Optimal attitude control of satellites by artificial neural networks: a pilot study

A pilot study is described on the practical application of artificial neural networks. The limit cycle of the attitude control of a satellite is selected as the test case. One of the sources of the limit cycle is a position dependent error in the observed attitude. A Reinforcement Learning method is selected, which is able to adapt a controller such that a cost function is optimised. An estimate of the cost function is learned by a neural ‘critic’. In our approach, the estimated cost function is directly represented as a function of the parameters of a linear controller. The critic is implemented as a CMAC network. Results from simulations show that the method is able to find optimal parameters without unstable behaviour. In particular in the case of large discontinuities in the attitude measurements, the method shows a clear improvement compared to the conventional approach: the RMS attitude error decreases approximately 30%.