广义同余神经网络及BP神经网络的比较研究

Study on comparison for GCNNs and BPNN

从神经网络的结构、激励函数、权值调整算法等方面对三种广义同余神经网络(generalized congruenceneural network,GCNN)及传统BP神经网络(back propagation neural network,BPNN)的异同点进行了比较和研究。通过对正弦函数的逼近性能比较,表明最新改进的第三种GCNN既继承了前两种GCNN收敛速度快的优点,又具有传统BPNN稳定性好的优点;既克服了前两种GCNN不稳定性的缺点,又克服了传统BPNN收敛速度慢的缺点。采用分段线性激励函数有利于GCNN的推广应用。

This paper compared the difference and similarity for three types of GCNN and BPNN in neural networks＇ structure, activation function, weight adjustment algorithm, etc. The comparisons for approximation performance to sine function show that the 3rd improved GCNN （GCNN3） inherits fast convergent speed in the previous two GCNNs, and is as good as BPNN in stability. GCNN3 overcomes the instability shortage of the previous two GCNNs, and overcomes the slow convergence shortage of BPNN. Piecewise linear activation function in GCNN will in favor of its expanding applications.