helpingstar/gym-snakegame

🕹️ Snake game-based reinforcement learning environment using Gymnasium

gym-snakegame

A gymnasium-based RL environment for learning the snake game.

snakegame_10_clear.mp4

You can make a game with a grid size larger than 5x5.

Action Space	Observation Space
Discrete(4)	(board_size, board_size)

Installation

git clone https://github.com/helpingstar/gym-snakegame.git
cd gym-snakegame
pip install -e .

Usage

import gym_snakegame
import gymnasium as gym

env = gym.make(
    "gym_snakegame/SnakeGame-v0", board_size=10, n_channel=1, n_target=1, render_mode='human'
)

obs, info = env.reset()
for i in range(10000):
    action = env.action_space.sample()
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        obs, info = env.reset()
env.close()

parameter

• board_size : The size of a square board. The board has the shape (board_size, board_size).
• n_target : The number of targets placed on a board.
• n_channel : The number of channels of the observation. The observation has the shape (n_channel, board_size, board_size).

Observation

Observation Space : Box(0.0, board_size ** 2 + 1, (n_channel, board_size, board_size), uint32)

Channel

• 1 : Like the rendering format, all information is expressed in one channel.
• 2 : The snake and item channels are divided.
  • 0 : snake
  • 1 : item
• 4 : Channels are divided in the following order.
  • 0 : snake's head
  • 1 : snake's body
  • 2 : snake's tail
  • 3 : item

How to represent each element.

• 0 : empty
• 1 ~ board_size ** 2 : snake body
  • 1 : head
  • largest number : tail
• board_size ** 2 + 1 : target

Action

Action Space : Discrete(4)

• 0 : down
• 1 : right
• 2 : up
• 3 : left

Invalid Action

When the Agent takes an Invalid Action, the Agent continues in the direction it was previously going.