Implementation of a deep Q-network agent, that can learn successful policies directly from pixels of the screen as input. Trained for 1.8 million frames of the game, the agent was able to achieve an average score of 40 with upto 20 minutes of gameplay on its own.
There are 2 directories for the envs: Breakout and CartPole. The code in the directories was used to train the agent. The Demo code for Breakout is in file Demo-Breakout.ipynb.