Electronic Theses and Dissertations (Masters)

Permanent URI for this collectionhttps://hdl.handle.net/10539/38006

Browse

Search Results

Now showing 1 - 10 of 42
  • Thumbnail Image
    Item
    Parameter-Efficient Fine-Tuning of Pre-trained Large Language Models for Financial Text Analysis
    (University of the Witwatersrand, Johannesburg, 2024-07) Langa, Kelly Kiba; Bau, Hairong; Okuboyejo, Olaperi
    The recent advancements in natural language processing (NLP) have been largely fueled by the emergence of large language models (LLMs), which excel in capturing the complex semantic and syntactic structures of natural language. These models have revolutionized NLP tasks by leveraging transfer learning, where pre-trained LLMs are fine-tuned on domain-specific datasets. Financial sentiment analysis poses unique challenges due to the intricate nature of financial language, often necessitating more sophisticated approaches beyond what traditional sentiment analysis methods offer. Fine-tuning LLMs holds potential for improving modeling performance within the financial domain, but the computational expense of the standard full fine-tuning poses a challenge. This study investigates the efficacy of Parameter-Efficient Fine-Tuning (PEFT) methods for fine-tuning LLMs to specific tasks, with a focus on sentiment analysis in the financial domain. Through extensive analysis of PEFT methods, including Low-Rank Adaptation (LoRA), prompt tuning, prefix tuning, and adapters, several critical insights have emerged. The results demonstrate that by employing PEFT methods, performance levels that match or surpass those of full fine-tuning can be achieved. Particularly, adapting the Open Pre-trained Transformers (OPT) model with LoRA achieved the highest modeling performance, with an accuracy of 89%, while utilizing 0.19% of the model’s total parameters. This highlights the high modularity of PEFT methods, necessitating minimal storage sizes for trainable parameters, ranging from 0.1MB to 7MB for the OPT model. Despite slower convergence rates than full fine-tuning, PEFT methods resulted in substantial reductions in Graphics Processing Unit (GPU) memory consumption, with savings of up to 80%. Small-scale fine-tuned LLMs outperformed large-scale general-purpose LLMs such as ChatGPT, emphasizing the importance of domain-specific fine-tuning. Model head fine-tuning fell short compared to PEFT methods, suggesting additional benefits from training more layers. Compared to state-of-the-art non-LLM-based deep learning models, Long Short-Term Memory (LSTM), LLMs demonstrated superiority achieving a 17% increase in accuracy, thereby validating their higher implementation costs.
  • Thumbnail Image
    Item
    The Application of Attribution Methods to Explain an End-To-End Model For Self-Driving Cars
    (University of the Witwatersrand, Johannesburg, 2024-09) Chalom, Jason Max; Klein, Richard
    There has been significant development in producing autonomous vehicles but a growing concern is in understanding how these systems work, and why certain decisions were made. This has a direct impact on the safety of people who may come into contact with these systems. This research reproduced the experimental setup for an end-to-end system by Bojarski et al. [2016b]. End-to-end self-driving AI structures are built on top of black-box machine learning techniques. The source code can be found here: https://github.com/TRex22/ masters-gradsum. An allure of end-to-end structures is that they need very little human input once trained on large datasets and therefore have a much lower barrier to entry, but they are also harder to understand and interpret as a result. Bojarski et al. [2016b] defined a reduced problem space setup for a self-driving vehicle. This task only has a forward-facing camera which generates RGB images as input, and only the vehicle’s steering angle is the output. This work expanded the setup to include six CNN model architectures over the single model used by Bojarski et al. [2016b] to compare the behaviours, outputs and performance of the varying architectures. There have been recent developments in applying attribution methods to deep neural networks in order to understand the relationship between the features present in the input data and the output. GradCAM is an example of an attribution technique which has been validated by Adebayo et al. [2018]. We devised an attribution analysis scheme called GradSUM which is applied to the models throughout their training and evaluation phases in order to explain what features of the input data are being extracted by the models. This scheme uses GradCAM and uses segmentation maps to correlate inputted semantic information using the resultant gradient maps. This produces a model profile for an epoch which can then be used to analyse that epoch. Six models were trained, and their performance compared using their MSE loss. An autonomy metric (MoA) common in literature was also used. This metric tracks where a human has to take over to stop a dangerous situation. The models produced good loss results. Two model architectures were constructed to be simple in order to compare against the more complex models. These performed well on the loss and MoA metrics for the training data subset but performed poorly on other data. They were used as a control to make sure that the proposed GradSUM scheme would adequately help compare emergent behaviours between architectures. Using GradSUM on the architectures, we found that three out of the six models were able to learn meaningful contextual information. The other three models did not learn anything meaningful. The two trained simple models’ overall activation percentages were also close to zero, indicating these simple model architectures did not learn enough useful information or became over-trained on features not useful to safely driving a vehicle.
