Pusat Pengajian Sains Komputer - Tesis

Browse

Now showing 1 - 5 of 494

Near-Miss Traffic Trajectory Detection Based On Deep Learning
(2025-02)
Yang, Lu
Computer vision-based methods have indeed been widely employed for monitoring road traffic conditions. Traffic safety is a critical concern in urban environments, with near-miss events serving as valuable indicators of potential accidents. In this research, an innovative framework is proposed that combines yolov7 with transformer-based structures and segmentation techniques for robust object detection, tracking, and near-miss event analysis in traffic scenarios. Utilizing the real-time object detection capabilities of yolov7, it is augmented through the integration of transformer architectures. This enhancement enables the capture of longrange dependencies and contextual information, thereby improving accuracy in object recognition and localization. Additionally, segmentation methods are employed to delineate objects within the scene, further refining the detection box to better fit the target object.
An Improved Hierarchical Deep Reinforcement Learning For Complex Imperfect-Information Card Games
(2025-03)
Luo, Qian
Deep reinforcement learning (drl) has achieved significant breakthroughs in a variety of games, both with perfect and imperfect information, such as go, texas hold’em, and starcraft ii. However, doudizhu and big2 are classic complex card games with imperfect information and are popular in asia. They present new challenges for ai in competition, cooperation, inferring imperfect information, handling large state-action spaces, and training with sparse rewards. The deep monte carlo (dmc) method for these card games achieves significant success but still faces three key research problems: slowlearning speed, high loss during learning, and performance optimization. The primary objective of this research is to enhance the performance for these complex imperfect-information card games with a hierarchical deep reinforcement learning (hdrl) framework. Specifically, this main goal is divided into three sub-research objectives: improving learning efficiency indmctraining through oracle guiding, enhancing learning stability with adaptive deep monte carlo (admc), and improving the performance of proximal policy optimization (ppo) using relative advantage reward shaping (rars).
Improving The Feature Selection, Multi-Class Classification, And Imbalanced Dataset Of Breast Cancer
(2025-02)
Rahman, Emad Abd Al
The fact that breast cancer is prevalent among women makes it a serious problem in global health. Its early detection is crucial for improving treatment options and increasing survival rates, yet the complexity of diagnosing and determining the most effective treatment plan presents significant challenges. Recent years have seen the rise of ai and ml techniques as powerful tools in the fight against breast cancer, opening new possibilities for improving detection and treatment methods. This research aims to address key challenges in the application of ai/ml to breast cancer treatment planning, including imbalanced datasets, suboptimal feature selection methods, and the complexity of multi-class classification tasks. The study justifies its focus by addressing the unmet need for improved computational tools that can personalize and optimize treatment strategies for breast cancer patients. We aim to enhance model performance on imbalanced datasets by improving feature selection procedures, refine multi-class classification models, and to develop predictive models for personalized treatment planning. Successful completion of the thesis's objectives will allow it to make a substantial contribution to the creation of optimal treatment plans for individual patients.
Cloud Computing Adoption Model With The Moderating Role Of Perceived Trust In North-Esthern Nigerian Academic Libraries
(2025-03)
Yakubu, Aliyu Shehu
The role of cloud computing in enhancing service efficiency in academic libraries has been widely acknowledged. However, its adoption in nigerian academic libraries, particularly in the north-east, remains low or absent, leading to dissatisfaction with information service delivery. This lack of adoption has resulted in non-scalable services, data loss, limited partnerships, and technological obsolescence, all of which impact library efficiency and clients’ academic performance. This study investigates the factors influencing the adoption of cloud computing in north-eastern nigerian academic libraries, guided by the diffusion of innovation theory (doi) and the technological-organizational-environmental (toe) framework. Using a mixedmethods approach, 192 questionnaires were analyzed alongside interviews with 16 respondents to triangulate findings. Partial least square-structural equation modelling (pls-sem) was employed to test hypothesized relationships. Results show that all independent variables significantly influence the intention to adopt cloud computing, with government regulation exerting the highest influence (t-value: 6.725) and training/education the lowest (t-value: 1.925).
Optimized Location Dependent Data Retrieval Approach For Internet Of Things Based On Named Data Networking
(2025-02)
Ali, Aboodi Ahed Hussein
The internet of things (iot) demands efficient, adaptable, and scalable data retrieval mechanisms to meet the dual requirements of data-oriented and location-dependent host-oriented scenarios. Named data networking (ndn) offers a promising alternative to traditional ip-based architectures by focusing on content rather than host-based communication. However, existing ndn-based solutions face challenges in resource-constrained environments, including limited support for location-dependent data delivery and retrieval, inefficiencies in multicast forwarding, and high transmission overhead. This research introduces e-ndn, an enhanced ndn architecture tailored for wireless resource-constrained iot environments. E-ndn integrates three core modules: (1) the dlh naming scheme and local-first forwarding, which combines hierarchical location-enabled naming with proximity-aware interest suppression and backup forwarding procedures for enhanced reliability; (2) wildcard-based naming and forwarding, which optimizes multicast data retrieval by consolidating interest requests, reducing redundancy, and enabling more proximate location targeting; and (3) the path-selection module, which dynamically optimizes routing based on node proximity and capabilities, while defining and enabling broadcast domain limits to mitigate congestion.

Browse

Recent Submissions