Abstract
This paper proposes a method for controller approximation via neural network in the presence of parametric perturbations. The neural network is based on long short-term memory blocks and is trained to approximate a numerical optimal control law, solved for different parameter values. Using this approach, the obtained approximate control law learns to generate the control inputs based on different optimal control solutions for different parameters: as compared to training the neural network only based on the optimal control law defined for the nominal parameters, the overall system performance greatly improves when parameter variations are present, and does not degrade when the nominal parameters are used for testing. The proposed approach is validated experimentally on an inverted pendulum with dual-axis reaction wheels.