Is Bellman Equation Enough for Learning Control?