Deep Q-Learning on Snake game, implemented with PyTorch Lightning and Kivy visualization.
The results shown below were created with a model consisting of 2 hidden layers, both with 256 nodes, 11 input nodes, and 3 output nodes.
As an input binary vector was used, including 4 values for current snake movement direction, 4 values for food orientation based on snakes head, and 3 values for immediate danger around snakes head.
The output corresponds to 3 actions: turn left, go straight, turn right.
Untrained model:
After 2.5 min of training (10 000 iterations):
