Computers beating top human players -- good or bad?

[Mivo]

A few years ago when I learned about Go, one of the aspects that attracted me to the game was that even amateur players could easily beat computers. I actually don't know why that mattered on an "emotional" level (human pride?), but from a practical perspective it was quite nice to be able to play online without always wondering if your opponent is really a bot. Deep Blue's success and the taking off of software like Fritz and Rybka had partly killed my interest in chess.

Lately I've been playing and "studying" Backgammon, a game where computers fundamentally changed the game theory in the past 15-20 years and easily beat the best human players. In a way, I really quite like that because it means I always have a world class opponent available to me and I can learn from "the best", but on the flipside my desire to play anyone online is limited. It's trivial to cheat. So, I'm pretty torn on whether I like or dislike the playing strength of computers. It's awesome in one way, and discouraging and motivation-killing in another.

But back to Go! It's often said that playing against strong players is one of the better, if not the best, way to improve. Backgammon experts say the same, actually. The trouble with that is that most of us don't have a very strong amateur (5D+) to play frequently Go with, let alone an actual pro. So we play against other people at our own level and may progress (much?) slower than we could if we had an always-available strong sparing partner. Wouldn't it be very beneficial to have computers playing on the level of pros? Is the Backgammon analogy even applicable? Backgammon is a less complex game and you can usually figure out why a strong move is strong. It's not as transparent in Go.

We're not there yet, though. Unlike with Backgammon, computers are still weak compared to very good amateurs, and the top Go engines play a fairly artificial and not always fun style. I'm undecided whether it would be good for a SDK to play against a 5D bot, for example, though I wonder what people think about that. But Monte Carlo won't forever be the go-to technique and I think at some point (years, decades?) there may (or inevitably will) be engines utilizing neural networks and learning from their own mistakes, at which point they may catch up with humans and quite possibly revolutionize Go theory. (Tengen as the best opening!)

How do you feel about this? Would it remove some of what makes Go special? Or would it take the game to a new level? Would people improve at a faster rate than now? Good, bad, or both?

[hailthorn011]

A few years ago when I learned about Go, one of the aspects that attracted me to the game was that even amateur players could easily beat computers. I actually don't know why that mattered on an "emotional" level (human pride?), but from a practical perspective it was quite nice to be able to play online without always wondering if your opponent is really a bot. Deep Blue's success and the taking off of software like Fritz and Rybka had partly killed my interest in chess.

Lately I've been playing and "studying" Backgammon, a game where computers fundamentally changed the game theory in the past 15-20 years and easily beat the best human players. In a way, I really quite like that because it means I always have a world class opponent available to me and I can learn from "the best", but on the flipside my desire to play anyone online is limited. It's trivial to cheat. So, I'm pretty torn on whether I like or dislike the playing strength of computers. It's awesome in one way, and discouraging and motivation-killing in another.

But back to Go! It's often said that playing against strong players is one of the better, if not the best, way to improve. Backgammon experts say the same, actually. The trouble with that is that most of us don't have a very strong amateur (5D+) to play frequently Go with, let alone an actual pro. So we play against other people at our own level and may progress (much?) slower than we could if we had an always-available strong sparing partner. Wouldn't it be very beneficial to have computers playing on the level of pros? Is the Backgammon analogy even applicable? Backgammon is a less complex game and you can usually figure out why a strong move is strong. It's not as transparent in Go.

We're not there yet, though. Unlike with Backgammon, computers are still weak compared to very good amateurs, and the top Go engines play a fairly artificial and not always fun style. I'm undecided whether it would be good for a SDK to play against a 5D bot, for example, though I wonder what people think about that. But Monte Carlo won't forever be the go-to technique and I think at some point (years, decades?) there may (or inevitably will) be engines utilizing neural networks and learning from their own mistakes, at which point they may catch up with humans and quite possibly revolutionize Go theory. (Tengen as the best opening!)

How do you feel about this? Would it remove some of what makes Go special? Or would it take the game to a new level? Would people improve at a faster rate than now? Good, bad, or both?

I don't see a problem with it if bots eventually surpass humans. Sure it's a fun fact to point out that strong humans still beat the best bots when you're trying to bring in new folk to the game, but once you're already playing, it's not like you really care too much about it. And that really wasn't something that drew me to the game. When I first heard that, I thought, "cool," and moved on.

Personally speaking, I think it would be cool to be able to play a pro level bot whenever I wanted.

[palapiku]

Would it remove some of what makes Go special? Or would it take the game to a new level? Would people improve at a faster rate than now? Good, bad, or both?

yes; yes; yes; both.

[RobertJasiek]

Currently, proper high dan training by playing is possible only in East Asia; online games are not proper training: too fast, stupid rating system, not played serious enough. So programs of at least insei strength could be a great training help for other countries.

Programs beating the world top players will not remove the game's fun for me. I enjoy Go because of Go - not because of the non-existence of too strong programs.

Programs with 10p playing strength do not revolutionise go theory by themselves. For some more time, it will be humans' study of programs' play that can alter go theory. Already now, the real revolutions of go theory come from us go theory researchers; we provide reasoning and explanation where there was guesswork or intuition. Of course, also researchers will profit from stronger computers' play, which they can analyse more methodically than most players. For the researchers, computers are friends because they are an additional source of information.

[palapiku]

Already now, the real revolutions of go theory come from us go theory researchers

sigged

[SCWillson]

I think it is unnecessary to have a vastly stronger opponent to improve; two roughly equal players can elevate each other - indeed, at the higher levels of pro go players they have to because there are no better players.

As an aside, the fact a car can beat a human in a race does not diminish the relevance if running. So to eith go.

[Mike Novack]

And you have left something out?

