Planning

We propose a long-horizon background-planning algorithm for online RL. This used compressed subgoal models (abstract in state & time) for faster long-term decision making & smarter value propagation.