A hybrid neural network learning algorithm based on global search (differential evolution algorithm)and local search (pattern search)is put forward,which uses look-up and approximation method to optimize sigmoid function.At the experimental stage,curve approxi-mation and yarn picture classification are adopted whereas comparison and validation are performed on algorithmic efficiency and performance against the basic differential evolution algorithm (ODE)and the regeneratable dynamic differential evolution algorithm (RDDE)to elaborate the effectiveness of the algorithm.In the end,speed test comparison is carried out between integer weight and float weight neural networks, which tells that the computing speed of an integer weight neural network is much faster than a float weight neural network.Therefore a neural network trained by the proposed algorithm is more suitable for a reduced and faster embedded system.%提出基于全局搜索（差分进化算法）和局部搜索（模式搜索）的混合型神经网络学习算法（DEPS），并采用查找逼近法对sigmoid函数进行优化。实验部分采用曲线逼近和纱线图片分类两个实验，并与基本差分进化算法（ODE）和可再生动态差分进化算法（RDDE）在算法效率和性能进行对比、验证，说明算法的有效性。最后对整型和浮点型神经网络进行速度测试比较，说明整型权值神经网络在计算速度上远远快于浮点型权值神经网络。经算法训练后的神经网络更适合于结构精简、速度快的嵌入式系统。
展开▼
