Now that the dust has settled after AlphaGo's victory over Yi Se-tol, we are seeing some clear strands of thought emerge. Before these strands thicken too much, it may be worth questioning some of the assumptions behind this new thinking.

First, can AlphaGo teach us new things about the fuseki? It would certainly be foolish to discount the possibility. But we've been here before with Shin Fuseki. Within weeks of the first successful trials of that, even top pros were experimenting with it, and becoming ever more outlandish. But the fad faded within a couple of years and nowadays it's hard to pinpoint anything that Shin Fuseki taught us, except that it's probably OK to play weird moves. So in that narrow sense it deepened the magic of go, Kirby, but it still didn't tell us how to cast any spells. Since AlphaGo can't tell us yet what proportion of toad's liver and how many eyes of newts we need to add to our potions, it too may have deepened the magic of go but none of us is going to qualify to get into Hogwarts because of it.

Second, maybe Yi Se-tol was the wrong choice. Perhaps he was from the Slytherin house and we needed someone from Gryffindor?

For some of us the magic of go started to fade not with the advent of comuter go but once time limits were slashed and and an unending steam of teeny-boppers on a sugar rush were able to occupy the heights of go for five minutes at a time. This was the rise of Slytherins.

To me it was never any coincidence that Yi Se-tol's rise (and Yi Ch'ang-ho's decline, and Japan's in general) coincided with shortened time limits. I have already mentioned elsewhere Wang Xi's view that Korean go became characterised by reliance on assessment of probabilities (whereas Japan is characterised by soba go). That's another way of saying they rely on intuition, or instinct, and now we have Yi Se-tol himself confirming that that is not only how he has been playing but that is also how he lost to AlphaGo.

So can we argue that a Gryffindor kind of player would have done better? Surely we can entertain the possibility. If we simplify a little and assume that AlphaGo has two components - instinct, or neural networks, based on seeing millions of patterns and excellent calculation based on Monte-Carlo statistics - was Yi Se-tol the strongest choice of opponent? We know from chess that the single greatest strength of computers is that they don't make mistakes in calculation. Unlike humans, they don't have brain farts. The shorter the time limits in chess, the more brain farts humans have - and computers still have none. So why should go be any different? At least if you choose Slytherins to represent humankind.

But actually go is different from chess. Chess is almost 100% calculation. Every game for a chess computer is therefore a home game. But the element of calculation in go is perhaps much less. One of the main ways in which calculation does work well in go is in the endgame, and we have already seen that go computers use Monte Carlo algorithms to great effect here, though without precision - they just make moves that statistcally ensure a win by 0.5 point. The possible margin of victory becomes irrelevant. yet even here we know, from O Meien's book on the endgame, that humans are just as capable of doing something very similar. That is, a pro like Yi Ch'ang-ho can, almost algorithmically, calculate the outcome of a game from the early boundary-play stage onwards within a small margin of error. The only problem is that this method takes a bit of time - too much, apparently for games with modern time limits.

We can also surmise that most top pros can match AlphaGo for instinct nearly all the time, but lack the mechanical reliability of computers. Wouldn't more thinking time increase the reliability of pros? Furthermore, would more time not bring into play possible advantages for the pros: the ability to think rationally instead of relying on instinct/experience, or the ability to be (occasionally) even more precise? Or, would soba-go work better than probability go? After all, the premise of soba-go is that you are always making moves that are demonstrably never bad. They don't lead to quick victories, spectacular attacks or even clear initiatives. But they don't lead to quick losses either. It's just that soba-go too needs time.

How much time? I think we can find a good approximation from the games of Go Seigen. I don't imagine many people would dispute that he was a truly outstanding player. One of his characteristics was that he played much, much faster than his Japanese opponents. Now Go, being from China, was a half-blood prince, with a touch of muggle in him. That meant that, unlike his Japanese peers he didn't feel the need to show respect by playing slower than his teachers. He just played at the speed he felt was natural to find the right answers on the go board. Since he had plenty of time available to start with - 13 hours typically - he never had to stretch the elastic in his knickers too much. We can deduce therefore that the optimum time limits for a player of his level and particular abilities (a Gryffindor?) was somewhere between 5 and 8 hours, but probably much closer to 5 hours.

Would longer time limits benefit AlphaGo, too? I'm not certain that the answer is an automatic yes. Or that, if there is any improvement, that it would be noticeable at pro levels. I find it easy to believe that less time for the computer with more time for a pro would make a big difference in the pros favour, but more time for both may well still favour the pro. I think this view may even be supported by chess. There, computers are strong enough now to give grandmasters a handicap, but provided the human has a decent time limit you can't yet say that the gap in go terms has stretched to more than 2 or 3 stones. In blitz chess, of course, it's a waste of time (that is the magic word) for a human even to get the board and pieces out.

