamental audio processing tasks that are used in a variety of applications. An important example is the auto-tuning of vocals in which pitch estimation is applied, deviations from a desired target pitch are calculated, and the pitch of input vocal signal is modified to match the target pitch. Most existing approaches to auto-tuning are based on traditional digital signal processing (DSP) techniques for both the pitch detection and the pitch modification of the signal. In this thesis, the use of Convolutional Neural Networks (CNNs) is explored as a possible replacement of traditional DSP methods for pitch estimation, pitch modification as well as end-to-end autotuning. CNNs can model complex intput and output relationships and are more efficient than deep learning methods that take into account time/sequence information such as Long Term/Short Term (LSTM) networks and Recurrent Neural Networks (RNNs). The results show the potential of this approach as well as some of the challenges that need to be overcome. The experimental results indicate that larger data sets can result in better accuracy but they also tend to bring in more noise.
Real-time gesture-based sound control system
(2024) Khazaei, Mahya; Tzanetakis, George
This thesis presents a real-time, human-in-the-loop music control and manipulation system that dynamically adapts audio outputs based on the analysis of human movement captured via live-stream video. This project creates a responsive link between visual and auditory stimuli, fostering an interactive experience where dancers not only respond to music but dynamically influence it through their movements. The system enhances live performances, interactive installations, and personal entertainment, creating an immersive experience where users’ movements directly shape the music in real time. This project demonstrates how machine learning and signal processing techniques can create responsive audio-visual systems that evolve with each movement, bridging human interaction and machine response in a seamless loop. The system leverages computer vision techniques and machine learning tools to track and interpret the motion of individuals dancing or moving, enabling them to participate actively in shaping audio adjustments, such as tempo, pitch, effects, and playback sequence in real time. Constantly improving through ongoing training, the system allows users to generalize models for user-independent use by providing varied samples; around 50–80 samples are typically sufficient to label a simple gesture. Through an integrated pipeline of gesture training, cue mapping, and audio manipulation, this human-centered system continuously adapts to user input. Gestures are trained as signals from human to model, mapped to sound control commands, and then used to naturally manipulate audio elements.
Policy-value concordance for deep actor-critic reinforcement learning algorithms
(2024) Buro, Jonas; Haworth, Brandon
Designing general agents to optimize sequential decision-making underneath uncertainty has long been central to artificial intelligence research. Recent advances in deep reinforcement learning (RL) have made progress in this pursuit, achieving superhuman performance in a collection of challenging and visually complex domains, in a tabula rasa fashion without embedding human domain knowledge. Although making progress towards designing general problem-solving agents, these methods require significant amounts of data to learn effective decision-making policies relative to humans, preventing their application to most real-world problems for which no simulator exists. It is clear that the question of how to best learn models intended for downstream purposes such as planning in this context remains unresolved. Motivated by this gap in the literature, we propose a novel learning objective for RL algorithms with deep actor-critic architectures, with the goal of further investigating the efficacy of such methods as autonomous general problem solvers. These algorithms employ artificial neural networks as parameterized policy and value functions, which guide their decision-making processes. Our approach introduces a learning signal that explicitly captures desirable properties of the policy function in terms of the value function from the perspective of a downstream reward-maximizing agent. Specifically, the signal encourages the policy to favour actions in a manner that is concordant with the relative ordering of value function estimates during training. We hypothesize that when correctly balanced with other learning objectives, RL algorithms incorporating our method will converge to comparable strength policies using less real-world data relative to their original instantiations. To empirically investigate this hypothesis, we incorporate our technique with state-of-the-art RL algorithms, ranging from simple policy gradient actor-critic methods to more complex model-based architectures, and deploy them on standard deep RL benchmark tasks, and then perform statistical analysis on their performance data.
