Undergraduate Student Research

Permanent URI for this community

https://dspace.library.uvic.ca/handle/1828/5899

Browse

Now showing 1 - 20 of 33

Analysing the Performance of Cloud Gaming over a Low-Earth Orbit Satellite Network
(2023-09-18) Tolouei, Pouria
Starlink, as it continues to grow, brings the idea of a robust global network coverage one step closer to reality. Even though Starlink's performance has been a breakthrough for satellite internet by utilising low-earth orbit satellites, frequent satellite handovers are required to keep up with the high mobility of the satellites. Starlink performs these handovers every 15 seconds at synchronised 57, 12, 27, and 42 seconds after each minute. While this is effective for maintaining the connection, it leads to fluctuations in the latency which is not efficient for latency-sensitive applications such as cloud gaming. This study investigates the effect of the fast-changing delay of Starlink on cloud gaming. Through automated gameplay and data collection, the quantitative data analysis revealed that the Starlink network experiences higher and faster changing latency, less stable bandwidth, and more packet loss compared to a more traditional network. Furthermore, the satellite handovers cause frequent and predictable performance drops, especially in terms of input latency and packet loss which lead to stutters and input lag. While the fundamental structure of the Starlink network cannot be changed, the satellite handover can be anticipated by applications due to their fixed and synchronised nature. So, working toward a solution for satellite handovers will be the future of this work.
Analyzing the Accuracy of Agent Representations in Crowd Simulations
(2022-09-08) Shatzel, Liam
A crowd simulation is a computer-based simulation of a large number of characters. Crowd simulations are used, with increasing frequency, in fields such as movies, games, building design, and evacuation scenarios. For example, in emergency evacuation scenarios, crowd simulations can be used to predict the behavior of humans in order to test the safety of a building design. The focus of this research is to analyze the accuracy of representations of virtual humans (agents) within crowd simulations. Typically, disks are used to represent agents within crowd simulations, but these can oversimplify the area that an agent occupies. Having accurate representations of agents is necessary for a simulated crowd to reflect reality. In this work we focus on a comparative analysis, contrasting overestimates and underestimates of standard disk representations. Through this research, we provide quantitative data on the oversimplification that disk-based models present. This data can be used as a starting point for exploring better representations of humans in motion within simulated crowds.
Analyzing the broadband divide in the underserved Southern Gulf Islands
(University of Victoria, 2025) Archer, Finnegan
Broadband internet access has rapidly become essential for facilitating social, educational, and economic development. Yet rural communities lacking the economies of scale to justify the upfront costs of traditional cabled connections face a digital divide. To address this, line-of-sight (LOS) technologies help alleviate infrastructure deployment expenses, although they come with unique performance challenges. Starlink, SpaceX's low-Earth-orbit (LEO) satellite service, connects user terminals to overhead satellites and can outperform the Universal Service Objective (USO). Similarly, cellular networks such as those provided by Rogers and Telus use radio signals optimized for coverage and reliability. Our research explores the real-world performance of these solutions in the Southern Gulf Islands using a vehicular setup equipped with a Starlink Mini dish, an external GPS module, a dashboard camera, and cellular modems carrying Telus and Rogers SIMs. Metrics including latency, iperf3 throughput, GPS location, and cellular SNR were gathered alongside footage to assess performance. Quantitative data analysis demonstrated that regional terrain features, such as dense tree coverage, heavily affect line-of-sight performance.
Applications of graph neural networks in simulation vs. animation
(University Of Victoria, 2025) Shatzel, Liam
Graph neural networks (GNNs) provide a method, using graphs, to organize data which is not typically suited for a matrix or linear structure, allowing it to be learned in a neural network. GNNs have applications in drug discovery, recommendation systems, social networks, and physics simulations. This research focuses on analyzing the applications of GNNs for physical systems from an animation and simulation viewpoint. Simulation is the modelling of dynamic systems using physics, whereas animation is a sequence of images played to depict motion. In this work we first find the best hyperparameters for each use-case, then compare the mean squared error, model size, and rollout speed of each. Exploring the effectiveness of GNNs, when applied to animation and simulation, demonstrates their capabilities in media and physics. This also serves as a starting point for the applicability of GNNs to different domains.
Auditing genAI reasoning in qualitative analysis: A prompt-level case study
(University of Victoria, 2026) Baidwan, Deepkhushi
