An artificial intelligence able to solve a Rubik's Cube in 1,2s (and 20 movements)

Every day, an AI is able to meet a new challenge and this time, researchers have set their sights on the Rubik's Cube. 
The task seems easy for a machine since the speed of execution of the puzzle depends primarily on the number of solutions that can be in memory and the choice of the right algorithm. Nevertheless, the results obtained by this AI remain particularly impressive.

Nice score, but not a record

The current record held by a human is 3.47 seconds and the number of movements to solve the puzzle at such a speed is generally about fifty. Researchers at the University of Irvine, California, managed to solve the Rubik's Cube in just about twenty movements, and about 1.2 seconds, through the artificial intelligence DeepCubeA. 

Hold on, this is not a record for a machine! In fact, last year MIT had developed a robot capable of solving the cube in only 0.38 seconds. However, it is important to note that the robot developed by MIT was not equipped with artificial intelligence.

We do not yet understand the way of thinking of AI

The researchers began by proposing to the AI ​​a simulated version of the cube as it must be once solved. They mixed it up and the machine trained to solve the puzzle for two days, improving its techniques step by step, with more and more complex algorithms. 

According to a publication in Nature, the researchers had provided DeepCubeA with 10 million combinations and forced him to solve the puzzle in 30 moves or less. The AI ​​was then tested 1,000 times. She managed to solve the Rubik's Cube each time and managed to do it in the number of moves outsourced about 60% of the time.

Lire aussi : 
Une Intelligence Artificielle envoie 5 champions de Poker au tapis en même temps

But curiously, researchers are not yet fixated on how DeepCubeA has built its strategy. A Rubik's Cube can be mixed in billions of different combinations and, although the researchers have shown the machine what the cube should look like, the reasoning that led to this result remains to be studied. 

According to Pierre Baldi, co-author of the publication and professor at the Department of Computer Science at the University of Irvine, the algorithm could have wider applications. "  How can we create an AI that is smarter, more powerful and able to reason, understand and plan? This work represents one more step in this difficult task. "