Despite years of research, artificial intelligence does not surpass traditional algorithms solving Rubik's cubes.

Researchers at the University of California have created artificial intelligence, DeepCubeA, capable of solving a Rubik's cube in just 1.2 seconds. And the figure is even more impressive when we know that the world speed record in this task, in the hands of Shenyang Du, stands at 3.47 seconds.

Despite years of research, artificial intelligence does not surpass traditional algorithms solving Rubik's cubes.
Despite years of research, artificial intelligence does not surpass traditional algorithms solving Rubik's cubes.


DeepCubeA was trained using reinforcement learning techniques: it competes against itself thousands of times, without previously having individual instructions, to find a way to find a solution to the problem by minimizing the cost of it.

However, the real news lies not in this difference, but in the fact that this AI system is still three times as slow as min2phase, the fastest algorithm in this field that MIT developed last year and that, oblivious to the use of neural networks, beats AI using a traditional calculation method.

So a simple Rubik's cube is enough to put in doubt that artificial intelligence is always the best option to carry out computational tasks. But how is this?

"Solving Rubik's cube brings machines closer to thinking and planning."
This three-dimensional puzzle, created in 1974 by the Hungarian inventor Erno Rubik and quickly became a bestselling toy that would sell the 350 million units, represents a complex challenge for artificial intelligence due to the particular nature of the problem it poses.

It is because there are 43.252.003.274.489.856.000 possible combinations in a Rubik's cube, but only a particular state to reach a goal (that the six faces of the cube are chromatically homogeneous).

And since it is extraordinarily complicated to reach this state merely by random movements, it becomes a difficult problem to solve optimally using machine learning, as explained by the developers of DeepCubeA in an article published this week in Nature Machine Intelligence.