Imagine a fireball moving through a plane. Depending on which zone it is in, it has a different behavior: at first it moves freely, when it changes its zone it seems to be subject to gravity and if it continues moving it behaves as if it were attached to a spring. A physicist could study his movements and discover what physical laws are influencing her. But artificial intelligence can also do it. Two MIT researchers have designed 40 planes (like the one that is running our ball) governed by different physical laws, and have put an algorithm to work: its objective is to discover what are the laws that govern each of the worlds.
To achieve this, Tailin Wu and Max Tegmark, the two responsible for the investigation, have educated the algorithm to follow the same mental processes that physicists usually use to understand the world. They have given him instructions to comply with some basic rules, the same ones that led to Isaac Newton to describe gravity as a force or Galileo Galilei to identify the oscillation of a pendulum. Galileo observed at length a lamp that swung in the cathedral of Pisa and clocked the oscillation with its pulsations. He concluded that the period of oscillation was constant and independent of its amplitude. But to make this discovery, Galileo had to ignore the other details present in the cathedral: the resistance of the air, the temperature, the light.
Is it possible to design an artificial intelligence system that develops theories as Galileo did, paying attention only to the necessary information? This was precisely the goal of AI Physicist, the intelligent system designed by MIT that copies the model of the Italian physicist. So far, artificial intelligence systems have discovered patterns in information and have even derived certain laws of physics, according to MIT Technology. But in these cases, the AI always studied a set of data in its entirety.
To think like a physicist, the algorithm has learned to develop theories that describe a small part of the data set. He has also learned to apply the principle of the Ockham Razor, which gives priority to the simplest explanations. Another objective pursued by physicists is look for ways to unify theories and it is something that also applies to the algorithm. Finally, one of the aspects that separates machines from humans is their capacity to reuse what has already been learned. So the AI Physicist of Wu and Tegmark remembers the solutions learned and tries to use them with future problems.
Using these four basic rules, the algorithm has been pretty good at identifying when the fireball was under the influence of gravity or when it was moving freely. That is to say, he has been able to identify which physical law was acting on him at each moment in 90% of the scenarios. "The machine they have designed is capable of detecting the behavior of the ball by designing" theories ", so it is very effective in predicting how the ball will move, even though it is somewhat complex", explains Juan MR Parrondo , professor of the department of structure of the matter, thermal and electronic physics of the UCM.
- Your limitations despite success
Does it mean then that the next great discovery of physics will come from the hand of an artificial intelligence? It seems that the machines they are far from being the new Albert Einstein or the new Marie Curie. "The machine works very well because the problem is specially designed for the algorithm to be very effective." In addition, Carlos Díaz-Guerra, associate professor of the department of materials physics of the UCM, points out a detail that must not be forgotten and is cautious. He warns that this experiment is not yet published in any scientific journal, "which would mean that it has not yet passed the review process by other scientists prior to its publication."
Even so, Parrondo does not doubt the potential of this technology beyond this study: "I think it can be very useful to find patterns in experiments that produce a large number of data, such as in particle accelerator collisions. In fact, AI is already used partially to analyze data of that type, "explains the physicist. Ensures that these strategies can help find new particles in an experiment, but not new laws or new ways of describing physical reality. "I think the advance this article represents is more an advance for AI itself, than for AI applied to physics." The capacity of these artificial intelligence systems is far from that of humans like Galileo.