https://www.um.edu.mt/library/oar/handle/123456789/10779 2026-04-04T16:56:40Z 2026-04-04T16:56:40Z https://www.um.edu.mt/library/oar/handle/123456789/142036 2025-12-09T11:05:38Z 2024-01-01T00:00:00Z

Title: Multi-vehicle ride-pooling system using reinforcement learning Abstract: Ride-pooling services have surged in popularity in recent years due to them being more efficient, convenient, and cost-effective than alternative traditional methods such as taxis. Leveraging mobile technologies, ride pooling services use the location of drivers and customers to assign a shared vehicle to passengers travelling in the same direction, reducing the number of vehicles on the road and, therefore, helping reduce traffic congestion. Such services require algorithms that dynamically match passengers with nearby drivers and optimise routes. Ride-pooling can be considered a variant of the Vehicle Routing Problem (VRP), which is a combinatorial NP-hard problem as it involves finding an efficient set of routes for a fleet of vehicles to serve customers while satisfying constraints such as customers’ time windows and vehicle capacity. Literature shows how the use of various metaheuristic algorithms to solve the VRP, such as tabu search (TS), can provide good solutions in terms of quality but suffer from scalability. With the recent advances in artificial intelligence, Reinforcement Learning (RL) is also being applied to capture the dynamic and stochastic nature of the VRP. In this research, we propose to use RL to solve the Multi-Vehicle Routing for Ride-Pooling Problem (MVRRPP) and generate solutions faster than the traditional metaheuristic methods. This algorithm aims to optimise passenger allocation to vehicles and vehicle routing while minimising the overall waiting time, travel time and total driving distance. We first implemented a baseline algorithm that uses TS with an initial solution consisting of equally distributed customers along the routes to solve the MVRRPP and establish its performance. We then model the MVRRPP as an RL problem and solve it using the REINFORCE algorithm with a dynamic attention model consisting of a dynamic encoder-decoder architecture. The performance of this model was compared with the results achieved using TS. Finally, we evaluate the effect of using an RL solution as input for the TS algorithm. The results of the three models showed that TS found higher-quality solutions than RL and TS with RL; however, its computational complexity resulted in longer computation times when solving large problem instances. Using RL involved a trade-off between solution quality and computation time, where it was quicker to find a solution even for problems on a large scale. On the other hand, performance using TS with RL showed minimal improvement except for reducing the distance travelled by vehicles, which suggests that the RL solution, used as the initial solution for TS, was in a region of the search space that had an inferior local optimum then the initial solution used in the TS without RL. The choice between TS and RL for solving the MVRRPP depends on the application’s requirements and the problem’s complexity. Description: M.Sc.(Melit.)

2024-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/141991 2025-12-05T10:17:45Z 2025-01-01T00:00:00Z

Title: AI‐Driven gesture recognition with smart gloves Abstract: This research presents the development of an AI‐driven gesture recognition system aimed to enhance Human‐Computer Interaction through the use of smart gloves. Many emerging applications, such as virtual reality, robotics, and assistive technologies, require detailed motion capture of the hand in three dimensions. Traditional input devices are not designed to capture such motion, whereas wearable solutions like smart gloves offer a practical means of collecting complex motion data for gesture interpretation. This study proposes a system capable of interpreting dynamic hand gestures captured using smart gloves. A custom dataset was collected using Rokoko smart gloves, recording 14 gesture classes from 14 subjects. Time‐series data captured from the smart gloves was preprocessed, and a range of feature extraction methods, including statistical, frequency‐domain, and motion‐based techniques, were applied. Experimental results were carried out to determine which features or combination of features gives the best result. Dimensionality reduction methods, namely Principal Component Analysis and Autoencoders, were examined to optimise the feature space and reduce complexity. A number of classification models were implemented and compared, including Support Vector Machines, K‐Nearest Neighbours, Hidden Markov Models, as well as, deep learning approaches such as CNN‐LSTM networks. Experimental results showed that while most models achieved high accuracy on validation data, up to 93.64%, performance significantly decreased when tested on data from unseen subjects, dropping to 20.39‐28.93%. This highlights the challenge of inter‐subject generalisation. To mitigate this, personalised models were implemented, showing good performance improvements. The SVM classifiers achieved accuracy results ranging from 67.9% to 92.9%, and the majority of precision, recall, and F1 scores exceeding 85%, while CNN‐LSTM models achieved an accuracy above 95% consistently. Precision, recall, and F1‐score values also remained high. This work contributes to the field of gesture recognition by systematically evaluating feature engineering and modelling techniques on multichannel time‐series data. It underscores the importance of personalised learning strategies and provides insight into the practical limitations of real‐world deployment, such as latency and subject variability. Future work may explore domain adaptation, multimodal sensing, and real‐time implementation to further advance robust gesture‐based interfaces. Description: M.Sc.(Melit.)

