Comparison of CMAC architectures for neural network based control

Two control system architectures using CMAC (cerebellar model articulation controller) neural networks are compared. The first method uses CMAC to learn the inverse dynamics of the plant. The network predicts the control signal required during the next cycle by associating the current system state with previously trained states. The CMAC controller functions in conjunction with a traditional fixed gain controller to improve performance as the networks learns. The second method uses the CMAC network in a model reference structure. The network weights are adjusted as a function of the tracking error between the desired response and the system response. In this structure the network learns the relationship between previously experienced errors and the correct control signal. The methods are compared for speed of learning, tracking performance, noise rejection properties, robustness, and closed-loop stability.<>