On top of all that, chess computers haven't really come up with anything useful for teaching humans how to play better in general terms. So we can't yet expect go computers to do us any favours either.

Before we roll over and submit completely, I'd like to see a Gryffindor player pitted against a computer - SobaGo vs AlphaGo - at decent time limits (6 hours each?). Of course that is just another way of saying I really just have questions rather than answers. But one thing I feel I do know already is that, for me at least, AlphaGo has not destroyed the good magic in go. AlphaGo has simply confirmed that the real Voldemort was Mickey Mouse.

Set against that though is the report that for his (not)Jubango with Gu Li, Lee Sedol wanted longer time limits than Gu Li did. In the end Gu agreed to 4 hours each. Maybe Lee wanted even more but Gu certainly wanted less originally and they settled on 4. From the games I think we can see Lee Sedol benefited more from the extra time, particularly with reversals in the endgame. So whilst Lee's aggressive fighting style and balancing of probabilities might serve him well in faster games, for serious games he does like more time. Did I read something about him wishing he had more time against AlphaGo?


I think he was the right choice from a DeepMind PR perspective as the pre-eminent player of the last decade. As for the human with the best chance to beat AlphaGo, maybe not, but we'd really need more games against AlphaGo to judge that. JF, who do you see as a Gryffindor player? I think Lee Changho in his prime ~15 years ago could well do better against AlphaGo, but who of current players? Does Gryffindor = soba? What top players today play Soba-style, maybe Yoda Norimoto, but he's rather old? I recall him negotiating well against Park Junghwan recently but he didn't have the stamina to win. The worry I would have with soba style is what the professional judges as a market price exchange is actually a loss of 0.1 points or whatever (I know you don't like numbers, I just mean a very slight loss), which against humans is a worthy approach of playing a marathon game and looking to exploit slight mistakes of the opponent later, but I fear AlphaGo will just accumulate those slight advantages, not make mistakes later, and win as it appears to have excellent positional judgement.

From your teeny-bopper (and other!) comments I take it you are not a fan of Ke Jie, who seems to be the most popular choice of who should face AlphaGo next (including the CEO of MLily, sponsor of the (not)Jubango). But some of his victories against Lee Sedol showed a rather mature willingness to trade and settle for adequate results, a little soba-like perhaps?


But the only win Lee got was because of a brain fart by the computer, a mistake in reading in a complicated tactical situation, far below the usual high quality of its play.

Is this actually true? Maybe it is for "nearly all the time", but those several times a game that AlphaGo played brilliant unexpected moves are the key moves that made it win. Lee's comments seem to say it played moves he didn't expect, his intuition let him down. For example that invasion of the right side in game 1 (but some pros, such as Zhou Ruiyang, did see it).

I do certainly agree that more time will benefit humans more than computers: if you halve the time a computer has but give it a twice as powerful machine it's basically the same strength, but you can't give a player more brains for parallel processing to achieve the same in less time. A 6 hour game would certainly be interesting, but against who? Maybe a team of pros?

Against AlphaGo, Fan Hui won 2 of his 5 informal games with shorter time setting than his formal games, but I agree that more time will favor pro more than AlphaGo. I think the keyword to win against AlphaGo is to set your mind to 'informal mode'. But that would also irrevalant as AlphaGo will improve further, just like computer chess.

Probably. There would need to be a go AI version packaged up to do "analytics", rather than a competition version. See below.



Possibly: depends how you think the match might have been "improved". To the extent that Yi was the "obvious" choice 15 months ago, it isn't clear to me why he shouldn't have been asked.



I forget where I read, some time ago, the opinion that quick go between pros has the function of discovering "who is the stronger player". In any case, there is probably some validity in equating "strong" there with a sporting assessment, not one relating to go as a traditional art. Before concluding that AI go is lacking in artistry, we really ought though to allow the training cycle to take its course. AlphaGo is still a wunderkind, and its go must be taken as forced rhubarb, rather than a mature product of any sort.



I recall he said he spent much of his time counting: finding the good play wasn't what took time. Otake thinks 3 hours each is good, to take an "intuitive" pro who sets high standards also in underlying quality of play.



Actually, how to do that is my own hobbyhorse, rather than trying to defend the basic territory of human potential as players. It's a problem with several facets, for go. For backgammon, which has been in the position of AI superiority for two decades now, it is clear that AI is the research tool top players need to improve. Go is more complex in terms of taxonomy of positions, by perhaps three orders of magnitude.

