基于局部收敛权阵进化的BP神经网络MapReduce训练

发布时间：2018-04-18 03:29

本文选题：BP神经网络 + 遗传算法　；参考：《计算机工程与科学》2016年12期

【摘要】：为提高大样本集情况下BP神经网络的训练效率,提出了一种基于局部收敛权阵进化的BP神经网络MapReduce训练方法,以各Map任务基于其输入数据分片训练产生的局部收敛权阵作为初始种群,在Reduce任务中通过种群进化,选取适应度最高的权阵作为Map任务下一轮训练的初始权阵,直至该权阵对所有输入数据分片收敛。实验结果表明,与现有方法相比,该方法可有效避免MapReduce训练BP神经网络时容易陷入局部收敛的问题,并大大减少训练时间。
[Abstract]:In order to improve the training efficiency of BP neural network with large sample set, a BP neural network MapReduce training method based on local convergence weight matrix evolution is proposed.The local convergence weight matrix generated by each Map task based on its input data slice training is taken as the initial population. Through population evolution in the Reduce task, the highest fitness weight matrix is selected as the initial weight matrix for the next round of training of the Map task.Until the weight matrix is piecewise convergent to all input data.The experimental results show that the proposed method can effectively avoid the problem of local convergence and greatly reduce the training time when MapReduce trains BP neural networks.
【作者单位】：西北师范大学计算机科学与工程学院;
【基金】：国家自然科学基金(61462076) 甘肃省自然科学基金(1104GKCA023) 甘肃省科技攻关项目(1208RJZA134) 西北师范大学青年教师科研提升计划(NWNU-LKQN-12-30)
【分类号】：TP18

【相似文献】