基于延时反向传播算法的低复杂度滤波器研究

Low-Complexity Filter Based on Back-propagation Through Time Algorithm

为了进一步减少管状双线性递归神经网络的计算复杂度,在管状双线性递归神经网络中采用了延时反向传播算法。延时反向传播算法使用了阶次微分,误差函数对权值微分进行后向计算。后向计算顺序降低了初始化要求,减弱了网络对初始化条件敏感性并降低了计算的复杂度。该网络采用了模块化设计,各个模块以并行的方式执行任务,改善了计算效率。基于管状双线性递归神经网络的结构与神经元的数学模型,提出了具体的延时反向传播算法实现方案。同时进行了仿真来评估滤波器在非线性系统辨识方面的性能。实验结果表明基于延时反向传播算法的管状双线性递归神经网络提供了相当好的性能。

In order to further reduce the computational complexity of the pipelined bilinear recurrent neural network （ P - BLRNN）. The back - propagation through time （BPTT） algorithm is used into P- BLRNN. BPTT algorithm employs the order differential, and the error function calculates the weight differential to the back. The sequences of calculating to the back reduce the requirement of initialization. Weaken the network to initialization conditions sensitivity and reduces the computational complexity. Based on the structure and the mathematical model of neurons of P - BLRNN, the suitable implementation of the BPTT algorithm is presented. Simulations are carried out to evaluate the performance of nonlinear system identification based on the BPTr algorithm. Experimental re- suits show that the presented neural network filter provides better performance.