Maze Extrapolation

Table of Contents

• Introduction
• Features
• Installation
• Usage
• Models

Introduction

Maze Extrapolation is a research project dedicated to experimenting with neural networks designed to extrapolate on maze-solving tasks. We test reccurent and implicit networks on out-of-distribution mazes and explore the latent dynamics of the models using topological data analysis.

Read our paper: On Logical Extrapolation for Mazes with Recurrent and Implicit Networks.

The BibTeX citation is:

@article{knutson_maze-extrapolation_2024,
  title = {On Logical Extrapolation for Mazes with Recurrent and Implicit Networks},
  author = {
    Brandon Knutson and 
    Amandin Chyba Rabeendran and 
    Michael Ivanitskiy and 
    Jordan Pettyjohn and 
    Cecilia Diniz-Behn and 
    Samy Wu Fung and 
    Daniel McKenzie
  },
  journal = {arXiv preprint arXiv:2410.03020},
  year = {2024},
  url = {https://arxiv.org/abs/2410.03020},
}

Features

• Extrapolation accuracy quantification for models trained on maze images to produce solution path images.
• Contains a model from Bansal et. al..
• Modular architecture for easy integration of new models, just inherit from BaseNet

Installation

1. Clone this repository:

   git clone [email protected]:mines-opt-ml/maze_extrapolation.git

2. Navigate to the project directory:

   cd maze_extrapolation

3. Create a virtual environment (recommended):

   python -m venv venv

4. Activate the virtual environment:

   • On Windows:

     venv\Scripts\activate

   • On macOS/Linux:

     source venv/bin/activate

5. Install the required dependencies:

   pip install -r requirements.txt

6. (Optional) To deactivate the virtual environment when you're done:

   deactivate

Usage

To evaluate the models:

• Extrapolation accuracy testing:

  python -m src.test

• Topological data analysis:

  python -m src.analyze
  

See the notebooks folder for demos.