I think this is the best bet to crush a professional. the only caveat is that professionals learn as well, so although it is possible to train against an old version of some AI to find its flaws, a professional will most likely patch up their previous mistakes. Unless we find a way to train AI in a way where they don’t have to play games, but instead study games, I doubt that this is feasible.
1 Like
Go is a deterministic, perfect information game, no mind reading is necessary. For every possible board position and side to move, there is a singular “perfect move” (or possibly a number of moves that are equally perfect.) A perfect player would alway make that move. By definition, nothing can do better than this hypothetical perfect player.
It seems that what’s being asked here is whether a perfect player can kill every single enemy stone on a 19x19 game verses current human pros. We know it can on 5x5 (if it’s black.) From my understanding, 7x7 Go has been weekly solved and it has been established that even perfect play cannot eliminate all enemy stones (given an equally perfect opponent.) Komi, on 7x7 was established at 7 points (Chinese) if I recall.
3 Likes
Perfect player will do move that has 100% chance of winning. But it assumes that opponent is perfect too. So it will not do self atari for example because it knows that perfect opponent will capture.
But, if perfect player has time machine and knows that opponent is not perfect, it may play self atari because it will be 100% sure that opponent will not capture. So end result with more points becomes possible.
(perfect player+time machine) vs human will win by more points EACH TIME than (perfect player) vs human
So I created definition of more cool “perfect player”
therefore
…what I am saying is correct
Well, it is not a “perfect player” as they play worse moves to get a higher score. Against an opponent who would know how to counter it, they might very well lose the game. The premise is that they can fool their opponent into reacting in a suboptimal way to what is effectively an overplay.
1 Like
Nope, this theory doesn’t work. Using a time machine once, maybe. But constantly using a time machine, over and over again, in a short period of time, seems utterly unfeasible 
. And that is exactly what would be required.
Because the moment the perfect player using the time machine played a move that they did not the first time around, every move after that, by both players, would be different. Thereby requiring an additional use of the time machine to be certain of what was going to happen next.
I can’t imagine anybody could use a time machine over and over again. Why? Look at science fiction and take your pic from a multitude of possibilities. A more sound example would be a perfect player that possessed foresight, clairvoyance, precognition, telepathy, or any other ability that would allow them to successfully anticipate a players next move or line of thinking 
- List item
2 Likes
Against different opponent it will play differently. If opponent is perfect player, (perfect player+time machine) will play EQUAL to (perfect player)
perfect player bruteforced all possible go games to see game tree to victory. Why not also bruteforce all possible time lines to get BIGGER game tree to victory against certain human?
You cannot design a machine that does this… Mind reading and time travelling is a good metaphor to explain your idea, but it is not a practical tool that we could implement in AI, as far as I know.
2 Likes
that was explanation of TRUE ceiling.
And way to do something like this is
I hope you realize that is an assumption. For all we know, professionals are close enough to perfect play that there aren’t 50 points to win.
even Ke Jie lost on 2 stones handicap already,
but my idea isn’t about winrate. Its about winning by N points no matter how low winrate is.
Problem of current bots is that in simulation they play against equally strong bot. They need to simulate play against weaker bot instead. But weak Zero bot is still far from perfect simulation of human. So better to train human simulation part on human games, and Play Part as strong as possible.
Also, imagine if bot was trained on 4 stones handicap ONLY. Maybe even Ke Kie will lose.
Then it’s 361 points (area scoring). The pro just has to have an exceptional bad day (The chance that this will happen is >0 )
But on the other hand, if the winrate doesn’t matter, it’s more about how bad a pro can be if they have a bad day and not how “good” the opponent is.
2 Likes
margin of victory can be regulated until its 50% winrate. Question is how big it may be now and 100 years in future.
A standard game of Go has only two* outcomes: either black wins or white wins.
With only two distinct outcomes, it is easier to reason about perfect play and preferences over potentially random outcomes, since the distribution boils down to one number: the probability of winning.
Note (about “randomness”): While, of course, Go is inherently deterministic, randomness can be introduced into the outcome of the game through the uncertainty/unpredictability of your opponent’s behavior. For example, judging the likelihoods of whether the opponent would fall for one trick play or another.