  • Thumbnail Image
    Item
    Addressing Ambiguity in Human Robot Interaction using Compositional Reinforcement Learning and User Preference
    (University of the Witwatersrand, Johannesburg, 2024-09) Rajab, Jenalea Norma; Rosman, Benjamin; James, Steven
    The ability for social robots to integrate naturally with the lives of humans has many advantages in industry and assisted services. For effective Human Robot Interaction (HRI), social robots require communication abilities to understand an instruction from the user and perform tasks accordingly. Verbal communication is an intuitive natural interaction for non-expert users but it can also be a source of ambiguity, especially when there is also ambiguity in the environment (i.e. similar objects to be retrieved). Addressing ambiguity for task inference in HRI is an unsolved problem. Current approaches, that have been implemented in collaborative robots, include asking for clarifications from the user. Related research shows the promising results of using user preference in HRI, but no work has been found where user preference is employed specifically to address ambiguity in conjunction with clarifying questions. Additionally, these methods do not leverage knowledge learned from previous interactions with the environment and the life-long learning capabilities of Reinforcement Learning (RL) agents. Based on the related work and shortfalls, we propose a framework to address ambiguity in HRI (resulting from natural language instructions), that leverages the compositionality of learned object-attribute base-tasks in conjunction with user preference and clarifying questions for adaptive task inference. Evaluating our method in the BabyAI domain, we extensively test all components of our system and determine that our framework provides a viable solution for addressing the problem of ambiguity in HRI. We experimentally prove that our method improves user experience by decreasing the number of clarifying questions asked, while maintaining a high level of accuracy.
  • Thumbnail Image
    Item
    Information and Knowledge Discovery from Undergraduate STEM datasets: Educational Data Mining at a South African University
    (University of the Witwatersrand, Johannesburg, 2024-10) Netshitungulu, Ndikhwine; Mchunu, Mike
    South Africa is experiencing an acute shortage of highly trained and skilled professionals with much-needed qualifications in the disciplines of Science, Technology, Engineering and Mathematics (STEM). There is a need to accelerate the training and development of professionals in these critical, scarce-skill areas. Our universities and other institutions play an important role in supporting and facilitating the development of these much-needed professionals. The primary focus of this data mining based study was to examine a number of factors related to first time first year BSc. students enrolled in critical-skill STEM degree programmes at the University of the Witwatersrand (also known as Wits). These factors included ethnicity, gender, funding status, English language proficiency, prior computer programming experience, mathematical ability and student profiling. Wits and other Higher Education Institutions (HEIs) collect and accumulate large amounts of data about their students. Using different data mining techniques this data can be processed, analysed and leveraged to obtain meaningful and useful information about these students. This study was motivated by the need to obtain information from the longitudinal data (2015-2019) collected about first time first year students enrolled in STEM degree disciplines at Wits University. In order to conduct the study in a more structured manner, we followed the Cross-Industry Standard Process for Data Mining (CRISP-DM) methodology, which is increasingly being adopted as a project implementation standard by data mining practitioners working in different disciplines, including education. We applied different statistical and data mining techniques against our various datasets. These included Analysis of Variance (ANOVA), classification, feature selection, prediction, clustering and association rule mining techniques. Several important results emerged from our study. The result from the two-way ANOVA experiment showed the interaction between gender and ethnicity not to be significant. However, the ethnicity factor had a statistically significant effect on the APS. We also found the interaction between ethnicity and funding status to be significant. Individually, these