Postgraduate

On-going Postgraduate Research Projects Supervising:

Ms. Umanda Dikwatta, Measuring the Degree of Violence in Social Media Image Posts Using Machine Learning Algorithms
PhD Degree in Computer Science (2019-present)
Supervisor(s): TGI Fernando
Ms. Sankani Heenkenda, Developing Machine Learning Models for Translating Sinhala and Pali Inscriptions to English Language in the Sri Lankan Context
PhD Degree in Computer Science (2020-present)
Supervisor(s): TGI Fernando
Ms. MC Weerawardana (2020-present). Automatic Detection of Computer Generated Videos using Machine Learning Algorithms
PhD Degree in Computer Science
Supervisor(s): TGI Fernando
Ms. SCM de S Sirisuriya (2021-present). Automatic Routing in 2D/3D Environments using Deep Learning Algorithms
PhD Degree in Computer Science
Supervisor(s): TGI Fernando and MKA Ariyaratne
Mr. RPS Kathriarachchi (2022-present). A Machine Learning Model to Identify Human Verbal Deceits Based on Verbal Cues and Facial Expressions
PhD Degree in Computer Science
Supervisor(s): TGI Fernando and MKA Ariyaratna
AJP Samarawickrama (2023-present). Stockbroker Automation with Machine Learning and Natural Language Processing Techniques: An Application to the Sri Lankan Stock Market
PhD Degree in Computer Science
Supervisor(s): TGI Fernando

Postgraduate Research Projects Supervised

Dr. WMKS Ilmini, Personality Traits Assessment Using Facial Features with Deep Learning Algorithms
PhD Degree in Computer Science (2024)
Supervisor(s): TGI Fernando

Abstract:
This study aims to find an accurate and efficient deep learning architecture to predict the apparent personality (AP). Furthermore, it aims to understand the features that affect the AP and study the possibility of generating artificial images with different personality traits. The experiments in this study were carried out using the CVPR’17 dataset. To find an efficient technique to predict the AP using CNN and CNN-RNN techniques, pre-trained deep learning models were employed to design the predictive models, and the performance of the models was evaluated using MSE and time. The results of the CNN APD models show that the VGG16 model reported the lowest MSE for all personality traits in the test dataset. Furthermore, the MobileNet model took a comparably lower time than other models in training, validation, and testing. Then CNN-RNN models were designed to predict AP with scene data. First, the models were developed with selected pre-trained models followed by an LSTM layer. The models were then evaluated based on their accuracy and computational cost. Results show that the highest mean accuracy is reported by the ResNet152 based model (0.9077), followed by VGG19 (0.9036); VGGFace recorded the lowest (0.8962). Statistical test results proved no significant evidence to conclude that VGG19 and ResNet152-based models performed differently. The findings reveal that satisfactory accuracy is obtained with a limited number of frames extracted from videos since models achieved more than 90% accuracy for all traits. Next, we applied CAM, Grad-CAM, Guided Grad-CAM, and LIME XAI techniques in APD models to find the prominent facial features that affect personality predictions. CNNmodels were designed with VGGFace, and VGG19 pre-trained deep architectures and trained with background removed frames captured from the videos. Then, Grad-CAM and Guided Grad-CAM were applied to the test dataset and calculated the heatmap. Subsequently, the ‘SelectKBest’ feature selection method was used to analyse the XAI techniques’ results. Results convey that facial features affect personality prediction. Since two models activate different image regions when predicting the output, it is challenging to differentiate facial features based on personality traits. Hence, decided to use the model agnostic XAI technique, LIME. LIME disclosed that facial features negatively and positively affected the output but lacks evidence to definitively link facial features with personality. CAM is used to explain the outputs of the CNN-RNN APD models. Models were trained using raw frames extracted from videos. Next, the bitwise intersection between the heatmap and images without background was measured to identify how many features from the human body affected the network output. ResNet152 identified more human data than the VGG19 model by giving 46-51% scores. The employed models behave differently in identifying the prominent features which affect the models’ output based on the input. Finally, the GAN technique was used to find the possibility of conversion between personality scores. We employed the CycleGAN technique, a widely recognized method for unpaired image-to-image translation, to convert images between different domains. A CycleGAN model is trained with background-removed frames from the videos. The results are more like to match the brightness of the pixels rather than the facial features. We have reached the conclusion that there is improvement in this area, particularly with the incorporation of additional bench-marked datasets.
Ms. NTK Naulla, Predicting the next word of a Sinhala word series using Machine Learning
MSc Degree in Computer Science (2020)
Supervisor(s): TGI Fernando

