A month ago, Demis Hassabis, the founder of DeepMind, said that the research on Go would be pleasantly surprised soon. On October 28th, Nature magazine will introduce an artificial intelligence named AlphaGo developed by Google's artificial intelligence company DeepMind in the form of a cover paper, which has defeated the European Go champion and will play against the world champion Li Shizhen in March. This program uses two deep neural networks, strategy network and value network, which greatly reduces the complexity of the search space to be considered. The former reduces the breadth of search, while the latter reduces the depth of search, much like the thinking of intuitive quick locking strategy when people play Go.

So given a little time, people like Wu Qingyuan and Li Changgao at their peak (even if they keep learning) can't do without computers? (I'm referring to a PC equipped in 20 15, not a server cluster, similar to an ordinary computer running Pocket Fritz 4. )

Today (3- 12-20 16), AlphaGo has led Lee Sedol by 3:0.

This is not too unexpected. I remember someone said ten years ago that this problem could be solved within ten years. Maybe he had some ideas by then.

Google has really contributed a lot to solving this problem. I think many seemingly equally difficult problems are not insoluble, but whether we are willing to solve them and how much energy we are willing to spend on them. I think this inspiration is very important, especially when new technologies such as deep web appear, there are many places where new breakthroughs can be made as long as they are simply applied.

To tell you the truth, I read AlphaGo's thinking, which is not much different from my previous thinking. I read a paper on using convolutional neural networks to play chess on October 20 15 65438+ 10. I felt suddenly enlightened, and I thought of an improved idea (the program in the paper actually has obvious defects, and removing the defects is a more perfect idea). The real theoretical breakthrough is the paper, its writing. Google's contribution is to improve and practice the theory better, and they are more capable of solving such problems, unlike the program in that paper that uses pure neural networks, so it is difficult to reach the top level.

It is worth reflecting on why Go, as an oriental game, doesn't solve this problem for us. I think someone in China must have seen the solution, because amateurs like me can see something.