This repository shows how Autonomous Learning Library can be used to build new reinforcement learning agents. In particular, it contains a model based agent that predicts future frames and uses them to guide decision making.

First, you'll need the latest version of Pytorch. If you wish to view Tensorboard logs, you'll also need to grab a copy of that (it also comes with tensorflow). Then, you'll need to install the autonomous-learning-library along with the Atari environments:

pip install autonomous-learning-library[atari]

Unfortunately, the current IP holders for the Atari library made it more difficult to acquire a license and use the ROMs than it used to be. If you have a license to use the ROMs, you can try AutoROM.

You can run the agent as well as a baseline DQN agent using:

python main.py Pong

You can track progress using:

tensorboard --logdir runs

Once the script has finished (could take a long time, especially if you do not have a fast GPU!), you can see the final results using:

python plot.py

For us, the above instructions produced the following results:

As you can see, this agent isn't very good! On the other hand, the purpose of this agent was not performance, but to demonstrate the utility of the autonomous-learning-library in developing new agents not included in the original library. Maybe you can come up with ways of improving this agent!