Neural networks can be viewed as a universal approximator for nonlinear functions , but the multi - layer feed - forward neural network which be used usually is a static state network in nature , it is disagree with the real - time identification for dynamic system . moreover , recurrent neural networks can simulate the state memory mechanism of dynamic system , so it can be utilized as the model of dynamic time delay system 神经网络具有逼近任意连续非线性函数的能力,但常用的多层前馈式反传网络本质上是一种静态网络,不适合动态系统的实时辨识,而递归神经网络能够实现对动态系统状态记忆机制的模拟,因此更适合于作为动态时延系统的模型。
This method took the robot actual poses and corresponding joint errors as inputs and outputs of a feed - forward neural network respectively , so as to achieve the real - time joint errors in arbitrary pose through the neural network , and pose accuracy was improved only through correcting the joints angles 这种标定方法把机器人实际位姿和相应的关节角误差分别作为前馈神经网络的输入和输出来训练网络,从而获得机器人任意位姿时的关节角误差值,通过修改关节值来提高机器人的位姿精度。
The first algorithm based on feed - forward neural network , this algorithm is developed by the combination of conventional cma and feed - forward neural network . according to the cma blind equalization thought - way , a new cost function is proposed so that apply the feed - forward neural network to blind equalization 第一类是基于前馈神经网络( fnn )盲均衡算法,这种算法是将传统的cma算法与神经网络相结合提出的,根据传统算法的均衡思想而重定义了代价函数,使其能用到神经网络当中。