How far are we from stockfish version of go?

[Sneegurd]

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

If that is relevant if compared with computers... Humans have the strategy, computers win.

TCEC this year was good fun and chat again, even though superfinal was a bit boring...

[mschribr]

Be careful what you wish for. Today all chess analysis is done by computer. Computers analyze chess positions faster, deeper and with less mistakes. The computer can analyze in a few minutes what a team of chess players can do in weeks. The world chess champions prepare by watching what the computer does. See the video by 5 time world chess champion Anand, http://www.youtube.com/watch?v=JSOw1Yk_RQU
At about 27 minutes into the video Anand says while preparing for one world chess championship the computer told him that black was winning in a specific position. He used the computer analysis and won the game. Years later preparing for another world chess championship he was using a better program running on a faster computer. The computer told him white was winning for the same position. He used this new analysis and again won. It is scary how weak humans are in chess compared to computers.

[gowan]

So far it has been possible for human players to understand the computer recommendations and "explanations". I wonder how long it will be before humans won't even understand reasons behind the computer recommendations. Remember when you were first learning go and you couldn't understand why professionals played the moves they did? I wonder when computer play will be that far above the play of modern top pros. Soon the best computer program will defeat the top pros with the pros taking a significant handicap. I don't see why the programs couldn't get a lot stronger than that eventually. I recall Cho Chikun saying, at a time when he was the top player in the world, that he felt there was an vast, immense amount he did not understand about go.

[gowan]

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

If that is relevant if compared with computers... Humans have the strategy, computers win.

TCEC this year was good fun and chat again, even though superfinal was a bit boring...

Winning is overrated. Machines are superior to humans in so many ways. Take a 1500 meter race between and an automatic driverless car and a human on foot. Does anyone feel diminished by that? If you point out that the machine and the human are not doing the same thing, it won't be long before an autonomous robot moving using two "legs" will be able to "run" faster that any human will ever be able to. At my age (near 80) I do not play or study go for competition, rather to broaden my understanding and appreciation, and for interaction with friends. We can do this enjoyable activity regardless of how strong computers might be.

[Mike Novack]

Winning is overrated. Machines are superior to humans in so many ways. Take a 1500 meter race between and an automatic driverless car and a human on foot.

But there is more to that. You apparently have never seen a rodeo stunt race between a human and a horse. << hint -- humans have a much shorter turning radius >>

It all depends on the course! I can easily envision a 1500 meter course that the vehicle COULD traverse (let's not be totally unfair) but with such difficulty and such low net speed that the human would do it faster.

[Uberdude]

[Kirby]

Winning is overrated. Machines are superior to humans in so many ways. Take a 1500 meter race between and an automatic driverless car and a human on foot.

But there is more to that. You apparently have never seen a rodeo stunt race between a human and a horse. << hint -- humans have a much shorter turning radius >>

It all depends on the course! I can easily envision a 1500 meter course that the vehicle COULD traverse (let's not be totally unfair) but with such difficulty and such low net speed that the human would do it faster.

I've been thinking about this general topic from various contexts. For example, language study is one of my hobbies, and it's somewhat of a novelty to know some Korean or Japanese when I run into native speakers. But machine language translation is rapidly improving. Already you can get the general gist of a website's meaning by running it through a language translation service. There's still work to be done, but the day will come when machines can translate between various languages more quickly and accurately than humans can. Probably not too long after that, the day will come when this quality of translation is available to humans in real-time through whatever devices they are carrying.

What, then, is the value of studying a foreign language? Considering the practical utility of spending my time studying Korean and Japanese, it's not very valuable as machines become more adept. Does that make the study worthless? I've come to the conclusion that it's not worthless, but not very practically useful in the classical sense. If I want to interpret meaning from a text, or eventually, in real time from a speaker, it will eventually be much simpler and much more accurate to use an electronic device. The value, I believe, is by internalizing knowledge. By internalizing knowledge, I change who I am as a human being. By studying more Japanese, for example, I understand Japanese a little bit more. And maybe in some ways, I absorb some of the Japanese culture. My friend who doesn't study Japanese will eventually just as easily be able to translate Japanese texts or communicate with Japanese people. But they won't have the internalized Japanese knowledge and culture that can be obtained through internalizing the language.

