Alphago, as an artificial intelligence, is significantly different from IBM's deep blue supercomputer, which defeated Kasparov, the chess master in the last century, and the contemporary Apple Siri and Google now.
To understand alphago, first we need to understand what is behind alphago.
Behind it is a neural network system developed by deepmind, a British AI company acquired by Google in 2014. This system is different from dark blue. It is not a supercomputer, but a neural network system connected by many data centers as nodes, each node has multiple supercomputers Just like human brain, it is composed of 5-10 billion neurons, which is why this machine learning architecture is called neural network.
You can understand alphago as the artificial intelligence played by Johnny Middleton Depp in the transition, and the supercomputer it controls, just like the human beings controlled by artificial intelligence in the movie, serves a kind of hive mind together.
In transcendental hacker, Martin, a worker controlled by artificial intelligence. Martin doesn't think, but what he sees will be captured directly by AI
Alphago is an example developed for playing go on this neural network system. However, although its function has been described by its name, the neural network system behind alphago is suitable for any intelligent competitive sports.
The basis of this system is called convolutional neural network (CNN), which has excellent performance in large-scale image processing in the past. It is often used in artificial intelligence image recognition, such as Google's image search and Baidu's image recognition function. This also explains why alphago is based on convolutional neural network. After all, the principle of winning in go is:
The two players alternately place black and white pieces at the intersection of the board grid. After the drop, the chess piece cannot move. In the process of playing chess, people eat the seeds in the surrounding area and decide the outcome by the size of the surrounding area.
AlphaGo Logo / DeepMind
The system behind alphago also draws on a technique called deep Q-learning (dqn). The inspiration of reinforcement learning comes from the behaviorism theory in psychology, that is, how the organism, under the stimulation of the reward or punishment given by the environment, gradually forms the expectation of the stimulation and produces the habitual behavior that can obtain the maximum benefit. Not only that, alphaalphago also has a very good performance in judging the value function of the current situation and deciding the policy function of the next step, far more than the last go program Monte Carlo tree search algorithm which can match the human players.
The dqn adopted by alphago is a kind of reinforcement learning model with wide adaptability. To put it bluntly, you don't need to modify the code. You let it play go, and you let it play Super Mary and space invader on the red and white machine, and it won't be born by hand. As an artificial intelligence based on convolution neural network and using reinforcement learning model, alphago has a strong learning ability. It often starts a new project and can gain more strength than the most powerful player in the world by playing a few games.
In 2014, deepmind, which has been acquired by Google, tested the performance of artificial intelligence developed by itself with five Atari games, such as pong, brick, space invader, undersea rescuer and beam rider. The result shows that after two or three games, the control ability of neural network has far exceeded that of any known game expert in the world.
Deepmind uses the same set of artificial intelligence to test all kinds of intelligent competitive events without adjusting the code, and has made outstanding achievements, which is enough to prove how strong the learning ability of alphago sitting in front of Li Shishi today is.
Li Shishi holds the black and alphago holds the white. About 85 minutes into the rest phase
Prior to this, deepmind has carried out numerous virtual game training, as well as the experience of defeating the European go champion fan Hui Duan 2 last year, which makes alphago have trained top-level game skills, which is likely to be higher than any known go expert in the world.
The possible complexity of go
When the game has started for 40 minutes and both sides have spent about 20 minutes, the chess game has shown that Li Shishi started to attack alphago in the middle and upper position of the board, but alphago did not retreat or start a new battlefield. The most uncomfortable thing about the game between human and alphago is that alphago can't see the chess path.
Despite the changes on the chessboard, alphago and Li Shishi were not on the same line as each other long before this war. As for Li Shishi's comments on alphago and his own chess share between Erzi and Jean Xian, I'm afraid the first game is enough for him to regret.
Alphago is just a tool for deepmind to prove itself. You can also understand this confrontation with Li Shishi as Google's public relations strategy.
In 2014, the company once wrote on its official website: deepmind is committed to studying deep learning to truly understand wisdom. But for deepmind and Google, building alphago and other artificial intelligence neural networks is not the end.
Combining machine learning with neuroscience, a 'general-purpose learning algorithm' is created. Through this algorithm, deepmind and Google hope to be able to 'stereotype' intelligence and understand what intelligence is, so as to better help people understand the brain. Demis hassabis, one of deepmind's co founders, once wrote:
The best way to understand the most mysterious principle of human thinking is to extract wisdom by algorithm.
attempting to distil intelligence into an algorithmic construct may prove to be the best path to understanding some of the enduring mysteries of our minds.
Before Google acquired deepmind, one of the terms of the acquisition was that Google had to set up an AI ethics committee. Therefore, at this stage, people don't have to worry about such AI eventually killing or dominating human beings. But at least, artificial intelligence is doomed to defeat human beings in such intelligent sports as go.
As a game with a huge decision tree, go was originally suitable for human brain thinking, not for machine computing. But the direction of deepmind AI is to imitate human brain thinking and use neural network to 'reproduce' wisdom.