节点文献
MIMD系统上成批训练BP算法程序的并行划分
THE DIVISION FOR PARALLELIZATION ON MIMD OF THE BP LEARNING PROGRAM IN BATCH
【摘要】 本文对成批训练BP算法原有串行程序进行了数据相关性分析,提出了三种并行划分方法,在此基础上实现了以EP-860并行计算机系统为运行环境的并行计算程序.文章对三种划分方法从通信复杂性、处理机的空闲等待率等几个方面作了比较,最后给出了不同划分方法的使用场合.
【Abstract】 In this paper the authors analyse the data dependency of sequential program by BP algorithm of the cumulative weight adjustment. Three division methods derived by us are discussed. Based on this, we present the implementation of the parallel programs to the original for the multi-processor system EP-860. After comparing the three methods in several aspects such as communication complexity, idle waiting rate of processors, and so on, we describe their good points and shortcomings in detail. At the last, the paper gives the characters and the situations suited to each of three methods.
【Key words】 BP Algorithm; Neural Network; Division for Parallelization; Efficiency of Parallel Computing;
- 【文献出处】 模式识别与人工智能 ,Pattern Recognition and Artificial Intelligence , 编辑部邮箱 ,1998年01期
- 【分类号】TP18
- 【被引频次】1
- 【下载频次】43