But I think we won't know whether we have been "duped", until serious efforts have been made to address the question John raises here. For example, building a new joseki tool powered by AI assessment. One early application, for me, of the GoGoD database was the chance to look differently at "modular" go (in particular 19x10 sides). What the backgammon players are able to do is to check plays against "loss of equity", and pick out systematic errors in thinking. There is at least a chance we may be looking at a new era of go study, likewise.

John, what you refer to as "Mickey Mouse" Go was, indeed, magical to me. But I see that you can take this opportunity to, again, belittle the quality of modern pro games.

Of course, I have no way of knowing how AlphaGo would play on a game with longer time limits. But I suspect that "Mickey Mouse" Go fared as well as 6-hour time limit Go would have for team human.

FWIW, Lee Sedol has mentioned regrets in his play, and seems to believe that he could beat that version of AlphaGo given another chance to play it. And he didn't seem to be concerned about any sort of "Mickey Mouse" time limits

_________________
be immersed

When I first saw the announcement I thought they could've picked a better player for the match-up, but I didn't base my decision on any sort of deep insight, just intuition and a vague feeling that Yi-Se-Tol's go is the sort of Go that a computer would do very well against. While humorous in nature, John's post makes some good points and I too would like to see a match with longer time limits and perhaps a less brittle opponent.

We can't really fault Yi-Se-Tol too much since he is one of the strongest players ever, but it would be interesting to get some more info and for alphago to play some more matches with different players and different (longer) time limits. It was undoubtedly a publicity stunt from google, as a big company like that never does something unless it has some benefit to it. I remain optimistic that strong AI can eventually help us find good moves and review games in depth, but all that remains to be seen I suppose.

The fascination with time limits or playing styles is lost on me. I think AlphaGo is fluid enough to batter any top pro, no matter what his style might be. If you transport Go Seigen into 2016, make the time limits long enough for them to play for a week, and give Go a team of seconds, then perhaps he would win. Perhaps the Edo period can be revisited when an Arthur C Clarke future becomes reality.

Eric Dai thought AlphaGo's play was quite ugly, and that it couldn't teach much. I can respect that opinion but at the same time disagree entirely with it. AlphaGo played the most thought provoking Go I've seen in years, so I don't feel tricked in any way by the new breed of DeepNeuralNet + UCT. They provide us with a much stronger and more realistic PC opponent. Hopefully they will bring new analysis tools to us soon. Gone are the days when the state of the art was "Guess where the pro plays next".

_________________
North Lecale

I think longer time limits are worth a try for future human / computer matches. It wasn't hard to see that Lee was feeling time pressure in some games. There is a hypothesis that for humans longer time implies better moves, but I'm not sure there is universal consensus on that once you get beyond a certain threshold. In both chess and go, there are many stories of psychological fatigue longer games resulting in absurd play.

I think that John's post highlights that, while the advent of strong computers represents a cultural shift in go world, it is not the first such shift to occur. Go has weathered the changes from state sponsored competition to private tournaments, from long time limits to fast play, from Japanese dominance to Korean supremacy to Chinese hegemony (and, perhaps, back again), from close knit go clubs to Internet anonymity. With each major change, something is lost but something is also gained. Some fans are left behind, but new people continue to discover this beautiful game.

AlphaGo marks a significant moment in the history of go, but it's too early to say just what that means. The culture surrounding the game will inevitably change. Some will like it. Others won't. But the game will endure.

John's post does highlight that AlphaGo is not the first cultural shift in the Go world to occur. That part, I agree with.

But his post also tries to suggest that the old cultural of longer time limits (e.g. 6 hour games) would yield a superior result against AlphaGo:


He's already expressed that he favors longer time limits, and that today's "Mickey Mouse" time limits are not to his liking. His post here is an attempt to use Lee Sedol's loss to AlphaGo (which, by the way, I find to be disrespectful) as a means to support his claim that the longer time limits used from older games is somehow superior to the Go produced by today's "Mickey Mouse" time controls, as he would refer to them. We've argued about long time controls vs. short time controls in the past, and the claim here that a pro having longer time controls would beat AlphaGo is arbitrary - there is no evidence to suggest this, and on the contrary, AlphaGo even utilized a neural network to train its time management skills.

I, too, would welcome a matchup between AlphaGo and a pro using longer time controls. When AlphaGo wins, it would at least show that "Mickey Mouse" is no "Voldemort".