GenAI tools are increasingly used to assist qualitative researchers with tasks like transcript coding and theme generation. However, most evaluations focus only on whether GenAI outputs match human-produced codes, not on how the model arrived at those outputs. This study audits the reasoning process of NotebookLM, a retrieval-augmented generation (RAG) tool, across a structured 12-prompt sequence applied to eight interview transcripts on open pedagogy and institutional constraints. The audit logged 20 analytic decisions across two coding pathways (inductive and deductive), tracking triggers, rationale, and consequences at each stage. Four key observations emerged: prompt wording directly shaped coding behavior; supporting evidence was retrieved after coding decisions rather than driving them (a ""provenance gap""); ambiguous passages were automatically resolved without surfacing alternatives; and outputs visually resembled rigorous audit reports without being independently verifiable. The central finding is that GenAI outputs can perform methodological accountability without genuinely providing it, with significant implications for peer review, replication, and research integrity.
Because I Am A Woman: An Investigation of Gender Biases In Engineering Fields
(2022-06-29) Devathasan, Kezia
This project endeavoured to better understand the social factors that contribute to the success of female identifying students in Computer Science (CS) and Engineering. Extensive previous research has presented statistics on the proportion of women and other marginalized genders which drop out of pursuing careers or education in CS and Engineering, though little work has been done to accurately identify the reasons behind this attrition. The research that has attempted to identify these reasons has been recently challenged by other authors, who claim that hypotheses such as “The Leaky Pipeline” do not accurately represent the experiences of women in technology. A series of interviews and focus groups with undergraduate students in Engineering and Computer Science programs was held to discuss the reasons for potential differences. The focus groups provided insight into the types of biases faced by female identifying students in a Computer Science or Engineering field, and how these biases might impact their decision to pursue a career or education in this field. Major themes discovered included social barriers, barriers faced specifically in the classroom setting, impactful fear, and having to face microaggressions. Findings from this study contribute to making educational spaces in STEM more inclusive for female identifying students.
Computer Classification of News
(University of Victoria, 2024) Liu, Siyi
Facing the growing demand for psychological services, if we want computers to provide psychological services to humans, the first step is to identify human emotions accurately. This research is about how to enable computers to accurately identify text content, analyze it, and make judgments. In this process, the torchtext.datasets.AG_NEWS dataset was selected. AG_NEWS has four large classes ("World", "Sports", "Business", "Sci/Tech"). TfidfVectorizer was used to judge the importance of each word in a sentence. Common words like articles frequently appear in every sentence, so they are ignored. The importance of other words for sentence classification is judged based on their frequency of appearance – the higher the frequency of a word, the less weight it carries. Support Vector Machines are then used to optimize accuracy. The research showed AI tools can predict human emotions based on the interpretation of the text. If AI tools can accurately predict human emotions, this builds a foundation for the machines to respond appropriately to people needing psychological services.
Designing a Standardized Benchmark for OpenEXR
(University Of Victoria, 2025) Randhawa, Gurtej
This research proposes a new, domain-specific benchmark for evaluating OpenEXR performance in high-dynamic-range (HDR) image workflows. Building upon established benchmarking criteria (relevance, repeatability, verifiability, domain specificity) drawn from TPC-C guidelines and other benchmarking principles, we collaborated with Academy Software Foundation (ASWF) to identify five core EXR use cases: Simple Read, Playback, Compositing, Rendering, and Format Mismatch. Each test suite targets essential tasks in real-world VFX pipelines such as loading full images, reading sub-rectangles in daily review sessions, managing scanlines during compositing, accessing tiles for rendering, and coping with mismatched file layouts. Our methodology ensures meaningful, reproducible metrics (e.g., load time, memory usage) aligned with production realities. By incorporating open-source tools, public datasets, and community feedback, this benchmark aims to become a transparent, living standard that drives innovation, reduces storage costs, and improves overall efficiency in VFX pipelines.
Detecting Fake Users on Social Media with Neo4j and Random Forest Classifier
(2020-06-08) Zhao, Yichun
Fake news is defined by the Ethical Journalism Network as “deliberately fabricated and published” information intended “to deceive and mislead others.” It manipulates the ignorant into false beliefs and causes negative societal impacts. Fake social media users are perceived as popular, and they spread fake news by making it look real. The objective of this research project is to improve the accuracy of detecting fake users in the previous study by A. Mehrotra, M. Sarreddy and S. Singh, by using different centrality measures supported by the Neo4j graph database and two more datasets. The machine learning algorithm - random forest classifier, which uses the centrality measures as its features, detects fake users on social media with reasonable results.
Developments in The Cubing Problem
