ADAPTIVE ATTITUDE CONTROL USING TREE-STRUCTURED WAVELET NETWORKS

摘要

To be a practical alternative to more conventional control strategies, neurocontrol designs must minimize the amount of computation required to achieve accurate tracking. Recent progress in this direction has been made by using dynamically structured networks of wavelet basis functions [3,24], but the resulting controllers may still require an unacceptably large amount of computation in high dimensional state spaces. Working within the stability framework developed in [24] we use MARS [10] inspired techniques to extend the wavelet network designs in [3,24]. In addition to dynamically choosing scale and translation parameters, the controller developed also varies its dependence on individual components of the measured state vector, attempting to construct low dimensional approximations to the functions required to accurately track a reference trajectory. A specific application to spacecraft attitude control is presented, and the performance of the controller is illustrated with simulations.