为克服“胜者全得”对传网络的缺陷,提出使用基于软竞争机制的对传网络.这样增加了网络的训练复杂度.为此,把竞争层中隐单元的输出作为未观察到的缺省随机变量,使用EM算法对基于软竞争的对传网络进行训练,降低训练复杂度,加快网络的收敛速度.在实现EM算法的M步时,根据基于软竞争机制对传网络竞争层的特点,对EM算法实现进行改进,没有使用常用的迭代重新加权最小二乘算法,而利用样本加权平均求取隐层单元的权值向量,使EM算法更加简单易行,收敛速度快.仿真实验结果表明,基于软竞争机制的对传网络具有很好的泛化性能,特别在模式分类上具有很好的实际应用价值.
The paper is concerned with the counterpropagation network (CPN) based on the softcompetition mechanism instead of that of the winner-take-all to overcome the later' s faults, and the complexity of the network is raised too. By regarding the output of the hidden node in the competition layer as the unknown default random variable, the CPN based on the soft-competition mechanism is trained by the EM algorithm in order to decrease the complexity and accelerate the velocity of convergence of the network. In the M step of the EM algorithm, the algorithm in common use is modified according to the characterstic of the competition layer of the improved CPN, replacing the iteratively reweighted leastsquares method with the sample reweighted average method. The results of simulating experiments show that the modified network has a better performance while the training time is greatly reduced. The practicability and convergence of the network is improved rapidly, being particularly useful in the field of classification of patterns.