Abstract:
Advancement of technology has led humans to make predictions before its occurrence. Text, stock market, weather predictors are some of them. Next word predictors have been embedded in many applications for the ease of the user. Gmail is a day-to-day application that uses this technology. Apart from that, the keyboard apps of current smartphones use certain technologies to suggest the next word to the user, based on the user experience and built-in data. Since these applications are used by people all over the world, the need to use these technologies in different languages other than English language has been crucial. Sinhala language is used widely in Sri Lanka, it is being one of the two official languages used in the country. The aim of this research is to build up a next word predictor in Sinhala using a Recurrent Neural Network in Machine Learning. Due to the inability to build a single RNN to cover a whole language with multiple domains for a masters program research, this research is restricted to a dataset consisting of news reports of sports websites/news sites (similar to articles to that of a newspaper) published in the Sinhala language, that are related to Sri Lankan Cricket comprising of cricket related articles on the three categories namely, domestic, school and international. The dataset was preprocessed in advance to remove most of the unwanted and repeated proper nouns based on a criteria and randomly broken down to training, validation and testing data. A Keras model is built with an aim of obtaining an accuracy above 0.7 and a minimum loss while it is run several times to get the average accuracy and loss. The learning curves of different optimizers were tested to find the best optimizer along with different learning rates and momentum values to get the best model. Dropouts were added after each LSTM layer to avoid overfitting of the dataset. The transliteration of the Sinhala text dataset was run using the same Keras model built to see if there exists a difference when using the Sinhala text to that of the transliterated form (use of familiar Latin letters). A Flask application is built that displays the next most possible word (top five words with the highest probability), when a user inputs a single word or two words in Sinhala text.
Ms. WPJ Pemarathne, Electrical Cable Optimization System for Single-storey Building Using Ant Colony Optimization
MPhil Degree in Computer Science (2020)
Supervisor(s): TGI Fernando

Abstract:

This dissertation presents the applicability of Ant Colony System Algorithm to optimize the electrical cabling of a single-storey building. Designing the cable and wiring layouts is one of complicated and tedious processes in building construction. The algorithm is initially modified and designed to optimize the path between the starting and target points in a single wall. Modification was done to the heuristic function of the state transition rule of the ACS. When designing points in the grid of the walls, the BS 7671 standard (IET Wiring Regulations) was followed. Then the parameter tuning has been done to achieve the best possible solution. As the next step, obstacles have been introduced to the environment and the algorithm has been modified to avoid the obstacles. Multi Objective Ant Colony System Algorithm for Electrical Wire Routing (MOACS-EWR) have been designed to optimize three objectives length, number of bends and the angles of the bends thorough stages. Modifications were done to the global pheromone updating rule of ACS and two heuristic functions were designed. Applied the modified algorithm using two approaches. First approach, path is optimized considering all possible paths from the starting to target points. Second approach, path is optimized considering only the horizontal and vertical paths. The algorithm is designed through five experiments and the performance has been compared with the Multi-Objective Ant Colony System Algorithm to cover multiple points and has been designed to optimize the path between the starting, intermediate and target points. This problem is designed using two approaches; in the first approach MOACS-EWR algorithm is applied to the paths which follow the required points to reach the target and then selecting the best path out of those. In the next approach, we have applied a modified local pheromone updating rule with an extra amount of pheromones to the edges which contain the intermediate point. Then the shortest tour with the intermediate points is selected. Both approaches have been compared for the performance using several experiments. Results were compared statistically and the approach two outperforms approach one. Then the MOACS-EWR algorithm has been applied to optimize the electrical wiring paths in a single room. The method has been applied to four rooms and then the performance has been tested. The new algorithm offered remarkable results and was able to find the best directions to reach the target through an optimized path. Finally, the algorithm has been applied to optimize cable and wiring layouts in a single-storey building. Throughout the modifications and experiments carried out, we can prove that the modified algorithms are efficient and applicable to optimize electrical cable and wiring layouts.
Dr. MKA Ariyaratne, Firefly Algorithm Based Self-Tuning Algorithm to Solve Systems of Nonlinear Equations
PhD Degree in Computer Science (2018)
Supervisor(s): TGI Fernando and S Weerakoon

