Remember "go master" AlphaGo? This time it was very thorough

AI Li Shishi Wuzhen games

youxishiguangVGTIME· 2017-10-19 13:48:22

also remember the artificial intelligence AlphaGo developed by Google's DeepMind team?

its fame in March 2016 with the go master Li Shishi fight, let the world know its name; followed by the name of "Master", in January 2017 on the network in 30 seconds Kuaiqi game winning streak global dozens of chess master, was ranked first in the world including kija; then in May 2017 held in Wuzhen on the game kija, to play AlphaGo but had lost all three matches.

AlphaGo and Li Shishi on

so far, once considered the human could not be overcome by artificial intelligence with weiqi AlphaGo, this project has been standing on the peak, it is duguqiubai.

, but it failed anyway, and it was a 100:0 fiasco.

who is going to make it impossible to win a game?

the answer is DeepMind's next generation of artificial intelligence: AlphaGo Zero.

" published today in the journal "nature", the DeepMind team introduced they beat Li Shishi AlphaGo in 2016, because it is the first one in the history of the project on the go to beat the human machine, there is no doubt that this is an artificial intelligence technology new milepost. And after that, they developed a new generation of artificial intelligence: AlphaGo Zero. If

, the generation of AlphaGo is the analysis of current human left tens of thousands of games and game to go through the learning rules and the winning way, then the use of reinforcement learning technology (reinforcement learning) AlphaGo Zero is more simple, it only needs to know the rules of the game, you can begin to self learning. Demis Hassabis, CEO of

DeepMind, said, "the way it's learning is to play with itself, starting with random games that don't have any rules. In this process, it quickly surpassed the human level, and eventually defeated the previous generation of AlphaGo mentioned in the paper with 100:0's success. "

" is more powerful than the previous generation of AlphaGo because he is not limited to human knowledge. It can learn from scratch, until it beats the strongest go player in the world: AlphaGo himself. "

AlphaGo "Zero self reinforcement learning

after 72 hours, but millions of times of self playing after AlphaGo Zero only use 4 TPU beat 48 TPU AlphaGo Lee, also mentioned above is the 2016 defeat of Li Shishi's version, at this point it against AlphaGo Master scores as an overwhelming 89:11, Master is mentioned above beat kija version; and after 40 days of self play, it ultimately to score 100 straight victories beyond the AlphaGo Master.

in addition, the researchers also found that due to no longer limited to the human existing data and game theory, AlphaGo Zero game very strange, often people never have the variant of chess method. Just for a few dozen days, it not only mastered the technique of go which has been accumulated for hundreds of years, but also had a new understanding and development.

Demis Hassabis says: "the strongest version of the current, AlphaGo Zero shows us that even without the use of human data, we can make great progress even with fewer computing resources.". Of course, our ultimate goal is to use these technologies to solve real-world problems, such as protein folding or new material design. If we can make the same progress on these issues, it will improve everyone's lives and make great progress in human knowledge. "The once unimaginable future of

seems to be a step closer to us, and now it's curious to know what the next field of artificial intelligence will be

source: Caixin, gizmodo

The lastest articles of youxishiguangVGTIME