Google’s AlphaGo Artificial Intelligence program has outdone itself. If in 2016 he surprised us after defeating the best human players of the ancient Chinese board game ‘Go’ -in which intuition prevails to win-, now he presents himself with a completely self-taught new version: he is smarter than ever and has learned in three days what humanity has cost thousands of years and, furthermore, without any human help.
And it is that the updated version of AlphaGo is completely self-taught, an important step towards the emergence of machines that achieve superhuman capabilities “without human contributions,” according to those responsible for the program in the journal Nature.
The game of ‘Go’ is the most complex two-person challenge ever invented and Alphago Zero has mastered it in three days. He created his own groundbreaking moves to overshadow all the hard-earned knowledge for us. Ultimately, AlphaGo Zero won by 100 games to 0, defeating overwhelmingly.
“AlphaGo Zero not only rediscovered the common patterns and openings that humans tend to play … it ultimately discarded them, preferring its own variants that humans don’t even know about or play at this time,” said David Silver, Principal Investigator for Alpha Go. .
Zero’s predecessor AlphaGo was awesome, but the new self-taught version has redefined his training arsenal by completely eradicating human teachings from his upbringing. He picked up the rules of the game and, without instructions, the system learned how to play, devised a strategy, and improved as he competed against himself , starting with a “completely random game” to find out how the reward was won.
Unlike its predecessors, AlphaGo Zero “is no longer limited by the limits of human knowledge,” wrote Silver and DeepMind CEO Demis Hassabis in a blog.
In the study, the researchers expose how that surprising self-reliance has ‘sharpened’ Zero’s intelligence to the point of devastating: in 100 games there was not a single contrary victory. Not even one.
Even more surprising, that hack came after just three days of AlphaGo Zero’s self-play training, in which he distilled the equivalent of thousands of years of human knowledge of the game.
Aside from being self-taught, the team behind AlphaGo Zero attributes their mastery of Go to a unique enhanced neural network (as if it were a human brain) and more advanced training simulations.
In comparison, AlphaGo Zero had four data processing units and AlphaGo 48; AlphaGo Zero played 4.9 million training games over three days and AlphaGo 30 million games over several months.
But just because AI is advancing at such an astonishing rate doesn’t necessarily mean that Zero is smarter or more capable of humans in other fields far from this complicated board game invented more than 2,500 years ago.
“However, this is not the beginning of any end because AlphaGo Zero, like all other successful AIs so far, is extremely limited in what it knows and what it can do compared to humans and even other animals.” clarified Satinder Singh of the University of Michigan (USA).
They may not be able to do everything humans can do right now, but they can do so many other things that we can’t … welcome to the future.
Reference: Mastering the game of Go without human knowledge, Nature (2017). nature.com/articles/doi:10.1038/nature24270
Image credit: Deepmind