Abstract:

This thesis focuses on applying nature inspired algorithms for finding roots of systems of nonlinear equations. The developments have been done to solve single variable nonlinear equations, systems of nonlinear equations and in applying a self-tuning framework on tuning the parameters of the algorithms that are used for problem solving. Fields including Engineering, Mathematics, Chemistry, Computer Science and Economics often encounter applications of univariate as well as systems of nonlinear equations. Providing solutions for such is challenging and the common method of solving them is the use of numerical methods. Numerical methods often have requirements to be fulfilled to begin with the process of finding approximations. The use of different optimization techniques in such situations have been widely applied in all fields of Engineering as the capabilities of computers continue to increase. The remarkable performance of nature inspired algorithms over other optimization techniques encourages researchers to apply them to various optimization problems. Recently developed algorithms like firefly algorithm, bat algorithm and artificial bee colony have shown their success over many difficult optimization tasks where other optimization techniques fail. From the initial study, strengths and weaknesses of such algorithms were identified. Particularly, the firefly algorithm was identified as a suitable algorithm for the problem. This was later improved to solve univariate nonlinear equations having complex roots. The main consideration was paid on two important tasks; finding almost all real and complex roots within a reasonable range and omit the necessity of the continuity and differentiability of the functions which is essential for many numerical methods. While applying the firefly algorithm as the suitable meta-heuristic algorithm, modifications to the original algorithm have been proposed also to identify solutions simultaneously (through archiving) and to identify the poorly performed populations (through a counter variable). Once completing a moving round by all fireflies, better ones (whose fitness is measured against a predefined threshold) are noted and are put into an archive. Poor populations (which have not contributed to the archive) are identified at a predefined point and new fireflies are introduced to the population in a random manner. This random replacements enhance the exploration property. The proposed new firefly algorithm is named as Modified Firefly Algorithm (MODFA) to solve nonlinear equations. Finding roots of a univariate nonlinear equation can be considered as a single objective problem, while finding roots of a nonlinear system can handle in either; as single objective or multi-objective. The current approach for handling multi-objective optimization problems is to employ the concept of Pareto optimality. But with the MODFA, finding roots of a system of nonlinear equations is handled as a single objective optimization problem. The new concept of archiving is introduced and with that, within a single run of the algorithm, many solutions can be obtained simultaneously. Another concept used here is a self-tuning framework which is used to tune the parameters of the used nature inspired algorithms. The purpose of using it for the study is to let the users use the algorithm without having knowledge about algorithm specific parameters. This is named as Self Tuning Modified Firefly Algorithm (STMODFA). The performance of the newly proposed algorithm, is evaluated by comparing it with other algorithms such as genetic algorithms, particle swarm optimization Algorithm, differential evolution , harmony seach, cuckoo search algorithm. According to the results obtained with these, it has been shown that the MODFA demonstrates better performance than the other nature inspired algorithms. It is hardly seen the ability of finding roots simultaneously by other algorithms. Since representations can be easily handled, all implementations were implemented using MATLAB. For almost all univariate and systems of nonlinear equations, the accuracy of an approximation is set as 10^-2. Because the study focuses more on finding as many roots as possible within a single run. To increase the accuracy, later in this research, the concept of hybridization has been introduced. Hybrids of MODFA have been built with numerical as well as natural optimization techniques. This concept gave successful results enabling MODFA to find roots simultaneously with a high accuracy around 10^-12.

