Model Free Optimal Control Approach for UAVs

Abstract

This thesis discusses use of model free control algorithms for application on an UAV. The work contrasts use of LQR based methods to conventional model free learning based on neural net approximators which do not provide guarantees of optimal solution. The model free control methods discussed here are based on discrete time linear systems with LQR based costs that have proven convergence to the optimal solution. We discuss Q-learning algorithm for LQR and a policy gradient methods with a variation. We see applications in simulations for a noisy measurement case with sub-optimal controller as well as a scenario where the dynamics of system has been altered due to disturbances or manipulation.