Application of Asynchronous Parallel Pattern Search Algorithm with ANN to Improve the Accuracy Rate of Identifying Learning Disabilities

• Main Authors: Chiou Wei-Wen, 邱委彣
• Other Authors: Wu Tung-Kuang
• Format: Others
• Language: zh-TW
• Published: 2010
• Online Access: http://ndltd.ncl.edu.tw/handle/88070213841075875694

碩士 === 國立彰化師範大學 === 資訊管理學系所 === 98 === In the past, the identification of students with learning disabilities (LDs) are mainly carried out by artificial means. But this method causes a big burden in the heavy teaching and service special education to teachers. In recent years, some studies use an evolutionary algorithm with Artificial Neural Network (ANN) to assist the identification of students with LDs. However, this method requires a lot of computing time. According to past researches, the use of Asynchronous Parallel Pattern Search (APPS) algorithm with classification algorithm could reduce the classification model training and validation time. Therefore, this study combined with grid-based computing resources in many computers, using APPS to search for the essential three major parameters of ANN: Learning rate, Momentum and Hidden nodes. In addition, this experiment is based on dual-core CPU computers. In order to take full advantage of performance, we added the concept of multi-threading to increase the search directions to avoid controlling by a single-threading APPS only. Finally, the experimental results showed that APPS with ANN could significantly reduce the computing time indeed, and APPS for Multi- Threaded MPI Modified version (MTMM) is better than Single-Threaded MPI version (STM). In single machine, the Average Correct Identification Rate (Avg_CIR) had been improved from 0.29% to 1.10%. In MPI environment, the Avg_CIR had been improved from 0.79% to 3.30%. In addition, compared the results of this study with previous studies’ results, with the same data sets, data attributes, and the same number of computers, the Average computing Time (Avg_Time) could be significantly reduced from 12.14% to 31.90%, and the Avg_CIR could be further enhanced.