STOCHASTIC TUNING OF A SPACECRAFT CONTROLLER USING NEURAL NETWORKS

There have been many demonstrations of the advantages of using neural networks in control systems. Networks, such as the MLP, offer a level of adaptability and non-linearity, both of which are required in some control systems. However, for spacecraft attitude control, high levels of dependability are also required. This poses serious questions for the acceptability of neural networks. This paper describes a suggested control system which uses two MLP networks for the control of thruste rs on the SOHO spacecraft. However, rather than applying the networks directly, they form part of a stochastic parameter-selection system wh ich is used to adapt a conventional (PD) control system. It ir suggest ed that using neural networks indirectly in this way better guarantees the dependability/reliability of the control system.