ventus550 / deepmac

Deep Q-learning pacman agent

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Deepmac

Deep Q-Learning agent playing Pac-Man for the Neural Networks course. The final model is demonstrably improving over time but the training requires thousands of episodes thus being computationally far too expensive. That being said agent easily outperforms random agents. As of this moment the most successful approach was to run double Q-Learning with convolutional neural networks on memory-sampled transition batches while scheduling gradual decrease of epsilon (in ɛ-greedy strategy).

About

Deep Q-learning pacman agent

License:MIT License


Languages

Language:Jupyter Notebook 98.3%Language:Python 1.7%