(Original title: After one year of cultivation, which capabilities of AlphaGo have made rapid progress?)
The first game of the Wuzhen Weiqi Summit was over. It took four and a half hours. Ke Jie's 9th Division defeated AlphaGo and AlphaGo won a quarter of a quarter. The result of this calm wave must have been predicted by many people, but everyone from one At the beginning, we knew that the focus was not on how Ke Jie lost, but how much AlphaGo had evolved. It was still a long time before the Deep Mind team presented papers to explain how strong AlphaGo is. So we can actually watch from the scene. See some clues.
Chess speed
When the first game was played halfway down, many people in the chess group boiled. It was found that the AlphaGo's speed was very fast. The speed of each step fell between 30S and 40S. Lei Feng network AI science and technology commentary said: "(This game) is like euthanasia, lose it unconsciously, (as compared to last year) the machine is falling fast but the basic error is not."
In the overall situation, AlphaGo's lame performance is calm and decisive, and Ke Jie's fall situation is more changeable. After more than four hours of the game, AlphaGo won a quarter.
Chess
AlphaGo 1.0 time (when he played against Li Shishi in March last year), its chess game is summarized in that it is good at remembering the game + enhancing learning + computing power. You can also see the shadow of human chess wind, etc. AlphaGo disguised as Master60. During the winning streak, many chess masters, including Nie Weiping, started to evaluate their chess style. “It's difficult to speculate about Master's style... It's always expressionless, it's never affected by emotions, it's always a rhythm, it's about winning you.†And today At the end of the first game, Sogou Wang Xiaochuan already knew that AlphaGo 2.0 chess game was “completely separated from human experienceâ€. “AlphaGo 2.0 is out of the machine to imitate people, and the style of playing chess is also It will be completely out of the person's style, in the game with Ke Jie, we will continue to appear unexpected moves, and these moves in the textbooks will be considered low-level mistakes or completely unreasonable, but where a normal player is It wouldn't be this way to play, but every newcomer would be so misplaced to play this game, and AlphaGo 2.0 will continue to create such a situation. The key is he is right."
In the end, however, Ke Jie said that he used a good example in his post-meeting interview to illustrate AlphaGo’s unpredictable style of play: (when reviewing the game, he also broke the 54th hand), he said Definitely: "It's shocking that this cannot be done in human ways. After it is broken, it puts its feet on the ground and becomes thicker, with two birds with one stone."
algorithm
The essence of the algorithm used by AlphaGo 1.0 (when he played against Li Shishi in March last year), according to FB Tian Yuandong's reading of the Nature Paper, evaluating the chess game from the valuation network, choosing the strategy network, and adopting Fast rollout. Accelerate and use Monte Carlo search to connect these three parts together to form a complete system.
This time, although it was known that active forums such as forums had begun to speculate on the algorithm used by the AlphaGo 2.0 for cattle fork, with hardware miscalculations, the truth was still announced by Deep Mind personally. The visual algorithm predicts that this should be based on Its stand-alone version of 10 GPU + TPU help forecast more reliable.
Reduced hardware requirements
According to a paper published in the January 2016 issue of the DeepMind staff, the distributed version (AlphaGo Distributed) uses 1202 CPUs and 176 GPUs.
But what kind of hardware configuration AlphaGo used in this game, DeepMind said in the press conference after the meeting, this game's AlphaGo is a new version, it improved the new algorithm, the main progress is less than ten times the amount of computation, self The game is stronger. Hassabis said that during the game, the program was operating on a single machine. This was different from last year when it was distributed. This time there are more powerful algorithms that are simpler and better to work with and get faster. In addition, Hassabis said that computing power can be "acquired in the Google Cloud, using a TPU, ten processing units, less than ten times the amount of calculations." Simply put, this AlphaGo is a stand-alone version.
Iteration speed
Hassabi mentioned in a speech at the University of Cambridge in this link that it took three months to train an AlphaGo from zero and now it only takes one week (“We also optimized the performance. It used to take 3 month to train a New version of AlphaGo from scratch. Now we can do it in one week.â€).
If you say that the version during the battle against Li Shishi is V18 iterative version, disguised as Master60 winning streak and announced the completion of the upgrade is the V25 version, the current AlphaGo is V version, will not be a week when training? This is what we are going to personally ask the DeepMind team tomorrow.
From single player to team player
Delightful friends must have noticed that Ke Jie and AlphaGo had a match match, and there was a match match on the morning of May 26th. Both sides of the match - Coulee + AlphaGoV Laugh + AlphaGo, ie the two sides of the match were played by a player and AlphaGo respectively. Composition, the player cooperates with AlphaGo to compete. On the afternoon of May 26th, Chen Yaohua, Zhou Ruiyang, Yan Yan, Shi Yue, Tang Weixing and AlphaGo had a 5-on-1 team competition.
The only reason why Google dares to set up the game system is that AlphaGo now knows not only the part of the game that people play, but also the part of AI that is played by AI, and it can be integrated, so it dares to challenge the match. A highly-coordinated competition system of the companions dared to challenge the team competition to test their own "fighting power" limit. The most terrible part of AI is not how strong its individual strength is. It is that it has begun to understand and “interact†with its surroundings. Is this kind of integration more terrifying than the man-machine battle in March last year?
summary:
The last time AlphaGo was disguised as a Master test, Ke Jie had already lost to Master once. He had not seen AI Go's powerful but he did so with the calm performance of Ke Jie at the scene today and his late night emotions. We have reason to believe that this is not a more powerful player than AI Go players and human Go players. The DeepMind team must have some unanswered answers that need this game to give a positive verification and then go further in that direction. .
Outdoor Full Color LED Display P4.81 Rental,which made from die cast cabinet, it is easy maintenance. Outdoor Rental LED Display using the distribution and modular design to improve the stability of the LED Screen control system. P4.81 LED Display widely used in Outdoor Rental Projects, like Performances, Forums, Weddings, Conferences etc.
1. High refresh rate 1200Hz ,best scan method for energy saving, perfect 16bits grey level, exclusive technology on driving IC to remove ghost shadow and more pattern, good effect on photographing.
2. Thin: It is die-casting aluminum cabinet, thus it is featured by high strength, toughness, precision and it is hard to be deformed. The edge is just 33mm comparing to the 160mm of common cabinet.
3. Fast: There is fast locking device in four corners. It takes only 10 seconds to assemble one cabinet. It is fast and precise assembling.
P4.81 Outdoor Screen,P4.81 Rental Outdoor,Led Display Outdoor P4.81,Outdoor Hd Rental Led Display
Shenzhen Jongsun Electronic Technology Co., Ltd. , https://www.jongsunled.com