On Wednesday (18) the makers of the computer that beat the best players in a strategy board game that requires intuition, reported that the computer is even smarter.
The scientific journal Nature reported that the current version of AlphaGo is completely self-dictated, which is an extremely important step towards the emergence of machines with superhuman abilities, that is, without any human assistance.
AlphaGo Zero won 100 matches played. The Artificial Intelligence (AI) system was christened AlphaGo Zero and it taught itself in just a few days to master the Chinese board game known as "Go". The system invented its own original moves to eclipse all the acumen humans have acquired in Go over thousands of years.
After three days of self-training, the machine was tested against AlphaGo, its forerunner which had previously beaten top human champions.
According to AlphaGo Principal Researcher Dvaid Silver "AlphaGo Zero has not only rediscovered the common patterns and openings that humans tend to play with... ".
Played with white and black stones on a board, the Chinese game is over 3.000 years old. In March 2016, AlphaGo made world headlines with their 4-1 victory over Go champion Lee Se-Dol 18 times.
Experts at the time said that Lee's defeat showed that AI was progressing faster than previously thought, who called for standards to ensure that AI always remains under human control.
Already in May of this year, the AlphaGo Master update beat the world champion Ke Jie in the three matches played. AlphaGo Zero did not learn from humans, not even by playing against them, unlike its predecessors, who trained with data from thousands of human games before practicing, say researchers at British artificial intelligence company DeepMind, which develops the system. , acquired by Google.
Silver to explain the advance said "All previous versions of AlphaGo were told: 'Well, in this position the human expert played this particular move, and in this other position the human expert played here'".
Which didn't happen with the AlphaGo Zero. It was programmed to respond to a reward, a plus point for every win versus a minus point for every loss. Starting from the Go rules alone and with no instructions, the system learned the game, devised a strategy, and improved as it competed against itself, starting with a "completely random game" to find out how the reward is earned.
The strategy used is known as "reinforcement learning", a process of trial and error. Silver and DeepMind CEO Demis Hassabis commented that "unlike its predecessors, AlphaGo Zero is no longer constrained by the limits of human knowledge".
As such, AlphaGo Zero surprisingly utilized a single machine, a "neural network" that mimics the human brain, compared to the multi-machine "brain" that Lee beat.
Satinder Singh of the University of Michigan, in a commentary also published by Nature, highlighted "The findings suggested that AI based on reinforcement learning performs better than those that rely on human knowledge."
"However, this is not the beginning of any end because AlphaGo Zero, like every other successful AI so far, is extremely limited in what it knows and what it can do compared to humans and even other animals." completed.
Anders Sandberg of the University of Oxford said, "AlphaGo Zero's ability to learn on its own can "seem frighteningly autonomous."
Sandberg told AFP, "But there is an important difference between the multipurpose intelligence that humans have and the specialized intelligence of computer software. What DeepMind has demonstrated in recent years is that it is possible to make software that can be turned into experts in different domains. , but they don't become generally intelligent".
