http://www.wired.com/2016/01/in-a-hu...the-game-of-go

Big news as Go is a far, far more complex game than chess (a game tree complexity of 10^360 vs 10^123 of chess). Couldn't find any comparisons in game complexity of go vs. poker (go is a game of complete information), maybe someone knowledgeable in this field can share his thoughts.

The news comes in the same week that a homemade robot solved a Rubik cube in 1 second.

I was more impressed to find out that the human world record for the rubik's cube is under 5 seconds.

[ ] top player
[ ] grandmaster
[ ] breakthrough

They do plan to play a public match against Lee See Dol though who checks all the boxes and is still a top active professional player. I can't wait to see if they have a chance as everyone's guess was that it's at least a decade until AI wins against top pros.

I used to work with Demis Hassabis, the Deepmind CEO, at Bullfrog Productions back in the 90s. His first AI work was on the game Theme Park.
He's done alright for himself since those days!

He's also been known to dabble in poker as well.

http://pokerdb.thehendonmob.com/player.php?a=r&n=42073

This. I'm no authority on Go ranking, but I'm pretty sure that calling a 2 dan player a "Top Player" is quite a stretch.

In fact, I'd think that the 2014 result they talk about when a computer given a 4 stone handicap beat 9 dan player Norimoto Yoda would be a bigger accomplishment.

Also, just occurred to me that this thread has nothing to do with poker, so off it goes...

He's a 2dan professional, so (probably) stronger than a 7-dan amateur. Still nowhere near the top players in the world - though I don't think he'd be a 4-stone underdog to the top players - but it's probably fair to compare him to a chess grandmaster.

No, this is incorrect. It makes more sense to think of Fan Hui as an 8d Amateur IMO. Professional ranks for people who live outside of Asia often dont make sense. (He lives in France). Professional ranks aren't strength-based. Fan is something like the 700th-strongest player in the world.

Anyway, a much clearer estimate is that Lee Sedol is 2-3 stones stronger than Fan Hui and this bot is stronger than Fan. It's of professional strength and is giving 4 stones to Crazy Stone (the bot that beat Yoda with 4 stones) and winning 90%.

Ah, interesting. I thought 9 dan vs. 2 dan would be more than a 4 stone advantage. If not, then this is indeed a step forward.

And I wonder if a 4 stone handicap actually improves a computer's play by more than 4 stones. What I mean by that is a computer is doing more in the way of simply computing outcomes than a human would be, and a 4 stone handicap might simplify that more for a computer than it would for a human. Not sure if that even made sense the way it did in my head.

They say a little knowledge is a dangerous thing - sounds like I may have caught myself in that trap. I wasn't aware of how much the professional ranks broke down when compared geographically.

The bot is going to play Lee Sedol for a million dollars in March! ( http://www.usgo.org/news/2016/01/alp...or-ai-advance/ )

I think the most understandable analogy would be between a Golf "teaching pro" and a "touring pro". Or, like chess, a Go rank is a sort of title like Grandmaster in that you can never lose your 9p rank even if you can't play at that level anymore.

There are also European and American pros (but Fan is a Chinese pro who lives in France) but they are much weaker than Asian pros. The whole system is a mess, honestly.

Yeah, I was wrong. While he is way below top pros he is quite strong and got smacked down pretty hard.
I can't wait for the showdown vs Lee See Dol!

I saw an article on this earlier today and was wondering whether he really was a "top player", and how big of a "breakthrough" this was (or wasn't) in reality. I don't know enough about Go to have any idea, so planned to start a thread, but I see I was beaten to the punch. Having read the posts so far, my impression is that the following is about the right way to fill in the checkboxes?

I am doubtful about "Top player" (he's maybe world #700 or something and is one of a handful of the strongest players in Europe. He is super strong but toward the bottom end of professional strength) and "Grandmaster" (I'm not sure there is a good analogy because the associations make x # of pros every year regardless of relative strength of the year), but I would definitely tick "breakthrough."

This is the first leap made by the 3rd gen of Go AI. The first gen was like what works for chess, basically brute force and clever pruning. This got machines to like 2-3d. Next they changed the brute force to Monte Carlo simulation with pruning and that is what the current best commerically available bots (Crazy Stone, Zen) use to play at about 6d. This google bot (and the one facebook are running, but it is only at 5d right now) are both neural network bots and to get one to 3 stones stronger is a clear breakthrough.

Very impressive, def a breakthrough (whatever that is). The only weird thing is that they decided to release it before actually playing "Kasparov". Was there too much Zuck-pressure?

From a PR perspective, it probably generates more hype for the match if people know the bot has already soundly beaten a known professional player. Beating any current professional 5-0 is itself very noteworthy. Even the weaker professionals are usually very, very good. (At least the ones from China/Japan/Korea, since they have to officially qualify for limited spots to be called a professional.)

I don't really see any reason to keep this quiet. If the bot loses to Lee Sedol, they wouldn't want to say "Oh, by the way, we did at least beat this other, weaker pro a couple months ago." Kinda kills the story.

