RNN learning controller trained by ADRC

In this paper, the problem of training a recurrent neural network (RNN) controller to approximate an active disturbance rejection control (ADRC) is studied. Two learning methods, online learning and offline learning, are proposed. For the online setting, the RNN learning controller is trained by the error of the control quantity compared to that of ADRC in real time, while for the offline setting, the RNN controller is trained by the error of the whole state trajectory compared to that of ADRC obtained in advance. With randomized weight initialization, success rates of trained online and offline RNN learning controllers are compared by simulation. Furthermore, the robustness property of the trained RNN controller is analyzed. A metric of robustness, which can be calculated from the weights of the trained RNN controller is also proposed.