There is still a long way to go before the computer programs challenge top human players.

The last big jump in performance of the programs was when they went from AIs to using MCTS. That was a major breakthrough and one that could not have been predicted in advance. It would not surprise me if fine tuning of MCTS programs might get them 1-2 stones stronger than they are now. It would also not surprise me if it turned out that we have already reached the limit of performance for the MCTS algorithm and that further advance will require a new breakthrough.

The strongest programs running on very powerful hardware can now win a high handicap game against a top level human player. Not even close to being able to play them even. The strongest programs running on very powerful hardware cannot even win against the stornger amateur players in an even game.

PS -- the programs aren't playing all that oddly compared to a human whose style is "loose". If you were presented with a number of games in some of which one of the players was a program and you were asked to identify I don't think you would be all that successful provided not all of the human players had a "tight" style.

[jts]

As an aside, the fact a car can beat a human in a race does not diminish the relevance if running. So to with go.

Hmm, about that...

[palapiku]

The last big jump in performance of the programs was when they went from AIs to using MCTS. That was a major breakthrough and one that could not have been predicted in advance. It would not surprise me if fine tuning of MCTS programs might get them 1-2 stones stronger than they are now. It would also not surprise me if it turned out that we have already reached the limit of performance for the MCTS algorithm and that further advance will require a new breakthrough.

When MCTS programs were introduced, they were mid-SDK, only a few stones stronger than the strongest existing bots. They have since more than doubled that gap. Zen became the first KGS 1 dan in 2009. Now we have KGS 6d bots, a tremendous improvement in just three years. And the huge jump from mid-SDK to 6d can't be explained by increased processing power alone. It just doesn't seem like go AI development has stagnated since the introduction of MCTS. Even back when Zen made 1d, people would just laugh if you told them we'd have KGS 6d bots in just three years. With this kind of history, it's hard to be pessimistic about the future of go AI.

[Mivo]

How do you guys feel about the suitability of the currently available 5D computer programs as sparing partners for DDKs and SDKs? The fun factor aside, would it be for 15, 9 or 5k player more efficient (in terms of improvement) to play against a 5D bot than to play against an evenly matched human opponent?

[oren]

How do you guys feel about the suitability of the currently available 5D computer programs as sparing partners for DDKs and SDKs? The fun factor aside, would it be for 15, 9 or 5k player more efficient (in terms of improvement) to play against a 5D bot than to play against an evenly matched human opponent?

I would say a better human opponent is best because you can get explanations and thoughts.

[jts]

How do you guys feel about the suitability of the currently available 5D computer programs as sparing partners for DDKs and SDKs? The fun factor aside, would it be for 15, 9 or 5k player more efficient (in terms of improvement) to play against a 5D bot than to play against an evenly matched human opponent?

Can I ask us to back up one step - does anyone really play for years against a bot that they have no chance of beating? I might play one or two games of chess against my laptop, but do people train to play chess against their computers? (I wouldn't be surprised either way, but given the normal psychology of beginners - they are eager for advice, guidance, etc, but very leery of playing anyone other than other beginners - I would be a bit surprised if losing every game is a sustainable study program.)

[Bill Spight]

How do you guys feel about the suitability of the currently available 5D computer programs as sparing partners for DDKs and SDKs? The fun factor aside, would it be for 15, 9 or 5k player more efficient (in terms of improvement) to play against a 5D bot than to play against an evenly matched human opponent?

Yes. In nearly all games if you want to advance, it is good to play against stronger players. They do not have to be bots, OC. Still, it is good to take the proper handicap.

[Mike Novack]

The program as opponent?

Well you need to do it properly. Any tool is useless if you don't know how to use it. Some of the objections indicate having a clear idea how to properly use this tool......

Zero chance to win --- that means you have not adjusted the either the playing strength or the handicap appropriately. You probably want these adjusted so that you are winning 10-20% of the games or 20-25% if the first is too discouraging for you. In other words, you want the program (strength setting and hnadicap) to be enough stronger than you so that you will not be learning "bad habits" of your current level* of play, so that mistakes typical of your level of play will be pubished and you learn not to make them, etc. but not so far above you that you have no chance to win the times you manage to "put it all together".

Note however that I'm not sure the MCTS programs cannot be adjusted that much if you are a beginner. Might be too eratic or too unlike a human in play. But at least some of these programs use an AI for the lower levels of playing strength.

Look, by all means, if you have available to you a human player that dregree strionger than you are willing to play lots of games against you at the indicated adjustment take advantage of that. But that is far less likely to be available to you than some suitable programs. So the real comparison should be between playing against a program under ideal strength difference to teach you vs playing against human opponents with not the right strength difference.


* You might retain bad habits typical of players a couple stones stronger than you are now, but as you improve and so the program strength/handicap adjusted accordingly, the (now) stronger program will punish these mistakes and you unlearn the bad habit. Note that this same problem exists playing human opponents.

[gowan]

I think current strong programs are very likely not to be able to give reasons for their move choices, especially the Monte Carlo method programs. All they "know" is that a certain move gives (or doesn't give) the best probability of winning. The very best way to improve is to find a really strong teacher (like a strong pro) who can adjust his/her play to the rank level of the student and is familiar with the mistakes commonly made by players at various levels. Such a teacher would play even games with the student, posing "problems" for the student at the appropriate level by creating situations in the game that require the student to practice specific technique(s). Some of this would be through deliberately making mistakes for the student to recognise and punish. And such a teacher can give reasons for moves such as "slow", "inefficient shape", "need to fight here", "inconsistent strategy", etc. With this sort of teaching the student will be far less likely to learn bad technique which would have to be unlearned later. Unlearning takes more time than learning correctly in the first place.