Learning to Solve Nonlinear Optimization Problem with Deep Reinforcement Learning

Nonlinear least-squares problems (NLS) are pop-ular in engineering and scientific fields. Traditional optimization methods such as Newton's method and Gaussian-Newton method (GN) suffer from the sensibility to initial values and the high computational complexity. In this paper, we propose LS-DDPG, a robust optimization method utilizing deep rein-forcement learning algorithms to solve nonlinear least-squares problems. The experiment results on synthetic data demonstrate that the proposed method outperforms Newton's method in terms of computation cost, convergence speed and initial values sensibility. In addition, LS-DDPG is utilized on model predictive control (MPC) problems for trajectory planning and tracking tasks in self-driving with longer prediction horizon and higher accuracy than baseline methods.