Fulfilling Formal Specifications ASAP by Model-free Reinforcement Learning