factors were statistically significantly related to the APS. Regarding English language proficiency, the difference between English Home Language (ENA) students and English First Additional Language (ENB) students was significant. Amongst the NSC Grade 12 subjects, mathematics had the most significant relationship with the Admission Point Score (APS). In two first semester Computer Science I courses, students with prior computer programming experience significantly outperformed their peers, who lacked computer programming experience. We also found that collectively, the minimum admission requirement subjects for admission to the first year STEM disciplines were suitable predictors of the overall, first year outcome. However, the result we obtained using these subjects as predictors of degree attainment was not encouraging. This wide ranging study has shown that data mining techniques can be used effectively in educational settings to leverage the data universities collect on their students, about whom insights and information can be obtained, information that can be used for well-informed decision making. Beyond our specific educational context, there is also much to learn from this study and its findings by other, higher educational institutions with STEM students in a situation similar to ours.
  • Thumbnail Image
    Item
    Formation Strategy Optimization Using Multi-Agent Reinforcement Learning
    (University of the Witwatersrand, Johannesburg, 2024-08) Njupoun, Abdel Mfougouon; Ingram, Branden; Rosman, Benjamin; Tasse, Geraud Nangue
    A long-lasting goal of artificial intelligence is to design agents capable of cooperative problem-solving. The RoboCup soccer competition provides a challenging environment for investigating the design of such intelligent and autonomous agents using machine learning techniques, such as Multi-Agent Reinforcement Learning (MARL). Cooperation is inherently difficult due to the need for agents to align their strategies, adapt to each other’s actions, and make decisions that benefit the collective goal over individual success. In this context, developing effective defensive strategies is particularly challenging. It requires agents to not only understand and anticipate the actions of opponents but also to coordinate with teammates in a dynamic environment where split-second decisions can determine the outcome of a game. This complexity is compounded by the unpredictable nature of the opponent’s strategies and the continuous adaptation required to counter them effectively. This research aims to investigate the application of Multi-Agent Reinforcement Learning in learning an effective defensive strategy in the RoboCup soccer competition. We use reward shaping to positively influence the behaviour of our simulated soccer players such that they can effectively defend against attacking soccer strategies. This reward function is then utilized to train agents and learn the optimal policy in centralized settings, namely Central Proximal Policy Optimization (CPPO). The training process involves exposing our agents to different fixed policies such as the Keepaway approach, where a team keeps the ball away from opponents, the Half-field offense strategy where the objective of the offense team is to strategically outplay the defense team to score goals, and the random direction changes aiming to mimic the unpredictability of human soccer, where players often change direction suddenly to evade defenders or create attacking opportunities. To evaluate our proposed strategy, we conduct a comparative analysis against established baselines like the NeuroHassle approach, which emphasizes early disruption of the opponent’s attack, and the Stable Marriage approach, focusing on optimal defender-attacker pairings. These methods, one based on reinforcement learning and the other on preference-based pairing, serve as benchmarks to gauge the effectiveness of our strategy in improving defensive gameplay. Evaluation metrics, including goals conceded and average distance between opponent players and our goal are used to analyze and identify the strengths and weaknesses of each approach. We evaluated our approach against the NeuroHassle and Stable Marriage methods by observing agent performance in keepaway, Half-field, and random direction changes offense scenarios. Utilizing those key metrics, we identify that our model demonstrated better defense strategies, offering insights for enhancing multi-agent systems in competitive environments like RoboCup soccer.
