针对传统对传网络(Counter Propagation Network,CPN)模型和学习算法存在的问题和不足,提出改进模型及竞争层的改进算法。在竞争层使用软竞争机制得到竞争层的输出,克服传统CPN使用胜者全得竞争机制的弊病,使竞争层中每一个神经元节点能充分发挥作用,参与网络的训练和权值的调整,提高竞争层中神经元的利用率,使网络能实现运用最少的神经元,达到要求的性能。从数值实验的对比看出,由于改进了网络模型和竞争算法,增强了CPN的模拟精度,CPN能更好地逼近模拟函数,提高了CPN的使用效率,网络性能得到了很大的提高。
This paper presents the improved model of CPN and algorithm in the competition layer, aiming at the problems and faults of the traditional CPN. The learning mechanism of soft - competition is used in the competition layer to get the output of the network and overcome the disadvantage of tradititonal CPN in which the winner - take - all mechanism is used. In the improved model, the neural node in the competition layer can do its best to join the train of the network and the adjustment of the power - value. The efficientcy of the neural nodes in the competition layer is improved sufficiently, so the network can work well with the least amount of neural node and come true the required capability of CPN. Because of the improvement of the model and the algorithm, the CPN can draw up the simulated function better and enhance the precision of the network, and boost up the efficiency of CPN. The experiment shows that the improvent of the model is effective and the efficiency of the network is sufficient.