Neuro-genetic adaptive attitude control

It has previously been demonstrated that for smooth dynamic systems, using relatively few sample points from a single trajectory, a neural network can be trained to perform very accurate short-term prediction over a large part of the phase space. In this paper, we exploit the capability of a Locally Predictive Network (LPN) to derive an adaptive control architecture for a satellite equipped with controllable, bidirectional thrusters on each of the three principal axes. It is assumed that a hardware implementation of the neural network is available. The inputs for the network are a small history of system states up to the present time and a set of current control inputs, the outputs are the next system state. Once the LPN has been trained successfully, at each time step a genetic algorithm searches the space of hypothetical control inputs. Given a set of control signals, the LPN is used to predict the state of the system at the next sample point. This enables the ‘fitness’ of each set of hypothetical control torques to be evaluated very rapidly. In effect, the genetic algorithm determines a satisfactory solution to the inverse kinematic problem in time to apply the solution (set of control torques) at the next control point. With the exception of the neuromodelling (which is repeated only when the system dynamics change), the whole process is then repeated. The results presented indicate that such an architecture is easily able to master the attitude control problem for arbitrary slew angles, with arbitrary a priori unknowndynamics and noise in the sensor system.