  • Thumbnail Image
    Item
    Ball tracking by object detection using deep neural network aided Kalman filtering in a real-time simulated RoboCup Soccer environment
    (University of the Witwatersrand, Johannesburg, 2024-08) Nagy, Marcell Douglas; Klein, Richard; Ranchod, Pravesh
    Computer vision has the ability to provide an abundance of environmental information to a robotic system, which makes perception and interaction with a dynamic world possible. In this research report the problem of real-time ball tracking in the context of simulated RoboCup soccer is considered. A tracking-by-detection framework is utilized to take advantage of the high performance of modern neural detection techniques as well as the use of neural assisted Kalman filtering for tracking which strikes a balance between the ease of interpretation of the traditional model-based approach and the ability to learn complex dynamics using neural networks. The results show that a modern keypoint based detector can outperform established real-time RoboCup ball detection techniques such as traditional Viola-Jones based detection as well as popular anchor based detection approaches, in the context of simulated RoboCup soccer. Further, a neural approach to Kalman filtering is able to outperform the popular extended Kalman filter for simulated RoboCup soccer kick tracking by implicitly learning the system dynamics.
  • Thumbnail Image
    Item
    High-Speed Obstacle Avoidance in Unstructured Environments
    (University of the Witwatersrand, Johannesburg, 2024-10) Naidoo, Ashton; Ranchod, Pravesh; Rosman, Benjamin
    Micro aerial vehicles (MAVs) have many applications in various environments, some of which include site surveying, mapping, search and rescue, inspection, delivery, and photography. Even though MAVs have an incredible amount of application space, most industry applications do not make use of autonomous flight. Instead, a human pilot flies the MAV, especially where navigation in complex environments is required. While autonomous navigation systems do exist in industry, the complexity of the environment can inhibit the system’s ability to traverse the environment. High-speed autonomous navigation of MAVs, in unstructured environments has received more attention in recent years, which has resulted in new more robust algorithms which can navigate complex unstructured environments quickly and successfully. In this research report, a simulation environment is created in a photo-realistic simulation engine in which high-speed obstacle avoidance algorithms can be evaluated. This is done so that an assessment can be made regarding how well these algorithms generalise as the complexity of the environment increases. Obstacle avoidance research often overlooks the degree of environmental complexity when presenting performance metrics. As a result, the performance metrics for obstacle avoidance algorithms can be vague due to the lack of detailed metrics that adequately capture environmental complexity. To this end, an environmental complexity scoring framework for unstructured environments is proposed. Environments are divided into ten levels which increase in environmental complexity score. The experiment allows the MAV to move through the simulation environment at various target speeds. As the MAV moves through each level, evaluation metrics such as average speed and collisions are logged. Evaluation of these metrics indicates that if the environmental complexity is constant, speed increases cause collision avoidance performance to degrade. Conversely, as the environmental complexity increases, lower speeds are required for successful collision avoidance. It is further shown that object complexity has an impact on environmental complexity and subsequently successful obstacle avoidance. The greater the object complexity the more obstacle avoidance performance is reduced. The project repository can be found here. Results indicate that obstacle avoidance algorithms have maximum speed and environmental complexity threshold scores, below which successful obstacle avoidance can occur.
