DR-NTU (Digital Repository of NTU)
Not a member yet
116018 research outputs found
Sort by
Scalable, generalizable, and offline methods for imperfect-information extensive-form games
No full textImperfect-information extensive-form games (IIEFGs) offer a versatile framework for modeling interactions between multiple players in stochastic and imperfect-information scenarios. Due to their superior modeling capabilities, IIEFGs can be applied in a wide range of fields, including economics and computer science. Given their broad application, IIEFGs have become a vital area of research in game theory. There are numerous algorithms developed to solve IIEFGs, ranging from traditional programming-based to advanced deep learning-based methods.
Although these algorithms have been proven successful, they still suffer from scalability and generalizability issues when solving large-scale games. Furthermore, these algorithms require continuous interaction with the game environment, which impedes the applications in many real-world scenarios where the environment may be impractical. To mitigate these issues, this thesis is devoted to improving the scalability and generalizability and extending offline learning to the IIEFGs setting.
Our first emphasis is on improving the scalability of existing algorithms to solve IIEFGs. We propose two scalable algorithms: CFR-MIX and PSRO with pre-trained strategies which are based on the counterfactual regret minimization (CFR) and policy space response oracle (PSRO) frameworks, respectively. CFR-MIX is designed to address the exponential combinatorial action space problem in a special type of IIEFGs known as team-adversary games. In the CFR-MIX algorithm, we first propose a new strategy representation that represents a joint action strategy using the individual strategies of all agents, maintaining cooperation between agents through a consistency relationship. To compute the equilibrium with the new strategy representation under the CFR framework, we transform the strategy consistency relationship into a consistency relationship between cumulative regret values. We then introduce a novel decomposition method for cumulative regret values to ensure this consistency. CFR-MIX algorithm employs a mixing layer to implement the decomposition method by estimating cumulative regret values of joint actions as a non-linear combination of the cumulative regret values of individual actions. Experimental results show that CFR-MIX significantly outperforms existing algorithms. PSRO with pre-trained strategies method is designed to solve the long-term decision-making issue in large-scale pursuit-evasion games (PEGs), the special type of two-player zero-sum IIEFGs. This algorithm integrates the pre-training and fine-tuning paradigm into the PSRO framework. Specifically, we first pre-train the pursuer's policy base model against many different strategies of the evader. Then we proceed with the PSRO loop and fine-tune the pre-trained policy to attain the pursuer's best responses. Empirical evaluation shows that our approach significantly outperforms baselines in terms of speed and scalability.
The next focus is on enhancing the generalizability of existing algorithms for solving PEGs. We propose a generalizable framework, Grasper, designed to generate pursuer policies tailored to specific PEGs. Grasper features a novel architecture with two components: a graph neural network (GNN) to encode PEGs into hidden vectors and a hypernetwork to generate pursuer policies based on these hidden vectors. We develop a three-stage training pipeline involving a pre-pretraining stage for training the GNN through the GraphMAE algorithm, a pre-training stage for training the hypernetwork by utilizing heuristic-guided multi-task learning, and a fine-tuning stage for fine-tuning the pursuer base model generated by the hypernetwork to
compute the best response strategy. Experimental results demonstrate that Grasper outperforms baselines in both solution quality and generalizability.
The final emphasis is on applying offline learning to game solving. We introduce the offline equilibrium finding (Offline EF) paradigm, which computes equilibrium strategies from offline datasets. Then we construct diverse datasets encompassing a range of games using several methods. These datasets serve as a basis for evaluating the performance of offline algorithms. Next, we design a novel framework, BOMB, which integrates behavior cloning and model-based methods along with a novel parameter estimation method. The model-based method adapts any online EF algorithm to the offline setting. Then, our theoretical analysis provides performance guarantees of BOMB across various datasets. Experiments demonstrate BOMB’s superior efficiency over offline RL methods in computing equilibrium strategies.
