基于综合目标函数的神经网络多新息辨识算法

Multi-innovation identification algorithm of neural network based on generalized objective function

为提高动态神经网络学习算法的辨识精度及抗噪性能,提出一种基于综合目标函数的多新息辨识算法。该算法基于多新息理论在最小均方误差目标函数中引入一辅助项构造综合目标函数,利用该目标函数进行网络输出层权值的训练,并采用牛顿法推导出输出层权值的递推计算公式。与已有二阶学习算法相比,新算法鲁棒性强,收敛速度快,辨识精度高。仿真结果验证了算法的有效性。

To improve the identification accuracy and robustness to noise of dynamic neural network learning algorithm, multi-in- novation identification algorithm based on a generalized objective function was presented. The generalized function based on multi-innovation theory was constructed by combining an anxiliary constraint term with the least mean square error. The weight matrix of output layer was trained using the generalized function. The recursive equations for training weight matrix of output layer were derived using Newton iterative algorithm. Compared with the existed second-order learning algorithm, this algorithm has stronger robustness, better convergent rate and accuracy. Simulation results show the efficiency of the new algorithm.