Increasing Transparency of Reinforcement Learning using Shielding for Human Preferences and Explanations