DQN
This is an implementation of the classic DQN algorithm using Theano.
Requirements
- Python 3.6
- Theano 0.8.2
- OpenCV
- OpenAI Gym
- Numpy
Trained weights are provided for the following games: Boxing, Breakout, Pong and SpaceInvaders.
To train the model, please use:
python DQN.py --env [environment name]
To see it play:
python play.py --env [environment name]
The pickle files must be named after the game (i.e. breakout.pkl). You can also change that directly in the script. As mentionned in this blog from OpenAI, DQN, as well as more recent algorithms, are subject to adversarial attacks. To see the effects on this implementation, choose:
python play.py --env [environment name] --adv 1
Results
