Snake implemented on pygame meant to be used by human and AI agents
Snake game that can be controlled by human input and AI agents (DQN). Who's best? :snake: :robot:
Table of Contents
- 1. Getting Started (for human players)
- 2. Getting Started (using AI agents)
- 3. Contributing
- 4. License
- 5. Acknowledgments
Let's get the game up and running on your computer, with the instructions below. You can play the game and compare to the repos benchmark, which includes AI and humans (you can include yourself by a pull request to the file scores.json).
To play the game you need Python 3.4+. If you installed Anaconda the only package you need to download is pygame. Before installing it, make sure your Anaconda installation is up-to-date using the command (conda update conda anaconda)and updating all packages (conda update --all). To install pygame, use:
$ conda install -c cogsci pygame
It's highly recommended to use Anaconda to manage your Python packages and environments. If you chose not to, make sure you run requirements_human.txt, using:
$ python install setup.py
$ python3 install setup.py
You can download download the source code or clone the repository to your computer.
To clone the repository, open bash or command prompt, cd to the chosen directory and run the following code:
$ git clone https://github.com/Neves4/SnakeAI.git
To download the repo, just follow along the gif (click 'Clone or Download' and then 'Download ZIP').
The GUI allows you to choose between single games and the benchmark mode. It's also possible to choose between difficulty levels.
If using the repository files, change directory to the root, then to the game folder and use:
$ python snake.py [-h]
An example gameplay for a single player match is shown below.
In the benchmark mode, you will play through 10 games and your mean score/steps are going to be recorded and you can add to the leaderboards. Pull request changing the benchmark file (located in here) or open an issue with your score.
This game uses similar usage structure and methods to OpenAI's gym and you can easily integrate it with any agent, written in Pytorch, Tensorflow, Theano or Keras.
It's recommended that you use colab-rl, a repository that integrates state-of-the-art algorithms with games, because it already implements the agents and the game, making the process of quick prototyping much easier.
In this section, we're going to show the useful methods and properties and also demonstrate how to use in a real case
Below are listed some useful properties of the game class.
>>> print(game.nb_actions) 5 # number of actions. >>> print(game.food_pos) (6, 5) # current position of food. >>> print(game.steps) 10 # current number of steps in a given episode. >>> print(game.snake.length) 4 # current length of the snake in a given episode.
The methods you could use to integrate with any AI agent are:
>>> state = game.reset() # Reset the game and returns the game state right after resetting. >>> state = game.state() # Get the current game state. >>> game.food_pos = game.generate_food() # Update the food position. >>> state, reward, done, info = game.step(numerical_action) # Play a numerical_action, obtaining state, reward, over and info. >>> game.render() # Render the game in a pygame window.
To use with AI agents, you need to integrate the game with the AI agent. An example usage is:
from snake-on-pygame import Game from ai_agent import your_model # import your AI agent of choice game = Game(player = "ROBOT", board_size = board_size, local_state = local_state, relative_pos = RELATIVE_POS) state = game.reset() model = your_model() while not game.game_over: # Main loop, until game_over game.food_pos = game.generate_food() model.choose_action(state) # CHOOSE ACTION BASED ON MODEL/AI AGENT state, reward, done, info = game.step(action) mode.train(state, reward, game_over, done) print(info) model.test()
The above code is an example usage for one episode. If you want more episodes, wrap the while loop in a loop for nb_epochs (you choose).
Using snake-on-pygame with colab-rl (click here) is very straightforward and you can also experiment with hyperparameters on state-of-the-art algorithms.
A detailed usage is described on the repo's main README, but for short, after cloning it you can just execute the run_dqn.py script with:
$ python run_dqn.py [-h]
And you can read more about all the possible arguments in the file/repo. An trained DQN model, on a 10 x 10 board with no customization is represented on the below GIF.
Please read CONTRIBUTING.md for details on this repo's code of conduct, and the process for submitting pull requests.
This project is licensed under the MIT License - see the LICENSE file for details.
@farizrahman4u - For his qlearning4k snake code, I used it as the base of this repo's code.
@chuyangliu - Also for his snake code, which implemented the relative actions.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.