Learning to play Pacman (and other Atari games) using DQN methods (vanilla, double DQN and prioritized experience replay).
Training can be initialized locally or remotely on a cloud environment.
If you are completely in control of the machine the training is running on, you can run:
python train.py -c [config] [options]
Various configuration files are stored in config/.
You can specify by a flag which algorithm the agent will use to the training script.
no flag: train using vanilla DQN
--per: train using Prioritized Experience Replay
--doubledqn: train using Double DQN
Training runs will log to data/logs/training.log by default.
I use nohup in order to prevent termination of a SSH session.
Output will be redirected to a file nohup.out by default.
The & will start the process in the background.
nohup python train.py -c [config] [options] &
You can track the training progress using:
tail -f data/logs/training.log
The training process knows to shutdown softly, while saving all progress it has made so far and archiving it to a file named results.tar.gz.
In order to achieve soft shutdown, kill the Python process with SIGKILL (signal 2):
pkill -2 [pid of nohup'd Python process]
To clean up the data directory, run cleanup.sh. Note that this script deletes everything in data/ and recreates the directory entirely, along with some necessary files.
./cleanup.sh
To format code and documentation, run:
./format.py