FTRL-WRR: Learning-based two-path scheduler for LEO networks
(2024) Li, Daoping; Pan, Jianping
Multipath QUIC is inspired by the resource pooling principle, aiming to make a collection of resources behave as a single pool. However, current multipath schedulers tend to prioritize specific metrics like Round-Trip Time (RTT) or congestion window, often overlooking strategies that enhance overall resource usage and reduce flow completion time. This can lead to resource underutilization in high dynamic settings, such as those involving Low Earth Orbit (LEO) satellites. Addressing this challenge requires efficient traffic allocation to maximize bandwidth utilization. In this thesis, we verify that the relationship between traffic distribution and throughput in a two-path scenario resembles a quasi-concave function. Accordingly, we formulate the traffic allocation across two paths as a 1-dimensional optimization problem. To solve the two-path scheduling problem in dynamic environments, we introduce the FTRL-WRR algorithm. This approach integrates a Follow The Regularized Leader (FTRL) learner, ADWIN2 distribution change detector, and Weighted Round Robin (WRR) scheduler to enhance bandwidth utilization. We validate the effectiveness of the algorithm through extensive emulation and real-world testbed experiments, demonstrating consistent reduction in completion time across a range of scenarios. Additionally, we discuss the algorithm's limitations and suggest directions for future research.
Exploring text-based support for designing weave drafts
(2024) Nayar, Chehak; Somanath, Sowmya
We present the design and evaluation of Textere — a tool that helps weavers use text inputs to design weave drafts for weaving. Our research lies at the intersection of two areas of research: (i) text-based design tools, and (ii) design tools in weaving. Text-based design tools have been explored by researchers in various domains like garment design, 3D modeling, and data visualization, showing benefits for expanding creative possibilities, enabling rapid prototyping, and making design processes more accessible for a broader range of users. Motivated by such benefits, in our research we explore how text-based tools can help with designing weave drafts. Weaving is a design and production activity, wherein weavers map ideas, inspiration, or client requirements to visual elements like pattern, color, and weave structures to design a weave draft. The drafts are then physically produced using a loom. Weave drafts are designed before production, to convey what the appearance of the final product will look like. Design tools in weaving use different modalities, like audio and tactile, to make the design process more accessible, creative, and efficient— benefits that design tools in other domains have achieved using text support. Several text-based scenarios in weaving, require interpretation of words from text inputs. Yet, current text-based techniques and design tools in weaving are limited to mapping individual alphabets to specific weaving elements, or incorporating text as is in the weave draft. We extend this research space by exploring how weave drafts can be designed using meaning or interpretations of words. We developed Textere, a text-based tool for designing weave drafts using the open source AdaCAD weaving platform. Using Textere, weavers can map text inputs to visual elements such as color, weave structure, and patterns based on meanings and interpretations. We curated the text-to-visual mappings used in our system from existing user studies in research, that describe how people associate words to visual elements. To evaluate Textere, we first used the evaluation-by-demonstration method, to produce four physical woven samples designed using our tool. Further, we conducted a qualitative study with 12 weavers to evaluate opportunities and limitations of using Textere, by comparing workflows to the tool we extended, AdaCAD, with no explicit text-to-visual support. From our study we learned about the strengths and limitations of Textere. Informed by our results, we further discuss how text-based design tools like Textere can enable reflective decision making, generation and broadening of ideas, gaining different perspectives on what visual elements represent, and contribute to an ecosystem of tools for designing weave drafts. This thesis makes three contributions: i) a novel tool for designing weave drafts using text inputs, ii) empirical findings on the benefits and limitations of text-based interactions for designing weave drafts, and iii) a set of design implications for future text-based design technologies.
