thanks to the work of Christopher Clark and Amos Storkey at the University of Edinburgh in Scotland. These guys have applied the same machine learning techniques that have transformed face recognition algorithms to the problem of finding the next move in a game of Go. And the results leave little hope that humans will continue to dominate this game.
Pointer from Tyler Cowen. As I read the article, they have been following a strategy very similar to what I proposed six months ago.
If they are as close to success as the article indicates, then the world of Go is about to be completely upended. With Othello or Chess, most of what is knowable had already been articulated by humans by the time that computers came along. At least as far as Othello is concerned, computers did not come up with any new strategy or tactics. They just got more skillful than humans at making the best choice in close situations. With Go, my guess is that there may still be a lot left to be discovered about the game. If so, then computers will soon be in a position to make the discoveries. Even if there is nothing new to be discovered, once computers start making fewer mistakes than humans, human Go players will soon be studying computer games.
I suggest you look at the comments of the Tech Review article. In short, they are not nearly as close to success as the hyped-up article suggests. The program that is profiled is not (even close to) as strong as the state-of-the-art non-neural-network based Go programs, which are in turn not (even close to) as strong as a strong professional Go player. (For example, the greatest achievement in computer go so far is that one program beat a 9dan professional *with a four move handicap*, i.e. the computer started the game with 4 free stones on the board. While in theory the ranking system might suggest that mean the software is the equivalent of a 5dan pro, in fact I don’t think that’s really the case at the professional level. Probably the software is at the level of a very strong amateur player.)
Warning to the computers, anything that becomes a job loses all the fun.
Sam, agreed about the professional level. A professional dan is different than an amateur dan. The amateur ranks are 1d-8d, and the pro ranks are 1p-9p. The stone-per-dan rule only applies to amateur dans; higher-level players are closer than that.
It’s a crude estimate, but beating a 9p with a 4-stone handicap suggests that you are in the vicinity of 7d. That is, you can beat just about any amateur, but will be beaten by any pro.
http://senseis.xmp.net/?Shodan
Note that even 1d is rather an accomplishment for an amateur. 7d is in the stratosphere for anyone that has to also hold down a job.
The paper is much more modest than the PR article.
For one thing, the neural network played against the other Go programs when the other programs were set to stunted levels (drastically restricting their search time, for example).
The paper said their estimate for the strength of their network was 4k or 5k, based on the games with the known-stunted programs.
My take is that the neural network has demonstrated some skill, but it is quite far away from professional play. It probably has to be a hybrid with other approaches to get anywhere close.
While the post is about the strategy, my general critiques about this research are two I think. First is that it seems gimmicky.. I forget who said it bUT someone pointed out how they aren’t really playing the game like the human brain does. So to that extent it doesn’t help learn about brains. And I’m not sure how much it translates into building cars or robot surgery. Secondly, they built the pattern database using the tens of thousands of hours put in by the expert human players who produced the games populating the database. It is not clear exactly what the innovation is.