Will humans ever beat AlphaGo Master?

[negapesuo]

Or even Fine Arts, or any other AI that humans are currently incapable of getting even a single win against. I'd be nice to know that humans as a whole are getting better at the game, instead of simply going through fads of different play styles and joseki.

[ez4u]

The computations (reading) required are beyond us in real time. So even if we are able to make sense of what the bots are revealing, we won't be able to beat them head to head in the future. Indeed it will only get worse.

[Bill Spight]

Of course we will get better. It is not just that the bots are better at reading -- in fact, humans can sometimes read better than they do -- it is that they have better judgement. That's something that humans can learn.

OC, the bots will be getting better, too, as Dave points out.

[Knotwilg]

There may come a time where we need a definition for "human".

[Kirby]

At least today, I prefer to think of bots as tools and not as rivals.

Maybe same way that I'm not interested in picking a fight with a chainsaw.

[hyperpape]

The computations (reading) required are beyond us in real time. So even if we are able to make sense of what the bots are revealing, we won't be able to beat them head to head in the future. Indeed it will only get worse.

You're certainly right that the bots will improve faster than we humans. I took the question to be: if we freeze the bot's skill, will the best humans someday surpass it? That's a very interesting question.

[sorin]

If the question is "will humans beat a particular version of an AI that they currently have no chance against, like AlphaGo Master 2017": I think so, yes, if they can experiment a lot with it, top pros will eventually find weaknesses to exploit and be able to win.
However, that is not a very interesting thing to do. A more interesting question is "will humans reach a level where top pros can win against an AI playing at the AlphaGo Master 2017 level" (that is, not one particular program, but any program playing roughly at that level".

I would say that as long as human keep getting stronger, there are good chances that this will happen. This seems obvious, however the rate at which humans improve may be such that it can take a long time (or "forever") for that to happen.

Which brings an even more interesting question: do humans improve at Go, do newer generations play better than older ones? Intuitively the answer should be "yes" (since humans are good at learning from previous generations and improving things), but I cannot thing of any clear proof of that for Go. Increasing Elo ratings show an increase in strength, but that can be just an artifact of the Elo math, not reflect a real increase over time.
The fact that players from newer generations beat older generations in direct matches also seems to suggest an overall increase in strength, but that can be explained by players getting weaker as they age (not technically weaker, but less motivation to win, and less stamina to concentrate fully).

I hope someone has some good argument for answering "yes" to the last question

[jlt]

It should be possible to use a strong bot to evaluate the level of a player, and then compare Shusaku with modern top pros. If we find (say) a difference of 1.5 stones, then we can expect that humans will improve by the same order of magnitude in the future.

(Of course I don't know if this estimate is realistic.)

[Bill Spight]

Which brings an even more interesting question: do humans improve at Go, do newer generations play better than older ones? Intuitively the answer should be "yes" (since humans are good at learning from previous generations and improving things), but I cannot thing of any clear proof of that for Go.

Some years ago I addressed that question by looking at changes in joseki over time, from around 1700 to the present (pre-AI). Based on that, and a rough evaluation function, I made a guesstimate that top level play (but not that of players who stood out above the rest, such as Dosaku or Go Seigen) improved at about ½ stone per century. With AI, I would not be surprised if human pros improved by 2 stones in this century.

[sorin]

Some years ago I addressed that question by looking at changes in joseki over time, from around 1700 to the present (pre-AI). Based on that, and a rough evaluation function, I made a guesstimate that top level play (but not that of players who stood out above the rest, such as Dosaku or Go Seigen) improved at about ½ stone per century. With AI, I would not be surprised if human pros improved by 2 stones in this century.

Very interesting!! What is the assumption made, how do changes in joseki relate to increase in strength? How to tell between joseki change due to a fad, vs a real change in understanding?

[Bill Spight]

Some years ago I addressed that question by looking at changes in joseki over time, from around 1700 to the present (pre-AI). Based on that, and a rough evaluation function, I made a guesstimate that top level play (but not that of players who stood out above the rest, such as Dosaku or Go Seigen) improved at about ½ stone per century. With AI, I would not be surprised if human pros improved by 2 stones in this century.

Very interesting!! What is the assumption made, how do changes in joseki relate to increase in strength? How to tell between joseki change due to a fad, vs a real change in understanding?

