Google published Nature paper cover to crack puzzles Go

First, Yann LeCun explains why we go to study? He said play baby games that as a very difficult task, Go is a good example to verify the various combinations of learning skills, including pattern recognition, problem solving and planning, but also can be used to test a new idea of ​​tools, including machine learning, combined with reasoning and planning. Yann LeCun say, a scientist at Yuandong Facebook FAIR Go a few months ago began to study the project independently, he developed a called "dark forest" of robots. According to the paper, as described in the latest version of the robot will go convolution neural network and current robot classical methods - Monte Carlo Tree Search was integrated. Previously, worked in the 2014 Tokyo Go Challenge Cup with a slight advantage by allowing the child to overcome the human players of Crazy Stone is dependent on the Monte Carlo tree search, which is essentially a set of possible moves of all the results of each step analysis system. Therefore, some machines can be very proficient in chess, chess and other games. They see farther than the human players, it is possible to easily beat them. However, Go is not the case, there are too many possibilities chess to be considered. In any one round of chess, the average may have 35 kinds of moves. But the walk was able to go up to 250 species. And after the 250 possible moves, but also corresponds to an additional 250 possible, and so on. Thus, the number of Monte Carlo search to calculate every step of the walk brings all the results is impossible. Facebook can be seen from the research results, and by combining convolution trained neural network and Monte Carlo tree search, we can match the pattern on policy evaluation function was added this new feature. This will also benefit other applications outside of the game, for example, natural language generation, in reply can add spontaneity and diversity, but also capable of reasoning, and reasoning need is searching possible answers and pick out the most Excellent logic chain. Our interest is not to want to develop the world's best chess player, but this is our first AI Progress interesting exercise. Since last November, DarkForest already beat some human players and other robots in the open Go Go Server KGS. DarkForest first edition is entirely based on convolution neural network. By supervised training mode to "imitate" human players. We use a lot of human professional video game player database, and then enter the checkerboard pattern of the game to the convolution neural networks, in order to train it to predict human players next moves. This requires large-scale neural network convolution, its input is a complete annotated with 19x19 Go board, and the neural network output is a representative of the human professional players take every step checkerboard map Probability distributions. It takes full advantage of the convolution neural network pattern recognition ability, and the success of this capability in the image of object recognition, face recognition and voice recognition has long been proved. LeCun said that 1994 will be the idea of ​​convolution neural network chess dates back a long time ago, Nicol Schraudolph and collaborators on the NIPS published a paper, the convolution neural networks and reinforcement learning to combine applied Go research questions. But it was the understanding of these technologies is not deep enough, and then the computer limits the size and complexity of the convolution can be trained neural network. Recently, the University of Toronto doctoral play baby games student Chris Maddison and Google DeepMind researchers in ICLR 2015 jointly published a paper on the article mentioned, with the game video database convolution trained neural network can have on the walk in predicting outstanding which performed. Amos Storkey Edinburgh University team, published in the same paper ICML reflects the result. Many researchers have come to believe that perhaps the depth of learning and convolution neural network can really make a difference on the Go. Amos Storkey said: "Go is driven by a variety of patterns on the board, the depth of the neural network is very good to summarize the board from a variety of modes, and therefore very suitable to play Go." This is the first edition of the dark forest prompted Yuandong published on the KGS server reason, then, a more advanced version quickly climbed to the third place on the KGS server, much better than previous studies. The ranking is better than most open source, even human players, also take several years to reach this level. And its human players like chess because it's very much like the human player moves. But at the same time it is very good campaign strategy. Local wars and strategies related to the next win is sometimes necessary to explore very specific, rather than a simple pattern recognition. Obviously, by combining convolution neural network and Monte Carlo tree search procedure can enhance the ability of strategy. In the past five years, Computer Go program through Monte Carlo Tree Search made great progress. Monte Carlo tree search is a computer chess program in the tree search method is applied to a "random" version. INRIA team of French researchers presented the first Monte Carlo tree search. After that, this method quickly spread to several of the best teams in computer Go, and become the standard method to develop top Go robot needed. The new paper released today described the latest version DarkForest and is called Dark Forest 3, which uses a combination of convolution neural network and Monte Carlo tree search. This program has been in operation on the KGS server over a month, and achieved the rank fifth adult group. This ranking means that it has become one of the best 100 players before the United States, has entered among the world's top Go robots. Interestingly, this project is funded by a small team we only spent a few months time developed, did not put any Go expert resources (of course, in addition to video games database). This is a powerful machine learning a great proof. The next wise choice is to convolution neural network and Monte Carlo Tree Search and enhance learning together, like Nicol Schraudolph of groundbreaking research. Using enhanced learning advantages is that you can let the machine battle with himself many times in a row to play the game he trained himself. This idea can be traced back to Gerry Tesauro's "NeuroGammon," a 1990s computer backgammon program that will enhance the learning neural network and combines, and defeated the game's world champion. We know that several teams worldwide are actively studying the system. Our system is still in development. Finally, at the end of Yann LeCun very profound meaning: Facebook has always been the attitude of the research is "release early, release often" in order to interpret the world of open source software popular motto. Go to our robotic system has been operating on the KGS server, our paper was published earlier in arXiv.org. We believe that when the research team quickly exchange research results with each other and with each other to promote research-based research, science will progress faster. Go armaments battle Google and Facebook Google and Facebook being carried out in a cracking algorithm chess competition. Previously, Facebook a researcher Rob Fergue think, "Go advanced artificial intelligence is the goal." He also acknowledged, Facebook move at least in a small area it is in competition with Google.play Frozen Games Google's Go Study impressive. Today, Google and Facebook use depth study to identify the network faces in the picture; the computer can recognize our voice commands; can translate one language into another; sometimes even able to understand natural human language. These techniques are dependent on the depth of the neural network. If you will be more than enough photos about trees enter into, they can learn to recognize out of a tree. If you enter more than enough dialogue, they can learn how to make some decent dialogue. If you enter more than enough chess moves, they can learn to play Go. "Go is driven by a variety of patterns on the board, the depth of the neural network is very good to summarize the board from a variety of modes, and therefore very suitable to play Go." University of Edinburgh professor Amos Storkey representation. He is using neural networks to deal with Go depth issues, like Google and Facebook did. They believe that these neural network will eventually be able to narrow the gap between machines and humans. At the next go, even the highest segment of players can not check out all the results of each step of the walk brings. They tend to make decisions based on the disk. By means of deep learning, researchers can replicate this approach. The success moves the picture input to the neural network, thereby helping to master the machine moves every successful appearance. "This approach not want to find out the best moves, but to learn the human chess style, and then effectively replicate human players." Storkey said.