Finally there is a trend of using recurrent neural network as a top component of the Q-network. Perhaps we will see even more sophisticated RNNs like DNC and Recurrent Entity Networks applied here. Also we'll see meta-reinforcement learning applied to a curriculum of environments.