Variable Gain Gradient Descent-based Robust Reinforcement Learning for Optimal Tracking Control of Unknown Nonlinear System with Input-Constraints.