Re evolution of alpha dog
4hw.org: Google's AI go alpha dog battle with Ke Jie won an overwhelming victory by 3:0 in 2008, and won the first place in the world in the go world. Just after people thought that the AI dog was as invincible as a God, the Google team developed an upgraded version of Alfa yuan, and beat the alpha dog by 100:0, which caused users to panic about AI, AI Can it really replace the imperfect human?
On October 18, local time, the deepmind team of Google's artificial intelligence team published a paper in the journal Nature, announcing that a new version of alphago zero computer program can quickly self-study go without any human input, and defeat the previous generation alphago with a record of 100-0. This is another milestone in the history of AI.
The new version of "alpha go" has been king for 40 days from scratch
Google's AI 'alpha go' stands at the top of go after several world-renowned man-machine wars -- in 2016, Google's deepmind team released alphago, and beat the world go champion and Korean chess player Li Shishi with a 4:1 record, shocking the world. Since then, alphago has evolved the alphago master version, and defeated the world's first go player and Chinese chess player Ke Jie 3:0. After success, the deepmind team announced alphago's retirement.
Now it has surpassed itself in a new way: after a year's closing, deepmind launched the latest version of alphago zero. Without any human guidance, it learned go from scratch, and in only three days of self-study and training, it defeated alphago, which was 100-0. It is worth noting that the previous versions all refer to and learn a lot of go experience of human professional players, while the new version of alphago zero is self playing, that is, learning by playing chess with different versions of oneself.
The UK's deep thinking company has developed 'alpha go', and its team has published a paper in the new issue of nature, introducing the latest version of 'alpha go - zero' When the company's previously developed computers began to learn go, the old alpha go needed to learn millions of human chess scores first, and after several months of intensive training, so as to conduct self-training and achieve transcendence.
The fourth version, the latest "alpha go zero", has got rid of this limitation and completely "started from scratch". It plays with itself and makes rapid progress through better algorithms.
The learning principle of zero is as follows: the research team didn't learn any human chess manual for zero in advance, only told it the rules of go, let it play chess on the board, play chess with itself, learn lessons from experiments and failures, explore the rules, and improve its chess skills in the actual battle.
In addition to the learning principle, another difference from the "old dog" is that zero uses a single neural network.
Previous versions of alphago used two kinds of neural networks, one is used to predict the best move of the next move, and the other is used to calculate who is more likely to win based on these moves.
Zero combines the two networks and only lets the neural network predict the winner, so that it can get more efficient training and evaluation, just like a go expert can predict the result of the game.
In addition, zero no longer uses the fast and random walk method. The old dog can predict which player will win from the current situation by fast walking, while zero uses high-quality neural network to evaluate the game.
In this way, "old dog" is like a chess player who looks at three steps step by step, while zero is more like a go master who looks at the whole situation and knows what to do.
Three days after starting to learn go, 'alpha go zero' defeated 'alpha go Lee' 100-0; 40 days later, it defeated 'alpha go master', which is beyond the reach of all human masters. Researchers believe that this progress marks a great breakthrough in artificial intelligence, which means that artificial intelligence can better enter the field which is blank for it.
According to the paper published by deepmind team in nature, the biggest breakthrough of the latest version of alphago zero is to realize the 'whiteboard theory'. According to the whiteboard theory, a baby is a whiteboard, which can improve intelligence through learning and training. Turing, the pioneer of artificial intelligence (AI), believes that as long as you can make an AI similar to a child with a machine and then train it, you can get an AI similar to, or even beyond, human intelligence. Alphago zero, a self-taught man, has realized this theory.
What's more, with the deepening of training, zero not only greatly improves its chess skills, but also is no longer limited by human cognition. It can discover new knowledge, independently discover the rules of the game, and develop new strategies. These creative moves of go imitate and completely surpass the new skills alphago uses when playing with Li Shishi and Ke Jie.
Hassabis, CEO of deep thinking, said he hoped that this progress in artificial intelligence could be used to analyze protein structure, design new materials and other fields, bringing positive and beneficial impact on people's lives.
According to David Silver, alphago's lead researcher, according to the guardian: 'it's more powerful than previous algorithms because it doesn't use any human data or human experience. We have eliminated the limitations of human knowledge, and it can create knowledge by itself. '