Browsing by Author "Sumathipala, P"

Now showing 1 - 6 of 6

Embargo
Emission Aware Multi-Metric Reinforcement Learning for Adaptive Traffic Signal Control in Sri Lanka
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Shangeeth A; Perera S.A.N; Munaweera K.D.M.P.; Sumathipala, P; Nawarathne, M
Urban intersections in Sri Lanka experience growing congestion and greenhouse gas emissions, while signal control remains fixed-time and insensitive to demand variability. An adaptive traffic signal control framework based on reinforcement learning (RL) is proposed. The control task is formulated as a Markov decision process with state features comprising queue length, waiting time, braking events, and estimated emissions. A Deep Q-Network (DQN) is trained in the SUMO-TraCI microsimulation under static and dynamic demand scenarios. Multiple reward formulations are examined, including emission-only and multi-metric designs with normalized, weighted components. Experiments show that emission-only rewards yield unstable learning, increased queuing, and no measurable emission reduction, indicating the limitations of single-metric optimization. A balanced reward combining queue length and braking achieves the most favorable trade-off, reducing average delay by up to 22% and CO2 emissions by 15% relative to fixed-time baselines. The framework is further instantiated on OpenStreetMap-derived models of Sri Lankan intersections (Borella, Kollupitiya), demonstrating applicability to local geometries and operating conditions. The results indicate that emission-aware RL controllers can outperform fixed-time operation and offer a viable pathway for deployment to support congestion mitigation and emission-reduction objectives.
Embargo
Integrating industrial technologies, tools and practices to the IT curriculum: an innovative course with .NET and java platforms
(acm.org, 2005-10-20) Athauda, R; Kodagoda, N; Wickramaratne, J; Sumathipala, P; Rupasinghe, L; Edirisighe, A; Gamage, A; De Silva, D
Exposure to state-of-art industry technologies, tools and practices by students provide CS/IT graduates highly desirable skills and marketability. A key expectation of the industry from their new cadre is a speedy integration into the business environment resulting in productive work. This usually requires having a sound technological background, a maturity to assess the environment and adapt quickly, and highly-developed soft skills to be productive in a team environment. Incorporating such experience and skills into a CS/IT curriculum is challenging and is still in its infancy stages. We undertook such as an endeavor in integrating .NET into the IT curriculum. Microsoft's .NET platform is becoming increasingly popular in the industry. Incorporating .NET into the undergraduate IT curriculum provides a plethora of skills and increases the employability of our graduates. We integrated .NET without a major revision to the existing curriculum by introducing an optional course in the final year (senior-level) of the IT undergraduate program. In addition to the .NET platform, the course covered the Java platform, which is similar in architecture to .NET. The course emulated an industry-based environment with real-world based assignments, focused on deliverables, used state-of-art IDEs and documentation, and pair programming to create a highly productive environment. The “soft skills” were integrated into the course with a project that implemented a virtual marketplace. Students in groups played different entities in the virtual marketplace and communicated with each other via Web Services. The project provided a virtual business environment and exposure to teamwork, collaboration, competition, negotiating, and creativity skills. Our first offering of the course in semester 1, 2005, attracted 128 students. The course created a highly productive environment throughout the semester. Students completed 7 assignments and the project within the 14-week semester. The initial results are encouraging and provide many insights to CS/IT departments planning to incorporate such courses.
Embargo
LISA : Enhance the explainability of medical images unifying current XAI techniques
(IEEE, 2022-07-18) Abeyagunasekera, S. H. P; Perera, Y; Chamara, K; Kaushalya, U; Sumathipala, P
This work proposed a unified approach to increase the explainability of the predictions made by Convolution Neural Networks (CNNs) on medical images using currently available Explainable Artificial Intelligent (XAI) techniques. This method in-cooperates multiple techniques such as LISA aka Local Interpretable Model Agnostic Explanations (LIME), integrated gradients, Anchors and Shapley Additive Explanations (SHAP) which is Shapley values-based approach to provide explanations for the predictions provided by Blackbox models. This unified method increases the confidence in the black-box model’s decision to be employed in crucial applications under the supervision of human specialists. In this work, a Chest X-ray (CXR) classification model for identifying Covid-19 patients is trained using transfer learning to illustrate the applicability of XAI techniques and the unified method (LISA) to explain model predictions. To derive predictions, an image-net based Inception V2 model is utilized as the transfer learning model.
