World #1 Go Player Ke Jie accepts Google Alpha Go Match..

[Dirt]

I was shocked to see Ke Jie try 3,3 moves early on. I'm not surprised it didn't work.

[Uri Blass]

Only by 0.5 point if I understood correctly. On the other hand AlphaGo is programmed to go for the surest win and not the win that delivers the most points (which may involve taking higher risks). So the fact that AlphaGo "only" won by 0.5 points may be misleading.

0.5 without komi of 7.5? That was my impression, but I don't know how to count very well. From what I understood, Ke Jie tried to gain points, so that Alpha would need to attack, but mishandled the opening and in fact after the opening Alpha had more (potential) points and Ke had to attack. There was no opportunity to attack Alpha in the whole game, and Alpha, gaining advantage from the opening, just played quietly for the win. Alpha still seems to go to the highest probability, not highest advantage, so potentially it may had 10-15 stones advantage in late midgame, but went for the safer 6-8 points. I will maybe post Crazy Stone Deep Learning (7 amateusr dan) 2 hour analysis, that shows that from this engine (Zenith 6 too) point of view, White (Alpha) had substantial advantage right from the opening after move 20, a thing which seems to be overlooked or muted by 9p pros commenting. It seems that even relatively weak engines compared to those pros (3-4 stones difference) are better at scoring in the initial stages of the game.

Ke here seems to have had no any chances, a pretty overwhelming win by Alphago.

thanks for the analysis. keep it going!


with chess most (although not all) experts seem to be saying even with a 32 piece database, a computer with knight handicap would not be able to defeat a grandmaster.

do you have an opinion what handicap would be needed for a 9 dan go player to defeat a computer with access to a solved database of go.?

Are you sure about your knight odds comment? It seems Larry Kaufman said earlier this year that Komodo already plays about equal with a few grandmasters at knight odds in blitz. Here's where I found his statement:
http://talkchess.com/forum/viewtopic.ph ... night+odds

Seems to me that it's only a matter of time before engines can beat grandmasters at knight odds in long time control games.

There is a big difference between blitz and long time control games and I do not believe it is possible to beat GM's with fide rating of 2500 with knight odd at LTC.

[Jesse Gersenson]

There is a big difference between blitz and long time control games and I do not believe it is possible to beat GM's with fide rating of 2500 with knight odd at LTC.

I would like to see a 2500 play Komodo at knight odds and LTC.

[JJJ]

There is a big difference between blitz and long time control games and I do not believe it is possible to beat GM's with fide rating of 2500 with knight odd at LTC.

I would like to see a 2500 play Komodo at knight odds and LTC.

So do I, it could be very interesting to see now Komodo is so much better.

[Laskos]

I was shocked to see Ke Jie try 3,3 moves early on. I'm not surprised it didn't work.

He mimicked "Master" (AlphaGo of 5 months ago with 60-0 record) playing that. Many pros now copy AlphaGo.

The game was indeed 0.5 win, just that error was 0. I seem to have missed the final counting. I got the full SGF file, it is interesting to see the endgame eval with Crazy Stone:

Margin of Win:


Probability of Win:

[Jesse Gersenson]

There is a big difference between blitz and long time control games and I do not believe it is possible to beat GM's with fide rating of 2500 with knight odd at LTC.

I would like to see a 2500 play Komodo at knight odds and LTC.

So do I, it could be very interesting to see now Komodo is so much better.

Komodo would get murdered giving knight odds to a 2500, but it'd be an instructive display.

[lkaufman]

There is a big difference between blitz and long time control games and I do not believe it is possible to beat GM's with fide rating of 2500 with knight odd at LTC.

I would like to see a 2500 play Komodo at knight odds and LTC.

So do I, it could be very interesting to see now Komodo is so much better.

Komodo would get murdered giving knight odds to a 2500, but it'd be an instructive display.

A couple years ago Komodo scored just 25% in each of three matches at 45' + 15" giving knight odds to players averaging about 2130 Elo strength (2230 USCF) in 12 games total. I'm planning to try again against players of this level to see if we can do better, but we haven't improved Komodo by 200 elo in this time so 50% would not be a reasonable expectation yet. However I may include some chess960 knight odds games, where I think that Komodo has better chances than at normal chess, because there is more to think about from the start. If we can beat low masters like this in chess 960 knight odds then we can try same against stronger opponents.

[Laskos]

Only by 0.5 point if I understood correctly. On the other hand AlphaGo is programmed to go for the surest win and not the win that delivers the most points (which may involve taking higher risks). So the fact that AlphaGo "only" won by 0.5 points may be misleading.