Keywords: Nature inspired algorithms, Firefly algorithm, Complex roots, Systems of nonlinear equations, Self-tuning framework, Archive, Root approximations.
Mr. Kasun Vimukthi Dissanayake Elibichchiya, Classification of Heart Beats to Identify and Alert Arrhythmia
MSc Degree in Computer Science (2018)
Supervisor(s): TGI Fernando

Abstract:

The heart generates electrical activities as a result of its contractions. This electrical activity can be picked up by external electrodes and be monitored. This representation of the electrical activity of the heart also known as Electrocardiogram data is used as a diagnostic tool in medicine excessively. With the recent technological advancements, Machine Learning is heavily applied in the medical domain as well as countless other disciplines. In the medical domain, Artificial Intelligence can be used not only to diagnose but also to predict the probability of a person having a specific disease in the feature. Even though there is much research conducted on the classification of the heart condition based on ECG data we only have few, which was conducted on predicting its condition in the near future. Therefore this work was carried out to first classify and then to predict the heart condition by using the ECG data, by using both Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). In order to carry out the research experiments, Keras an open source library which wraps around the TensorFlow backend was used to build and train the models. Also, the research was carried out using the MIT-BIH dataset as the base dataset for the experiments. Matrics like F1 score, confusion matrices, etc. were used along with the overall accuracy, When measuring the performance as the dataset was heavily biased. With that, the classifier portion was able to achieve 96% of accuracy in classifying the given ECG data to their respective classes. In the predictions, however, the results proved that any of the selected models are not up to the level for being used in any practical scenario. But it helped to understand that by using a Recurrent-Convolutional Neural Network (RCNN) hybrid architecture we can boost our accuracy with using minimum computational resources on predicting a time series.
Mr. SM Maddumahewa, A Digital Image Processing and Support Vector Machines Based Approach to Understand the Emission Levels of Diesel Vehicles
MSc Degree in Computer Science (2018)
Supervisor(s): TGI Fernando

Abstract:

Vehicular air pollution is a significant contributor to environmental pollution particularly with rising urbanization. Sri Lanka is also in the fray due to the considerable number of automobiles entering and exiting urban areas. Current registered vehicle fleet of Sri Lanka is over 5 million. With the rapid increment of the number of vehicles on the road, Diesel vehicles also enjoy an equal share of the same. Role of Diesel engine is to convert chemical force in diesel into mechanical force. Diesel fuel is a combination of hydrocarbons which would only produce carbon dioxide (CO₂) and water in the form of vapour (H₂O) during ideal combustion. Hence volumetric concentrations usually of 2-12% CO₂, 2-12% H₂O, 3-17% O₂ and 93- 59% N₂ ranges can be expected in the diesel emission. Nevertheless under usual conditions because of the incomplete burning of diesel, reactions between mixture components under high temperature and pressure, combustion of engine lubricating oil and oil additives and burning of non-hydrocarbon parts of diesel (sulfur compounds and fuel additives) various other by products are emitted upon burning of diesel fuel. Some of the common environmental pollutants of diesel emission include unburned hydrocarbons (HC), carbon monoxide (CO), nitrogen oxides (NO_x) or particulate matter (PM). These diesel emissions with pollutants could have undesirable impacts on health and environment. Upon exposure acute short-term symptoms like coughing , difficulty in breathing, dizziness, nausea, headache, tightness of chest, and irritation of the eyes/ noseand throat or long term exposure can lead to chronic and more serious health issues such as cardiovascular diseases and cancers. Furthermore Greenhouse effect and global worming are some common problems related to the environmental pollution. Emission test of diesel vehicles is performed under the international standard of SAE – J1667 (Society of Automobile Engineers). Testing is done by passing a special light ray in to the tail pipe of the vehicle, where the amount of light rays is disturbed by the exhausting emission is gauged. This is called the Absorption Coefficient or the K-factor, which should be below a limit value (currently eight in Sri Lanka) for a vehicle to pass the test. The Vehicle Emission Test (VET) program was introduced in Sri Lanka following the Supreme Court decision to increase air quality however emitted smoke measured by opacity meters and they are substantially expensive in nature and also requires skilled labour to operate, thus needs to be handled according to a defined procedure. This research discuss an approach to understand emission of diesel vehicle without using specific devise such as Opacity Meter, not requiring skilled labor to operate and minimizing handling hazels at a lesser cost using Digital Image Processing and Machine Learning concepts using Support Vector machine (SVM). “Scikit-learn” research library is used to build the Support vector machine model. “MATLAB R2013a” is used for Image Processing.
Ms. IDID Ariyasingha, Random Weight-based Ant Colony Optimization Algorithm for Multi-objective Optimization Problems
MPhil Degree in Computer Science (2011-2017).
Supervisor(s): TGI Fernando

