Life In 19x19

I just briefly went through all five games. My observations
- Fan Hui's openings are very traditional and peaceful. AI probably knows all these openings very well from past pro games. So there is nothing human player can gain at this stage. It remains to be seem how the AI reacts to the more 'modern' and 'imaginative' ways younger Chinese and Korean players play (for example how would the AI reacts to situations as in this game: http://www.go4go.net/go/games/sgfview/52548 (moves 28 through 68).
- Fan appears to make a number of mid-game mistakes. But so does the AI player (for example, one obvious mistake was move 60 in game 3).

macelee wrote:I just briefly went through all five games. My observations
- Fan Hui's openings are very traditional and peaceful. AI probably knows all these openings very well from past pro games. So there is nothing human player can gain at this stage. It remains to be seem how the AI reacts to the more 'modern' ways younger Chinese and Korean players play (for example how would the AI reacts to situations as in this game: http://www.go4go.net/go/games/sgfview/52548 (moves 28 through 68).
- Fan appears to make a number of mid-game mistakes. But so does the AI player (for example, one obvious mistake was move 60 in game 3).

It definitely seems to pull openings from pro games, which makes sense. It played the large avalanche perfectly, and the joseki after microchinese - split - attach that Fan Hui used both times. I also noticed AlphaGo played nirensei both games as black, making a sanrensei for one of them. It's interesting to me because black nirensei isn't that popular with pros today.

AlphaGo does vary its openings at least some, since it faced an identical position and chose the avalanche for one game and the hane for the other.

Would be interesting to see strong pros review this game.

This is probably a high insei level game in asia.

Some quote from the paper
https://storage.googleapis.com/deepmind ... ing-go.pdf

We trained the policy network pσ to classify positions according to expert moves played in the KGS data set. This data set contains 29.4 million positions from 160,000 games played by KGS 6 to 9 dan human players; 35.4% of the games are handicap games.

The final version of AlphaGo used 40 search threads, 48 CPUs, and 8 GPUs. We also implemented a distributed version of AlphaGo that exploited multiple machines, 40 search threads, 1202 CPUs and 176 GPUs.

The results of the tournament (see Figure 4,a) suggest that single machine AlphaGo is many dan ranks stronger than any previous Go program, winning 494 out of 495 games (99.8%) against other Go programs. To provide a greater challenge to AlphaGo, we also played games with 4 handicap stones (i.e. free moves for the opponent); AlphaGo won 77%, 86%, and 99% of handicap games against Crazy Stone, Zen and Pachi respectively. The distributed version of AlphaGo was significantly stronger, winning 77% of games against single machine AlphaGo and 100% of its games against other programs.

AlphaGo was using 48 CPUs + 8 GPUs. CrazyStone had 32 CPUs, Zen had 8 CPUs, Pachi had 16 CPUs. They were using 5 seconds per move.

macelee wrote:Fan appears to make a number of mid-game mistakes. But so does the AI player (for example, one obvious mistake was move 60 in game 3).

I wondered about that. Playing out ko threats is usually poor, especially when they lose endgame points. But "making the next threat bigger" can be justified. (There was a ko top right, and I don't suppose it was over the analysis horizon.)

But I doubt we're going to be able to get inside the head of this type of algorithm, with our usual concepts. Over the five games it appears to be at home in the middlegame.

I'd be very interested to see how it performs in complex ko fights. I suspect we'll get that chance.

Will Lee get to practice against it? He may discover its weaknesses. I suspect the kind of Go he usually plays to beat other humans may be quite different to the type of go that will do best against this AI.

Uberdude wrote:Will Lee get to practice against it? He may discover its weaknesses. I suspect the kind of Go he usually plays to beat other humans may be quite different to the type of go that will do best against this AI.

Hmmm. I would know more about pro psychology if I were five stones stronger.

Some pros wouldn't want to change their personal style just "to win". Striking comment from Ishida when he first beat Rin to become Honinbo, that he had had to adjust his style and that wasn't too satisfactory. An underdog thing, maybe.

Prepared variations? Quite in the Korean repertoire, and there are zillions of unorthodox ways to play.

Wacky openings or lots of ko* were some of the sort of things I was thinking of, but there could be other subtle changes in style which cannot readily be put into words or understood by weaker players. I recall one of the earlier neural net bots had a strange weakness in one game in that it had captured some stones with the crane's nest tesuji, but when its opponent made what should have been a futile attempt to escape the neural net it let them do so, presumably because the training data of expert games did not feature these futile attempts and so it didn't know how to continue correctly.

* I asked one of the people involved:
Andrew Simons: Does it have any expert knowledge or special approach to deal with ko? That's something many bots struggle with. And congratulations on a remarkable achievement!
Lucas Baker: Ko is actually not too hard given good search! But some expert knowledge programming in general is definitely required.

I find myself thinking that if Google can analyze go to this depth then just think what they can do with all the data they collect against you every day when you use any of their products.

This data set contains 29.4 million positions from 160,000 games played by KGS 6 to 9 dan human players

Is it just me or does that seem like a pretty bad dataset to try to beat Lee Sedol with?

To this layman, almost as surprising as the main news is the claim that

Without any lookahead search, the neural networks play Go at the level of state-of-the-art Monte-Carlo tree search programs that simulate thousands of random games of self-play.

Does this mean that such version plays virtually instantly, even on off-the-shelf hardware? It seems hard to believe that such a program, without any tree search, would beat the existing programs that took years to refine.

palapiku wrote:

This data set contains 29.4 million positions from 160,000 games played by KGS 6 to 9 dan human players

Is it just me or does that seem like a pretty bad dataset to try to beat Lee Sedol with?

This is really important, fellow go players. If AlphaGo wins, any player whose game is in that dataset can make the following claim:

"AlphaGo studied my games in order to beat Lee Sedol."

(You may already make this claim about Fan Hui, if you like.)

It will be at least a possibly true statement, in the same way that a member of a firing squad can claim that maybe it wasn't his/her own bullet that made the kill. That tesuji you played in that game a couple of years ago might just push the network over the critical threshold to find the winning move against Lee Sedol or anyone else AlpahaGo plays. (It could also be the mistake that was punished by your opponent, but thanks anyway.)

So, thanks, KGS 6-9d players! (Also, some of those datasets contain games that have at least 1 player that is in that range, so maybe weaker players are also included. Bettter check to be sure. )

How much did Google (a huge commercial venture) pay to use the KGS database (a private venture)? It should be enough to pay the KGS owner and developer (WMS) to retire and make all those requested improvements to the server. If Google can play Lee $1M, then it can pay KGS a portion of that for use of its data.

palapiku wrote:

This data set contains 29.4 million positions from 160,000 games played by KGS 6 to 9 dan human players

Is it just me or does that seem like a pretty bad dataset to try to beat Lee Sedol with?

A bit beyond my pay grade, but it seems like they are using the games as a "seed" of sorts to train with: https://news.ycombinator.com/item?id=10982243.