Model-free reinforcement learning (RL) works well when experience can be
collected cheaply and model-based RL is effective when system dynamics can be
modeled accurately. However, both assumptions can be violated in real world
problems such as robotics, where querying the system can be