The Thief Orienteering Problem (ThOP) is a discrete routing optimization problem that seeks to find the best possible sequence of items to collect along a feasible simple path connecting a pair of given terminals. ThOP is to be solved under strict limitations, including a time constraint. ThOP combines the following classic combinatorial optimization problems: the Knapsack Problem (KP) and the Orienteering Problem (OP). As such, each of the now subproblem's components become dependent. Existing dynamic programming approaches that can efficiently solve ThOP on directed acyclic graphs (DAGs) have motivated the study of graph transformations that extend these methods to broader classes of graphs. As supervised by Andrew Bloch-Hansen, this research continues to build on his previous work specifically considering ThOP on the following planar graph classes with maximum treewidth 2: 2-terminal series-parallel and outerplanar graphs. This research furthers the research by investigating ThOP on Halin graphs. Halin graphs are also planar, but have a maximum treewidth of 3, meaning, when compared to a tree graph structure, wherein only one unique simple path exists between any pair of vertices, a Halin graph is structurally more complex than either a series-parallel or outerplanar graph. Complementing the theoretical work, implementation efforts included coding exact and polynomial-time approximation scheme (PTAS) dynamic programming solvers, and supporting software infrastructure, for ThOP based off Bloch-Hansen's previous publications, in Python. This research was completed over the 2026 summer in parallel with other undergraduate students also supervised by Bloch-Hansen: Ayla Ventura, Yacob Mesfun, and Yanishka Gahlot. Overall, this research aimed to identify graph structures and families that admit efficient DAG transformations and thereby expand the applicability of dynamic programming methods for solving ThOP.

Many graph problems become easier to solve when the graph is a directed acyclic graph (DAG), but many real-world graph classes contain cycles. This presentation explores the problem of transforming cyclic graph classes into DAGs while preserving all valid traversals required for the Thief Orienteering Problem (ThOP). We examine transformation techniques based on edge orientation and vertex duplication. We also investigate how graph properties such as pathwidth may help explain the complexity of these transformations and predict their growth as graph structure becomes more complex.

We know what legendre transforms but I my work makes it easy to understand what they mean and why they work including some of its properties.

Automatic speech recognition (ASR) is a crucial accessibility technology for many populations, yet the people who rely on it the most tend to see the worst performance. Modern speaker-independent ASR systems such as OpenAI's Whisper present a tantalizing promise of high out-of-the-box performance for English language transcription, without the need for lengthy speaker enrollment processes common with previous generations of ASR technology. However, exceptional benchmark accuracy easily drops to 30% error rates or worse for anyone who isn't a young, able, white female American. Fine-tuning can improve speech recognition accuracy by improving coverage of under-represented speech features in the model, but existing processes for fine-tuning Whisper present a major technical barrier for laypeople seeking to improve ASR accuracy for their own voice and accessibility needs. 

I present a software implementation project that automates the technically challenging aspects of data gathering, preparation, and fine-tuning the Whisper tiny.en speech recognition model. Preliminary results show that for a single speaker, marked improvements in ASR accuracy are possible with a limited amount of audio data. 

Researching abstract math is not as complicated as it might sound. Really, research is as much, if not more, about learning as it is about writing new things. In this presentation, I will discuss my experience doing math research, a little about the conclusions of that research, as well as what (in my opinion) abstract math is all about. 

The multi-trip capacitated prize-collecting vehicle routing problem (MT-CPC-VRP) with depot release dates generalizes classical VRP by allowing multiple depot-to-depot trips per vehicle within a workday horizon, treating customer visits as optional and profit-maximizing, and restricting service by release date. This combination is common in same-day delivery and field-service settings but underrepresented in the routing literature.

This work builds an experiment framework for the MT-CPC-VRP and implements multiple solution methods: an exact CPLEX-based formulation, an exact bitmask dynamic-programming solver, and a heuristic label-setting algorithm that constructs feasible depot-to-depot trips while pruning dominated partial routes. An independent evaluator recomputes feasibility and profit from returned solutions rather than trusting solver-reported values.

These baselines establish the performance floor and ceiling against which an upcoming adaptive large neighborhood search (ALNS) metaheuristic will be benchmarked.

