This document summarizes and compares recent research on hierarchical reinforcement learning and model-based reinforcement learning using neural networks. It discusses work on reward augmentation, hierarchical actions, model-based video prediction, Value Iteration Networks, and the Predictron. Value Iteration Networks allow planning using value iteration in a neural network by treating the Bellman operator as a convolutional layer. The Predictron enables end-to-end learning and simulation of Markov reward processes without requiring interpretability of the state space. Future research directions include theoretical bounds for optimal policies on abstracted MDPs and learning smaller MDPs with hierarchical actions.