(University of Victoria, 2024) Janssen, Noah
In our research, we set out to make progress in either developing or proving it is likely impossible to develop a polynomial time algorithm to partition orthogonal polyhedra (cornerhedra) into the minimum number of boxes possible. For our purposes, a box is either a rectangle or square shape (i.e. a rectangular prism) made of UCs that is filled in. This is a special case of the problem of rectangular partition of polyhedrons which has applications in various fields and industries. In our research, we proved that cornerhedra can either have certain qualities or cannot. Examples of these are that some cornerhedra can be partitioned without splitting 2 layers and others cannot. We have also shown that the minimum number of boxes some cornerhedra can be partitioned into is greater than the number of outward-facing corners of that cornerhedron. Outside of this, we formulated conjectures about possible cornerhedra. An example of a conjecture would be whether there exists a cornerhedron such that when any two of its layers are split at least 2 new outward-facing corners are created.
Don't let that CPU sit idle: Hardware-aware heterogeneous general matrix multiplication (GEMM)
(University of Victoria, 2026) Warawa, Johnathan; Chester, Sean
Matrix multiplication is a fundamental operation used to train neural networks for machine learning. GPUs are well-optimized for several stages of this operation and are thus used to accelerate the work, however, GPUs must be "hosted" by CPUs that remain underutilized while the GPU works, burning cycles that could be put to use by a more sophisticated, heterogeneous algorithm that makes use of both the GPU and CPU at the same time. In this project, we increase the speed of these operations beyond what a single processor could accomplish by developing a heterogeneous algorithm which efficiently divides and interleaves these operations.
Effects of Similar Environments on Presence and Learning in Virtual Reality: Applications for Tsunami Preparedness
(2020-06-08) Johnson, Emma
Research surrounding the use of virtual reality for behaviour change is currently focused on desensitisation and influencing mood, but the potential of VR goes far beyond that. I propose to study how experiencing a profound or eye-opening situation in virtual reality can change the outlook and stance on emergency preparedness; while exploring how levels of presence and enviroment type change the effectiveness of anxiety inducing virtual environments for learning.
Enhancing relational databases with semantic search using word embeddings
(University of Victoria, 2026) Peng, Gerry
Relational databases can store and query structured data, but searching text is mostly limited to exact keyword matching. As a result, it can be difficult to retrieve conceptually related entries when different wording is used. This project explores how word embeddings represent semantic meaning in text and how they can be integrated into a relational database to support meaning based search. The project began by analyzing word embeddings to examine how semantic structure is captured in embedding spaces. Pretrained vector embeddings are compared with custom embeddings trained on a dataset of tweets related to the 2016 American election. Nearest neighbour analysis and vector arithmetic are used to observe how training data size and bias affect the resulting embeddings. Additionally, word embeddings are integrated into a movie database. Movie titles and plot descriptions are represented using word embeddings, and similarity comparisons retrieve movies based on semantic relevance rather than keyword matches. The results show embeddings trained on more general text corpora have more comprehensive semantic relationships, while embeddings trained on niche text perform well within their domain. Overall, this project shows word embeddings provide a practical way to extend traditional database systems with search that factors in word meaning.
Exploring clustering algorithms on spatial data in video games
(University of Victoria, 2025) Fitsner, Elise
This project looks at how unsupervised machine learning (specifically, clustering data using a Gaussian mixture model) can be used to try to identify different player archetypes in Counter-Strike: Global Offensive, using only data that relates to the spatial behaviour of the players. We found that within each side, a player's individual playstyle and the style of the round they were playing seem to have had more influence on the clustering than which map they were playing on, and we were able to identify four distinct styles of play on each side. In future work, this approach could be applied to areas such as cheat detection, improving video game map design, and quantifying effective play patterns to support skill development and training.
Exploring the implementation of gradients in vector graphics images through colour diffusion graphs
(2023-09-19) Goodwin, Linnea
Given that vector graphics aims to create infinitely scalable images, a problem arises when attempting to create a smooth gradient from one colour to another. The dominant method used to tackle this problem is essentially to simulate the diffusion of colour onto a triangulated mesh; the process would be similar to letting a drop of ink hit a wet napkin and spread out. The equation that governs diffusion is the Poisson equation, which allows for a function modifier to the spread of a material; i.e. a variable coefficient of diffusion across a surface. Previous research into the diffusion of colour to create vectorized gradients ignores this coefficient for a constant spread of colour. In this study, we recreate the spread of colour to form vectorized meshes with smooth gradients using PolyFEM and allow for the inclusion of a position-variable coefficient of spread. This research could advance the world of vector image generation and have potentially publishable applications.
