Re: [computer-go] SlugGo vs Many Faces, newest data

[David G Doshay <ddoshay@xxxxxxx>]

On Sep 7, 2004, at 3:56 PM, Tom Cooper wrote:

> I'd love to see slug go compete in the computer go ladder.
We have declined to play in the computer ladder for a few reasons. One 
reason is deciding if we want to dedicate a part of the cluster to it. 
At present we have not. The bigger reason is that the logic in SlugGo v 
0.9 is so obviously crude that to compete with it seems strange to me. 
I can easily see blunders in the way SlugGo plays that are due to the 
simplistic implementation. I would rather wait until I improve things 
at least to the point that the code implements my ideas the way I 
intended, rather than the way that was easy for an early 
implementation.

As I said at the start of all this, while encouraging, this is all 
preliminary. We have some specific improvements in mind that should 
improve speed and results.

>  Is it too slow at present?
Too slow is a relative term. We are not within the one hour limit that 
is the norm in computer Go tournaments.

We stopped taking timing data a while ago. The only data I have at hand 
is from some runs against GNU Go on our G4 cluster and our very first 
runs on the G5 cluster. This data is from source code that is earlier 
than the SlugGo v 0.9 code that played against Many Faces. GNU Go was 
playing at level 15. In the recent Olympiad GNU Go also played at level 
15.

Time is in seconds and is for 10 game matches.

G4 total time:
GNU Go	14,625	50,644	25,657	23,218	21,708	26,021
SlugGo	87,838	107,743	105,665	112,083	116,149	105,522

G4 longest single move:
GNU Go	134	19,798	183	102	193	510	
SlugGo	338	347		292	331	512	449

G5 total time:
GNU Go	3,250	2,634	2,438	2,382
SlugGo	68,219	56,467	52,671	50,903

G5 longest single move:
GNU Go	48	13	22	12
SlugGo	561	206	182	151

So, ignoring the outlier where GNU Go took 19,798 second to make one 
move,
SlugGo takes 4.9 times longer than GNU Go on the G4 towers, and 21.3 
times
as long on the G5 cluster nodes. The G5 cn's are 1.8 times as fast as 
the G4
towers for SlugGo, and an impressive 8.0 times as fast for GNU Go. We 
guess
that GNU Go must fit in the L2 cache on the G5, but have not 
investigated deeper.

Bottom line: an average of 95 minutes per game, with 2 hours also not 
unusual
on the G5 cn's. An average of 3 hours per game on the G4 towers.

Cheers,
David

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/