I believe that the same holds true in other pursuits. I can navigate to Wolfram Alpha with just a few keystrokes to calculate a complicated integral or to factor a large number. The answers I get are obtained faster and with more precision than I'm able to get with my current mathematical capability - and if I need only the answer to a math problem, the site provides all that I need. Yet, this form of knowledge is shallow; the computer tells me the answer and I trust it - but I haven't internalized the knowledge. It's not a part of me. There's a black box between the solution and the question - a shrouded cloud of mystery that's hidden from my view. The answer is not a part of me. It's something that's given to me.

The same can be extended to physical pursuits, such as running for exercise. A car may be able to drive faster than you can run. But the speed and the power are not a part of you. They don't increase your ability as a human being. They don't contribute to your health and function. Running for exercise does, on the other hand, contribute to yourself. You internalize new ability, health, and stamina which weren't there before. Whatever you choose to study, it is valuable. It shapes who you are, and it adds to yourself as a person. It reduces your reliance on external factors; if I'm a native speaker of French, I don't need a machine to tell me how to speak. The knowledge is part of who you are.

With this in mind, I can come back to my earlier question of studying Japanese or Korean. Does machine translation make my study worthless? Absolutely not. Pursuing this knowledge contributes to who I am as a person. It increases my capability as an individual, and expands who I am.

In the same way, AlphaGo and other computer programs, which are surpassing human skill in the game of go, do not take away from the capacity for go to increase my capability as a human being. By studying go, increasing my knowledge or reading power, or gaining experience from the game, I stretch and improve myself. I become more powerful and a better human being - even if the moves I end up playing are sub-optimal.

Exercise, study, personal improvement is always valuable. Even if we can rely on other people or machines to do things for us, these activities empower us and improve our overall capabilities as human beings.

That, in my opinion, is a worthwhile pursuit.

[Gotraskhalana]

While I concede that the original babelfish will certainly come in my lifetime, I think that you underestimate human capability and you totally underestimate non-vocabulary problems in intercultural communication.

I think that interactive maps are already at a level that translation will be in the next ten to twenty years, and I have seen people who did not know the city, did not try to get a general idea of its layout and simply following the phone from A to B use it very poorly.

[Jhyn]

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.

I'm pretty late to the party, but while I get your point to some extent, your argument does look a lot like this webcomic:



I think that we humans are really fond of explanations and proud of our ability to create explanations. With a beginner I could probably find two of three reasonable-looking explanations that a given move is both good, bad, brilliant and terrible, and a pro could do the same with me for most of the middlegame.

Explanations are good only for one thing: they let us play better, they let us grasp the immense complexity as well as we can. But we must not confuse the prey and its shadow (sorry for my French). If a computer - or a wizard - finds a move that looks objectively better than the current theory, this gives us a glimpse of the "real" Go - isn't it our job, as well as we can, to craft explanations that bring this move into a theory that we can comprehend? (as a go scientist getting access to a "realer" world than he used to)

If you end up rejecting a "true" move because you do not have an explanation for it yet, how are you any different than the mathematician that refused to give up on Cantor's Set Theory after it was proven inconsistent because it was so elegant? When we have a convenient lie, and the truth, we shouldn't forget that we accepted the lie for one reason - because it got us, in practice, closer to the truth - and that we must also abandon it when it has run its due course. Otherwise, are we really playing go? Are we not hoping that our dear theory will not collapse under its own weight, closing our eyes to the fact that the real world that supported it went away? Shouldn't we welcome this opportunity?

I think the process I'm describing is reiterated, at the scale of the individual, many times through the journey of learning go. I know that I regularly face the same dilemma many times when going through pro games, even commented ones - why do the pros play some moves that I would reject as awful at first sight? Where did my instinct go wrong? Can I really change myself so that I would consider this move without throwing away everything I know? Does this commentary not contradict something I was pretty confident about (and I'm sure I read somewhere)?