2025-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/141986 2025-12-05T10:05:45Z 2022-01-01T00:00:00Z

Title: Bayesian approaches for ligand-based virtual screening applications Abstract: The objective of computer aided drug design is to discover new drugs by carrying out algorithmic modelling of chemical interactions of bioactive molecules. Drug discovery is known to be a notoriously lengthy and costly procedure, therefore this sparks a great motivation for further research in the field to be carried out in order to simultaneously reduce the time elapsed during drug discovery and also produce effective products. Virtual screening is an umbrella term for a variety of ligand-based and structure-based tools which are used to search databases of chemical structures. Of particular interest to this study is ligand-based virtual screening and this uses known and active compounds to a specific target to screen molecules of unknown activity. We explore the way in which statistical approaches, specifically Bayesian statistics have been adopted for Ligand-based Virtual Screening. We implement two main similarity models, the Bayesian Inference Network and the Bayesian Belief Network, as well as explore model tuning avenues as an attempt to improve upon our results. The first crucial research question we seek to answer is if such statistical approaches provide better screening results when compared to conventional similarity scoring techniques, specifically the Tanimoto similarity metric. Indeed, through our research we show that the Bayesian similarity models developed through this dissertation do in fact improve screening results. Significant improvements in the ROC AUC is recorded when the Bayesian Inference Network and the Bayesian Belief Network are employed instead of the Tanimoto similarity metric, with maximum improvements of 15.52% and 15.19% respectively. Secondly, we aim to determine whether screening effectiveness is improved when multiple actives to a known target are used to rank a compound database as opposed to a single active. Through this research we suggest that for such Bayesian similarity models for Ligand-based Virtual Screening, a single active provides better results. Description: M.Sc.(Melit.)

2022-01-01T00:00:00Z https://www.um.edu.mt/library/oar/handle/123456789/141983 2025-12-05T09:53:57Z 2025-01-01T00:00:00Z

Title: Table selection using information retrieval techniques for table-agnostic Text-to-SQL Abstract: Text-to-SQL has been effectively addressed using various NLP approaches, enabling the translation of natural language queries into SQL queries. A common prerequisite for these implementations, however, is the availability of the database table during inference. This requirement can pose challenges in scenarios where the table is not readily accessible to users. This work is motivated by the ongoing development of a chatbot tool within a private company, aimed at streamlining database interactions for the users. To address the table accessibility limitation, this study leverages Information Retrieval techniques to implement table selection based solely on the natural language query. We finetune pre-trained models like BERT and GIST-NoInstruct using the ColBERT method. We train our models using data we curate in-house by employing established LLM-prompting techniques. We prepare individual training datasets using two negative sampling techniques: uniform distribution and weighted probability distribution. We also experiment with various data fusion techniques such as RRF, CombMNZ, and Linear Combination to combine results from multiple search strategies. Our approach outperforms baseline methods in table retrieval, while also providing a comparative analysis of various retrieval strategies. Description: M.Sc.(Melit.)

2025-01-01T00:00:00Z