In Jin Yong's novel "The Legend of the Condor Heroes", Zhou Botong "draws the left hand and draws the right hand", the left hand attacks the right hand, the right hand fights in time, entertains himself, and is invincible in the world.
In the real world, there is also such a "young boy". He has never seen a game, nor has he got a person to point out. From scratch, he entertained himself and learned himself. After only 40 days, he used to dominate the martial arts. .
This "young child", called AlphaGo Zero, defeated the same-handed "tea brother" of Alpha Go Master, the first master of humanity, at the Wuzhen Go Summit in May this year. However, this brother who read almost all of the human game and defeated the first master of humanity by 3 to 0 was defeated by the "teacher" Alpha Yuan from the 21st day of self-study.
On October 19th, the Google DeepMind team, which created the AlphaGo mythology, published the Mastering the game of Go without human knowledge in Nature magazine, and introduced the latest research results of the team, the birth of Alpha, which caused a sensation in the industry.
Although the teacher is out of the same door, but the teacher's housekeeping skills are fundamentally different.
“In the past, all versions of AlphaGo began with training with human data. They were told how the human master was in this place and how to do it in another place.†Dr. David Silver, head of the DeepMind Alpha Dog project, introduced in an interview. "And Alpha does not use any human data, it is entirely self-learning, practicing from self-game."
Dr. David Silver said that in the algorithms they designed, the opponents of Alpha, or sparring, were always tuned to the same level. “So it starts at the most basic level, starting from scratch, starting with random moves, but at every step of the learning process, its opponents are just calibrated to the current level of the match. At the beginning, these opponents are very weak. But then it is getting stronger and stronger."
This kind of learning is one of the most popular research areas of artificial intelligence today - Reinforcement learning.
Dr. Li Wei, a professor of electronics and computer engineering at Duke University in Kunshan and Duke University in the United States, told Ms. (News ) that the new intensive learning method used by the DeepMind team is from a nerve that has no knowledge of Go. The network starts and then combines with a powerful search algorithm. “Simple explanation is that it starts to don’t know what to do, just try it. After trying it, you see the result. If it is a positive result, you know it’s right, and vice versa. I know that I am doing something wrong. This is how it learns by itself."
In this process, the Alpha element becomes its own "teacher", and the neural network is constantly adjusted and updated to evaluate the next position and win or lose. The updated neural network is recombined with the search algorithm to create a new one. A powerful version, however, repeating this process again, system performance is improved by each iteration, making neural network predictions more and more accurate, and alpha elements are becoming more and more powerful.
It is worth mentioning that the previous version of the Alpha dog usually uses two neural networks that predict the next "policy network" and evaluate the "value network" of the game. The more powerful alpha element uses only one neural network, which is an integrated version of the two networks.
In this sense, "AlphaGo Zero" is translated into "Alpha Yuan" instead of the literal "Alpha Zero". "The connotation is richer and represents the starting point of human cognition - neurons." Professor Li Wei said.
The above research updates people's perceptions of machine learning. "People generally think that machine learning is about big data and massive calculations, but with alpha elements, we find that algorithms are more important than so-called calculations or data availability," Dr. David Silver said.
Professor Li Wei has long been focusing on manufacturing big data research. He believes that the most significant point of this research is that it proves that artificial intelligence may be able to get rid of dependence on human experience and assistance in some areas. "One of the major difficulties of artificial intelligence is that it requires a lot of manpower to label data samples, while the alpha element proves that artificial intelligence can solve problems through unsupervised data, which is human unlabeled data. â€
Some people imagine that similar deep reinforcement learning algorithms may be more easily applied to other areas where humans lack understanding or lack large amounts of labeled data.
However, there is much practical significance and which real-world areas can be applied. Professor Li Wei said that “there is still a future.†“The game of Go is itself a more limited application. Humans think that Go is very complicated, but not for machines. Difficult. Moreover, playing Go is just a form of entertainment, not counting the actual problems people encounter in life."
So why does Google's AI choose Go?
According to the "First Financial News" report, in history, the first classic game that the computer first mastered was the tic-tac-toe game, which was a research project of a doctoral student in 1952. Then in 1994, the computer program Chinook successfully challenged the game of western checkers. Three years later, the IBM Deep Blue supercomputer defeated world champion Gary Kasparov in the chess game.
In addition to board games, IBM's Watson system successfully challenged the old quiz show Jeopardy game in 2011. In 2014, Google's own algorithm, learned to play dozens of Atari games with only the initial pixel information.
But one game is still human beings representing the top level, and that is Go.
Dr. Demis Hassabis, founder and CEO of Google DeepMind, explained in AlphaGo against Li Shishi in 2016. Go, with more than 3,000 years of history, is the most complicated game ever invented by humans. For artificial intelligence, this is the most The big challenges at the cutting edge require intuition and calculation. To be proficient in playing Go, you need to combine pattern recognition and planning.
"Go's search space is boundless - one Gorg larger than the Go board (the order of magnitude, 10 to the 100th power, even more than the number of atoms in the universe)." Therefore, the traditional artificial intelligence method That is, "building a search tree for all possible steps" is almost impossible to achieve in a Go game.
The key to defeating the human AlphaGo system is to compress Go's huge search space into a controllable range. Dr. David Silver has previously stated that the role of the strategy network is to predict the next step and to narrow down the search to the most likely steps. Another neural network, "valuenetwork," is used to reduce the depth of the search tree, estimating the winner of the game every step of the way, rather than searching for all the ways to end the game.
Professor Li Wei expressed his delight at the breakthrough brought by Alpha, but he also mentioned that "Alpha is only proved in the game of Go, unsupervised learning is better than supervised learning." ', but it does not prove that this is the 'optimal' method, perhaps the combination of the two semi-supervised learning, that is, at different times and stages, combined with the advantages of supervised or unsupervised learning, can get better results. â€
Professor Li Wei said that the technology of artificial intelligence is far from what people think. "For example, the reCAPTCHA verification code (image or text) used in Internet login cannot be automatically recognized by machine learning algorithms," he said. In some ways, robots do do better than people, but they are not completely replaceable at the moment. “Only when research proves that an artificial intelligence technology can solve some practical problems and artificial pain points, it really counts as a major breakthrough.â€
Dr. Denis Simon, executive vice president of Kunshan Duke University and a Sino-US science and technology policy and relationship expert, said in an interview with the news that the success of Alpha in the field of Go shows that it has great potential. Alpha has achieved its own ability by playing against itself. Every time it becomes smarter, every game is more challenging. This repetitive, fully participatory learning enhances the ability of alpha to deal with higher levels of strategic and complex issues. But the downside is that this is a closed system. “How can the Alpha element achieve further growth beyond its own limitations? In other words, can it think outside the box?â€
The main pain points in obtaining customers in foreign trade are fierce market competition, high customer acquisition costs, low customer stickiness, and single marketing methods. With the changes in the market environment, traditional marketing methods have been unable to meet customer needs. Enterprises need innovative marketing methods and technical means to improve customer stickiness and loyalty, reduce customer acquisition costs, in order to obtain more business opportunities in the fierce market competition.
koko
YIWU BEAUTYPLUS ART NAIL CO., LTD. , https://www.cn-gangdao.com