神经网络在线投影算法及非线性建模应用

On-line Neural Network Projection Algorithm for Nonlinear Modeling

针对神经网络难以在线学习的缺点,把神经网络当作结构已知的非线性系统,权系数的学习看成非线性系统的参数估计,基于新估计准则的非线性系统在线参数估计投影算法,给出前馈神经网络的一种在线运行投影学习算法。理论上证明该算法的全局收敛性,讨论算法参数的物理意义和取值范围。通过2个非线性时变系统的神经网络建模应用的仿真,验证算法的全局收敛性和在线运行能力。

To solve the on-line learning for neural networks, the neural network is taken as a nonlinear system with known structure, and the learning of weight parameter is taken as a parameter estimation for nonlinear system. Then,the on-line projection algorithm of the forward neural network is presented by the projection algorithm of nonlinear system parameter estimation based on a new estimation criterion. The global convergence is proved in theory, and the physical meaning and changing area of parameters are discussed. The simulation results of neural network modeling for nonlinear time-varying systems show the global convergence and on-line learning ability.