Multi-agent footstep steering with deep reinforcement learning
(2024) Peng, Kun; Haworth, Brandon
Crowd simulation plays a crucial role in a wide range of fields, from digital media to urban planning. However, traditional particle-based algorithms often lack essential information to present realistic human bipedal locomotion. This research aims to propose a more realistic and efficient steering model for crowd simulation by combining Multi-Agent Reinforcement Learning (MARL) with bipedal locomotion modelling. This study explores the advantages of MARL and analyzes a mathematical approach to simplifying complex bipedal locomotion. The approach utilizes the Proximal Policy Optimization algorithm and trains the model in adjustable randomized maze-like environments. Assessment results of the model indicate that the model learns goal-reaching behaviours and learns to avoid static and dynamic obstacles. Furthermore, the agents can simulate complex steering behaviours such as side-stepping and turning-like behaviours with two feet. This research contributes to the advancement of the field of crowd simulation through a flexible and realistic approach to modelling human steering behaviours in complex and dynamic environments.
Enhancing fact-checking in large language models: Cost-effective claim verification through first-order logic reformulation
(2024) Asghari, Sara; Thomo, Alex; Srinivasan, Venkatesh
In the realm of Large Language Models (LLMs), the ability to accurately perform Fact Checking (FC) tasks, which involves verifying complex claims against challenging evidence from multiple sources, remains a crucial yet under-explored area. Our study presents a comprehensive benchmarking of various LLMs, including GPT-4, on this critical task. We utilize a modern, challenging dataset designed explicitly for fact-checking, HOVER, which comprises thousands of evidence-claim pairs covering diverse aspects of life, history, and entertainment. This dataset differs from common datasets that evaluate the reading comprehension capabilities of LLMs, which are primarily composed of sets of question-and-answer pairs. Our findings demonstrate that GPT-4 not only decisively surpasses the current state-of-the-art (SOTA) models in FC tasks but also shows that other, open-source, LLMs (e.g. Mixtral and Llama-3) exhibit close-to-SOTA performance out-of-the-box. This implies that simply presenting these models with the evidence text and claim allows them to infer the claim’s veracity effectively. We contrast this with the existing SOTA methods, which involve complex, multi-step solutions, including the use of multiple LLMs to verify claims – a process that necessitates continuous updates and local execution, making it less accessible for regular users. Furthermore, we explore the impact of claim formulation on the FC task’s effectiveness. By converting complex claims into first-order logic (FOL) and then back into natural language, we observe improved performance in some LLMs, particularly with more challenging dataset subsets. This method, although utilizing GPT-4 for the FOL breakdown, serves as a practical guideline for users: more formally structured claims yield more reliable responses.
Edge server placement considering resilience in mobile edge computing networks
(2024) Begum, Syeda Mahfuza; Pan, Jianping
In today’s rapidly evolving communication landscape, the demand for exceptional Quality of Service (QoS) and Quality of Experience (QoE) in communication networks has reached unprecedented levels. This surge in demand can be attributed to the explosive growth and pervasive deployment of Internet infrastructure. Emerging technologies and novel applications underscore the urgency for a network architecture that not only delivers speed and efficiency but also boasts scalability and resilience beyond the capabilities of traditional cloud computing networks. Mobile Edge Computing (MEC) stands as a promising solution to address these challenges. By deploying Edge Servers (ESs) in close proximity to end-user devices, MEC enables the offloading of delay-sensitive and computationally intensive workloads from mobile applications. This deployment, in turn, mitigates latency issues and enhances the QoE for mobile users. However, the reliability of Edge Server Placement (ESP) within MEC networks is of paramount importance. While several studies have explored the ESP problem in MEC networks, they often focus on two main objectives: minimizing Edge Server (ES) access delay and optimizing workload distribution. However, one critical aspect has been relatively under-emphasized: the resiliency of ESP. The failure or malfunction of ESs, stemming from various challenges, can disrupt operations and degrade the overall QoS/QoE of the network. In this study, we tackle the ESP problem in MEC networks from a distinctive perspective. Our focal point is to minimize ES access delay, efficiently balance workloads, and significantly enhance network resilience. To achieve these objectives, our innovative algorithm employs a dual strategy. First, we utilize the robust K-medoids clustering algorithm for ESP, optimizing the architectural layout of MEC networks. Second, we introduce a bespoke heuristic algorithm designed to allocate multiple ESs to each Base Station (BS), thereby fortifying network resilience. This approach not only adheres to various constraints but also ensures uninterrupted services, even in the face of server failures, while consistently meeting key performance indicators. Experimental results, based on real-world data, prove the effectiveness of our algorithm. It not only reduces access delay and workload imbalances but also ensures responsive performance and uninterrupted services, even in scenarios involving ES failures.
