Imitate the Good and Avoid the Bad: An Incremental Approach to Safe Reinforcement Learning