An artificial intelligence (AI) program that made history by defeating a human champion in the ancient Chinese game of Go has risen to new heights by beating its former self.
Google's AlphaGo Zero stormed ahead of human play levels to wipe out the program's famous earlier version, winning 100 games and losing none.
It took three days of self-play training to overcome the AlphaGo incarnation that shocked the Go community by seeing off 18-time world champion Lee Sedol in 2016.
After 40 days, AlphaGo Zero grew even more powerful until it outperformed "Master", another updated version of the program which has beaten some of the world's best players including current number one Ke Jie.
Demis Hassabis, co-founder and chief executive of DeepMind, Google's UK-based artificial intelligence company, said: "It's amazing to see just how far AlphaGo has come in only two years.
"AlphaGo Zero is now the strongest version of our program and shows how much progress we can make even with less computing power and zero use of human data.
"Ultimately we want to harness algorithmic breakthroughs like this to help solve all sorts of pressing real world problems, like protein folding or designing new materials.
"If we can make the same progress on these problems that we have with AlphaGo, it has the potential to drive forward human understanding and positively impact all of our lives."
Go is a fiendishly difficult strategy game invented in ancient China more than 2,500 years ago.
Players move black or white pieces called "stones" on a board marked out as a grid. The object is to gain control of more territory than your opponent.
There are far more potential moves than in chess and Go is said to be the most complex board game ever devised, making it an ideal testing ground for artificial intelligence.
AlphaGo Zero uses a novel form of reinforcement learning in which it becomes its own teacher.
The system starts off with a neural network that knows nothing about the game of Go. It then plays games against itself, by combining the neural network with a powerful search algorithm.
As it plays, the neural network is tuned and updated to predict moves and likely winners.
In just a few days, and almost five million games of self-play, AlphaGo Zero could outperform both humans and all previous versions of the program.
As it trained, it independently discovered game principles that took humans thousands of years to work out, and also developed new strategies, the team reported in the journal Nature.
Why are you making commenting on The National only available to subscribers?
We know there are thousands of National readers who want to debate, argue and go back and forth in the comments section of our stories. We’ve got the most informed readers in Scotland, asking each other the big questions about the future of our country.
Unfortunately, though, these important debates are being spoiled by a vocal minority of trolls who aren’t really interested in the issues, try to derail the conversations, register under fake names, and post vile abuse.
So that’s why we’ve decided to make the ability to comment only available to our paying subscribers. That way, all the trolls who post abuse on our website will have to pay if they want to join the debate – and risk a permanent ban from the account that they subscribe with.
The conversation will go back to what it should be about – people who care passionately about the issues, but disagree constructively on what we should do about them. Let’s get that debate started!
Callum Baird, Editor of The National
Comments: Our rules
We want our comments to be a lively and valuable part of our community - a place where readers can debate and engage with the most important local issues. The ability to comment on our stories is a privilege, not a right, however, and that privilege may be withdrawn if it is abused or misused.
Please report any comments that break our rules.
Read the rules here