Learning to balance the inverted pendulum using neural networks

A learning architecture for training a neural network controller to provide the appropriate force to balance an inverted pendulum is proposed. The control rule is given on the output space of the inverted pendulum, introducing a priori knowledge in the learning algorithm. The system uses two neural networks, one for the identification and the other for the controlling. The identification network learns to identify the plant dynamics while the controlling network simultaneously changes its characteristics using the result of the identification to control the actual plant. The training process is performed using the backpropagation learning algorithm. The inverted pendulum system is simulated to illustrate these ideas. The controller is able to balance the inverted pendulum and guide the cart to the center of a track. demonstrating the effectiveness of the learning method.<>