DalSpace Institutional Repository

DalSpace is a digital service that collects, preserves, and distributes digital material produced by the Dalhousie community.

To learn about content guidelines, policies, and how to deposit, view the Help documents.
Contact us to get started submitting content to DalSpace at dalspace@dal.ca

Communities in DSpace

Select a community to browse its collections.

Now showing 1 - 10 of 19

Recent Submissions

Access status: Open Access ,
Replay Buffer with Efficient Local Forgetting for Adapting to Local Environment Changes in Deep Model-Based Reinforcement Learning
(2026-06-13) MacDonald, Ruis; Not Applicable; Master of Computer Science; Faculty of Computer Science; Not Applicable; NA; Not Applicable; Hassan Sajjad; Gabriel Spadon; Nils Wilde; Janarthanan Rajendran; Sageev Oore
Reinforcement learning (RL) is a computational framework in which an agent learns to achieve a goal, represented by a numerical reward signal, by interacting with its environment. Model-free RL methods learn how to act directly from experience, whereas model-based RL (MBRL) methods learn an environment model that predicts the next state and reward from the agent’s state and action, then use simulated data from that model to improve behaviour. When the environment changes locally, an MBRL agent can update its model using experience from the changed region and use planning to adjust its behaviour across the full state space. MBRL methods typically train the environment model using a standard first-in-first-out (FIFO) replay buffer, which removes the globally oldest experience once full, regardless of where it comes from in the state space. After a local change, FIFO buffers can cause two problems: outdated pre-change experience can mix with new post-change experience in the changed region, causing conflicting updates, while experience from regions not recently visited can be discarded, reducing model accuracy there. A Local Forgetting (LoFo) replay buffer addresses this interference-forgetting dilemma by discarding the oldest experience from the same local region as each new experience rather than from the buffer globally. However, the current LoFo instantiation identifies local regions by comparing each new transition’s learned state embedding with all stored transitions, which is computationally expensive. We present LoFoV2, which uses hash-based region assignment and per-region FIFO buffers to preserve localized discarding without buffer-wide distance comparisons. LoFoV2 uses the same learned state embedding as LoFoV1 but assigns each new experience directly to an approximate local region using locality-sensitive hashing. We evaluate adaptation to local reward changes using Deep Dyna-Q and DreamerV2 on low- and high-dimensional environments. For Deep Dyna-Q, LoFoV2 improves computational efficiency while maintaining adaptivity. For DreamerV2, the current instantiation does not yet achieve the same adaptive performance; however, our experiments identify challenges involved in extending LoFoV2 to this setting and provide guidance for future work.
Access status: Open Access ,
2026 International Coastal Access Symposium Summary Report
(Dalhousie University, Marine Affairs Program, 2026-06-09) Winkler, Nicolas; Harrison, Hannah L.; Hull, Lily; Manuel, Patricia; Kofahl, Mike
Access status: Open Access ,
Combining multiple groundwater hydrograph analyses to characterize aquifer dynamics and drivers in complex hydrogeological settings
(Elsevier, 2026-08-01) Chai, Z.; LeRoux, N.K.; Jamieson, R.C.; Hill, A.M.R.; Somers, L.D.; Kurylyk, B.L.
Access status: Open Access ,
Deep Learning for Field-Based Cereal Phenomics
(2026-05-29) Ravichandran, Prabahar; Yes; Doctor of Philosophy; Department of Mechanical Engineering; Not Applicable; Dr. Hao Gan; Yes; Dr. Clifton Johnston; Dr. Vasantha Rupasinghe; Dr. Ya-Jun Pan; Dr. Young Ki Chang
Plant phenomics has become a critical bottleneck in crop improvement, as the capacity to generate high-quality phenotypic data has not kept pace with advances in genotyping. This thesis integrates deep learning into cereal phenomics through end-to-end, non-destructive pipelines that combine proximal and remote sensing with modern neural network architectures. Near-infrared spectroscopy (NIRS) coupled with convolutional neural networks (CNNs) is applied to rice for predicting amylose content, chalkiness, grain dimensions, and grain shape. Visible to near-infrared (VIS-NIR) hyperspectral imaging is investigated for wheat grain phenotyping using grain-level patches and one-, two-, and three-dimensional CNN architectures, where reformulating cultivar-independent prediction as a classification task improves robustness for moisture and protein. For field-based phenotyping, a deep learning framework estimates rice blast severity from canopy-level images acquired in the Universal Blast Nursery (UBN), and foundation and instance segmentation models are fine-tuned to delineate breeding plots from uncrewed aerial vehicle (UAV) imagery, supporting reliable plot-level trait extraction. The work also treats phenomics models as production systems: pipelines are containerised and deployed across workstations, high-performance computing, and cloud platforms, with systematic evaluation of inference latency and system-level bottlenecks. Together, these contributions demonstrate that deep learning can convert plant phenomics from a limiting factor into a scalable, decision-ready component of modern cereal breeding programs.
Access status: Open Access ,
Narratives of Multispecies Assemblages: Human Engagement with Roadkilled Animals
(2026) Perez-Haynes, Saylor
Access status: Open Access ,
Special or Restricted? Autistic People's Experiences of Special Interests
(2026-05) Aydogan, Clover