A Forest of Code: Visualizing the Release Information of the Linux Kernel
(2016-04-12) Wilde, Evan
With an average of over 900 top-level merges into the Linux kernel per release, maintenance of older versions of the kernel becomes nearly impossible. Maintainers must be able to understand how changes to the current vision of the kernel fit into older versions of the kernel. This presents the need for a tool to provide meaningful explanation of what is happening in the kernel. Our goal is to design a web-based system capable of visualizing the commit and release information of the Linux kernel in a meaningful way.
From clicks to constructs: Valorant player profiles and the link to motivation, experience, and wellbeing
(University of Victoria, 2026) Duco, Jim
Unpacking the complex relationship between players' in-game behaviours and psychosocial factors is necessary to differentiate when gaming leads to benefits versus harms, which can inform public policy and research. This study explores how different patterns of play in Valorant relate to player motivation, experience, and wellbeing. Using publicly available game data from 1,974 players and self-reported psychosocial survey data from a subset of 160 participants, we apply a clustering approach to identify distinct player profiles. Exploratory Factor Analysis (EFA) reduces behavioural variables into 14 latent factors, which are then analyzed using Latent Profile Analysis (LPA) to reveal eight gameplay profiles characterized by differences in playstyle, performance, and social behaviour. The profiles are subsequently linked to measures of psychological needs, motivation, wellbeing, and potential problematic gaming indicators. Results suggest that gaming outcomes are shaped more by how, when, why, and with whom individuals play rather than by the game itself. The findings highlight the diversity of player experiences and demonstrate how behavioural clustering can generate insights about gaming and wellbeing without requiring fully labelled datasets, providing a foundation for future causal and longitudinal research.
Hydrophobicity at the molecular scale: Characterizing the aqueous-polystyrene interface with vibrational sum-frequency generation spectroscopy
(University Of Victoria, 2025) Bevington, Arden; Uddin, Md. Mosfeq; Hore, Dennis K.
Interfaces between hydrophobic polymer materials and water are ubiquitous in our everyday lives, from plastic water bottles and raincoats to cutting edge materials and drug delivery capsules. Surprisingly, the molecular-level interactions that occur at their interfaces with water are not well understood. Of particular interest is the origin of the surface charge that has been observed at these interfaces. In this work we investigate water structure and surface charge at the interface with a common hydrophobic polymer, polystyrene, using variable-angle vibrational sum-frequency generation spectroscopy (SFG). SFG is an inherently surface-specific spectroscopic technique that allows us to characterize molecular vibrations at the interface. Variable-angle SFG reveals that water molecules at the surface have a net orientation with their hydrogen atoms pointed towards the surface and show an increase in the magnitude of the surface charge as ionic strength of the solution is increased.
Improving Typing Experiences Through the Use of a Keyboard Interface With Integrated Gesture Recognition
(2023-03-17) Norrie, Samantha
The current keyboard typing experience is inefficient due to the need to rely on external technologies such as mice or trackpads. The Keyboard Interface With Integrated Gesture Recognition (KIWIGR) aims to solve this issue by implementing common word processing actions into keyboard gestures. The gestures explored in this research project aid users with document navigation as well as text highlighting. A sensor tablet and a machine learning model were used to implement these gestures into a keyboard-like system. The current prototype of the KIWIGR uses image amplification, image combination, and the aforementioned machine learning model to help predict keyboard gestures.
Insights into the Nature of Drum Machine Sounds
(2017-04-10) Shier, Jordie
Since the early 1970s programmable drum machines have been used extensively in the creation and performance of popular music. Recordings of individual drum sounds from these drum machines are widely available online and the use of such audio is common practice in music production. This work seeks to shed light on the sonic qualities of drum machine sounds using a sample set of 2026 kick and 2204 snare drums representing over 200 different drum machines. A set of audio feature extraction algorithms from the field of Music Information Retrieval (MIR) is used to create a feature vector describing each sample. The primary components of variation found are Spectral Centroid and Loudness for both kick and snare drums. This is introductory work in the analysis of drum machine samples and the results may prove useful for future work in the field of Intelligent Music Production (IMP) as well as in the development of music production pedagogical methods.

Browse

Browsing Undergraduate Student Research by Department "Department of Computer Science"

Results Per Page

Sort Options