I'm kind of busy now, but I realize that I may have left an incorrect impression. So just a quick response, for now. Tenuki played a large part in my thinking, so I am not just talking about joseki but fuseki, as well. When to tenuki? Generally when there is a larger play elsewhere.

When I was starting to play go, I always answered a pincer. I assumed that my reply was larger than the pincer. But in general, plays get smaller in an area as more stones are played, unless they are played randomly or riskily. But, as I now know, top players in the 17th century typically did not answer a pincer, even a one space pincer. That's how they started many games: 3-4, 5-3, pincer, tenuki. Over time they abandoned the pincer in favor of a play in an open corner. How much did they gain by doing so? Around 1 pt., eh? If their play improved by 1 pt., on average, every 15 years or so, they improved around ½ stone per century. (This is not the only piece of evidence I base my guess on.)

Now, AlphaGo pincered about half as often as human pros. So it seems like humans are still pincering too often. Just by imitating AlphaGo they could pick up maybe 1 pt. in 5 years. Not just from not pincering so often, OC, but in a number of ways. That's more than one stone per century.

[John Fairbairn]

... top players in the 17th century typically did not answer a pincer, even a one space pincer. That's how they started many games: 3-4, 5-3, pincer, tenuki. Over time they abandoned the pincer in favor of a play in an open corner...

Now, AlphaGo pincered about half as often as human pros. So it seems like humans are still pincering too often.

Bill, as some Meijinal support for your thesis, you may like to consider this (Black has just played 7 at the triangle):



This opening first appeared when Genjo and Chitoku played it as teenagers in 1797. In almost every appearance of it after that one or other of them was involved. There was clearly something troubling about the opening because there were 12 occurrences but with EIGHT different choices for White 8.

Most notably, Genjo and Chitoku used it in their last two games together, so that in all they played it against each other three times (Chitoku as White in the first instance, Genjo as White in the last two when they were at their peak).

The tight pincer at White A was the first choice (Chitoku's) and it became the most popular move (four cases).

But more than a decade on, Genjo chose the much "weaker" pincer at B (i.e. I am assuming he felt considerably less of a pincering effect). The three years later he eschewed the pincer altogether and played kosumi at C.

For reference, my LZ is suggesting just three very closely matched candidates and not one is a pincer. Significantly, I think, in addition to the press at D5 and the approach at P17, it is suggesting F17, very similar to Genjo's final idea. C-G were of course playing without komi so we can expect some slight differences anyway. Choosing Genjo's move actually reduces the win ratio (55%) by much less than a percentage point.

[gowan]

Following Bill's heuristic analysis it seems that humans might reach AlphaGo level in two or three hundred years If the goal is to be stronger than all existing AI players I think it will never happen. Just as humans can improve, at whatever speed, it would seem that AI will also improve, strange as that might seem. I see no reason why the neural net learning AI approach could never be bettered. And what about future development of quantum computing?

[sorin]

top players in the 17th century typically did not answer a pincer, even a one space pincer. That's how they started many games: 3-4, 5-3, pincer, tenuki. Over time they abandoned the pincer in favor of a play in an open corner. How much did they gain by doing so? Around 1 pt., eh? If their play improved by 1 pt., on average, every 15 years or so, they improved around ½ stone per century. (This is not the only piece of evidence I base my guess on.)

Now, AlphaGo pincered about half as often as human pros. So it seems like humans are still pincering too often. Just by imitating AlphaGo they could pick up maybe 1 pt. in 5 years. Not just from not pincering so often, OC, but in a number of ways. That's more than one stone per century.

I guess pincers are just a special case of a more general group of situations where humans are likely to misjudge, which is "fighting situations". Maybe it has to do with human psychology, such as "fighting spirit", "don't lose face", "make sense of previous moves", "better die than play that", etc.

AIs have no psychology, which is a big strength.

[Bill Spight]

AIs have no psychology, which is a big strength.

I'm with Arthur C. Clarke, who patented the geosynchronous satellite many years before one was actually put into orbit, on this one. Sufficiently powerful AIs have a psychology, we just don't know what it is. It's alien in each case, but not necessarily dissimilar to human psychology. Current AIs do not explain themselves, so we have to study their psychology like we do other non-human psychology.

I'm sorry, Dave.