Approaches to cube solving
Before the aforementioned paper was published, there were two major directions for solving the Rubik's Cube:
- By using group theory, it is possible to significantly reduce the state space to be checked. One of the most popular solutions using this approach is Kociemba's algorithm (https://en.wikipedia.org/wiki/Optimal_solutions_for_Rubik%27s_Cube#Kociemba's_algorithm).
- By using brute force search accompanied by manually crafted heuristics, we can direct the search in the most promising direction. A vivid example of this is Korf's algorithm (https://en.wikipedia.org/wiki/Optimal_solutions_for_Rubik%27s_Cube#Korf's_algorithm), which uses A* search with a large database of patterns to cut out bad directions.
The paper introduced a third approach: by training the neural network (NN) on lots of randomly shuffled cubes, it is possible to get the policy that will show us the direction to take toward the solved state. The...
