摘要
本文给出一种改进的BP网络,网络中各个神经元的状态只取1或-1,这是考虑到其用数字电路实现的方便性.在此基础上,同时提出了初始条件选取的方法以增加训练速度,这种初始条件的选取也最有可能使网络收敛到全局最佳.尽管网络中神经元的状态取1或-1,但在训练中还是利用Sigmoid函数tanh(·).网络的输入既可以是元素为1或-1的模式也可以是任何连续变量所构成的模式.计算机仿真结果表明,用同样的训练算法,利用本文提出的方法选取初始条件比随机选取初始条件收敛要快的多,与传统结构相比,该网络具有更好的性能.
This paper presents an improved backpropagation(BP) network, where the state of each element chosen is either 1 or-1 for(he convenience of digital implementation.In order to increase the training speed, the adoption of initial conditions is suggested, which also increases the possibility of the network's convergence to the global optimum.Though the element state chosen is hard-limiting, sigmoid function tanh(·) is still used in training.The input of the network could be either hard-limiting quantizers or any differentiable nonlinearities.Computer simulation results show that for the same training algorithm, by using our adopted initial conditions the operation converges much faster than by random initial conditions.The prcperties of this network is better than those of the original structure.
出处
《北方交通大学学报》
CSCD
北大核心
1992年第1期22-27,共6页
Journal of Northern Jiaotong University
关键词
BP网络
神经元网络
复合控制
networks
feedback complement
complex control/neural networks
BP networks