To conclude, this doctoral thesis addresses key challenges in solving imperfect-information extensive-form games, focusing on scalability, generalizability, and offline learning. Through innovative algorithms like CFR-MIX, PSRO with pre-trained strategies, Grasper, and BOMB, it advances the state of the art, offering efficient and practical solutions for both theoretical and real-world applications.Doctor of Philosoph
The role of post-pyrolysis carbon dioxide capture in hydrogen recovery from waste-derived pyrolysis gas
No full textThe study elucidated the role of post-pyrolysis CO2 removal using a CaO sorbent on upgrading pyrolysis gas into H2-rich gas. The pyrolysis gas was obtained from various waste-derived feedstocks (municipal sewage sludge, refused derived fuel (RDF), pine sawdust (biomass), and marine litter) by pyrolysis (600 °C), and underwent the treatment with CaO (600 °C) followed by the thermolytic decomposition at 1300 °C. The results show that CaO utilization does not change significantly H2 yield and higher heating value of gas on a feedstock mass basis but increases H2 purity from 49.6–83.3 to 63.7–94.4 vol% and H2/CO ratio from 1.1–5.7 to 1.9–19.8 of the product gas across all feedstocks. Normalized rate constants K for CO2 capture by CaO, determined by the grain model, varied between 0.0001 and 0.0006 min−1, revealing the feedstock-specific effectiveness of CaO. The higher CaO consumption rates were observed in case of RDF and biomass compared to sludge and marine litter. This could be attributed to the coking and faster carbonation of CaO caused by the composition of pyrolytic products. The obtained results emphasize the potential of integrating CaO sorbent into pyrolysis-based processes for the production of decarbonized H2-rich gas with higher purity. Moreover, the use of normalized kinetic parameters provides a straightforward method for the selection of feedstocks suitable for decarbonization of pyrolysis gas using CaO sorbent. The predominant factor affecting the carbonation kinetics of the CaO sorbent was found to be the CO2 flow rates in their respective pyrolysis gases.National Research Foundation (NRF)Public Utilities Board (PUB)The authors extend their gratitude to the World Wide Fund for Nature (Singapore) Limited for supplying a feedstock essential for this study. We also thank the DII Collaborative Graduate Program and the THERS Interdisciplinary Frontier Next Generation Researcher, Nagoya University for their financial support. This work was supported by the National Research Foundation, Singapore, and PUB, Singapore’s National Water Agency, under the RIE2025 Urban Solutions and Sustainability (USS) (Water) Center of Excellence (CoE) Program awarded to the Nanyang Environment & Water Research Institute (NEWRI), Nanyang Technological University, Singapore. The views expressed herein are those of the authors and do not necessarily reflect the views of the National Research Foundation, Singapore, or PUB, Singapore’s National Water Agency
Humanitarian engagement with Myanmar in the wake of the 2021 coup
No full textThe article investigates humanitarian engagement with Myanmar in the wake of the 2021 coup. It opens by examining humanitarian action in Myanmar before 2010, during the 2010s, and after 2021. Drawing on key stakeholder interviews conducted in 2023, it then presents internal and external perspectives on humanitarianism in Myanmar in the 2020s. While the Russian invasion of Ukraine in 2022 and the outbreak of war in Gaza in 2023 generated significant attention from donor nations, humanitarian needs in Myanmar, encompassing 18.6 million citizens, including 3.4 million internally displaced persons, go largely unmet. The argument from international society is that regional states should work towards a ‘Myanmar-led’ solution. Challenges faced by local and international actors, however, raise questions about the possibility of a solution that bridges the divide between these two humanitarian communities and is also acceptable to all groups in Myanmar. The article concludes by proposing that initiatives at the regional and global levels coalesce, overcome actor territoriality and generate political will by putting affected communities at the centre of humanitarian engagement to overcome current limitations and barriers to action
Implementation of minimum output variance filtered reference least mean square algorithm with optimal time-varying penalty factor estimate to overcome output saturation
No full textThe minimum output variance filtered reference least mean square (MOV-FxLMS) algorithm can effectively prevent the instability of active noise control (ANC) systems caused by the output saturation of the secondary source. The penalty factor, a critical parameter in MOV-FxLMS algorithm, is usually determined by trial and error, and its inaccurate estimate degrades algorithm's performance. Previous studies prove that estimating the optimal penalty factor requires prior knowledge of the disturbance. In practice, the penalty factor varies with the acoustic environment and primary source. Hence, this paper proposes an optimal time-varying penalty factor estimate method, which can track the variation of the disturbance and primary noise and assist the MOV-FxLMS algorithm in achieving the optimal control with output constraint. Moreover, the proposed algorithm also efficiently reduces computations and storage capacity requirements compared to other algorithms. The numerical simulation not only demonstrates that the proposed algorithm can react to noise variations but also reduces the influence of uncorrelated signals at the error sensor. Furthermore, the real-time experiment on a noise duct demonstrates the effectiveness of the proposed algorithm for the output saturation problem, exhibiting practical significance
A recommender system for employee recruitment
No full textWith the rapid and constant influx of data in today’s increasing digital world, job seeking has also evolved where applicants can apply for numerous positions with just a few clicks of a button within a few minutes. Coupled with an increasingly mobile world, this has significantly increased competition for jobs, with applicants eyeing both local and international roles. Consequently, Human Resources face an immense challenge in efficiently filtering and evaluating the overwhelming number of resumes they receive. Furthermore, Human Resources may lack the necessary domain knowledge to accurately assess an applicant’s qualifications in highly specialised fields.