If we talk about hypothetical perfect play toward “maximizing the margin of victory”, then we are really dealing with a considerably different game (even though, of course, it shares obvious structural similarities). Such a game is different in that it has many more distinct outcomes, e.g., B+0.5, W+0.5, B+1.5, W+1.5, B+2.5, W+2.5, … etc. Now, even formulating preferences over outcomes is tricky, since one must consider preferences between distributions over a set of many outcomes. Maximization of the expected score** is only just one possible way of considering preferences, and it does not capture all of possible the nuances of different risk vs reward behavior/desires.
Randomness in how the opponent may behave in pesky, but unavoidable. Any proposal to consider telepathy or time travel to remove that uncertainty is even more fundamentally and unfairly changing the game.
*Okay, okay, we could have a draw as third outcome if use integer komi. Okay, okay, okay, let’s assume that we’re not using Japanese rules (to eliminate the possibility of other pesky outcomes like “no result” or “both players lose”).
**The expected score should not be confused with or assumed to be a proxy for the expected utility. While a utility function should exist (if we accept the von Neumann-Morgenstern axioms, it is not necessarily simply the score, and would instead depend on the risk-aversion/appetite of the players.
2 Likes
i think AI is already pretty darn strong
2 stones vs pros online and AI is still winning like 95%+
on fox server, i often see fineart play moves that seem “normal” yet in the end just own pros left and right , it’s like those moves had some magic touch to them that turn normal moves into godlike moves
in the end, it’s like fineart plays random tenuki, then all of a sudden all the moves add up getting sente and lots of free forcing moves, utilizing aji to the max, dominating the whole game
often when pro wins a game is because those pro somehow manage a kill
i don’t think this applies, because fineart is utterly destroying pros on fox thru attacks, sente, pretty much all rounders, sure they might win by points in the end, but the process of winning was because everything adds up
perhaps maybe they play much safer, much more for points in an even game, but when giving handicap stones, they utterly destroy them, mentally and physically
1 Like
Answering the spirit of the original question asked, I suspect there’s a point in the game where Pro players know that the AI has “won” and is no longer playing moves which aren’t completely safe “win” moves. You can see this in some commentary, when the Pro player notices the AI playing safe moves rather than more edge moves or provocative moves.
If you can watch the commentary and see how the AI plays (or follow along a game and see when the computer feel’s it’s ahead by 90-95%) then you can assume it would be at this point that it could start to gain significant advantage by playing bolder, less safe moves to gain more territory than required simply to win.
In essence, the AI switching to “teaching mode” when it’s secured the win and a Pro player can sense this. I think you can find your answer in some of the commentary from AlphaGo games where the analysis mentions the AI is now playing very safe moves.
2 Likes
Until now, almost every FineArt game has been 1 1/2 handicap. (2 stones and 7.5 komi) Yesterday a different version of FineArt started playing 2 stone games. Maybe we’ll see the handicap raise for those Fox training games and begin to slowly answer this question.
3 Likes
2 stones handicap for who ? I.A start with 2 stone ahead or its the true player who need a little advantage ?
Returning this thread a little later, it’s still a very open and interesting question.
Over the last year the field of superhuman bots has become very crowded: FineArt, DeepZen, Golaxy, ELF, KataGo, Leela Zero etc. And, ofc, we have no idea how any of these compare to AlphaGo Zero.
I don’t follow bots much, but it seems to me that it’s not even clear how far the AI have already come, since professionals have too much pride to play games with four or five handicap games. It’s a shame that Go originated in “face-saving cultures” since I’m sure that plays a role – in contrast, I’m sure Carlsen wouldn’t be afraid to play against Stockfish on Twitch at odds of rook and knight if it meant a laugh.
I’d suggest that AI have reached Human+3 level at this point, maybe Human+4. In terms of handicap stones, it is obvious that there is a finite amount of progress that can be made. Winning percentage against a fixed opponent, eg. AlphaGo Master, will also eventually top out at 100%.
1 Like
Are you sure?
4 Likes
To cut this point out of the textwall:
I think the only way to measure the strength of an AI is to measure its winning percentage in a certain set (let’s say 1,000 games) against a designated opponent of fixed ability. That opponent is in turn chosen by having a winning percentage of 90 – 95% over the bot below it in the chain, and so on down to the bottom.
-
Bot #4 (winning 800 of 1,000 games against Bot #2) should be considered as stronger than Bot #3 (winning 700 of 1,000 games against Bot #2).
-
We know that Bot #2 is reliable because it won 950 of 1,000 games against Bot #1.
-
And we know Bot #1 is reliable because we tested it first against professionals.