Embargo
Machine Learning-Based Prediction of Settled Water Turbidity for Optimizing Alum Dosage in Drinking Water Treatment
(Association for Computing Machinery, Inc, 2026-06-17) Cooray, S; Fernando, D; Jeyachandran, P; Jayakodi, D; Sumathipala, P; Fousdeen, S; Nawarathne, M
Efficient chemical dosing is essential in drinking water treatment plants to ensure effective turbidity removal while minimizing chemical consumption and operational costs. Conventional alum dosing methods rely on manual jar tests and operator experience, which are time-consuming and often inadequate for handling rapid variations in raw water quality. Key parameters such as turbidity, pH, and electrical conductivity significantly influence the coagulation process, making accurate dosage determination a complex and dynamic challenge. This study proposes a machine learning-based approach to support optimal alum dosage selection by predicting settled water turbidity using historical treatment data. The dataset includes raw water turbidity, pH, conductivity, and applied alum dosage, which are used to train regression models capable of capturing nonlinear relationships between input parameters and treatment performance. The study used 7,975 hourly operational records collected from Ambathale Water Treatment Plant during January-December 2024, and model evaluation was performed using a time-aware 80:20 chronological train-test split. Multiple algorithms, including Gradient Boosting, Random Forest, XGBoost, and LightGBM, were explored alongside feature engineering techniques to improve model effectiveness. Experimental results demonstrate that the Random Forest model achieves superior performance, with an R² score of 0.7909, RMSE of 0.5197, and MAE of 0.3166, indicating strong predictive reliability under varying water quality conditions. The trained model enables simulation of different dosing scenarios, allowing identification of the alum dosage that produces the lowest predicted turbidity under varying raw water conditions.The proposed approach provides a practical decision- support solution for improving treatment efficiency, reducing chemical overuse, and supporting more reliable operation of drinking water treatment processes.
Embargo
Optimizing LLMs for Context-Sensitive Text Summarization: Insights from PLOS and SciSummNet
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Nawarathne U.M.M.P.K; Subasinghe, S; Rathnayake, T; Sumathipala, P
Text summarization is an important task in natural language processing (NLP), as it allows vital information to be extracted from large amounts of textual input. This work investigates the efficacy of fine tuning large language models (LLMs) for summarization tasks by comparing their performance to that of untuned alternatives. Using the PLOS dataset consisted of biomedical papers and SciSummNet dataset consisted of computing linguistics related scientific papers, LLM models such as Falconsai-T5 text summarization, T5-small, Bart and Bart-large-CNN were tested using ROUGE, BLEU, and METEOR metrics to measure their summarization abilities. The results show that fine-tuning improves model performance significantly across the evaluation metrics. Fine tuned models performed well on the SciSummNet dataset, with Bart-tuned model getting the highest scores across multiple measures, demonstrating the efficacy of targeted adaptation where the obtained results for ROUGE-1-0.6808, ROUGE-2 0.6487, BLEU - 0.4163, and METEOR - 0.5083. The analysis underscores the importance of aligning model architectures and datasets to achieve optimal results. The structured biomedical content of the PLOS dataset highlighted the vital necessity of semantic preservation, as evidenced by specially the higher METEOR scores achieved by fine tuned models. The SciSummNet dataset, which focuses on computing and linguistic publications, demonstrated the ability of the models to navigate abstract concepts and complex technical vocabulary. This study demonstrates the transformative power of fine-tuning in customizing LLMs resulting in significant increases in summarization accuracy.
Embargo
Unveiling EEG Emotional Patterns during Interactive Engagement Activities: A Performance Comparison of Machine Learning Models
(Institute of Electrical and Electronics Engineers Inc., 2025-12-09) Rathnayake, T; Subasinghe, S; Nawarathne, M; Sumathipala, P
This paper presents a comparative evaluation of machine learning and deep learning models for emotion recognition from electroencephalography (EEG) signals recorded during interactive engagement activities. EEG data were collected using a Muse headband as participants performed engagement activities designed to elicit five emotions: Afraid, Happy, Calm, Neutral, and Sensitive. After preprocessing and feature extraction, eight machine learning and deep learning models were trained. Based on the experiment, Random Forest emerged as the best-performing classifier, achieving ~95.6% accuracy with balanced precision-recall, while gradient boosting and SVM also shown comparative results. The models demonstrated potential for real-time streaming despite slightly lower accuracy, highlighting scalability with larger datasets. These findings demonstrate the feasibility of consumer-grade EEG devices, enhanced with multimodal features, for robust emotion recognition in realistic engagement activities. The results underscore the effectiveness of multimodal fusion and engagement-based design, confirming that lightweight EEG systems can support practical, real-time affective computing applications in education, adaptive interfaces, and mental well-being monitoring.