Alpha Go to play Ke Jie in May

https://arstechnica.com/information-technology/2017/04/deepm...

It's a three game match.

Hassabis: "Instead of diminishing the game, as some feared, artificial intelligence has actually made human players stronger and more creative. It’s humbling to see how pros and amateurs alike, who have pored over every detail of AlphaGo’s innovative game play, have actually learned new knowledge and strategies about perhaps the most studied and contemplated game in history."

I read in some leaks earlier to the official reports, that the version of AlphaGo playing will be one that has learned the game by itself from scratch without learning from human games. There is no mentioning of that in the official news, so that may not be true; but it would be interesting. At the very least I don't expect the version that has learned from human play to lose even one of the games; after the online games that AlphaGo won 60:0 against pro players that seems unlikely, although then again those were games with like 15 seconds per move which should be a disadvantage for the human.

I'm excited!

http://www.usgo.org/news/2017/04/world-1-ke-jie-9p-to-take-on-alphago-in-china/

From an AI perspective, it makes sense to have a decent starting point, so it makes sense to start with pro games, but OTOH, it also makes sense to remove the possibility that we humans have a blind spot (for example maybe our joseki are actually not seeing the "big picture"/whole-board and are too focused on the corners... but as far as the next match goes, It matters little to me if AlphaGo has "learned" from Pro/human games or if "the version of AlphaGo playing will be one that has learned the game by itself from scratch without learning from human games."

Do the good players here agree with Brady Daniels analysis of two similar joseki games of AlphaGo?

https://www.youtube.com/watch?v=6rB2cYOeppQ

In the second game are the ko threats as big as he says?

Carroll, i've no idea how big they're exactly, but it's at least plausible to me that they're big...

Very interesting video anyway, I've learned something from it, even if I'll probably never play Go... :)

Do you mean the one around 35'15'' ? I think he's correct :)

Well, Brady Daniels tells nice stories about the games, quite entertaining. It helps to learn a certain way of thinking during the games but to really understand what is happening one should read ahead like AlphaGo.

I like AlphaGo games but I don't think humans can imitate it. I mean we can play the same moves but the games will still be messy, humans don't know as well as AlphaGo when they are ahead and for sure don't know so well the value of moves, not even professionals. I really think AlphaGo can now give at least two stones handicap to anybody,

He is correct about the ko :-)

Crelo, but is AlphaGo play *correct* or it is only "working" in self-play and again human players?

Probably nobody knows but anyway the question is interesting. I think probably it's indeed the right way to play and it should be copied.

To put the same idea in another way, counting territory accurately in Go is clearly necessary for near perfect play. But the same can't be said for counting "centipawns" in chess.

Lazyplayer, there is no way to tell if the AI plays correct (perfect) moves or is just slightly better than us. There is also the danger to trust the AI too much, maybe there realy are better moves. We must get used to this new world. :-)

Do you think the AI is capable to beat the best players in a long game?

Something like 1 month per player.

I would believe it only when I gonna see it. Now, I just believe the AI is capable to use the fact that humans gonna make mistakes cause of too short time.

Florian, but in reality it's very hard for an human to remember the analysis already done. We've very bad short term memory. If you solve this problem then indeed humans should still be able to win when given enough time, i guess. Well, maybe you would need more time than their lifetimes... :)

Then, if you prefer, you can do a match like this:

150 of the bests humans are playing the same game against the AI, during a year. I guess it is enough.

The goal is just to know if really the AI is playing better then humans in general. If there is no hope anymore. Cause finally it is the real question: Is the AI playing better than humans in the general sense.

I don't believe it, they will need to show me.

Well Florian, I refer you back to the quote at the beginning of this thread. Pros are learning "new knowledge and strategy" from studying Alpha Go's games. In the video, Hassabis briefly shows a couple of techniques that had previously been under-appreciated. If that's not playing better than humans in the general sense, then what is?

Of course the AI is playing better, because it wins :-) AlphaGo is not an exception anymore, its software structure can be reproduced.

https://qz.com/936654/googles-alpha-go-now-has-a-serious-game-playing-rival-with-tencents-jueyi-or-fineart/

Anyway, the AI is still a human creation, we should feel proud, actually we will learn more this way.

David, be sure that I don't want to say that there is nothing to learn from AI or that AI is NEVER playing better than humans.

This is completely different.

My question is different. The fact is that AI in general are not playing "really" better with a lot of time to play, in a 2 hours or 3 hours game (per players), they are already playing very very well.

A human is doing mistakes in a 2 hours game, these mistakes could be corrected in a 1 month game, the AI will also play better in a 1 month game but you can be sure that AI won't play better as humans will.

Crelo is agreeing with me, even if he does not realize it.

He said that AI can give no more than 2 stones to best players in a 2 hours game, this almost implies that in a one month game humans can beat the AI.

I am afraid the computer will advance more in one month than humans.