I think AlphaGo team is a bit too greedy and rigid going after ELO. That hurts the style of play, and allows for some false conclusions, like "only 0.5 points". It doesn't deliver for a patzer like me the magnitude of the Win. And not only me, I see better players misinformed about the course of the game, at least I had 7d engines to analyze in real-time. The commentators seemed to be clueless on who is winning in the first 80 moves, and by that time the game was decided. Surprising, giving they are good pros, and have better clues than 7d engine. Also, this fake kindness of Hassabis from DeepMind tweeting "AlphaGo won only by 0.5 points" is misleading. What's their matter, couldn't they decrease winning chances against Ke Jie from 98% to 96% by going after points and not likelihood of Win? Just 3 games, it's not statistics, probably anyway all 3 will go its side. But for most, one thing is to see 30 points advantage which even a patzer like me can easily understand, and another 0.5 points with 100% certainty. Also, it's hard to me to infer if AlphaGo could have invaded there and there, completely destroying Ke Jie by 50 points and showing that it's different category players. Sure, for this, re-training is needed, but after they got 60-0 result 5 months ago, why they bother with likelihoods instead of points?

I got the complete game record in SGF, yes, you are right, the win was 0.5 points, I just failed seeing and counting it to the end. But the evolution of the game according to Crazy Stone is that the game by move 60-80 was decided, while there were comments of pros who said that "Ke Jie is standing well against AlphaGo" at that time.
The complete game record (SGF) 4 hour analysis:

White margins in points:


Likelihood (t-value) of White Win:

[Laskos]

This thing which destroyed Ke Jie runs on a machine 10 times smaller than that against Lee Sedol, a large single machine with like 4 GPU. I would dare to say that AlphaGo on home PC is stronger than top Go pros. A remarkable achievement. They also have new approaches built in, ready to be published in "Nature" after the match. I guess in 3-4 years we will have Go on Smartphone beating top Go players. Outstanding progress, this seemed very far away only 2 years ago.

I wouldn't be so sure about that. The difference between Titan Xp and Adreno 540 is far greater than difference between i7-6950X and Snapdragon 835.

In many benchamrks the latest Samsungs are only 5 time slower than a fully equipped Intel 7700K. The progress in Go was 2000+ ELO points in the last 1.5 years. Add another 1000 in 3-4 years and some moderate progress of Samsung, and you are there. Let's see, but I cannot wait to buy the new Zen for PC, or even better, AlphaGo if they make it commercial. I already bought a strong Nvidia card. Stupid hobbies, as a player I am a patzer in everything .

As I said i7-6950X vs Snapdragom 835 is like 5x for chess (SF benchmark).
However TitanXp is 30x compared to Adreno 540 and that's in games, and in actual ML difference is easily 50x. And AlphaGo hardware atm has 4 TitanXp's. That's close to 200x and AlphaGo strength lies almost exclusively in GPU power.

No, if you follow "Nature" paper, 60-70% (especially tactically, MCTS) of gain comes from CPU. This year we will have on a home PC engines, either AlphaGo (stronger) or Zen, having the level of top Pro players. My PC is 7d, my phone with Crazy Stone is 4d. The distance is not that big, and it will probably come down to yearly improvement, which seems important, maybe 5-6 stones in the following 3-4 years. Especially that DeepMind team speaks of new approaches which will be published in "Nature" again. Others will implement them too in the following years. I fully expect that Go will be stronger on smartphone than most Pro players in a matter of 3-4 years. The situation could become similar to Chess. Smartphones are already strong, beating 99% of human casual players (4d is well beyond not me, but strong Chines amateurs).

[Milos]

I think AlphaGo team is a bit too greedy and rigid going after ELO. That hurts the style of play, and allows for some false conclusions, like "only 0.5 points". It doesn't deliver for a patzer like me the magnitude of the Win. And not only me, I see better players misinformed about the course of the game, at least I had 7d engines to analyze in real-time. The commentators seemed to be clueless on who is winning in the first 80 moves, and by that time the game was decided. Surprising, giving they are good pros, and have better clues than 7d engine. Also, this fake kindness of Hassabis from DeepMind tweeting "AlphaGo won only by 0.5 points" is misleading. What's their matter, couldn't they decrease winning chances against Ke Jie from 98% to 96% by going after points and not likelihood of Win? Just 3 games, it's not statistics, probably anyway all 3 will go its side. But for most, one thing is to see 30 points advantage which even a patzer like me can easily understand, and another 0.5 points with 100% certainty. Also, it's hard to me to infer if AlphaGo could have invaded there and there, completely destroying Ke Jie by 50 points and showing that it's different category players. Sure, for this, re-training is needed, but after they got 60-0 result 5 months ago, why they bother with likelihoods instead of points?

It is obvious they don't want to embarrass Ke Jie, but instead make it look as competitive as possible and in the same time still win 3:0. I wouldn't be surprised if this was even written in some contract.