This project is not finished yet.

TODO:

1. Implement saving and restoring trained neural networks [done]
2. Implement logging [done]
3. Actually train the neural network using cloud computing

Solving the Rubik's Cube Without Human Knowledge

https://arxiv.org/abs/1805.07470

In this article Stephen McAleer, Forest Agostinelli, Alexander Shmakov, Pierre Baldi describe how they approached a problem of solving a Rubik's Cube by a computer without supervision. General idea was to apply methods from reinforcement learning and use neural networks as functions approximating value of a given state as well as decision functions what move to make. Main challenge to overcome was to account for the fact that randomly doing moves would not result in a solved cube even after a long time. That's why authors trained these networks using what they have named "Autodidactic iteration", that is starting from simple positions (cube being only a few moves away from solved) and moving to more complicated cases when the network is already trained.