Feature: One man's Go program looks to remake AlphaGo Zero - and beyond

BRUSSELS, April 8 (Xinhua) -- A computer program that plays the board game of Go and upsets Chinese Go professionals carries the promise beyond remaking AlphaGo Zero, which has outsmarted any human Go players.

Leela Zero, developed by 35-year-old Belgian electronics engineer Gian-Carlo Pascutto, beat in recent days Chinese professionals on popular Go platform Yike Weiqi.

"It's beginning to feel like AlphaGo," said Fu Qixuan, co-founder and CEO of Yike, referring to the program by tech powerhouse Google Deepmind.

But in the case of Leela Zero, it is the work of one man - who barely plays Go himself - plus contribution of computing power by volunteers from across the globe, compared to a concerted team effort by a resourceful Internet giant.

LEARNING FROM SCRATCH

Go is a game widely viewed as a grand challenge for artificial intelligence, and the Google Deepmind team's success was a game-changer and wake-up call in recent years.

Leela Zero is a fairly faithful reimplementation of the system described in the AlphaGo Zero paper "Mastering the Game of Go without Human Knowledge" published in the academic journal Nature, Pascutto said, adding that "for all intents and purposes, it is an open source AlphaGo Zero."

Revealed in October 2017, AlphaGo Zero was the first computer program that learns to play simply by playing games against itself, starting from completely random play. In doing so, it quickly surpassed human level of play by using a novel form of reinforcement learning, in which AlphaGo Zero becomes its own teacher.

In other words, the system starts off with a neural network that knows nothing about Go. It then plays games against itself, by combining this neural network with a powerful search algorithm. As it plays, the neural network is tuned and updated to predict moves, as well as the eventual winner of the games.

DISTRIBUTED COMPUTING

Similar to AlphaGo Zero, Leela Zero also used what is known as a Monte Carlo tree search and a deep residual convolutional neural network stack, with no human provided knowledge.

But there is still a catch. Network weights, something crucial to the improvement of the computer program's strength, has to be computed and Pascutto does not possess the computing powers of Google Deepmind.

Instead of relying solely on his own computing hardwares, which by Pascutto's first estimation would take 1,700 years to reach a level similar to AlphaGo Zero, he made Leela Zero - together with the source code - freely available online.

Starting November 10, 2017, anonymous Internet users have contributed their own computing capacity to the cause: in general about 500 people connect every day, letting their own hardwares do Leela Zero's self-training. At any given moment, the number of people connected has never been under 200.

Still, Pascutto said that lagged behind Google Deepmind's army of chips.

"To give some comparison, Alpha Zero used about 5000 special purpose chips each worth about 4 high end graphics cards. So the distributed project is probably several hundred times slower." Pascutto said.

EXPLORING POSSIBILITIES

But four months into the distributed computing, Leela Zero already began to show good results.

"I did not have any idea what to expect, but I am very happy with the current results: it has beaten strong professionals," Pascutto said.

Fu, of Go platform Yike Weiqi, agreed.

"It's a grand experiment, with Gian-Carlo Pascutto laying the foundation and Internet users' passion helping Leela Zero grow. Its openness and public participation is amazing," Fu said.

Fu's platform, which enables humans to play against each other, accommodated Leela Zero via what is known as the "Go Text Protocol" and has arranged a number of professional players against the program.

While Leela Zero has crushed many of them - including national champions of China, it also made some very amateurish mistakes in the games.

"This also happens without self-training, but obviously as the program has no human knowledge programmed in to 'fix the holes' in its knowledge, they are more pronounced," Pascutto said.

He said he explicitly chose to exclude human knowledge "because this allows seeing where the move choices of the program match those of the best humans. Any differences there should be very interesting for Go players."

WIDER OPENNESS

Working thousands of miles away from East Asia, where Go was invented and considered one of the four essential arts of the cultured aristocratic Chinese scholars in antiquity, Pascutto is a native of Ninove, roughly halfway between Brussels and Ghent.

He was actually not a Go enthusiast and last played the game more than a decade ago. But he did want to shed light on the ancient game - perhaps more than other computer programmers do.

After Google Deepmind retired AlphaGo from playing with humans, there are other computer programs available, but they haven't revealed much about their coding.

Pascutto believes his work has meaning not only in remaking AlphaGo, which he described as "maybe close enough."

"As these (other) teams don't publish what they do, we do not know about their methods. In the end, if the result is not widely available, what purpose does it serve? Deepmind's program beat the best humans, but what use is it to the Go players? They have gotten only a few games to study," he said.

"I have made everything open so others can carry on the work," Pascutto said. Enditem