Inferring network topology for distributed machine learning model training
(2024) An, Renjun; Wu, Kui
With the application of distributed machine learning in various industries, there is an increasing demand for model training using cloud computing resources. However, many cloud computing service providers refuse to provide end-users with information about the underlying network topology for commercial and security reasons. Due to this opaqueness, it is challenging to arrange the computation modules in different Virtual Machines (VMs) to achieve the best resource utilization efficiency. To address this problem, we propose an algorithm called Flow Tracking (FT), which uses external measurements to infer the internal structure of a general graph. Compared to the state-of-the-art topology inference algorithms, FT achieves the most accurate topology measured in four different metrics. Notably, FT achieves 100% reconstruction of the underlying topology under the shortest-path routing strategy of the underlying network. Experimentally, resource allocation using the inferred topology improves the model training efficiency significantly compared to random allocation.
Self-admitted scientific debt: Navigating cross-domain challenges in scientific software
(2024) Awon, Ahmed Musa; Ernst, Neil
Scientific software development faces unique cross-domain challenges, requiring expertise from both scientific and software engineering disciplines. These challenges often manifest as technical debt, specifically in the form of Self-Admitted Technical Debt (SATD). While technical debt is a well-recognized issue in software engineering, its impact within scientific software remains underexplored. In particular, the integration of domain-specific scientific knowledge with robust software engineering practices presents ongoing difficulties. This work investigates these cross-domain challenges in scientific software in various fields—including high-energy physics, astronomy, molecular biology, climate modeling, and applied mathematics—through SATD analysis. We examined 28,680 code comments from nine open-source scientific projects, identifying 11 types of technical debt. Among them, we introduced a novel category termed Scientific Debt, representing the issues that arise when integrating scientific findings with software development. We identified five key indicators of SD: assumptions, missing edge cases, accuracy challenges, translation challenges, and the incorporation of new scientific discoveries. Our findings reveal that Scientific Debt accumulates at a significantly higher rate than it is resolved, with the Missing Edge Cases indicator being the most frequently addressed. To further support the management of this debt, we explore the potential of Large Language Models (LLMs) in identifying and predicting cross-domain challenges. Our preliminary investigation suggests that LLMs could help detect issues requiring both scientific and software expertise, offering a promising direction for future efforts to manage and mitigate Scientific Debt.
Vectron: A dynamic programming auto vectorization framework
(2024) Naser Moghaddasi, Sourena; Numanagić, Ibrahim
Dynamic programming (DP) is a fundamental algorithmic strategy that decomposes large problems into manageable subproblems. It is a cornerstone of many important computational methods in diverse fields, especially in the field of computational genomics, where it is used for sequence comparison. However, as the scale of the data keeps increasing, these algorithms are becoming a major computational bottleneck, and there is a need for strategies that can improve their performance. Here, we present Vectron, a novel auto-vectorization suite that targets array-based DP implementations written in Python and converts them to efficient vectorized counterparts that can efficiently process multiple problem instances in parallel. Leveraging Single Instruction Multiple Data (SIMD) capabilities in modern CPUs, along with Graphics Processing Units (GPUs), Vectron delivers significant speedups, ranging from 10% to more than 20x, over the conventional C++ implementations and manually vectorized and domain-specific state-of-the-art implementations, without necessitating large algorithm or code changes. Vectron's generality enables automatic vectorization of any array-based DP algorithm and, as a result, presents an attractive solution to optimization challenges inherent to DP algorithms.

Browse

Recent Submissions