Abstract: Most real-world optimisation problems are concerned with multiple objectives and they are very difficult to optimise simultaneously. Researchers have proposed several ant colony optimisation algorithms for solving multiple objective problems over the last few decades. Therefore, this thesis concentrates on analysing, improving and developing ant colony optimisation algorithms for solving multiple objectives.Initially, the thesis reviews the recently proposed ant colony optimisation algorithms which have been introduced for optimising multiple objectives simultaneously. Then, it analyses the performances of these multi-objective ant colony optimisation (MO-ACO) algorithms when applied to some combinatorial optimisation problems. Therefore, at the beginning of the research, performances of the MOACO algorithms are analysed by applying them to several benchmark instances of the travelling salesman problem (TSP). It considers various objectives, such as two, three and four objectives, and changes the number of ants and number of iterations for understanding their effects on the performances of MOACO algorithms. The results of the detailed analysis have shown that some MOACO algorithms achieve better performances. Also, their performance slightly depends on the number of ants, the number of objectives and the number of iterations used in the colony. Single objective optimisation problems are considered in most of the algorithms when solving the job shop scheduling problem in the literature. However, some ant colony optimisation algorithms for solving the job shop scheduling problem with multiple objectives have been proposed in recent years, because real-world applications are concerned with multiple objectives. Hence, this thesis analyses the performance of some recent multi-objective ant colony optimisation algorithms by applying them to sixteen benchmark problem instances of the job shop scheduling problem on up to 20 jobs * 5 machines. Also, it considers two, three, and four objectives and it optimises four criteria – makespan, mean flow time, mean tardiness, and mean machine idle time – simultaneously. Furthermore, different numbers of ants are used in a colony, to see their effects on the performance of the algorithms. The results obtained have shown that the performance of some multi-objective ant colony optimisation algorithms depends on the number of objectives and the number of ants. This thesis proposes a new ant colony optimisation (ACO) algorithm named the random weight-based ant colony optimisation algorithm (RWACO) to solve multiple objectives simultaneously. The RWACO algorithm is based on the ant colony system (ACS) algorithm and uses randomly generated weights for each objective associated with heuristic information which is called the random weight-based method. The performance of the newly proposed algorithm, RWACO, is evaluated by comparing it with more recent MOACO algorithms when applied to the three combinatorial optimisation problems: the travelling salesman problem (TSP), the job shop scheduling problem (JSSP) and the quadratic assignment problem (QAP). According to the results obtained with these studies, it has been shown that the RWACO algorithm achieves better performances than all the other MOACO algorithms considered in these studies. The random weight-based method, which has been introduced for the RWACO algorithm, is applied to the Pareto-strength ant colony optimisation algorithm (PSACO) to examine its effects on the performance of the PSACO algorithm. The performance is evaluated by applying it to the travelling salesman problem. The experimental results have shown that the PSACO algorithm performs better when the random weight-based method is applied.
Mr. AJP Samarawickrama, A Recurrent Neural Network Approach in Predicting Daily Stock Prices: An Application to the Sri Lankan Stock Market
MSc Degree in Computer Science (2017)
Supervisor(s): TGI Fernando

Abstract: Many studies have been carried out to predict stock prices using different Artificial Neural Network (ANN) models during past time in different countries. Recurrent Neural Networks (RNNs) is a sub-field of neural networks that use feedback connections. Several types of RNN models have been used by researchers in predicting financial time series. This study was conducted in order to develop models to predict daily stock prices of selected listed companies of Colombo Stock Exchange (CSE) based on Recurrent Neural Network (RNN) approach and to measure the accuracy of the models developed and identify the shortcomings of the models if present. Feed Forward, Simple Recurrent Neural Network, Gated Recurrent Unit and Long Short-Term Memory architectures were used to build models. Each network has 6 input neurons and 1 output neuron. Closing, High and Low prices of past two days were selected as input variables for each company. Stock prices from 2002/01/01 to 2013/06/30 selected for the study. Most recent data was selected for the test data set, next recent set was selected for validation and the other set was selected for training. Keras package was used as the software to build and train neural networks. When considering both iterative errors and forecasting errors feedforward networks produce the highest and lowest errors. The forecasting accuracy of the best feedforward networks is approximately 99%. SRNN and LSTM networks generally produce lower errors compared with feedforward networks but in some occasions, the error is higher than feedforward networks. Compared to other two networks, GRU networks are having comparatively higher forecasting errors.
Mr. DAA Deepal, Convolutional Neural Network Approach for the Detection of Lung Cancer in Chest X-Ray Images
MSc Degree in Computer Science (2016)
Supervisor(s): TGI Fernando