_________________
be immersed

Yeah, I sort of sidestepped the time limit / style portion of the discussion.

I don't have a strong preference for long or short games, so I certainly wouldn't equate short time limits with a force of evil wreaking havoc upon the game. But I can see John's point that changing time limits has induced a shift in style among modern players, and I can certainly understand how that might be distasteful. Players who can manage complexity relatively quickly can gain an advantage even when they can't out plan their opponents. Lee Sedol's recently published comment about his reliance upon intuition in situations where calculation would be possible is, at least partially, a result of this shift in the game. He doesn't necessarily out think all of his opponents, but his "feel" for the game has allowed him to create situations where he comes out ahead. (This is obviously backed up by very, very strong reading skills; I don't mean to imply that he would be a poor player under different circumstances.) I think this is also evident in his match with Gu Li, where he would often get behind in the opening but create a complex situation that allowed him to fight his way back throughout the middle game.

Is it bad that time limits changed the game? For people who appreciated the unique beauty of the games of old, yes. For those who appreciate the fighting style that shows up in modern games, no. That's my point: every cultural shift will be welcomed by some and mourned by others. I do thoroughly appreciate the work John has done to bring the history of the game and its different forms of beauty over the years to an English speaking audience. Cultural shifts are, by and large, neither good nor bad. But anytime we reflexively declare newer to be better and overlook our past, something of great value is lost.

I don't have any good answers about how AlphaGo will affect the style of play or whether longer time limits would help humans compete better. I'm neither an expert historian nor a strong enough player to judge which style is likely to give a computer the most challenge. I am fairly certain that no style will triumph for long; the day is soon coming when a computer will be indisputably better than any human opponent. But humans (and the game) will adapt to the changes that brings.

I actually believe humans do have the advantage over machines with shorter time limits (and there is evidence to suggest this -- Fan Hui games), because we can still evaluate these complex positions much faster than computers. However, this "dominance" won't last much longer, though.

AlphaGo was supposed to be several stones weaker back then, which was probably why Fan won some quick games against it. However, he said he got completely demoralised playing it, which probably explains why he didn't win anything on slower time limits. So overall, fairly flimsy evidence.

_________________
North Lecale

I disagree. Machines will always be better than humans at fast play, unless we are at the limit of the machines strength and you don't have enough money for a bigger computer (or the task doesn't parallelise, AlphaGo does). The reason is humans have an inherent minimum time-scale for decision making, of the order of seconds. Making strategic plans and reading in Go takes seconds or minutes, not milliseconds or nanoseconds. Yes some strong players can read faster than others, maybe at a glance you can read some moves in 1/10 of a second, but not 1/100. There are certain pattern-matchy, intuition, neural net sort of things going on in the brain (even if Robert Jasiek doesn't think so) which might be sub second, and I can play KGS blitz at one second a move and maybe still be dan level (which is kind of amazing, much like neural net only no MCTS bots), but I don't think anyone can play at 1/100 of a second a move, and not only for the physical problem of playing a stone / clicking a mouse.

This idea of an inherent time-scale to an organism is rather interesting, I recall a recent study that looked at the flicker-frequency threshold of various animals (i.e. a light flashing 60 times a second appears steady to a human, that's how the monitor you are reading this on works) and flies had a much higher one than humans, and elephants lower. So flies probably perceive (if we can use that word for their thought processes) time passing more slowly than us, which is why they are so hard to swat. See http://www.scientificamerican.com/artic ... ion-world/ .

Anyway, if you want to look for a similar 'inherent time-scale' for computers it would be their clock frequency, which these days is a few GHz (billion times a second) (there are some limits to this based on the speed of light/electrical signals and processor size). Each operation is a simple addition or similar, so doing something useful like selecting a Go move will take many billions/zillions of operations. So if that task takes 60 seconds on some computer, it can essentially be thought of as X billion operations. If those operations can be parallelised (which they can, AlphaGo vs Lee Sedol was a distributed version) they you could pretty much do the same task in 1 second on 60 computers. There are inefficiencies of parralelisation so maybe it would be 100 not 60 computers, and joining them all together has challenges/loses, but the basic point remains: if a computer with a long time can do something, a bigger computer with a short time can do the same thing*. So if AlphaGo is not currently beating pros at 1 second a move, that's just because Google didn't use enough of their data centre on AlphaGo.

* This is different to saying more power will make the computer stronger for ever, there can be (and indeed are mentioned in the AlphaGo Nature paper) diminishing returns of computing power to playing strength. If you have a 13 dan bot at 5 minutes a move, you can have a 13 dan bot at 1 second a move if you get a bigger computer, but not necessarily a 14 dan bot.

