Life In 19x19

Something that can run on moderate software and can beat all professionals. Sure we have this new zen but it's not as strong as it's chess counterpart and it still requires incredible hardware.

Since DeepBlue it took around ~10 years for a good desktop chess engines to appear and only now ~20 years later can we truly say that we have engines that can no longer be defeated by any grandmaster.

I simply can't wait for something similar to lichess with go engine in the browser itself, which can analyze your games quickly.

I find it always exciting when I see that during analysis of some game even dan players can get into complete disagreement about a specific position / continuation. With a "stockfish like" Go engine such discussions may become more boring in future.

From my previous time in chess I remember that there were certain players who were very fond of their installed engines and let them automatically analyze all of their tournament games. However, after all, this did not help them becoming stronger.

Nevertheless the topic of strong engines on moderate hardware is of course highly interesting.

I've been playing a little bit of chess recently, and followed the world championship. It seems like there was room for the spectators to look beyond just accepting what the engine told them, if only because the engines did not always have super-decisive recommendations.

As far as I know, computer programs like Alphgo and Zen don't have any concept of go strategy and no principles that they follow. Even Stockfish can't reasonably be claimed to play perfect chess, nor can Alphago or Zen be claimed to play perfect go. It makes sense to me that a strong human player might have a better grasp of strategy or principles than that exhibited by Alphago or Zen. Probably Alphago or Zen can say it considers a certain move the best one in the position but it can't explain why. Its reason is because it says so. Humans play go not just to win but to understand and enjoy. Therefore a human commentator can be very valuable to humans playing go. I think it's the case that human commentators of chess games are still popular; they haven't been thrown on the junk pile.

Krama wrote:Something that can run on moderate software and can beat all professionals. Sure we have this new zen but it's not as strong as it's chess counterpart and it still requires incredible hardware.

I would say currently about an order of magnitude. But that's using MY definition of "moderate hardware" << I assume you meant hardware, not software >>

Look like no one directly answer your question yet
I think we'll reach Rybka level (sorry for mentioning this engine) in 2018 (could be Zenith Go 8) and StockFish level within 2020 (never lose to any human even with unlimited undo chances, can give an exciting handicaps game to top pro).
Let's say my definition of moderate hardware is $2000 desktop buy in that year. (But personally I think it's too expensive even for videogaming)
A little bit too optimistic, but we're living in an acceleration age.

stockfish?

DrStraw wrote:stockfish?

Yes, stockfish.

I think the value of AI development is not how strongly they can play games like chess or go but rather how can they contribute to the good of our planet. Alphago's developers are thinking about other applications of their methods. It is a mistake to be concerned only about how well software plays games.

Regarding the value of human commentators, I think that humans still play a significant role, even in chess.

A good move for a computer program is not necessarily a good move for a human player. There is the argument that a "good move" is a good move no matter what. But on the other hand, a move is considered "good" by a computer because of its own mechanisms of evaluating positions. In other words, a move is "good" for a computer because the computer is able to use that move effectively.

A human looking at the move may see that the computer recommends it, but has no idea of the strategy behind that move. And in fact, the move may only be subtly good, and bad in many cases if you don't know how to follow it up or respond to various responses by the opponent. In some ways, it's similar to playing out a joseki that you don't understand.

Joseki is cool, but if you don't get the reasons behind the moves, you can get into trouble pretty quickly.

To summarize my opinion: a suboptimal move played with an understanding of its meaning is superior to an optimal move that you don't understand.

The same phenomenon already existed before computers were good at go: pro games. You can review a pro game without commentary, and you have several examples of good moves. Using software like Kombilo you can analyze a database of go games and try to find patterns. Yes, it's pretty educational. But without understanding why a pro played a certain way, or the subtleties of their strategy, just copying the moves is not very good.

This is where the role of commentators come in. Human players can explain meaning behind moves and give insight into strategy.

I think strong computer programs can provide educational analysis features, and it will be a nice supplement to study materials we already have (e.g. GoGoD + Kombilo). Pro game databases and strong programs can give us some idea of how a "good player" would play in a given position. That's great.

But the strategy and meaning behind it all remains in the domain of human players.

hyperpape wrote:

DrStraw wrote:stockfish?

Yes, stockfish.

I also had to google wikipede (but had suspected something like this). Here’s the missing link

Zen is quite strong enough for analyzing games in its current commercial version.

Kim Myungwan (9p) said in the AGA stream Zen would be a could training tool for him.
If it is good enough for Myungwan, it is good enough for me

And it works indeed! After years of stagnation I improved my play from 5k to 2k in a few weeks when I started to review my lost games with Zen.

Memorizing pro games every day and starting a tsumego routine may have helped as well

And we should not treat simply being given a "best move" without explanation as valueless. True, it is much easier if the variations that would follow, both this move and the (perhaps better at first glance) alternatives that are being rejected are all laid out for us.

But most of the time, just seeing that move which we didn't even consider is enough. As evidence of this, consider a game where we do have access to the thinking of the human players (say the 2 vs 2 game of "The Go Consultants"). Several times in the game one side or the other made a move that the other team hadn't considered, but in all those cases WHY that overlooked move was good was obvious to other team as soon as they sw it.

This is really a matter of how many levels difference in strength are involved. Learning from just seeing the better move is going to work a LOT better if just a few levels of strength involved rather than many levels difference.

For what it's worth we go players have had many many games with good moves to study, namely uncommented pro games. And how many times to we see writing on these forums that they find playing through uncommented pro games not useful because they can't understand the moves. In our own games it is more useful to have comments such as

"This move is good because the thickness it creates outweighs the territory it gives your opponent"; or

"White didn't play that move because it loses sente 10 moves later"; or

"You lost that game because your response to your opponent's amashi strategy ended with your having amarigatachi"; etc.,

rather than "Black N7 was bad" or "White M15 was an excellent move"

gowan wrote:For what it's worth we go players have had many many games with good moves to study, namely uncommented pro games. And how many times to we see writing on these forums that they find playing through uncommented pro games not useful because they can't understand the moves. In our own games it is more useful to have comments such as

"This move is good because the thickness it creates outweighs the territory it gives your opponent"; or

"White didn't play that move because it loses sente 10 moves later"; or

"You lost that game because your response to your opponent's amashi strategy ended with your having amarigatachi"; etc.,

rather than "Black N7 was bad" or "White M15 was an excellent move"

This is true, but there's a big difference between seeing a game filled with good moves and having someone point out that a particular move that I made is a poor choice. I suspect that when I review my own games the losing move often often occurs several moves before the move I am able to identify as bad. It would be invaluable to have the same strong player tell me where I went wrong every single game; I'm fairly certain I could use that feedback to improve my play.

I could get something like this kind of feedback with Crazystone right now (truthfully, it would probably be strong enough at my level), but the pace of development in Go playing programs is so fast that I'm waiting for the dust to settle before buying something. I'm particularly hopeful there will be an application with a native Linux client released, though I may not be willing to wait quite that long. I suspect I'll need to upgrade my hardware before using one of the new deep learning programs, too. I don't have a video card that can take advantage of the recent developments right now.