To address these challenges, this project explores the development of a recommender system for employee recruitment, focusing on job title prediction based on the resume. Natural Language Processing techniques and Machine Learning models were used on an online dataset to classify resumes into their relevant roles. Models such as Random Forest, Logistic Regression, Support Vector Classifier and k-Nearest Neighbours were implemented and evaluated. Hyperparameter tuning, feature selection and varying dataset size were also done to assess their impact on the model accuracy.
This project demonstrated that Machine Learning models can be an effective approach for job classification across a range of roles. However, incorporating additional factors could further enhance the comprehensiveness of the model’s assessment.Bachelor's degre
Retrieval-augmented generation (RAG) for precedent search and case law retrieval in legal domain
No full textThis final-year project focuses on designing and implementing a Retrieval-Augmented Generation (RAG) system to address inefficiencies of traditional search for relevant case law from extensive legal documents in the legal domain. This system integrates a Large Language Model (LLM) with a retrieval mechanism using Facebook AI Similarity Search (FAISS), that fetches relevant legal documents from a knowledge base built using a collection of U.S. legal opinions from the CourtListener platform. The additional retrieval component of RAG ensures that the generated responses are both factually grounded and contextually rich, which may be difficult to achieve by a standalone LLM.
The proposed RAG system caters to legal professionals, including lawyers, judges, legal researchers, and paralegals, who require efficient retrieval of case law. By implementing a RAG system for precedent search, this project aims to improve the reliability of AI-assisted case law retrieval and reduce workloads across the legal domain.Bachelor's degre
Weather data visualisation
No full textWeather applications have become an integral part of daily life, providing users with real- time weather updates and forecasts. This project explores the development and evaluation of a geospatial web application designed to visualise weather data and predict rainfall effectively. By integrating real-time and historical weather data from the National Environment Agency (NEA) of Singapore, the application allows users to monitor weather conditions dynamically.
Through an exploratory analysis, the project examines the effectiveness of different visualisation methods for static and dynamic weather data. Findings from user studies indicate that colour schemes are more effective for visualising static weather data, while multi-dimensional symbols better convey changes in dynamic data. Additionally, a machine learning analysis suggests a relationship between wind speed and humidity in short-term rainfall predictions, though further research is required to enhance predictive accuracy.
