Tuesday , June 15 2021

A neural network learned to solve Rubik's cube

We've talked about that already the cube of rubik, it's a fascination for mathematicians, for their complexity, for the ability to study this kind of three-dimensional hobby.

Mathematicians have already shown the properties of the cube, the complexity and the way they are solved. The interesting thing, however, is that youA neural network has learned to solve the cube by himself.

Technology challenge

Researchers from California University, Irving, have made an intelligent system that can, on average, resolve the cube, in 1.2 seconds and use about 20 movements.

This is 2 seconds faster than man holding the world record, with 3.47 seconds. It can be said that normal people can solve it, know the basic algorithms, in about 50 seconds.

But the neural network, called DeepCubeA, It doesn't have the world record in & # 39; the resolution of & # 39; a cube between machines.

Other researchers showed up last year a robot that could solve the cube in 0.38 seconds, using the algorithm min2phase, from MIT, that's just three times the algorithm used by the neural network.

DeepCubeA learned through himself

While other algorithms were specifically designed to solve the cube, DeepCubeA had to work out its own solution, Your own way.

More interesting is the fact that the researchers are not exactly sure how DeepCubeA realized how to make it a color on each of & # 39; s faces.

Over there several million possible combinations for the cube, but only a complete state.

And while the researchers told him what the definitive state of the cube was, DeepCubeA found for itself how to reach this solution and today it is not fully understood how to develop its strategy.

One of the reasons for this may be a neural network is finally a mathematical model, symbolic, that is expressed by comparison. Moving from there to the conceptual part now seems to be a problem in neural networks.

Personal training

The researchers, for this study, she started with a simulated version of Rubik's cube and stirred it.

DeepCubeA trained him, feed it It took two dayss, improve their skills in an attempt to solve the most difficult combinations.

According to an article published in Nature, the researchers gave DeepCubeA 10 million combinations and asked him to fix it in 30 than fewer moves.

De IA was tested with about a thousand combinations. The system succeeded every time in solving the cube and did so in the minimum stipulated number of moves in 60% of & # 39; e try.

The algorithm can solve other games like the one "Game of 15" and Sokoban.

DeepCubeA uses an artificial neural network that emulates what human neurons do, with teaching techniques, where To detect and teach learned patterns with input given by humans.

It assumed an enhanced learning mechanism, which means that what i learned was raising the state of problems from a target state without having any knowledge about the domain.

Researchers They had previously published an article with another hobby approach, called Approximate interaction policy.

It has to be noticed DeepCubeA's neural network was not specifically designed to solve Rubik's cubes, The algorithm has broad implications:

"How can we create advanced AI that is more robust and capable of reasoning, understanding and planning?" Pierre Baldi, professor of computer science and senior author of the study, said: "This work is a step towards this much more complex goal".

Source link