A Study Of Differential Evolution Extreme Learning Machine Based On Hadoop

DE-ELM is characterized with good generation performance and high classification accuracy of machine learning algorithms, and has received extensive attention in the industry. However, with the data explosion, the traditional DE-ELM is hard to meet the needs of massive and high-dimensional data processing. The open source Hadoop(one kind of the cloud computing platforms) is a platform with low cost, high error-tolerance rate and strong adaptability to processing massive and high-dimensional data, which provides effective means to solve the above problems. Therefore, the study of how to paralyze traditional DE-ELM and deploy it to Hadoop are of great significance.Based on Hadoop platform, the Map Reduce of DE-ELM is studied and the premature convergence problem of this algorithm is improved. The main works are as follows:The calculation of massive and high-dimensional data processing is complex and slow, and the study realizes MapReduce of DE-ELM based on Hadoop platform, which enhances the processing ability of the massive and high-dimensional data and its computation complexity. The main idea of the algorithm is: since the most complex part of MR-DE-ELM is the operation of large-scale matrix multiplication operator and large-scale matrix transposition operator, according to matrix multiplication, the calculation of every element is independent to each other. Through parallel computing, the large-scale matrix multiplication is converted to 2 steps: Vector inner product and vector sum. Through proper setting of the elements’ key-Value, we can realize the transposition of large scale matrix. MR-DE-ELM enhances the processing efficiency of the massive and high-dimensional data and its computation complexity.The paper analyzes MR-DE-ELM and the premature convergence problem of DE leads to the issue that the classification accuracy remain to be further improved in current situation. This paper proposes the Map Reduce of DE-ELM based on Dual-populations and Dual-strategy, which improves the classification accuracy. The main ideas of the algorithm are as the follow: firstly, divides the population into two sub-populations in the course of evolution, and each sub-population sets different mutation strategies and crossover operator in the period of mutation and crossover. Then, evolves each sub-population in parallel and independently without mutually interference; sets evolutionary generations and determine whether there exists the exchange of information between sub-populations. If the information exchanges, compares the two sub-population and seek out the best individual and replace the worst one in the other sub-population. Finally, sets the Maximum Generation maxGas the stop condition of the algorithm. MR-DpsDE-ELM improves the phenomenon of premature convergence and increases the classification accuracy.