准对角递归神经网络及其算法的研究

提出一种准对角递归神经网络(QDRNN)结构及学习算法。此QDRNN结构上与对角递归神经网络(DRNN)相似，保留了DRNN结构简单的优点，以减小计算量，同时增加了相邻递归神经元之间的关联，可以直接应用BP学习算法进行训练。进一步，引入递推预报误差(RPE)学习算法，并且证明了其稳定性。仿真结果表明，QDRNN比DRNN具有更好的非线性逼近能力，而运算时间却增加甚微，DRNN的学习算法稍加变化即可应用。

Study on a Quasi-Diagonal Recurrent Neural Network and Its Algorithm

A structure and training algorithm for quasi-diagonal recurrent neural network(QDRNN) is presented. The QDRNN is similar as diagonal recurrent neural network(DRNN) in the structure, namely, the simple structure of the DRNN is retained to decrease computational requirement, but the connective right between the adjacent recurrent neutrons are increased. The BP training algorithm may be used directly in the QDRNN. Furthermore, the recursive prediction error(RPE) learning algorithm for QDRNN is introduced, whose stability is demonstrated. The theoretical analysis and simulation results show that the QDRNN has better approximate ability than the DRNN, and the computing time is increased slightly. Especially, all training algorithm of DRNN may be utilized in QDRNN only by little transformation.