Improved neural network control of inverted pendulums

Abstract

Nowadays, neural network controllers (NNCs) are getting more and more prevalent because they are able to handle unknown systems by learning them and adapt to their changing behaviour. The family of robust fixed point transformations (RFPT) has been partly developed to solve control tasks without knowing the exact parameters of a controlled system. When disturbances effect a plant or the neural network controller is not trained properly RFPT integrated to the controller is suitable to reduce the problems caused by the model approximation and make the controller robust to the unknown external forces. In this paper, a novel combination of neural networks and robust fixed point transformations is introduced to balance an inverted pendulum on the top of a cart of changing nominal position. The results show that the inaccuracies caused by the disturbances can be reduced significantly when RFPT is used in the control process.