Towards General-Purpose Model-Free Reinforcement Learning