When you give an AI significantly more time (assuming the computer resources are fully available) it can use all of that time; when you give humans more time they need to sleep & eat during some of that time... and if the AI programming were adjusted for the larger amount of time (again assuming the computer resources are fully available) think about how deeply the AI could probe into variations (I'm thinking about "brute force" checking of move choices made by the core process).  I'm with Crelo at 2017-05-06 ---> I think the computer AI would advance more than humans with the extra game time.

https://phys.org/news/2017-05-ready-rematch-machine-ancient-game.html

It is tomorrow but I did not find if it will be streamed, anyone knows?

they will be streamed

https://events.google.com/alphago2017/

https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A

The question that is the most interesting to me and also arises a couple of times here is "if AlphaGo plays better in general / if AlphaGo plays correctly". The meaning of this is basically if AlphaGo plays better openings or just recovers in endgames.

Experiment that should be done is:

1) play N moves human vs AlphaGo

2) finish that game AlphaGo vs AlphaGo

3) assume that the winner played better opening

4) try another N maybe

It seems that nobody has spotted any particular weaknesses in AlphaGo's opening play - or they would have exploited them. On occasion it plays surprising moves, but it does not mean they are bad, only that humans have not fully understood their benefits before. A top pro does not give away many points in the endgame anyway, so it needs to play good openings in order to win consistently.

On Sensei's Library they have started to collect joseki (local patterns) popularized by AlphaGo (I only looked at this briefly).

I don't know a thing about Go, but I assume that AlphaGo plays perfectly near ending when result is easier to compute. The only weakness should be first N moves, otherwise it may play perfectly all along.

Arek, i guess it depends on the game. Probably it's like in Hex, where there are some very hard positions with many stones, but not that many.

And alphago "prudent" style probably allows it to play perfect when the position is favorable enough for one side.

Arek - I think you may be underestimating the top pros. If AlphaGo made weak opening moves, they would attack the weaknesses mercilessly. Instead, they are analyzing its moves and considering to emulate the new strategies. It is true that finding correct moves towards the end may be easier for the computer, but then it is also easier for the humans.

lazyplayer - AlphaGo probably plays best when it is only slightly ahead. If it is far ahead, it may be overly timid in order to secure the win rather than maximizing the score. When it is behind, it may try "wild" moves in order to try to shake off the opponent.

Tasmanian, how do you know that these "timid" moves aren't in some sense the "best"?

I think it's a rational strategy, and humans can't do it because they play more according to local patterns instead of playing according to the global state of the board.

The objective of the game is to conquer as much territory+prisoners/area as possible. So the best move, if you could analyze any position perfectly, is the one that maximizes the score. But without the ability to analyze perfectly, the concept of risk enters the picture, and it makes sense to keep it small if you are ahead (whether you are a human or a computer program). Humans certainly consider risk when playing go. A safe play could be to secure your group/territory while allowing your opponent to do the same with theirs, while a risky play could be to dive into a big fight.

Eheh, no, objective is to have more points than opponent. Wide margin doesn't matter.

Yes they consider risk but they've less flexibility. They often have to play what they know is good in general (that is, good for balanced positions).

So if the game is almost over, there are no uncertain complications left, you are ahead by 40 points, you can kill one of your opponent's groups for 30 more points, or you can fill in an eye in one of your own groups, thus commit suicide and lose 30 points, these two moves are equally good?

Tasmanian, depends how you define "good"...

Here we were debating perfect play near the end of the game, so it makes more sense to define "good" as "it makes sure that you preserve your win".

Tasmanian, anyway maybe in the pursuit of safety, AlphaGo objectively gives away the win. Maybe. I'm waiting for commentary from people that actually are good at Go! :)

You do that, and let me know if they disagree with anything I said here.

Tasmanian, let me clarify two things:

1) I don't underestimate pros - if anything maybe I overestimate AlphaGo. I say that AlphaGo either plays perfectly throughout or perfectly only near the end. It doesn't say anything about the pros.

2) I don't say AlphaGo plays weakly in opening - I just say first N moves which may be first 300 moves if you just pick N = 300. My claim is that in a game that last K moves and K > N, AlphaGo plays better between Nth and Kth moves then between 1st and Nth.

Hence it would be interesting to run AlphaGo vs AlphaGo starting from a position after N moves, for example of the second match against Ke Jie. Commentators said that it was close in the beginning.

As a Go player, I don't think that the assumption holds. What I mean is that I'm fairly confident that the advantage in opening can't be determined by finishing the game with AlphaGo vs AlphaGo. Analysing with N > 50 also doesn't really seem like it would give reliable enough data to understand the opening advantage.

I'm saying this because I doubt that simply categorizing whether the AI is better in either the opening, the endgame, or in both is accurate enough. A more likely scenario would be that both humans and the AI have strenghts and weaknesses in all phases of the game - even in the endgame. The reading depth of top AI is not enough to crack the endgame reliably quite yet. Typically both top pros and AI play the last 40 or so moves nearly perfectly, rarely making a mistake of more than 0.75 points during the course of the endgame.

