kyo-takano/efficientcube

Uses self-supervised learning on backward state trajectories from the goal state rather than forward exploration, enabling DNNs to solve Rubik's Cube, 15 Puzzle, and Lights Out with fewer training iterations. Includes pretrained TorchScript models and integrates beam search decoding with configurable width to balance solution optimality against computational cost. Provides standalone Jupyter notebooks for training and inference on Colab/Kaggle, plus a Python package with straightforward API for applying scrambles and retrieving solution paths.

No commits in the last 6 months.