Requires: python 3.5 pytorch 0.4 Important files: DQN.py pacmanDQN_Agents.py
To test the DQN network, launch: python3 pacman.py -p PacmanDQN -n 200 -x 100 -l smallGrid
To train the DQN network, launch: python3 pacman.py -p PacmanDQN -n 3000 -x 2900 -l smallGrid
Where: -n = number of episodes -x = episodes used for training (graphics = off)
Remarks: the game files had to be updated for python3 (print was not working) the model has already been trained and wins most of the time the model has been optimized, it requires less then 30 000 episodes to converge
To test training, change: model_trained = False in pacmanDQN_Agents.py (line 26)
I used the Pacman game engine provided by the UC Berkley Intro to AI project: http://ai.berkeley.edu/reinforcement.html
I'd like to cite: https://github.com/tychovdo/PacmanDQN His implementation in Tensorflow helped me configure the Neural Network architecture.