New AI Tackles Math Problems That Take Millions of Steps

Researchers have created an artificially intelligent system that does just the opposite of survival at the moment. But it just doesn’t think a few steps ahead – it thinks a few million steps ahead.

A team led by mathematician Sergey Gukov from the California Institute of Technology (Caltech) has created a new type of machine-learning algorithm that has been created to solve mathematics, which requires a long series of steps. Like a Truly Long series of steps; We’re talking about a million steps or more.

Specifically, AI was able to make progress with a complex problem Andrews – Cartis estimateWhich has stumped mathematicians for decades. The estimate basically asks: Is certain math puzzle can always be resolved using the approved steps set, such as re -sorted or undoing steps?

To that end, the new Caltech program “rare and tough to find the long sequences of the steps that are” the first author of the study and a mathematician at the University of Rootgars Ali Shehper said in a Caltek StatementThe “It’s like looking for your path through the Earth -shaped labyrinth. These are the long way you have to test and there is only one way here that works. “

Has been posted in a print study Arxative Updated last August and Tuesday, Shehper and his colleagues have used their newly developed AI so how they are involved in abstract algebra to solve the families of the problems related to Andrew -Cartis estimate. Obviously, they did not solve the guess themselves. Although it may seem like anticolamic, researchers denied the possible counter -examples on the assumption. Denying the counter example does not necessarily make the original assumption true, it strengthens it.

Shehpa explained, “Denying some counter -examples confidents the validity of our original assumptions and helps us to create insights in the main problem,” shehpa explained. “It gives us a new way to think about it.” Gukov compares mathematics problems with Rubic’s cubes.

“Can you take this scrambled, complex rubik cube and get it back to its original position? You have to check these long sequences of this long step and you don’t know if you are in the right direction at the very end, “he explained.

So how does AI do that? Basically, thinking outside the box. After learning a reinforcement, researchers first trained AI with growing difficult tasks after feeding mathematics easily. “It tries to take several steps and is rewarded to solve the problems,” said Shehper. “We still encourage the program to do more with some curiosity. In the end, it develops new strategies that people are better than what they can do. This is the magic of learning reinforcement ””

The algorithm eventually learned to create long sequences of unexpected steps, which researchers called the “Super Moves”. In contrast, the output of the chatzipt is much more annoying.

“If you ask Chatzept to write a letter, it will bring something in common. It is less likely to bring something unique and extremely real. This is a good parrot, “Gukov said.” Our program is good to come with Outoliers. “

I can think of at least one external event that would be really convenient to predict any AI: Financial crash. Although the current machine learning programs have not been able to achieve this level of prognostic sophistication, researchers have assumed that their methods could one day contribute to that kind of intelligent forecast.

“Basically, our program knows how to learn,” Gukov explained. “It’s thinking out of the box.” He also added that the party “made significant improvements in a region of decades old mathematics.” What is more, Gukov and his colleagues have been given priority that does not require a lot of computing power, which makes their work accessible to other academics with smaller size computers.

Although the practical applications of this achievement may not be clear in our daily life, their work is favorable to machine-learning algorithms to solve humanity problems (not to destroy our civilization).