Some reinforcement learning algorithms to play snake game taken from Sutton's book

$ python snake.py --help

1. Monte Carlo
2. one-step Q-learning
3. n-step SARSA

To stop training press Ctrl-C and Q.pkl file will be saved in current directory. Then it can be used to continue training or to follow learned policy with visualization.

python snake.py --train --x=5 --y=5 --algo=mc

It is enough to train non-growing snake on a 5x5 grid to be able to use it on arbitrary large grid.
The more you train the better it becomes. Test it:

python snake.py --x=10 --y=10

python snake.py --train --x=5 --y=5 --grow --algo=sarsa --step=4

Additional 9 boolean indicators are added to a snake's state for each cell around a head, indicating if cell belongs to a snake. So snake is myopic in terms of what it can see. Adding all grid cells is not tractable due to enourmouse ammount of possible states.

python snake.py --x=5 --y=5 --grow --delay=0.3

Left to right: Monte Carlo, 1-step SARSA, 4-step SARSA, 1-step Q-learning