  • Thumbnail Image
    Item
    Optimisation of Kick Latency for Enhanced Performance of Robots in the RoboCup Three-Dimensional League through Proximal Policy Optimisation (PPO)
    (University of the Witwatersrand, Johannesburg, 2024-07) Nekhumbe, Humbulani Colbert; Ranchod, Pravesh
    This study aimed to enhance the kicking ability of Nao robots in the three-dimensional RoboCup simulation by addressing a crucial challenge observed in the University of Witwatersrand RoboCup team. The focal challenge revolved around a noticeable delay and slow movement manifested by the robot during ball kicks, leading to vulnerabilities in ball possession against opposing teams. To surmount this challenge, the implementation of Proximal Policy Optimisation (PPO), a methodology pioneered by OpenAI, was advocated. The precise objective was to optimise kick parameters, with a primary emphasis on curtailing kick latency. This optimisation aimed to ensure swift and accurate execution across various kicking scenarios, encompassing actions like propelling the ball into the opponent’s territory to bolster ball possession and thwart adversary manoeuvres. Harnessing the iterative advancements embedded in PPO, the successor to Trust Region Policy Optimisation (TRPO), the endeavour was to refine the kicking behaviour of Nao robots. This optimisation process significantly reduced the observed kick delay, and this made the robot more agile and effective at competing in the complex three-dimensional RoboCup simulation environment. The study’s outcomes highlighted substantial progress in reducing kick latency and improving the adaptability of robotic soccer players, opening up possibilities for further exploration in reinforcement learning for autonomous agents.
  • Thumbnail Image
    Item
    Magnetic field strength estimations for the main phases of solar cycles 13-24 using magnetohydrodynamic Rossby waves in the lower tachocline
    (University of the Witwatersrand, Johannesburg, 2024-09) Morris, Tania Mari; Duba, Thama
    The magnetic field strength (MFS) estimates used by the existing space weather prediction models (SWPMs) are inaccurate. Consequently, it has been indicated that there is a need to find solutions to rectify the wrong assumption that the magnetic field remains constant in strength and location throughout the solar cycle. This study explores a solution to this problem by increasing the granularity and accuracy of the previous MFS estimations by calculating them for the main phases of the solar cycle (solar minimum and maximum) per hemisphere for solar cycles 13-24. A dispersion relation of the fast magnetohydrodynamic (MHD) Rossby wave was derived analytically in spherical coordinates that included a toroidal magnetic field and latitudinal differential rotation, which adequately captures the dynamics of the lower tachocline. Secondly, a change to the methodology of calculating the MFS that utilises the established connection between the observed Rieger-type periodicity (RTP) in solar activity of 150-190 days and the fast MHD Rossby wave in the lower tachocline was proposed. Furthermore, a new magnetic field profile (MFP) of Bφ = B0 sin(6Θ) was introduced to improve the model results. This MFP has a maximum and minimum value at the same latitudes associated with sunspot appearances during these extreme solar cycle phases. Consequently, the MFS values were calculated at a latitude of 29◦ (solar min) and 16◦ (solar max) using the hemispheric RTP data. The average MFS (RTP) for the solar min and solar max in the dominant hemisphere was established to be 8 kG (212 days) and 76 kG (163 days), respectively. For the non-dominant hemisphere, the average MFS was established to be 5 kG (213 days) and 50 kG (183 days) for the solar min and max, respectively. The results of this study show a significant difference in the results based on latitude. The findings have also revealed that the periodicity of increased solar activity associated with a specific MFS is affected not only by the solar cycle strength and hemispheric asymmetry but also by the solar cycle phase (or latitude) considered. Additionally, we strongly argue that this study’s MFS results represent reality more closely than previously calculated results. Therefore, we propose that the MFS estimates reported in this study should be considered for the input to various existing space weather prediction models.
  • Thumbnail Image
    Item
    Applications of Recurrent Neural Networks in Modeling the COVID-19 Pandemic
    (University of the Witwatersrand, Johannesburg, 2024-03) Hayashi, Kentaro; Mellado, Bruce
    This study attempted to introduce moving averages and a feature selection method to the forecasting model, with the aim of improving the fluctuating values and unstable accuracy of the risk index developed by the University of Witwatersrand and iThemba LABS and used by the Gauteng Department of Health. It was confirmed that the introduction of moving averages improved the fluctuation of the values, with the seven-day moving average being the most effective. For feature selection, Correlation-based Feature Selection(CFS), the simplest of the filter methods with low computational complexity, was introduced as it is not possible to spend as much time as possible on daily operations due to providing information timely. The introduction of CFS was found to enable efficient feature selection.