Lee Sedol vs. AlphaGo [link to live feed]

[oreopoulos]

There is a HUGE difference between chess and GO in terms of what it takes to create a great engine.

a) Chess is played in a limited "space-board" with pieces of different skills that create "dynamics".
b) local events in a chess game affect the result of the struggle dramatically. This way its easier to guide search. This is extremely important.
A local struggle can define the result of the game in chess, making what happens in the rest of the game irrelevant.

c) in go local cannot radically define global. A local defeat in no way can define a loss

Go is not a search-first game. Its more of a transformation game.
You should be able to find what patternA will evolve to. How patternA and patternB can play along. Let call that meta-search. You have to "adjust" both the local fight and also how that local fight can be combined with possible patterns from an infinite playbook pattern

[diep]

The situation is easy as i see it:

In GO the huge branching factor and number of moves for the average game is so huge, that no computer or human can base or improve his play, improving that. I.e improving search in GO does not do anything(compared to evaluation function).
So evaluation is THE MOST crucial factor by far!

oh comeon. Assuming you have a reasonable evaluation search is everything in Go of course.

From the 361 moves you can HARD FORWARD PRUNE kind of 300 moves already.

Well since you decided to talk not for this world but only for an imaginary one, why not take it further?
Let's assume having an evaluation function that prunes all moves every time and always choses the best move. So search is irrelevant.



**Oh and wait. If you are able in the first place to prune 300 out of 361 moves then your evaluation is not reasonably good, it's extra super good! So your super play will still be mainly because of evaluation function and not search.

You are making fun of it. But this is EXACTLY what the top go programs until recently did do.

They hard forward pruned. I tried to talk several programmers out of it, yet their nps simply was too low to get the needed 30+ search lines otherwise.

As for the go players:

He blundered bigtime game 2. Commentator 9-dan pro (just like Lee Sedol 9 dan) said about that: "he (Lee Sedol) played 2 slack moves in the middlegame that lost him the game".

You can look that comments up. Around minute 35 here: https://www.youtube.com/watch?v=yCALyQRN3hw

You do believe a 9-dan pro calling Lee Sedol playing slack moves?

But now reality ok. Reality of computer program. One of the games Lee Sedol gives away a territory at the top from lefttop to right top of around 70 points.

It's not 100% territory yet for alphago. Then he goes toy around the board and doesn't ATTACK.

Monte Carlo type programs you must put under pressure so that they fail low as then the horizon effect is total horrible with those programs as they don't know how to write a proper search...

You saw that in game 4 good example of how it extended a dead group and started giving away moves and how putting pressure had it play the move F11 which later caused it to miss a crucial tempo for the center fight.

If you give computer 70 points near the corners and edges then that's very easy to calculate for a machine how to make that its territory.

Not understanding your opponent is the dumbest of all mistakes...

p.s. i do expect a win in game 5 for Lee - now that he has fulfilled his contractual duty first 3 games he is free to play...

Also realize that i say already for years that if you DO make a proper search that's getting the 30 ply search lines without random monte carlo nonsense that could reduce that at a point you don't want it to get reduced, that it's easier to tactical outsearch go players than chess players.

Yet most likely they will can the program and it will disappear just like how it came there of course, after this match. So for a serious search algorithm for go there is no one gonna do effort to implement it.

You need a critical mass in terms of nodes searched...

[Daniel Shawul]

A few days ago you were saying the 2-dan guy (Fan Hui) was weak. Now that alphago has actually won a match against a top player, you come up with conspiracy theories. Where were you the first three games btw ? It is time for you to face reality for once and grasp the world doesn't revolve around you. You are completely ignorant of the state-of-the-art go before alphago, with your 30+ deep search nonsense. It is not search if you don't have good evaluation! I know this because I have actually tried it, learned my lesson, then implemented MCTS.

The way i see it many chess programmers ego including yours is seriously bruised, cause they thought brute force with a pile of processing power is the answer to everything (I think 42 is a better answer). That approach of deep blue has brought nothing to AI. You can sit back, relax and watch the rise of the machines with deep learning!

[Dirt]

You do believe a 9-dan pro calling Lee Sedol playing slack moves?

