Neural network architecture for online system identification and adaptively optimized control

A neural network architecture for simultaneous system identification and adaptively optimized control is presented. Basic ingredients consisting of tapped delay lines and standard analog neurons (with backpropagation) are combined within a modular/hierarchical architecture to implement an MRAC scheme having rigorous stability and convergence properties. Control adaptation is carried out in the presence of unknown (and possibly persistent) plant disturbances and instrumentation noise and requires no built-in modeling information. The overall adaptive neural controller is composed of a system replicator network and the control adaptor network. With no prior information, the system replicator autonomously builds an internal model of the plant, effectively identifying the plant within the closed loop. Simultaneously, this internal model is used by the control adaptor to adapt the online dynamic compensation so that the closed-loop input/output characteristics match those of the ideal refractive system. The efficacy of the proposed approach is illustrated by several numerical examples involving vibration control for precision space structures