The Research of Quadrotor Flight Control Based on Reinforcement Learning and ADP

This paper studies the application of Lookup-Table reinforcement learning method into the continuous state space control of quadrotor simulator and designs a attitude controller for the quadrotor simulator based on Q-learning; for the improvement of defects concerning difficulty in the learning algorithm's convergence and low efficiency in learning when Q-learning is faced with large-scale and continuous-space optimized decision, the method of kernel approximate dynamic programming is introduced, Kernel-based Least-Squares Policy Iteration (KLSPI) is proposed, and a controller for the quadrotor simulator is designed based on this algorithm. The experiment shows that the reinforcement learning control method is of fast convergence speed, small steady-state error, strong adaptive ability and good control effect; when dealing with the problem of continuous state space, the Least-Squares Policy Iteration can converge better strategies with fewer training data compared with the traditional method of discretizing state space first.