Experiments with using neural nets to solve a Rubik's Cube - read README first
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
models_26_lstm
models_26_lstm_tests
models_2_full
models_2_fulltwo
models_2_lstm
models_2_lstm_tests
models_2_rnn
models_3_full
models_3_fulltwo
models_3_lstm
models_3_lstm_tests
models_3_rnn
models_4_full
models_4_fulltwo
models_4_lstm
models_4_lstm_tests
models_4_rnn
models_5_full
models_5_fulltwo
models_5_lstm
models_5_lstm_tests
models_5_rnn
models_6_full
models_6_fulltwo
models_6_lstm
models_6_lstm_tests
models_6_rnn
models_7_full
models_7_fulltwo
models_7_lstm
models_7_lstm_tests
models_7_rnn
models_8_full
models_8_fulltwo
models_8_lstm
models_8_lstm_tests
models_8_rnn
models_9_full
models_9_fulltwo
models_9_lstm
models_9_lstm_tests
models_9_rnn
models_more_data
models_streaming
test_results
.gitattributes
.gitignore
26_lstm_10
26_lstm_20
26_lstm_30
9_lstm_10
9_lstm_20
9_lstm_30
9_lstm_40
Plotter.ipynb
README.md
averager.lua
boost_supervised_test.lua
move_accuracy.lua
parse_solve_percent.py
rl_refiner_rubiks.lua
rubiks.lua
rubiks_utils.lua
simple_rnn.lua
single_model_test.sh
test_models.sh
test_rubiks.lua
train_models.sh
train_rubiks.lua

README.md

READ BEFORE CLONING

This repo is enormous, several GB big. I decided to save all my models from all my experiments, some of which took up quite a bit of space because I didn't even try for space efficiency. Cloning this will take a long time and a lot of space. I'm not saying you can't do it, just know what you're getting into.

This is the code I'm using for a research side project. Basically, I'm throwing neural nets at a Rubik's Cube to see how well the neural net can learn the dynamics of a Rubik's Cube problem. Past research has shown the Rubik's Cube is solvable with iteratively deepened A* search, so this isn't about doing something believed to be difficult. I'm more interested in finding out how well a neural net can recreate/imitate search.

If I have time I'll update this README. This is highly unrefined research code and as of now I have no plans to clean it up or make it more readable to other people. Off the top of my head, you'll need Torch and the rnn library for Torch. (There's code for parallel data set generation that needs the parallel library, but turns out dataset generation is fast enough to make it a premature optimization, and you shouldn't need it for training.)