基于随机参数调整的改进反向传播学习算法

Modified BP Algorithm Based On Random Adjustment of Parameters

摘要:

针对BP(反向传播)神经网络学习易陷入局部极小的缺陷,提出了一种改进BP神经网络学习算法——RMBP算法.RMBP算法在学习参数调整中增加了随机性,使其方便地跳出局部极小点,并沿梯度下降方向到达全局极小点.异或问题的学习试验结果表明,RMBP算法较BP学习算法和其他常见的改进BP算法具有学习速度快、学习精度高、资源占用少的优势.最后,结合民航飞机实时飞行仿真系统研究,对一组飞机空气动力参数样本进行了学习,以说明RMBP算法的有效性.

Abstract:

A modified BP algorithm of neural network,random adjustment of parameters (RMBP) algorithm,is proposed to overcome the defect of easy going into local minimum of BP neural network.By adding randomization to the progress of parameters adjustment,RMBP helps BP neural network jump out of local minimum point easily,and get to global minimum point along gradient descent direction.Result of study experiment of XOR problem shows that RMBP has advantages of high study speed,high study precision and low resource occupation in comparison with other modified or unmodified BP algorithm.Finally,in combination with study of real-time flight simulation system of civil aviation airplane,a study is made of a group of aircraft aerodynamic force character parameters to verify the effectiveness of RMBP algorithm.