I am reminded of Wittgenstein's On Certainty: we have (momentarily) fixed beliefs of various strengths, which is a background necessary to live (play go). Therefore we can apply doubt to only a few fixed beliefs at the same time - strenghtening, weakening, throwing away - while the rest of the bedrock is not moving; even the doubt is only a product of these other fixed beliefs. This implies that absolute doubt (playing with pure reading) is in practice meaningless, and also that we have the mental ability to progressively adapt our thought process and build better theories.

[dfan]

You pick on the concept of "explanations" a lot, but Kirby didn't use that word.

If you play a move that is excellent with perfect play afterwards and horrible otherwise, and you don't see the followup, it is generally not a good decision to play it if you're trying to win this game. I think that is all that Kirby's statement is trying to say, and it seems pretty uncontroversial to me.

If you don't care so much about the result of this particular game, and playing correct moves that you don't understand yet helps you expand your horizons, then great.

[Bill Spight]

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.

As you know, I am hardly against understanding. However, this view downplays the value of imitation.

But without understanding why a pro played a certain way, or the subtleties of their strategy, just copying the moves is not very good.

As I said.

[Jhyn]

My point is that I believe that the idea that we actually understand our moves, in the sense that our (inner) explanation of why we play it corresponds to the reason why it is good, is largely an illusion (though a useful one). Bill Spight makes a similar point.

Have you never asked to a beginner why he played a certain (bad) move and faced a long explanation full of "I thought I would push him in this general direction" and "I thought that if he plays here and there I can kill his group"? Sometimes I want to answer "What you said is not even wrong". Does it really give a value to his move? Well, I think I agree with you that maybe. But the most important thing is the attitude of the beginner, that he is ready to change his beliefs, otherwise his clumsy explanations won't carry him far.

I do think that many moves of AlphaGo are too far from my current understanding that I can include them in my repertoire without breaking everything else. But this is true for most pro games, and many pieces of commentary "don't make sense" to me. If I always had a negative attitude towards these moves I think I would be weaker today. In practical terms it is much harder for me to learn from a computer (today), but I think our (individual and collective) attitude should be an hopeful expectation of betterment.

[Mike Novack]

I think the "language translation" problem is actually a useful analogy. And that cartoon.

"Is the explanation useful to us" is not just a function of who (or what in this case) is giving the explanation. It also depends non us.

How well we get the gist of a somewhat garbled machine translation of a language depends on us at our end figuring out what must have been meant in that context. It is in fact our great abilities in doing that (our OWN language would require this*) that makes the machine translation appear better than it is.

Consider our "expert" (human or machine) trying to explain to us why some move is good or bad.

"the move gains a couple points in sente"
"the move surrenders a ko threat"

Do you not see that which answer might be given for the same move could depend on the context of the game (how likely or unlikely to be a major ko). That there would be a whole lot of additional explanation required. Not just things like the balance of ko treats for each side but higher level things like "behind, so the game lost unless a major ko can be created an won"

[Bill Spight]

the balance of ko treats for each side

I'll have a ko treat, please.

[lightvector]

I'm generally more with Bill on imitation being valuable. It sometimes happens that a certain move just jumps out at me wanting to be played where when I try to read it I find it beyond my ability, and where I really don't know what will happen if I dive in and try it. Sometimes it's a move from a lesson or from a similar shape in pro game that I explicitly remember, sometimes it's something that I think I've seen before but don't recall exactly, and sometimes it's a move that I have absolutely no idea about, except that my intuition for no apparent reason is jumping up and down bringing the move to my attention.

Often I have a choice between this move and a "safer" move. But not only is the risky non-understood move a better opportunity for learning, when I do try it I surprisingly often get a good result, making it also on average not bad even just for winning that game.

The key thing is that it has to be a move that is instinctively shouting at me to be played, even if I have no conscious understanding or ability to justify why that move is good. If I just try a move randomly or that I vaguely remember from somewhere but that doesn't have that instinctive attention-grabbing nature, it's more often that it doesn't end well. I think this is because there are places where my shape instinct and intuition have reached somewhere currently beyond my ability to read or consciously explain, such that when they do strongly suggest a move, trusting them even without understanding is actually a good bet. And also I think one of the routes to obtaining that understanding is precisely to trust and try the move enough until it starts to make sense.