Crazystone with H4 beats Ishida

[tchan001]

http://mainichi.jp/english/english/news ... 7000c.html


Source of sgf: http://www.computer-go.info/h-c/

[logan]

mv 224, sweet

[ez4u]

In his blog, Takao Shinji was most impressed by . He wrote that humans would only think of cutting off the single stone with H4. J3 looks like an elementary mistake but if White connects at H4 with 53, White will play the atari at K2 and then capture everything with G6. After the plays at the bottom, Black remains securely ahead throughout the game, he concludes.

[Bill Spight]

One thing that struck me was how human like most of Crazystone's play was. I was especially struck by the play in the two top corners. I suspect that it is not just that the program has improved over the past few years.

In a way, it is a disappointment. One thing that I enjoyed about computer play was how frequently strong programs would tenuki or make seemingly vague plays that worked out later on. That was thought provoking, like meeting an alien intelligence.

[RobertJasiek]

Bill, yes. However, too many program inventions are pruned by human-like input in the code. This is not necessarily a sign of greater intelligence, but could be a sign of restricted intelligence on the altar of making the program as human-like as their programmers wish.

[tundra]

.

[jts]

One thing that struck me was how human like most of Crazystone's play was. I was especially struck by the play in the two top corners. I suspect that it is not just that the program has improved over the past few years.

In a way, it is a disappointment. One thing that I enjoyed about computer play was how frequently strong programs would tenuki or make seemingly vague plays that worked out later on. That was thought provoking, like meeting an alien intelligence.

I felt the same way (about human play), although I couldn't put it into words as precisely as you have.

[Boidhre]

What were the time settings for the game?

If it was blitz, the result is not so interesting, at least IMHO. But if the limits were more reasonable, say at least one hour of main time for each player, plus some form of overtime, it becomes much more interesting.

From here: http://jsb.cs.uec.ac.jp/~igo/eng/

30 mins absolute.

[Kirby]

Bill, yes. However, too many program inventions are pruned by human-like input in the code. This is not necessarily a sign of greater intelligence, but could be a sign of restricted intelligence on the altar of making the program as human-like as their programmers wish.

To have a measure of greater intelligence, it is necessary to specify what is meant by intelligence. I don't think there is a universal definition in the world of computer AI.

If we say that "intelligence" in a go program means that it is more likely to win a game, it seems clear that being more like a human does not necessarily imply this.

[Bill Spight]

A 64-core computer against a 64-year-old pro nicknamed "computer".

I wonder what threats CrazyStone saw at moves 126 and 224.

Maybe someone with experience with computer play can comment, but I think that I understand 126. I am pretty sure that CrazyStone builds a game tree with two parts. The top portion is like the game trees built by classical computer programs, an enumeration of pertinent variations that end in evaluated leaves. The bottom portion is (more or less) random play from each leaf to the end of the game, used to evaluate the leaf. (This portion may be ephemeral.) My guess is that some of these random playouts resulted in White making life or ko on the top side, which resulted in losses for Black. Black 126 reduced the probability of those losses, while eliminating the White sente on the edge. The same probably holds for 224, with a much lower probability of loss. Black 124 is also a safety first move.

[ez4u]

What were the time settings for the game?

If it was blitz, the result is not so interesting, at least IMHO. But if the limits were more reasonable, say at least one hour of main time for each player, plus some form of overtime, it becomes much more interesting.

From here: http://jsb.cs.uec.ac.jp/~igo/eng/

30 mins absolute.

Thirty minutes followed by single-period, 30-second byo yomi. From the Japanese tournament page.

[Rémi]

What were the time settings for the game?

If it was blitz, the result is not so interesting, at least IMHO. But if the limits were more reasonable, say at least one hour of main time for each player, plus some form of overtime, it becomes much more interesting.

From here: http://jsb.cs.uec.ac.jp/~igo/eng/

30 mins absolute.

This was the time control for the UEC Cup. The games against Ishida were played with a byo-yomi of 30 second per move. Crazy Stone was using 25 second per move of thinking time to leave 5 second for move input.

Rémi

[ez4u]

Hi Rémi! Could you comment on your experience? How was the tournament, the game against Ishida, etc.?

[Rémi]

Hi Rémi! Could you comment on your experience? How was the tournament, the game against Ishida, etc.?

The tournament was really nice. This year the organizers had more sponsors and funding, so they could offer to cover travel expenses of some participants. As a result, many western Go programmers attended (David Fotland, Martin Mueller, and Petr Baudis were there). All the top programs participated, except for Ginsei Igo (this year the organizers required the presence of programmers, which may be difficult for North Koreans). We had a lot of fun together.

You can see some photos of the final day there:
http://www.yss-aya.com/photo/20130316ue ... dex01.html

It was great to have the opportunity to play against Ishida. I operated Crazy Stone and was a bit nervous during the match. Below is the plot of the win rate:



The win rate dropped a bit when Ishida invaded at the top.

Rémi