Near-Optimal Finite-Time Feedback Controller Synthesis Using Supervised and Unsupervised Learning
Dongliang Zheng, Panagiotis Tsiotras · AIAA Scitech 2021 Forum · 2021
View Video Presentation: https://doi.org/10.2514/6.2021-1950.vid This paper proposes near-optimal controller synthesis for finite-time optimal control problems using ideas borrowed from the machine learning community. Based on a finite number of pre-computed optimal trajectories, a model of the optimal controller is built using both supervised (regression) and unsupervised (classification) learning. Neural networks are trained to approximate the mapping from state to control, resulting in a state-feedback controller. The resulting neural network controller (NNC) is used for online control prediction and generating near-optimal trajectories to steer the system from any initial state to the goal state. Two extensions are introduced compared with previous related works in the literature. First, the use of clustering divides the offline optimal trajectories into different groups and then builds suitable regression models for each group. Second, the use of time labels as an additional input to the regression model highlights the improvement enabled by the use of time-stamped data for finite-time optimal controller synthesis. Multiple simulations, including a 3D quadrotor with a 12-dimensional state-space, are provided to evaluate the performance of the proposed neural network controller. The proposed method generates near-optimal trajectories for nontrivial nonlinear systems and is suitable for real-time implementation.