Gym-Env-SPaRC 0.1.0

A Gymnasium environment for SPaRC puzzles

Gym Environment for SpaRC Project

---

Description

A custom Gymnasium environment for SPaRC. The Game and Dataset was develop by: https://sparc.gipplab.org/ . This project allows reinforcement learning (RL) agents and humans to interact/play the puzzles used in SPaRC. For how the puzzles work also look up https://sparc.gipplab.org/ .

---

Arguments when creating the Gym Env:

• puzzles(pd.DataFrame) The puzzles that should be used. The puzzles must come in the shape of https://sparc.gipplab.org/ .
• traceback (bool) When set to True it allows the Agent to move back on his path, if False it does not.
• render_mode(str) Optional, If render_mode='human' the Gym Environment will visualize every step using pygame.

---

Installation and Usage

how to use:

• clone the repository
• install the dependecies
• Run Final_Product.py to play as a human or customize Final_product.py how you want to use the gym

Packages with Versions:

• gymnasium=0.28.1
• numpy=1.26.4
• pygame=2.2.0
• yaml=0.2.5
• pandas=2.2.1

---

Environment Details

Action Space

• Discrete(4): Represents the four possible moves:
  • 0: Right
  • 1: Up
  • 2: Left
  • 3: Down

Observation Space

• Dict A Dictionary of:
  • base: Dict: A dictionary of 2D arrays representing the puzzle state:
    • "visited": Tracks visited cells.
    • "gaps": Represents gaps in the grid.
    • "agent_location": Current position of the agent.
    • "target_location": Goal position.
    • Additional keys for unique properties like "stars", "triangles", etc.
    • The 2D Arrays are of the shape of the puzzle and are One-hot Encoded
  • color: list: A 2D Array representing the colors of the properties.
    • 8 possible colors are represented with 1-8
    • 2D Array is of shape of the puzzle
  • additional_info: list A 2D Array with additiional Info about the puzzle
    • 2D Array is of shape of the puzzle
    • Possible additiona_info:
    • ID of the polyshape
    • Count of the Triangles

Reward System

• Sparse Rewards:
  • Outcome Reward:
    • +1: For solving the puzzle.
    • 0: For intermediate steps.
    • -1: for Failing.
  • Normal Reward:
    • +1: For solving the puzzle.
    • -1: For Failing.
    • +0.01: For staying on a solution path on each step.

Visualization

The environment uses pygame for rendering:

• Agent: Blue square.
• Target: Red square.
• Gaps: Dark Green cells.
• Visited cells: light Green cells.
• Unique Properties:
  • Stars: colored star.
  • Square: colored square.
  • Triangles: Colored triangles with counts.
  • Polyshapes: Colored polygons.
  • Ylop: Colored Polygons with marker ylop.
  • Dots: Small black circles.

Folder Structure

• Gym-TheWitnessGame/ # Custom environment implementation
  • gymnasium_env/ # Core environment logic
    • init.py # Environment initialization
    • gym_Witness.py # Core environment logic
    • register_env.py # Environment registration
  • Final_Product.py # Main script for human interaction
  • human_play.py # Function for human play
  • README.md # Project documentation

---

Acknowlegdments

Special thanks to Lars Benedikt Kaesberg ([e-mail]) and Jan Philip Wahle for giving me the opportunity to do this Project aswell as supervising the Project.