Online Feedback Error Learning Control for an Inverted Pendulum

This paper presents an application of the feedback error learning technique for online control of an inverted pendulum which has uncertain friction nonlinearity. In order to build up online adaptive learning control, i) the preliminary offline training and the scaling factor for the neural network to escape from the local minimum, and ii) twostage learning scheme are introduced. After some learning cycles, the vibrations of the inverted pendulum are completely ceased that the feedback error learning scheme acts as an adaptive controller to minimize the control error. This means that the neural network acquires the inverse dynamic model of the plant through learning, and then compensates the nonlinearity of the plant. The phase relationships of the control outputs between the conventional feedback controller and the adaptive neural network controller are clarified. It is also shown that this control system works well for a step reference signal after learning.