Customizing asynchronous parallel pattern search algorithm to improve ANN classifier for learning disabilities students identification

Abstract

Due to the implicit characteristics of learning disabilities (LD), the diagnosis of students with learning disabilities has been a difficult process that requires extensive man power and takes a long time. Through genetic-based parameters optimization, artificial neural network (ANN) classifier has proven to be a good predictor to the diagnosis of students with learning disabilities. In this study, we examine another optimization algorithm, the asynchronous parallel pattern search (APPS), to search for appropriate parameters in constructing ANN-based LD classifier. To fully take advantage of modern multi-cored CPU technologies and to further expand the potential search space, various modifications to both of the serial and parallel versions of the original APPS implementations have been developed. The outcomes show that APPS in its original implementation can be competitive to genetic algorithm in term of accuracy, while requiring much less execution time. Furthermore, with consecutive (two-step) applications of the modified APPS algorithm to fine-tune the ANN parameters, we have further improved the ANN-based LD identification accuracy as compared to our previous results using genetic algorithm.