maze

Gymnasium RL environment for maze game.

This game generates a maze, and the player controls the agent to reach the goal point. When the agent reaches the goal, the game ends and a new game begins (A new maze is generated.)

Overview	Description
Action Space	Discrete(5, start=-1)
Observation Space	"maze": MultiBinary([H, W, 4]), "agent": Box([0,0], [H_max - 1, W_max - 1], (2,), int), "target": Box([0,0], [H_max - 1, W_max - 1], (2,), int)
import	gymnasium.make("gym_maze/Maze-v0")

H_max, W_max: the maximum values for height and width. These values are determined when the environment is created. (See Arguments)

Installtion

To get started with this project, follow these steps to set up a Conda environment and install the nessessary dependencies.

# Clone the repository to your local machine.
git clone https://github.com/chointer/maze.git

# Create and Activate a Conda Environment.
conda create --name myenv python=3.11.9
conda activate myenv

# Install Dependencies
cd maze
pip install -r requriements.txt

Usage

import gymnasium as gym
import gym_maze

env = gym.make('gym_maze/Maze-v0', render_mode="human", height_range=[2, 5], width_range=[2, 5])

obs, info = env.reset()
for i in range(1000):
    action = env.action_space.sample()
    obs, reward, terminated, _, info = env.step(action)
    if terminated:
        env.reset()
env.close()

If you want to control the agent manually,

import gymnasium as gym
import gym_maze
from gym_maze.utils.play_discrete import play_discrete

env = gym.make('gym_maze/Maze-v0', render_mode="rgb_array", height_range=[5, 20], width_range=[5, 20])
play_discrete(env, noop=-1)

Maze Generation

This environment generates mazes using Growing Tree Algorithm.
The parameter of this algorithm, nr_ratio, can be controled in the reset function which creates mazes.

The maze size range is set when the environment is created. (See Arguments)
height_range and width_range take integer lists as boundaries for the height and width ranges, respectively. While the size of the maze can change each time it is generated, the size of the "maze" observation remains fixed as the largest maze size; (H_max, W_max, 4).

Action Space

Value	Meaning
-1	NOOP
0	UP
1	DWON
2	LEFT
3	RIGHT

Observation Space

• "maze"
  MultiBinary([H_max, W_max, 4])
  "maze" contains maze map information. If [h, w, d] element is False, agent cannot move in the d direction at the h-th row w-th column (blocked).

• "agent"
  Box([0,0], [H_max - 1, W_max - 1], (2,), int)
  "agent" contains the agent's position.

  Num	Observation	Min	Max
  0	Y position of the agent	0	H_max - 1
  1	X position of the agent	0	W_max - 1

• "target"
  Box([0,0], [H_max - 1, W_max - 1], (2,), int)
  "target" contains the target's position.

  Num	Observation	Min	Max
  0	Y position of the target	0	H_max - 1
  1	X position of the target	0	W_max - 1

Rewards

Each time the agent moves, it receives a reward of -0.01. When the agent reaches the goal, it receives a reward of 1 (the game is terminated).

Starting State

When the reset function is run, a maze is randomly generated, the positions of the agent and the goal are randomly set within the maze.

Arguments

import gymnasium as gym
gym.make('gym_maze/Maze-v0', height_range=[5, 10], width_range=[5, 10])

• height_range: The boundary for the height range of mazes.
• width_range: The boundary for the width range of mazes.

It is recommended to set the maximum height and width to 20 due to the rendering setting.
The minimum height and width must be greater than 1.