Rubik's Conquered: DeepCubeA Gives RL a Twist

I was wondering if RL can be used solve Rubic’s cube which has 10^19 states and has required complex algorithms. Sure enough it is already done! DeepCube used MCTS+5 layer FC network trained for 44hr X 3 GPU and recently DeepCubeA does even better: https://www.nature.com/articles/s42256-019-0070-z

Discussion