基于Levenberg-Marquardt算法的神经网络监督控制

Neural Network Supervised Control Based on Levenberg-Marquardt Algorithm

提出了基于Levenberg Marquardt(L M )算法的前向多层神经网络在线监督的控制方法 ,其算法是梯度下降法与高斯 牛顿法的结合 .对于训练次数及准确度 ,L M算法明显优于共轭梯度法及变学习率的BP(BackPropagation)算法 ,适用于在线学习与控制 .因此 ,利用L M算法的特点进行在线训练神经网络 ,以实现实时非线性控制 .仿真结果表明 ,该控制方法优于常规控制算法 ,明显改善了在未知负载扰动时 ,伺服系统的跟踪性能 ,显著地降低了跟踪误差 。

A multilayer neural network supervised online control strategy based on Levenberg Marquardt training algorithm is proposed for the tracking control problem of the electrohydraulic position servo systems subjected to constant and timevarying external load disturbances. The LevenbergMarquardt algorithm is the combination of the steepest decent algorithm with the GaussNewton algorithm. Compared with a conjugate gradient algorithm and a variable learning rate algorithm, the LevenbergMarquardt algorithm is much more efficient than either of them on the training steps and accuracy. Therefore, it can be applied to online control. The output of the system successfully tracked the specified sinusoidal after a relatively short online training period. The control strategy is used to adapt to uncertainties of disturbances and learns their inherent nonlinearities. Simulation results illustrate that a neurocontroller used in supervised control schemes can result in good robustness and tracking property.