Verifying PCTL Specifications on Markov Decision Processes via Reinforcement Learning