多層ネットワークによる倒立振子の安定化学習制御

抄録

We propose a learning control scheme using a multi-layered neural network such that the control performance is improved for unknown controlled objects by repeated trials. This system was applied to the stabilization of an inverted pendulum. The system consists of three subsystems ; a neurocontroller, a temporary target generator and an error evaluator. The temporary target generator yields reference angle of the inverted pendulum to control the cart position. The error evaluator generates a teaching signal for the neuro controller by comparing the system error with an error-reference model. The backpropagation algorithm was used for the learning of neural network. The learning time could be shortened through the pre-learning process by imbedding a priori knowledge in the neural network. We verified the effectiveness of the proposed method by computer simulation.