SnakeAI

This library implements an environment for the Snake Game. This can be used for playing, but also for training Reinforcement Learning agents.

Table of Contents

1. Playing Snake
2. Playing with an RL Agent
3. Training an agent

Playing Snake

To play the game, you just need to create a UserGame object and call the play method.

from snakeai.game.user_controller import UserGame
game = UserGame(20)
game.play()

The number specified in the UserGame definition is the dimension of the board (the length of its side). To change direction, you need to use the arrow keys. At the end of the game, press R to play again.

You can also choose whether to play in a new window or inside the command line, by specifying the mode when creating the UserGame object. This can be GUIMode.WINDOW or GUIMode.CLI.

from snakeai.game.user_controller import UserGame
from snakeai.game.constants import GUIMode
game = UserGame(20, mode=GUIMode.CLI)
game.play()

In this case, to play you need to write up, down, left, right.

Playing with a RL Agent

There are five different agents you can choose from, contained in the snakeai.agents package:

• recursive_agent.Recursive uses a simulated environment to find the best move
• tabular_agent.StateAgent uses the Sarsa algorithm. The states describe the board exactly
• simple_tabular_agent.SimpleStateAgent uses the Sarsa algrithm, with simplified states
• simple_mc_agent.SimpleMCAgent uses the Monte Carlo method, with simplified states
• simple_deep_agent.DQN uses a Deep Q Network, with simplified states.

The agents labelled as "simple" use as state an array composed of 12 boolean values. The StateAgent, instead, uses a complete description of the board.

To play with one of these agents, you can use the play method from the relevant module. For example, to play with the SimpleStateAgent (contained in the snakeai.agents.simple_tabular_agent module):

from snakeai.agents.simple_tabular_agent import play
play()

If you do not specify a path, the program will use a default model.

Training an agent

To train one of these agents, you can use the following code:

from snakeai.agents.simple_tabular_agent import  SimpleStateAgent
from snakeai.game.agent_controller import AgentGame

episodes = 100_000
game = AgentGame(dim=20, show=False)
agent = SimpleStateAgent(20)
for _ in range(episodes):
    game.play(agent)
    agent.reset()

Then you can save an agent using the save method. The path should end without any extension.

agent.save(path = "./models/my_agent")