Human-Computer Go challenge, starting this Saturday

[maproom]

A three-game match will be played on KGS between Franz-Josef Dickhut, German 6-dan, and Crazy Stone, 6-dan program by Rémi Coulom.

It will be played over three Saturdays, October 4th 11th and 18th, starting at 16:00 CEST, which I believe is 14:00 UTC. You will be able to watch the games as they are played on KGS, in the 'Computer Go' room. The players' KGS accounts are 'fj' and 'CrazyStone'. The games will be even, with 6.5 komi, and one hour each plus five 40-second byoyomi periods.

The event is sponsored by codecentric, and called the "codecentric go challenge 2014". You can read the transcript of interviews with Franz-Josef and Rémi at https://blog.codecentric.de/en/2014/10/ ... mi-coulom/ - both think Franz-Josef is more likely to win. Once each game has been played, a commentary will become available at https://go.codecentric.de/ .

Crazy Stone will be running on 20 cores, on which it can achieve about 40,000 playouts a second.

[S2W]

Sounds interesting. I have so many questions - here are a few:
1. If this is a special event - how did crazy stone get its rank?
2. Who decides on the number of cores/hardware/time? Does it factor into the rank?
3. How does crazy stone use its time (Does it play the same time each move, does it use all the time? Does it read during its opponents time?)

(Not trolling just curious)

[yoyoma]

Sounds interesting. I have so many questions - here are a few:
1. If this is a special event - how did crazy stone get its rank?
2. Who decides on the number of cores/hardware/time? Does it factor into the rank?
3. How does crazy stone use its time (Does it play the same time each move, does it use all the time? Does it read during its opponents time?)

(Not trolling just curious)

CrazyStone is 6d on KGS (in fast games -- 10 x 15s byoyomis): http://www.gokgs.com/gameArchives.jsp?u ... 14&month=3

I don't know the details of how they picked the hardware. The blog makes it sound like Remi is just using what he has available to him through his university. Yes it has a "very big strength difference" [quoting Remi]. I think CrazyStone that got 6d on KGS was probably already using a similar cluster of machines to get that rating though, so I wouldn't expect it to suddenly be much stronger. Also that 6d is on blitz games, the bots typically lose about 1 kgs stone when playing slower time controls.

I'm looking forward to the games, I hope CrazyStone can win at least 1. But I wouldn't be surprised if it lost all 3.

[maproom]

To answer the third question (yoyoma has given good answers to the first two): Crazy Stone does time management, just as human players do. It spends more time on mid-game moves than on end-game moves; the time it spends it influenced by how much it has left; it plays obvious moves faster than difficult ones. Thinking in the opponent's time is called "pondering" by computer Go people - I do not know whether Crazy Stone ponders.

[maproom]

Correction:

It's a five-game match. The first three games are scheduled for October 4th, 11th, 18th. The other two (which may not be needed) have not yet been scheduled.

[hyperpape]

And CrazyStone won the first game by 1.5.

[Uberdude]

CrazyStone was leading by a lot more but did the traditional Monte Carlo bot points haemorrhaging:

http://files.gokgs.com/games/2014/10/4/fj-CrazyStone.sgf

[Krama]

CS does that a lot actually. If it is leading and is certain in victory it will give you points.

Same thing happened against Yoda Norimoto when cs was winning by 10 points or so but in the endgame it won only by few points.

Yoda later said that he was grateful towards the bot since it reduced the winning margin.

[emdio]

These are CrazyStone's analysis on the game
http://remi.coulom.free.fr/CrazyStone/a ... index.html

BTW, it seems the first game was played on Rémi 8 core desktop computer.

[snorri]

These are CrazyStone's analysis on the game
http://remi.coulom.free.fr/CrazyStone/a ... index.html

BTW, it seems the first game was played on Rémi 8 core desktop computer.

Interesting. I look forward to the publication of "Crazy Stone: The Only Move"

See this evaluation of move 119.

It is amazing how little calculation is given to moves other than R10 at that point.

[Unusedname]

These are CrazyStone's analysis on the game
http://remi.coulom.free.fr/CrazyStone/a ... index.html

BTW, it seems the first game was played on Rémi 8 core desktop computer.

0 Q4 1024000 0.489239

[0.489239 = P(B Wins)]

Is this an estimation of the value of komi or does CrazyStone have some confidence issues?

4-4 was a fine move CStone
No need to doubt yourself.

edit: Also for those lazier than me

16:00 CEST is 7AM west coast. 10 AM east coast. (If I'm reading this chart correctly...)

[Cassandra]

0 Q4 1024000 0.489239

[0.489239 = P(B Wins)]

Is this an estimation of the value of komi or does CrazyStone have some confidence issues?

From the 1,024,000 playouts, which CrasyStone calculated, Black won 500,980.

Another example:
nnn Xm 100,000 0.75

From 100,000 playouts of CrasyStone, Black won 75,000.


From the programme's point of view, P(B Wins) is the probability of a Black win, with the given move played next.

[goTony]

Very interesting what is the current status? I find this match most interesting. I believe we are less than 5 years away from a new worlds best player..... Of course it makes me sad....

[daal]

Very interesting what is the current status?

Franz-Josef Dickhut is ahead 2 games to 1. You can look at the games and read a bit of commentary here: https://go.codecentric.de/

[goTony]

Very interesting what is the current status?

Franz-Josef Dickhut is ahead 2 games to 1. You can look at the games and read a bit of commentary here: https://go.codecentric.de/

Ty I will enjoy looking it up. Glad to see flesh and bones, is ahead.