报告题目 (Title)：Designing High-Dimensional Closed-Loop Optimal Control Using Deep Neural Networks(利用深度神经网络设计高维闭环最优控制)
报告人 (Speaker)： 胡卫
报告时间 (Time)：2024年11月22日（周五）16:00
报告地点 (Place)：校本部GJ406
邀请人(Inviter)：秦晓雪
报告摘要：Optimal control of high-dimensional nonlinear systems is a longstanding challenge, traditionally constrained by the curse of dimensionality in methods like the Hamilton-Jacobi-Bellman equation. A promising alternative leverages supervised learning, using open-loop optimal control solvers to generate training data and neural networks as high-dimensional function approximators. Despite some successes, this approach struggles with more complex problems due to covariance shift (distribution mismatch) between training and testing and data discontinuities. We address these by a strategic forward training method and a special neural network design exploiting the linear quadratic regulator (LQR). By integrating these techniques, we develop a robust closed-loop policy that operates effectively over a broad domain for a 7-DoF manipulator, achieving near globally optimal total costs.