| || || Neural networks (Computer science) -- Fiji|
| || || Intelligent student progress monitoring in an e-learning environment using neural network trained with memetic algorithm|
Author: Singh, Shaveen
Institution: University of the South Pacific.
Subject: Neural networks (Computer science) -- Fiji, Genetic programming (Computer science) -- Fiji, Genetic algorithms
Call No.: Pac QA 76 .87 .S56 2014
Copyright:20-40% of this thesis may be copied without the authors written permission
Abstract: Massive Open Online Courses (MOOCs) and other forms of digital learning have been widely heralded as practices that will serve to radically revolutionize education. However, offering such massive and open courses brings about its challenge and the need for monitoring large number of students. So much learner produced data is accumulated in the e-learning platform in the form of logs, which opens up opportunities to develop learning systems which can adapt to students abilities. This thesis explores the potential of modelling student behaviour in an e-learning environment (Moodle) using Machine Learning (ML) techniques, which could act as a personalised feedback mechanism for a students’ academic progression, and a means of courseware evaluation and maintenance. Eleven different offerings from six courses and four disciplines were experimented from School of Computing, Information and Mathematical Sciences (SCIMS) at University of the South Pacific (USP). Feature selection and attribute ranking technique was employed to identify courseware modules and user actions that had greater influence in accurately predicting student performance. The most suitable training profile for the Artificial Neural Network (ANN) model was determined for predicting ‘at-risk’ students around the faculty mid-semester reporting deadline (Week 8). The ANN model was implemented with a hybrid memetic learning algorithm. The architecture and parameters of the ANN was optimized using a past course offering and the learning model generated was then validated with the subsequent offering of the same course. An evaluation of the proposed system presented high prediction accuracy for identifying “at-risk” students for a good mix of courses experimented. The system also proved more reliable than the current mid-semester “at-risk” reporting practise implemented at USP. v To ensure that the model was able to adapt to the evolving nature of e-learning environment and improve with the availability of new learner profiles, an appropriate adaptive online learning strategy was also suggested for re-training the ANN. The hybrid meta-heuristic demonstrated that the incremental learning approach was able to cope with changes in data, and delivered trained neural networks that guaranteed faster convergence and finer performance without significantly increasing the learning cost.