Application of Artificial Neural Networks for Active Roll Control Based on Actor-Critic Reinforcement Learning

Abstract

AbstractThis work shows the application of artificial neural networks for the control task of the roll angle in passenger cars. The training of the artificial neural network is based on the specific actor-critic reinforcement learning training algorithm. It is implemented and trained utilizing the Python API for TensorFlow and set up in a co-simulation with the vehicle simulation realized in IPG CarMaker via MATLAB/Simulink to enable online learning. Subsequently it is validated in different representative driving maneuvers. For showing the practicability of the resulting neural controller it is also validated for different vehicle classes with respect to their corresponding structure, geometries and components. An analytical approach to adjust the resulting controller to various vehicle bodies dependent on physical correlations is presented.KeywordsArtificial neural networkMachine learningActor-criticReinforcement learningActive roll controlVehicle dynamics