For centuries, horses were key to the best forms of transportation. When cars were introduced, no doubt they were less reliable, slower, and had numerous disadvantages. And I am certain that some wistful buggy-makers were able to list reasons their mode was better, and could come up with racing track configurations that would favor equestrian contestants.

Longer playing times might be similar to lighter spoked wheels, altering playing styles might be similar to racing in dirt versus pavement. In the immediate future such things might help a little. In the medium and long term, they won't help at all. We are entering a new era.

No disrespect to JF intended, but The OP reads like something that could have been read in Horse and Buggy magazine a century ago.

_________________
- Brady
Want to see videos of low-dan mistakes and what to learn from them? Brady's Blunders

This is not completely relevant to the discussion, but this is 2016. We were supposed to have the flying cars already, where is my pro-level go commenting software for (Unix) laptops?

_________________
La victoire est un hasard, la défaite une nécessité.

a) Because nobody has written it.
b) Because nobody has handed you a run exec complied and linked for running under linux.

OK, I really want to talk about the first part but put the second part there to straighten out a misconception. Software isn't written for particular operating systems, hasn't been for longer than many of you have been alive (yes in the early days, not only operating systems but hardware taken into account, but probably few of you can remember things like spacing instructions on the drum of an IBM 650). When we say "available for Windows but not for linux" we usually mean you can easily get a complied and linked run exec (an executable). Strictly speaking, free software IS "available" provided you are able and willing to do that compile and link yourself or have a friend to do it for you or money to pay somebody to do it for you.

But like I said, the first part. What folks need to understand is that this is TWQ separate things, a program able to play at pro level and a program able to comment on a prod level game. The mistake here is to imagine that that these are closely related but they are not. In other words, shown a given move and being required to say what is good about that move (the comment part) is very different from finding and choosing that move in the first place (the playing part). The mistake is in thinking "of course would have to know what is good about a move to have selected it" in order to play because, for example, the MCTS programs do not << they have determined that it IS the move most likely to win the game, not WHY it is most likely to win the game >> The newer neural net programs (program plus training) are selecting "gut feeling of the distilled experience of the training database" and so again can't say why.

Pretend there were TWO pros involved. One of them is playing the game but cannot communicate with you (maybe doesn't speak English, etc. *) The other is looking at the moves made and explaining to you why that move was made.

* Or for a better metafor, is a severely handicapped savant, plays brilliantly but nobody can tell why.

I think AlphaGo is already teaching us new things about the fuseki, in that pros are trying idea from it in their games, for example the opening from game 5. I collected some examples in another thread: viewtopic.php?f=13&t=12953. This may just be a fad for a while, not some ground-breaking revolution in opening theory like "you can play 4-4 points, not just 3-4s you know", but these are early days yet. It might not be doling out expertly crafted natural language explanations for its moves (though maybe neural nets will do that in a few years/decades, with the caveat that neural nets have all sorts of complex and mysterious emergent properties and it may recognising patterns and learning concepts that we as human Go players haven't even recognised or named), but I thought JF was a proponent of the learn Go by example method, rather than the RJ style of trying to define and explain and spell-out everything. And as I think Jhyn is alluding to, it's only played 5 games!

I don't think that this does justice to the massive parallel processing that goes on in the brain when playing go. It is not just a question of pattern matching. There is real computation going on. There is analysis, there is synthesis, there are feedback loops. Decisions may emerge through some process of competition in the cerebral cortex. OC, we are unaware of nearly all of this. Consider visual processing. The last time I studied any of this, we had discovered that the brain constructs at least 18 different images before synthesizing what we think we see. Consider language processing. We occasionally recognize ambiguities in language, but as we make sense of speech we eliminate hundreds of ambiguities per second, without conscious effort. This is not just a matter of pattern recognition, as we have to choose among a number of patterns that fit what we hear, and construct a meaningful interpretation (usually one) as we go along. By comparison our conscious processing is linear, and painfully slow. But even our parallel processing is slow by comparison with computer processing. We may unconsciously process hundreds of possibilities while the computer is processing hundreds of thousands, or millions. If we weren't better at it, there would be no contest.

More thoughts at This 'n' That. ( viewtopic.php?f=12&t=12327&p=204275#p204275 )

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

My impression is that not only AlphaGo, but MCTS computer programs tend to have a cosmic style. Perhaps AlphaGo is showing us that that is not just an artifact of evaluation based upon randomized playouts. In any event, it is strong enough to make such plays respectable again.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.