Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 18·5 min read
Solvable Sokoban Without a Solver via Diffusion
TL;DR
A transformer-based discrete diffusion model was developed to solve Sokoban puzzles without traditional solvers, achieving a 77.4% solvability rate. Remarkably, 94.5% of unsolvable puzzles could be made solvable by removing just one wall.
✦ Why It Matters
Explore the model's capabilities to enhance puzzle generation in your projects today.
Key Takeaways
- Utilize the transformer-based model for generating Sokoban puzzles.
- Experiment with wall removal strategies to increase solvability.
- Access the publicly available model and instructions for your own use.
- Consider non-local interactions in puzzle design for better outcomes.
How It Works
The model employs a masked diffusion approach, where it learns to predict the placement of tiles based on any subset of the board rather than a fixed order. This flexibility allows it to effectively manage the complex interactions inherent in Sokoban puzzles, where a decision in one area can significantly impact the entire grid.