Deep Reinforcement Learning

This is a project designed to make the development, training and testing of DRL algorithms easy. Any new algorithms can be implemented as an Agent class using tensorflow. Then, they can be compared against each other under the exact same setting.

Agents are derived from the BaseAgent class, so that a particular algorithms can be implemented in very few lines. And if a new agent is very similar to an old one, you can just derive from the old one and re-implement only the new part.

The currently implemented agents are:

• DQN
• DoubleDQN
• DuelingDoubleDQN

Example usage: python train.py -agent DQN -device 1 -env_name Breakout-v0

Understand the runs

When you are testing many different algorithms in many different settings things can get confusing. DRL tries to organize everything you might need in the simplest manner.

Use tensorboard --logdir=log to view all the runs. They will be all organized by a number, the name of the agent and the name of the environment.

You can also see the arguments of each run in log/[run_name]/config.txt

Performance

I have very limited compute resources and they are being uses for my research. I would really appreciate if somebody wanted to help by running DQN in different environments.

DQN on breakout-v0 for 50,000,000 steps

This run clearly was still learning and needed to continue. When I have time I will test on the original Breakout._