AI Learns To Solve Rubik's Cube - Fast!

The latest neural network to impress is DeepCubeA from Forest Agostinelli, Stephen McAleer, Alexander Shmakov and Pierre Baldi of the University of California, Irvine. This is a deep neural network that learns a range of combinatorial puzzles - sliding block15, 24, 35, 48 puzzles, Lights Out, Sokoban and, of course, Rubik's cube. The network learns a reinforcment value function, but it does this "backwards". That is, it starts from a solution and randomly takes moves away from the goal. As it steps away from the goal, the moves and configurations become increasingly low in value, - i.e. they are moving away from the goal.