Reinforcement Learning with Function Approximation: From Linear to Nonlinear