The ball-and-beam problem is a well known benchmark for testing new control algorithms. We deal with the off-line training of neurocontrollers to balance the bal at a fixed arbitrary location on the beam. Resulting neurocontrollers are tested on our original hardware. We record a time series of positions of the ball, and it is the only signal permitted to use for identification and control. We utilize recurrent neural networks for all modules of our designs. We obtain a sufficiently accurate neural network identification model of the system using the parallel identification method. Two neurocontrol designs are discussed. The conventional approach is based on truncated backpropagation through time. Another design uses an adaptive critic approach, which is a form of approximate dynamic programming.
展开▼
