首页> 外文学位 >Distributed and multithreaded neural network algorithms for stock price learning.
【24h】

Distributed and multithreaded neural network algorithms for stock price learning.

机译:用于股票价格学习的分布式和多线程神经网络算法。

获取原文
获取原文并翻译 | 示例

摘要

In this thesis, we focus on the problem of learning the stock price movement using neural networks in distributed and shared memory environments. We parallelize the Backpropagation Neural Network (BPNN) algorithm on an 8 node Beowulf cluster with Message Passing Interface (MPI) and on an 8 processor Pentium III Symmetric Multi-Processor (SMP) machine using OpenMP. We have developed algorithms for two types of BPNN: neuron parallelism and training set parallelism on the distributed architecture. In neuron parallelism the hidden nodes are partitioned and distributed among the various processors while in training set parallelism, the input data (stock prices) is partitioned and distributed among the processors. On SMP, we exploit two different approaches for parallelizing neural network with multithreading. The first approach, loop-level parallelism is a fine-grained algorithm where the iterations of a loop are divided dynamically among the threads by the compiler. In the second approach, coarse-grained parallelism, the user intervenes and creates a limited number of threads where each thread is given an equal amount of workload. (Abstract shortened by UMI.).
机译:在本文中,我们着重研究在分布式和共享存储环境中使用神经网络学习股价走势的问题。我们在具有消息传递接口(MPI)的8节点Beowulf群集上和在使用OpenMP的8处理器Pentium III对称多处理器(SMP)机器上并行化反向传播神经网络(BPNN)算法。我们已经为两种类型的BPNN开发了算法:神经元并行性和分布式体系结构上的训练集并行性。在神经元并行化中,隐藏节点在各个处理器之间分配和分布,而在训练集并行性中,输入数据(股票价格)在处理器之间分配和分配。在SMP上,我们利用两种不同的方法来使神经网络与多线程并行化。第一种方法是循环级并行性,它是一种细粒度的算法,其中循环的迭代由编译器在线程之间动态分配。在第二种方法(粗粒度并行度)中,用户干预并创建了有限数量的线程,其中每个线程被赋予相等的工作量。 (摘要由UMI缩短。)。

著录项

  • 作者

    Rahman, Mohammad Rashedur.;

  • 作者单位

    University of Manitoba (Canada).;

  • 授予单位 University of Manitoba (Canada).;
  • 学科 Computer science.
  • 学位 M.Sc.
  • 年度 2003
  • 页码 66 p.
  • 总页数 66
  • 原文格式 PDF
  • 正文语种 eng
  • 中图分类
  • 关键词

相似文献

  • 外文文献
  • 中文文献
  • 专利
获取原文

客服邮箱:kefu@zhangqiaokeyan.com

京公网安备:11010802029741号 ICP备案号:京ICP备15016152号-6 六维联合信息科技 (北京) 有限公司©版权所有
  • 客服微信

  • 服务号