Reinforcement Learning for Architectural Space Planning

This project implements a reinforcement learning-based approach to architectural space planning. The system uses various RL algorithms (Value Iteration, Policy Iteration, and Deep RL) to optimize building layouts based on specified constraints and objectives.

Project Structure

.
├── src/
│   ├── environment.py    # Custom gym-like environment for space planning
│   ├── agent.py         # Implementation of RL agents
│   ├── train.py         # Training script and experiment management
│   └── visualize.py     # Visualization utilities (TODO)
├── results/             # Directory for experiment results
├── main.tex            # LaTeX documentation
├── requirements.txt    # Project dependencies
└── README.md          # This file

Setup

1. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

2. Install dependencies:

pip install -r requirements.txt

Usage

Training

To train the agents and run experiments:

python src/train.py

This will:

• Initialize the environment and agents
• Run training experiments for all configured algorithms
• Save results and generate visualization plots in the results/ directory

Configuration

The training configuration can be modified in src/train.py. Key parameters include:

• Environment parameters:

  • Grid size
  • Maximum steps per episode
  • Room size constraints

• Algorithm parameters:

  • Discount factor (gamma)
  • Convergence threshold (theta)
  • Neural network architecture (for Deep RL)
  • Number of training episodes

Algorithms

The project implements three different approaches:

1. Value Iteration

   • Suitable for smaller state spaces
   • Guarantees optimal policy
   • Faster convergence for simple layouts

2. Policy Iteration

   • Better performance on larger state spaces
   • More efficient policy updates
   • Good balance of exploration/exploitation

3. Deep RL (Actor-Critic)

   • Handles continuous state/action spaces
   • Scales better to complex layouts
   • More flexible reward structures

Environment

The environment (ArchitecturalEnvironment) implements a custom gym-like interface with:

• State space: Grid representation of the building layout
• Action space: Room creation, modification, and removal
• Reward function: Based on space utilization and design constraints

Results

Training results are saved in the results/ directory, including:

• Training statistics (JSON format)
• Performance comparison plots
• Training time comparisons
• Final layout visualizations

Contributing

1. Fork the repository
2. Create a feature branch
3. Commit your changes
4. Push to the branch
5. Create a Pull Request

License

MIT License - see LICENSE file for details