Pip install pytorch and fire, then change directory to src/ and run python3 ./main.py. This trains and saves the model and later evaluates the average number of moves necessary. Command line arguments can be listed with python3 ./main.py --help (inspect the code for meaning, lol).
If training takes too long, reduce the multiplier passed to models.getModel() in main.py or decrease the number of epochs. Loss in accuracy shouldn't be too bad.
Training data is generated on the fly, thus no dataset must be supplied.
This project solves the 3x3x3 Rubik's Cube with neural networks. The goal was to find the simplest/stupidest approach that actually works. Thus no game trees nor reinforcement learning like DeepCube or DeepCubeA.
Training is done by scrambling a solved cube anywhere from 1 to 20 times, and making the neural network predict to undo the last step using the cube state. Thus, if a scrambled cube is given, it is likely to predict moves that will unscramble the cube.
-
The cube is fully scrambled.
-
The trained neural network attempts to solve the cube ("undo" a scrambling move) for 100 steps, while checking it doesn't stumble on visited states.
-
If the cube hasn't been solved, it is randomized for 20 steps (these count as moves), to avoid converging to a local attractor.
-
Go to point 2.
Neural network details:
- Input: 6 * 6 * 9 (6 neurons per tile)
- Output: 12 (only 90 deg. rotations)
- Architecture: 6 * 6 * 9 -> 8192 -> 4096 -> 1024 -> 6 * 2 (similar to DeepCube)
accuracy@step vs epochs (log plot) (even steps are omitted for clarity).
The solver has so far solved 100% of cubes, however it takes on average 4600 moves to solve a fully scrambled cube.
Rubik's cube has roughly 4x10^19 states, but the nn. sees roughly 6x10^7 states during it's training. If the actions of the nn. were fully random on unseen states, the probability of stumbling on a visited state during the 4600 steps is p<10^-8.
Thus no, it learns some useful generalizable knowledge.