The new version of "AlphaGo" computer program can start self-learning Go from a blank state, and it is self-learning and self-learning, without any manual input of human chess. This new program is called "AlphaGo Zero", and its unbeaten record of 100-0 has dealt a severe blow to the old version of "Alpha Dog" that once wrote history.

The artificial intelligence "Alpha Dog" comes from Google's DeepMind company. As early as 20 16 and 1 June, Nature magazine reported that this program beat all other Go programs with a winning rate of 99.8% and beat the European Go champion with 5∶0.

Then in March of the same year, Alpha Dog challenged Li Shishi, the world champion of Weiqi. After five rounds of fierce fighting, human beings lost to artificial intelligence by 1∶4, which caused a great sensation. People's shock stems from the fact that on the board of Go 19× 19, there are even hundreds of moves in one move, and it is difficult for traditional artificial intelligence algorithms to estimate the situation and the next move.

But Alpha Dog is still gathering strength. In June 5438+10, a mysterious account owner announced that he was a new version of "Alpha Dog". This account once challenged Ke Jie, Park Jun-hwan, Yu Tai-shan and other top players on the well-known Go platform, and won, and then defeated Nie Weiping, the "chess saint", until the 60th set blew itself up and finally won the final.

3 days, "Afayuan" taught himself.

In the latest paper, David Silver, Demis Hassabis and their colleagues, who are in charge of the Alpha Dog Project of Deep Thinking Company in London, reported a new version of Alpha Dog software-Alpha Yuan.

The most striking thing about Afayuan lies in its self-study ability. Its learning starts from "zero" and is purely based on its own game. It does not need any guidance from the history of chess, nor does it need any inspiration from human experience.

For this new software, human input is limited to a chessboard and a set of chess pieces, and there is no human data. Afayuan only uses a neural network, which is trained to predict the program's own moves and the outcome of the game, so that Afayuan can make progress in every self-game.

After three days of training-including nearly 5 million rounds of self-play-Afayuan has been able to surpass human beings and defeat the previous version of Afayuan. David Silver said that Afa source is far more powerful than Afa source. It is no longer limited by human knowledge, but can discover new knowledge by itself.