environment-framework 0.3.0a1

Loose building blocks to create agent-environment loops.

🌐 Environment Framework

This repository contains the Python package environment-framework. The project aims to provide loose building blocks to manage the logic, observation, estimation and visualization of an agent-environment loop. It can be used to implement problems which might be solved with reinforcement learning or dynamic programming algorithms.

A wrapper around gymnasium is provided to connect to well-known frameworks in the field.

The wrapper for gymnasium uses the gymnasium>=0.26 API structure!

🤔 Why create this project?

The project emerges from a previous project of mine. It was used to separate the different elements of the projects agent-environment-loop.

🚀 Get Started

Installation

pip3 install environment-framework

👩‍🏫 GridWorld Example

The implemented example of GridWorld can also be found in a Jupyter notebook grid_world.ipynb.

pip3 install "environment-framework[extra]"
jupyter lab

class Action:
    UP = 0
    DOWN = 1
    RIGHT = 2
    LEFT = 3

class GridWorldGame:
    def __init__(self, size: int) -> None:
        self.size = size
        self.player_position = (0, 0)
        self.target_position = (0, 0)
        self.reset()

    @property
    def done(self) -> bool:
        return self.player_position == self.target_position

    @property
    def space(self) -> Space:
        return Discrete(4)

    def act(self, action: int, **_: Any) -> None:
        if action == Action.UP:
            self.player_position = (self.player_position[0], self.player_position[1] - 1)
        if action == Action.DOWN:
            self.player_position = (self.player_position[0], self.player_position[1] + 1)
        if action == Action.RIGHT:
            self.player_position = (self.player_position[0] + 1, self.player_position[1])
        if action == Action.LEFT:
            self.player_position = (self.player_position[0] - 1, self.player_position[1])
        corrected_x = max(0, min(self.size - 1, self.player_position[0]))
        corrected_y = max(0, min(self.size - 1, self.player_position[1]))
        self.player_position = (corrected_x, corrected_y)

    def reset(self) -> None:
        def get_random_position() -> int:
            return randint(0, self.size - 1)
        self.player_position = (get_random_position(), get_random_position())
        self.target_position = (get_random_position(), get_random_position())
        if self.done:
            self.reset()

class GridWorldObserver:
    def __init__(self, game: GridWorldGame) -> None:
        self.game = game

    @property
    def space(self) -> Space:
        return Box(shape=(4,), low=-math.inf, high=math.inf)

    def observe(self) -> NDArray:
        return np.array(
            [*self.game.player_position, *self.game.target_position],
            dtype=np.float32,
        )

class GridWorldEstimator:
    def __init__(self, game: GridWorldGame) -> None:
        self.game = game

    def estimate(self) -> float:
        return -1 + float(self.game.done)

class GridWorldVisualizer(PygameHumanVisualizer):
    BLUE = [0, 0, 255]
    GREEN = [0, 255, 0]

    def __init__(self, game: GridWorldGame) -> None:
        super().__init__(50)
        self.game = game

    def render_rgb(self) -> NDArray[np.uint8]:
        frame = [[[0 for k in range(3)] for j in range(self.game.size)] for i in range(self.game.size)]
        frame[self.game.player_position[1]][self.game.player_position[0]] = self.BLUE
        frame[self.game.target_position[1]][self.game.target_position[0]] = self.GREEN
        return np.array(frame, dtype=np.uint8)

class GridWorldLevel(Level):
    _game: GridWorldGame
    _observer: GridWorldObserver
    _estimator: GridWorldEstimator
    _visualizer: GridWorldVisualizer

    def reset(self) -> None:
        self._game.reset()

    def step(self, action: int) -> Any:
        self._game.act(action)

game = GridWorldGame(7)
level = GridWorldLevel(
    game,
    GridWorldObserver(game),
    GridWorldEstimator(game),
    GridWorldVisualizer(game),
)
simulator = Simulator(level, 50)
FPS = 4
DONE = False
while not DONE:
    action = simulator.action_space.sample()
    simulator.step(action)
    obs = simulator.observe()
    reward = simulator.estimate()
    simulator.render_human(FPS)
    DONE = simulator.truncated or simulator.done
simulator.close()

📃 Documentation

Some doc-strings are already added. Documentation is a work-in-progress and will be updated on a time by time basis.

💃🕺 Contribution

I welcome everybody contributing to this project. Please read the CONTRIBUTING.md for more information. Feel free to open an issue on the project if you have any further questions.

💻 Development

The repository provides tools for development using hatch.

All dependencies for the project also can be found in a requirements-file.

Install the development dependencies.

pip3 install -r requirements/dev.txt

or

pip3 install "environment-framework[dev]"

Tools

To run all development tools, type checking, linting and tests hatch is required.

make all

Type checking

Type checking with mypy.

make mypy

Linting

Linting with pylint.

make lint

Tests

Run tests with pytest.

make test

Update dependencies

Update python requirements with pip-compile.

make update

🧾 License

This repository is licensed under the MIT License.