论文部分内容阅读
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.
In this paper, an improved BP network is proposed, in which the state of each neuron in the network only takes 1 or -1, which is taken into account its convenience implemented by digital circuits. On this basis, the method of initial condition selection In order to increase the speed of training, this initial condition selection is also most likely to make the network converge to the global optimum.Although the neurons in the network take the state of 1 or -1, but still use Sigmoid function tanh (·) in the training network The input can be either the pattern of elements 1 or -1 or the pattern of any continuous variable.Computer simulation results show that using the same training algorithm and using the proposed method to select the initial conditions than the random selection of the initial conditions to converge Much faster, the network has better performance than the traditional architecture.