Model-based Reinforcement Learning from Signal Temporal Logic Specifications