Effective Air Traffic Control (ATC) communication is a required asset when piloting an aircraft, requiring in-depth knowledge of aviation phraseology, protocols, and aircraft instrumentation. Pilots must make time-sensitive decisions and be able to effectively communicate with ATC. Novice pilots and aviation students often experience difficulty under a high stress environment to execute these principles. Currently, outside of the airplane, opportunities for communication practice are limited, slowing student progress. 

Studies have shown that Simulation-Based Learning, a highly adopted practice across many fields, is effective at improving skill development. While standard elsewhere, this specific learning tool has limited implementations for Canadian aviation students and what does exist has restrictive utility to be tailored with relevant scenarios.

Thus, on behalf of the MRU aviation department and in collaboration with professors and students of multiple disciplines, ATC Assist is being developed as a learning first mobile application aimed to help improve the proficiency of aviation students conversing with ATC. 

This presentation will discuss the challenges associated with developing a learning application adapted to improve user proficiency at aviation communication. We will explore the design decisions that have been made thus far and the reasons for them. Finally, we will discuss future implementation and plans.

Algorithms used for solving NP complete problems often have to use heuristic or metaheuristic algorithms to find efficient solutions in a reasonable amount of time. Swarm Intelligence algorithms are one class of metaheuristic algorithms that utilize the behaviour of nature for algorithms that can solve difficult problems for more conventionally designed algorithms. The problem we worked with was the Thief Orienteering Problem, where there is a thief with a knapsack who is trying to maximize their profit by collecting the highest value items possible at each node or city, with there being travel costs to go between each node or city. One of the biggest challenges with this problem is the fact that as our weight increases by picking items up, our speed decreases, which leads to highly variable travel speed across edges. The swarm intelligence algorithm that was explored was the paper on “Efficiently solving the thief orienteering problem with a max-min ant colony optimization approach”, where the Thief Orienteering Problem is handled by a MAX-MIN ant colony system, which utilizes ant pheromones to allow the ants to iteratively discover the best routes possible, while a packing heuristic is used to determine the fitness of a route which would maximize the total profit. The authors had already created their own algorithm, with modifications being required to work with our graph class, which primarily utilizes DAGs. The importance of this work and research on The Thief Orienteering Problem is that we can compare a metaheuristic with a PTAS(polynomial time approximation scheme) with a strict time guarantee, to see if our PTAS algorithm can beat metaheuristic algorithms that mimic ant and ant colony behaviour to solve NP complete problems.

Server scheduling deals with minimizing response time -- that is, the interval between arrival and completion for a particular job -- while attempting to balance competing priorities such as seniority and job size. The scheduler First-Come, First-Served (FCFS), also known as First-in, First-Out (FIFO), prioritizes jobs by arrival time. On the other hand, Processor Sharing (PS), an ideal version of round-robin scheduling, divides processing capacity equally among all active jobs.

Because both FCFS and PS are easy to implement, these schedules form the basis for many real world scheduling systems.

Scheduling performance can be strengthened by job-size awareness in the scheduling process. That is, schedulers that usethe information of job sizes in their schedule can outperform schedulers that don't. In particular, the Fair Sojourn Protocol (FSP) guarantees that, when job sizes are known, no job completes later than it would under PS. To investigate the resulting performance trade-off, we used a C++ simulator provided by my advisor. The simulator's architecture is fully decoupled whereby its scheduling, workload generation, and logging are implemented as independent modules, allowing each component to be extended irrespective of the others.

Using this simulator, we evaluated the scheduler performance under varying conditions. We compared the mean response time and slowdown of FSP and Practical Size-Based Scheduling (PSBS) against those of PS across different job sizes.

In experiments without job-size estimation error, the scheduling policies behaved as expected. During the final stage of the project, we started the groundwork for evaluating how these policies perform when job-size estimates are inaccurate.

Split Federated Learning (SFL) trains a shared model across many clients and a server while keeping raw data local, by splitting the network at a cut layer. This talk covers three pieces of work. First, SplitBench, an open-source framework that brings the main split learning and split federated learning variants into one configurable platform, so algorithms can be compared fairly and reproducibly under identical conditions. Second, using that framework to study class imbalance on the DermaMNIST skin-lesion dataset, where increasing data heterogeneity drops overall accuracy only modestly but collapses minority-class macro-F1 from 65% to 27%, showing why class-balanced metrics matter. Third, extending the framework with per-client dynamic cut selection, where each client gets its own cut layer from its compute and network speed, so faster clients keep working instead of idling while the slowest client sets the round time, cutting training time at no cost to accuracy.