User feedback highlighted the need for improved granularity in visualisation and expanded data coverage. Limitations such as dataset constraints, accessibility concerns for colour-blind users, and the lack of A/B testing for visualisation methods were also identified. Future work includes refining machine learning models for extended rainfall forecasting and integrating broader meteorological datasets for more comprehensive weather analysis.Bachelor's degre
Carbene-catalyzed chirality-controlled site-selective acylation of saccharides
No full textAcylation stands as a fundamental process in both biological pathways and synthetic chemical reactions, with acylated saccharides and their derivatives holding diverse applications ranging from bioactive agents to synthetic building blocks. A longstanding objective in organic synthesis has been the site-selective acylation of saccharides without extensive pre-protection of alcohol units. In this study, we demonstrate that by simply altering the chirality of N-heterocyclic carbene (NHC) organic catalysts, the site-selectivity of saccharide acylation reactions can be effectively modulated. Our investigation reveals that this intriguing selectivity shift stems from a combination of factors, including chirality match/mismatch and inter- / intramolecular hydrogen bonding between the NHC catalyst and saccharide substrates. These findings provide valuable insights into catalyst design and reaction engineering, highlighting potential applications in glycoside analysis, such as fluorescent labelling, α/β identification, orthogonal reactions, and selective late-stage modifications.Ministry of Education (MOE)Nanyang Technological UniversityNational Research Foundation (NRF)Published versionThis work was supported by grants from Natural Science Foundation of China (No.22101266), Excellent Youth Program of Hennan Province (242300421119, Y.-G.L.) and International Postdoctoral Exchange Fellowship Program (Talent-Introduction Program, No. YJ20210304, Y.- G.L.) by the Office of China Postdoc Council, China Postdoctoral Science Foundation (No. 2022M722865, Y.-G.L.) and Zhengzhou University (2024ZDGGJS068, Y.-G.L.), Open Projects Fund of Shandong Key Laboratory of Carbohydrate Chemistry and Glycobiology, Shandong University(No. 2023CCG08, Y.-G.L.); We acknowledge funding support from the National Key Research and Development Program of China (2022YFD1700300 Y.R.C.), the National Natural Science Foundation of China (U23A20201, 22071036, Y.R.C.), the Frontiers Science Center for Asymmetric Synthesis and Medicinal Molecules, Department of Education, Guizhou Province [Qianjiaohe KY number (2020)004, Y.R.C.], the Natural Science Foundation of Guizhou University [Guida Tegang Hezi (2023)23, Y.R.C.], the Program of Introducing Talents of Discipline to Universities of China (111 Program, D20023, Y.R.C.) at Guizhou University, the Singapore National Research Foundation under its NRF Competitive Research Program (NRF-CRP22-2019-0002, Y.R.C.), the Singapore Ministry of Education under its MOE AcRF Tier 1 Award (RG70/ 21, RG84/22, Y.R.C.), MOE AcRF Tier 2 Award (MOE-T2EP10222-0006, Y.R.C.), and MOE AcRF Tier 3 Award (MOE2018-T3-1-003, Y.R.C.), a Chair Professorship Grant, and Nanyang Technological University. X.Z. acknowledges the funding support from the Chinese University of Hong Kong (CUHK) under the Vice-Chancellor Early Career Professorship Scheme Research Startup Fund (Project Code 4933634, X.Z.) and Research Startup Matching Support (Project Code 5501779, X.Z.)
Lady Lazarus
No full text'Lady Lazarus' documents an individual’s contest with the ungovernable nature of life, their futile struggles against such forces, and a rather unfortunate instinctual insistence upon existing. Through the medium of the moving image and stop motion animation, we witness as the protagonist's deliquescence rejects the notion of life, yet their lingering corporeality rejects death. Informed by personal struggles with chronic afflictions of the body and mind, the work is a quiet introspection of personal escapist propensities and the tumultuous human experience of death, birth and especially everything in between.Bachelor's degre
Attention-based contrastive self-supervised learning model for antibody sequence-structure co-design
No full textIn traditional antibody discovery methods, a large volume of biological samples of
antibodies and antigens is required for wet lab experiments, making the process both expensive and time-consuming. However, computational methods have the potential to significantly accelerate antibody discovery by predicting antibody-antigen binding affinity, thus optimizing antibody design. While machine learning techniques have been employed to predict these interactions, most existing models rely on sequence data, which lacks critical structural information necessary for more accurate predictions. Structural data, though more informative, remains limited in availability. This study proposes a contrastive self-supervised learning model that utilizes antibody-antigen structure data to predict binding affinity. Contrastive learning has shown promise in other fields for learning effective representations from unlabelled data by distinguishing between high-affinity and low-affinity complexes. Despite the scarcity of structural data, this approach aims to improve the accuracy of binding affinity predictions while reducing data requirements. The study encompasses data preprocessing, feature extraction, model development, and evaluation, with the goal of advancing the efficiency and scalability of antibody-antigen interaction prediction.Bachelor's degre