PPO - Reinforcement Learning

Snake-AI

Observation

Visual observation

• Visual observation. The AI is observing a 13x13 space around this head. Six on the left and right site and the Head in the middle.

Static observation

• Ray tracing along the yellow dashed lines. The AI is able to see himself, walls and the apple.
• Direction (red line) of the snake.
• apple (blue point) and tail compass. Indicates the relativ position according to the apple or the last part of the snake.
• step counter. If the snake doesn't eat an apple in a descried amount of steps the game ends.

Evaluation / Reward

+100 if the snake reaches the max length. | win

+2.5 if snake eats an apple.

-10 if the snake dies. | loss

PPO - Implementation

• Implementation with ray tune.

License

GPLv3 (General Public License 3)

Results

Dependencies

Install packages: pip install -r requirements.txt
Be careful with pytorch dependencies. They often might not work with. If so, install on your own.