I've seen many people saying Hikaru blundered in playing Nxg3. The effrontery!

[Laskos]

My experience of playing human chess masters is that I think I'm doing better than I expected, then suddenly a win for the opponent emerges.

If Crazy Stone genuinely had a good evaluation, it would be able to beat human opponents. Maybe it is weak at evaluating the "frameworks" that will eventually become territory?

It might also be related to some deeper tactics too, from what I saw, these "weak" (much stronger than me anyway) engines lose large fights too to strong humans, so it's not clear to me whether the general assessment of the position is to blame for their weakness.

AlphaGo lost to tesuji, so it seems I was about right. The evaluation can hardly help in unique long line fights, MCTS seems to be to blame. Let's see in the 5th game if it is a systematic weakness.

[Uri Blass]

I wonder what is the basis for alphago's rating

I think that it should be more than 12 elo better than Lee based on the results.
http://www.goratings.org/

1 Ke Jie ♂ cn 3621
2 Park Jungwhan ♂ kr 3569
3 Iyama Yuta ♂ jp 3546
4 Google AlphaGo 3533
5 Lee Sedol ♂ kr 3521

[Laskos]

I wonder what is the basis for alphago's rating

I think that it should be more than 12 elo better than Lee based on the results.
http://www.goratings.org/

1 Ke Jie ♂ cn 3621
2 Park Jungwhan ♂ kr 3569
3 Iyama Yuta ♂ jp 3546
4 Google AlphaGo 3533
5 Lee Sedol ♂ kr 3521

Maybe they included the 5-0 result against Fan Hui too? Bayeselo, depending on parameters, can be a bit weird with the 5-0 results against 600 elo points weaker opponents, it might say that against Fan Hui 5-0 means a performance of 3400. Against Lee Sedol 3650, so all in all roughly say 3525.

[Isaac]

I wonder what is the basis for alphago's rating

I think that it should be more than 12 elo better than Lee based on the results.
http://www.goratings.org/

1 Ke Jie ♂ cn 3621
2 Park Jungwhan ♂ kr 3569
3 Iyama Yuta ♂ jp 3546
4 Google AlphaGo 3533
5 Lee Sedol ♂ kr 3521

Maybe they included the 5-0 result against Fan Hui too? Bayeselo, depending on parameters, can be a bit weird with the 5-0 results against 600 elo points weaker opponents, it might say that against Fan Hui 5-0 means a performance of 3400. Against Lee Sedol 3650, so all in all roughly say 3525.

They did. Click on "alphago" in the rating list. It displays all the games that counted to establish the rating. They didn't consider the blitz games it lost against Fan Hui, only the 5 LTC ones. And they did not yet included the 5th game against Se Dol.
Also since it's using WHR system, Alphago's rating will change over time even if it doesn't play. Because Fan Hui and Se Dol's ratings will fluctuate either because they'll play or because their past opponent's will play or the opponent of their opponent, etc.

[Leto]

I wonder what is the basis for alphago's rating

I think that it should be more than 12 elo better than Lee based on the results.
http://www.goratings.org/

1 Ke Jie ♂ cn 3621
2 Park Jungwhan ♂ kr 3569
3 Iyama Yuta ♂ jp 3546
4 Google AlphaGo 3533
5 Lee Sedol ♂ kr 3521

Maybe they included the 5-0 result against Fan Hui too? Bayeselo, depending on parameters, can be a bit weird with the 5-0 results against 600 elo points weaker opponents, it might say that against Fan Hui 5-0 means a performance of 3400. Against Lee Sedol 3650, so all in all roughly say 3525.

They did. Click on "alphago" in the rating list. It displays all the games that counted to establish the rating. They didn't consider the blitz games it lost against Fan Hui, only the 5 LTC ones. And they did not yet included the 5th game against Se Dol.
Also since it's using WHR system, Alphago's rating will change over time even if it doesn't play. Because Fan Hui and Se Dol's ratings will fluctuate either because they'll play or because their past opponent's will play or the opponent of their opponent, etc.

The 5th game against Lee Sedol which was a win for the machine has now been included. AlphaGo is now the second highest rated Go player with a rating of 3586.

[whereagles]

AlphaGo was given an honorary 9th Dan certificate. Cute