As far as the bot iself, they need to find strong players for it to play to have any idea of its strength beyond "well, I guess it can beat anyone in the office." There is often a great difference in playing strength between pros and even very strong amateurs. Knowing the bot beat a pro 5-0 may have been what they needed to get $1 million in funding for the match vs Lee Sedol.

The story is also a bit scary for poker players. If a company with google's talentpool and money were dedicated, could these techniques create a strong NLHE bot faster than expected? Maybe not, the game could still be too hard without playing decision trees all the way through, but it's tough to be 100% sure.

super cool that you worked together, i always had a bit of a man crush on demis in the early nineties.


guys,

https://medium.com/backchannel/has-d...bec#.fjsdsv1eq

A much better article than Wired's follow-up:

http://www.wired.com/2016/01/googles...?mbid=nl_13116

Which begs the question - why?

I'm not saying it can't, or won't happen at some point, but I see no reason to believe it's going to happen just several weeks from now.

A chess computer first beat a grandmaster (Bent Larsen, former World Championship candidate and still in the world top 100) in 1989, but didn't beat a world champion till 1996. So it seems shocking that go computers could pass the roughly equivalent milestones just a few months apart.

But this doesn't tell the full story - there were quite a lot of man/machine tournaments in the 1980s and early 1990s, and the computers probably had 100+ competitive games against GMs. Even a 2000 strength player will probably defeat a GM one time in 100, and in 2009 go programs could play at low dan level (roughly equivalent imo), so I think the rise in go program strength from 2009-2016 is probably comparable to chess 1989-1996.

From a quick, google search, there are about 1500 grandmasters, and Fan Hui is supposedly ranked around #650-#700 in the world in Go. Assuming there are about equal numbers of go players and chess players in the world, that means Fan Hui is probably a bit stronger than the lower ranking chess grandmasters, relatively speaking.

So this is the very beginning of AI's ascent through the pro ranks, not the end. The quote saying that "Many experts believe that AlphaGo will win this heavyweight bout" doesn't seem right - most professional Go players seem to think the bot is a significant underdog at this point in time, unless it really improves a lot in 4.5 months.

The strongest Go bots before AlphaGo appeared were not close to the equivalent of grandmaster strength. AlphaGo was able to give those other bots four handicap stones (which is huge) and still win at least 77% of the time. The strongest other bots were around ~6 dan (amateur), whereas pro players converted to that amateur scale would probably be 9-11 dan.

Sounds like some pros were estimating that alphago would need about a whole stone (start black with reverse komi) to make the game fair versus the best professionals, but obviously it's very difficult to know for sure. It's possible they underestimate the AI since it's easier to see weaknesses than strengths. I would definitely wager that it'll take less than 8 years from now for AI to compete with/beat the very best human players though.

I know nothing of the world of Go, but this thread has been very illuminating. Easily 100x more informative and more interesting than any of the articles I've read about this result. Good job, team.

I used to study Go in Korea, so I think I'm more informed than most people here. (my peak rank was 7d on tygem, but I guess I was more like 5d or 6d there)

The gap between a low dan player and a professional is just HUGE, low dan players make a lot of stupid mistakes that will turn the game around in 1 move. You can't compare low dan players to GMs.

While Fan Hui is still a pro, he has not been playing competitively for a really long time. That said, beating Fan Hui is a far bigger accomplishment than beating Takemiya Masaki at 4 handicap. Takemiya Masaki used to be a strong player back in the 80s and 90s but nowadays he doesn't really play competitively anymore and even if you compare him at his peak, he would still be significantly weaker than the current top pros.

Saying Fan Hui is number 633rd in the world is also incorrect. I'm sure there are a lot of players studying to be professionals that are stronger than him. I'm sure in China alone, there are a lot more than 600 players that are stronger.

As for the ranking, it's really just a number. There will be amateur players that are stronger than old 9ps. And pretty much all the recent 1ps are stronger than the semi-retired 9ps.

I still somewhat follow Go news and all the top pros are thinking that AlphaGo is much weaker than Lee Sedol. I think the only chance that AlphaGo has is if it didn't play up to its full potential in the Fan Hui match. Which is possible.

Computers don't really care about winning by 50 points or by half a point, so it will play a move that seems much worse but it has a 0.001% higher chance of winning by half-point than the move that will win by 50 points.

With all that said I heard that Fan Hui beat AlphaGo twice in unofficial matches, if that's the case I can't see how it can possibly beat Lee Sedol in March.

If someone wants to bet on AlphaGo, I'll take Lee Sedol.

If someone is interested in watching the game, https://gogameguru.com/

Beating a 2 dan(?) is no small feat. I'd certainly call it a breakthrough and the AI will not look bad against better players by design (imo). Since it improves by reinforcement learning, essentially playing itself and adjusting connections accordingly, it should get still get better over time. It'll certainly be stronger in March than the version that accomplished this feat.
[since they know who they play they may also fork off a branch that just uses his games as initial training (as opposed to the tons of games they used beofore) and experiment with that a bit]

My prediction: AI will lose and win if they play again in 12 month or so.

1 BTC?