It's a small snake game build using pygame which plays by itself using Dueling DQN implemented with Pytorch.
If you try to run the game from zero it took more than 1 million iterations to get to this place (see image bellow).
My hardware to train the Agent is:
- Core i7 (8 cores)
- 16 Mb
- NVIDIA Geforce 1060 with 6 Gb
Run main.py if you want to train and watch the Agent!
Run play.py if you just want to watch the Agent!
Run train.py if you already run main.py and have some memory... the Agent will train on stored memory (this is not the common pratice).
The model was based on the original Dueling DQN presented in the paper. I used RMSProp and MSE as proposed in the same paper as the Optmizer and Loss.