The paper's results
The final result published in the paper is quite impressive. After 44 hours of training on a machine with three GPUs, the network learned how to solve cubes at the same level as (and sometimes better than) human-crafted solvers. The final model has been compared against the two solvers described earlier: the Kociemba two-stage solver and Korf. The method proposed in the paper is named DeepCube.
To compare efficiency, 640 randomly scrambled cubes were used in all the methods. The depth of the scramble was 1,000 moves. The time limit for the solution was an hour and both the DeepCube and Kociemba solvers were able to solve all of the cubes within the limit. The Kociemba solver is very fast, and its median solution time is just one second, but due to the hardcoded rules implemented in the method, its solutions are not always the shortest ones.
The DeepCube method took much more time, with the median time being about 10 minutes, but it was able to match the...