No discussion of reinforcement learning would be complete without at least a mention of the paper, Playing Atari with Deep Reinforcement Learning by Mnih et al. (https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf) then of DeepMind, now of Google. In this landmark paper, the authors used a convolutional neural network to train a deep Q network to play Atari 2600 games. They took the raw pixel output from the Atari 2600 games, scaled it down a bit, converted it to gray scale, and then used that as the state space input for the network. In order for the computer to understand the velocity and direction of the objects on screen, they used a four image buffer as an input to the deep Q network.

The authors were able to create an agent that was able to play seven Atari 2600 games with the exact same neural network architecture, and the agent was better than a human on three of those games. This was later extended to 49 games, the majority of which it was better at than a human. This paper was a really important step towards general AI, and it's really the foundation of much of the research currently happening in reinforcement learning.