Abstract: The chest X-rays are considered to be the most widely used technique within the health industry for the detection of lung cancer. Nevertheless, it is very difficult to identify lung nodules using raw chest X-ray images and analysis of such medical images has become a very complicated and tedious task. This study mainly concerned with Convolutional Neural Network Approach to identify whether the suspicious area is a nodule or non-nodule. The JSRT digital image of the chest X-ray database developed by the Japanese Society of Radiological Technology (JSRT) and was used to train and test these models. Further, the support vector machine and Multilayer perceptron were used for comparison with Convolutional Neural Network model. “Pylearn2” research library is used to build the Convolutional Neural Network model and Multilayer perceptron model. “Scikit-learn” research library is used to build the Support vector machine model. “MATLAB R2013a” is used to extract nodule and non-nodule locations from the original images and other Image processing parts.
Ms. HPS Nishani, Improved Newton’s Method to Solve Systems of Nonlinear Equations
MSc Degree in Industrial Mathematics (2011)
Supervisor(s): S Weerakoon, TGI Fernando and M Liyanage

Abstract: Improved Newton’s Method (INM) is a widely accepted third-order iterative method introduced in the late 90’s to solve nonlinear equations. It has become so popular among numerical analysts that it records more than 500 citations in recognized international journals. However, even after more than a decade of the initial introduction of INM, nobody took the challenge of extending the INM for systems of nonlinear equations. The objective of this research was to prove the third order convergence of the Improved Newton’s Method in solving systems of nonlinear equations. We were able to extend the Improved Newton’s Method to functions of several variables and provide a rigorous proof for the third-order convergence. This theory was supported by computational results using several systems of nonlinear equations. Computational algorithms were implemented using MATLAB.
Dr. RP Abeysooriya, Genetic Optimization of Cut Order Planning in Apparel Manufacturing
MSc Degree in Industrial Mathematics (2011)
Supervisor(s): TGI Fernando

Abstract: Cutting is one of the main value-adding processes of the apparel manufacturing process. It serves as the major input provider for the sewing process by feeding cut panels required to sew garments. Fabric cutting process acts as a second major cost contributor to the manufacturing process due to the high expenditure on the marker making fabric spreading and cutting, which is about 5-10% of the total manufacturing cost. Owing to this reason, apparel manufacturers heavily focused on reducing the cost incurred in the cutting department. The work plan of the cutting department is termed as Cut-order plan, which plans the entire cutting process; the number of markers needed, sizes to be included in each marker, the quantity of garments from each size need to be included in the marker and the number of fabric plies that will be cut from each marker. Researchers highlight that an effective cut-order plan results in reducing the above-mentioned cost factors of the cutting process, thereby reducing the entire manufacturing cost to a greater extent. This study aims at optimizing the cut-order planning process of garment manufacturing process. Genetic Algorithm (GA) principles are adopted in achieving this goal. A comprehensive literature study was carried out to understanding the cutting process and the theoretical background of cut-order planning. Furthermore, GA principles and their application possibility of cut order planning are studied in detail. Next, the optimization algorithm was developed based on GA principles and the computer-based program was developed to execute the algorithm, using MATLAB. With the aim of validating the developed algorithm, cut-order plan data was collected from several apparel manufacturing companies, and then the algorithm was validated using them. Same data was input to the software tools available for generating cut order plans, with the aim of comparing their results with the results obtained by the developed program. Finally, the output was finalized by doing the necessary changes based on the comparison.
Ms. CP Amarajeewa, Optical Character Recognition (OCR) of Sinhala Characters by Feature Analysis Followed by Matrix Matching
MSc Degree in Industrial Mathematics (2002)
Supervisor(s): GK Watugala, TGI Fernando and HKGdeZ Amarasekara
Mr. PAS Pathiraja, Modelling Sinhala Characters using Cubic Rational Bezier Curves
MSc Degree in Industrial Mathematics (2002)
Supervisor(s): GK Watugala, TGI Fernando and HKGdeZ Amarasekara