Learning-driven Nonlinear Optimal Control via Gaussian Process Regression