In my opinion, the answer to your question is that AlphaGo both plays better in general, and with a higher degree of correctness. I'm not a researcher on the subject so do take my words with a grain of salt. However, I don't think that the experiment setup you described would be able to answer the question convincingly enough; in my opinion relative skill in Go is not so binary that making conclusions from AlphaGo finishing board positions against itself would bring out the truth of the matter.

The summit in China is finished. I think we can tell Alphago is playing better opening and better middle game than Ke Jie. We don't know for sure about te end game because Alphago is playing for safety once is ahead.

50 games of Alphago against iteself were made public (http://www.alphago-games.com/). Impossible to understand them. My impression is that Alphago considers a 7.5p komi too big, Black is playing reckless moves, white seems more balanced, black won only 12 games.

Crelo, why "impossible to understand them"?

I would expect that looking at (reliably and consistently) high quality games is very instructive.

And i would expect the stronger the players, the larger komi... why they've chosen 7.5? maybe it was closest to balanced (for AlphaGo) that they could find?

It's also amusing that AlphaGo plays often 3-3 as answer to 4-4. It's like in hex, hehe. We just want corner so badly :D

Impossible to understand means I cannot follow the logic behind the moves, someof them are against the current theory. We will have to unlearn first and try to understand after. Even so, the humans might not be able to play Alphago style because of limited and imperfect computational power.

Komi has nothing to do with the players strenght, it is just a number of points to compensate white to play second.

It is true that komi increased in the last century because the advancements in the game theory. Korea and Japan are using a 6.5 komi, China is using 7.5. Because the chinese rules are counting areas instead of territories 6.5 is the same as 5.5 so they have to use 7.5

The match with Ke Jie was played with chinese rules so 7.5 komi.

Crelo, yes, i know. It seems 6.5 komi under area scoring is not significantly higher than 5.5 komi.

I guess the ideal would be 7 points in komi plus "Button Go" ;)

https://deepmind.com/blog/alphago-zero-learning-scratch/

"After just three days of self-play training, AlphaGo Zero emphatically defeated the previously published version of AlphaGo - which had itself defeated 18-time world champion Lee Sedol - by 100 games to 0. After 40 days of self training, AlphaGo Zero became even stronger, outperforming the version of AlphaGo known as “Master”, which has defeated the world's best players and world number one Ke Jie." - from that article

"It still required a huge amount of computing power — four of the specialized chips called tensor processing units, which Hassabis estimated to be US$25 million of hardware." -- from the blurb

The "old" AlphaGo seemed to win more often with white in self-play.  Perhaps the new version can tell us if the komi (7.5 with Chinese rules) is too high? (Strictly speaking, the correct komi should be an integer and correct play should give a draw, although many players probably would not like to open up for that.)

Tasmanian, if a perfectly played game at komi 6.5 results in a black win, and a perfectly played game at komi 7.5 results in a white win, are you saying komi should be set at 7 and drawn games should be allowed? This would be a significant rules change which on a practical level would make the game less interesting IMO. There are other games which do not allow draws, such as Hex for example. Should the rules be changed for Hex as well?

Even with such amazing improvements of Alpha Go Zero, no one is claiming to have solved Go. This is apparent from the fact that for a specific komi, with Zero playing both sides, sometimes black wins and sometimes white wins.

A rules change I would like to see would be to adopt "komi pie." Before the first stone is placed, one player sets the (non-integer) komi and the other player chooses which side to play. Get Alpha Go zero to set the komi by itself in this manner, and see what value it settles on after many games.

Yes. I don't personally agree that it would make the game less interesting, but I believe it is true that the go community would not want this change. I know it is not solved, but considering how strong the old version was, and given that this one is significantly stronger, its estimate of the optimal score should be somewhat reliable. We have a similar situation in Othello, where the estimated optimal score happens to be a draw.

Hex does not really allow for draws as it is not game with a score, just obtaining a goal that is necessarily achieved sooner or later, and the swap rule appears to even out the odds pretty well.

Sounds like a komi of 7.25 is the way to go. :))))))

I have never seen or heard of anyone using the swap rule in go btw. Could be interesting to see something other than corner openings. Like choose the first move along with tie breaker +0.5 komi for either player and then the second guy can swap or not.

It's really not necessary to change the komi rules though. Whatever player you end up being you can win, because your opponent will not play correctly and thats for sure. I wouldn't assume that a perfect game has ever even been played before.

Well of course no perfect game has been played yet. And komi pie is effectively implemented on some servers, KGS for example. One person sets the parameters of the game including the komi, and whoever decides to play this person sometimes gets to choose which side to play. Setting the komi is a valid way to slice the pie all by itself, with no need to place an intial stone.