My objective for this project was to replicate DeepMind's Deep Q-Network agent as published in [1].

This repo contains my implementation in Python using tensorflow and OpenAI's gym.

[1] Mnih et al. 2015. "Human-level control through Deep Reinforcement Learning". Nature 518.

python 2.7

tensorflow 1.0

gym 0.8.1

tqdm 4.11.2 (simply pip install tqdm if missing)

/code     code base
/logs     tensorboard logs and checkpoints
/models   saved models

The agent be trained on any atari environment by running

 python main.py --env [atari-environment] # e.g. Pong-v3

OpenAI has a list of all available atari games. The only requirement is that you specify preprocessing options in config.py for any new environment.

All hyperparameters such as the learning rate for the Q-network or the exploration rate of the agent can be changed in main.py.

The final checkpoints for all trained agents are saved in /models. Each agent can be loaded and its control performance evaluated (by default using 20 episodes).

To load an agent go to the folder ./code and run:

 python evaluate.py --env Pong-v3 --n_episodes 20          # Pong agent
 python evaluate.py --env Boxing-v3 --n_episodes 20        # Boxing agent

where --env can be any gym environment.