[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [computer-go] SlugGo v.s. Many Faces

On Aug 25, 2004, at 8:50 AM, wagjan@xxxxxxxxxxxxxxxxx wrote:


Hi,

I suppose that there is a small branching factor at all level, or at
least it is possible to branch at any level. But if there are 64
processor, then averages barnching factor can not be higher then
1.3 or searching time will be exponentional. But such a deep
search seems to be meaningless except if there is a *very* good
move selector. It would be more interested to take a bit wider
search instead. I personally like wider search near the root. Eg.
trying 4 moves at 1st ply and only two candidates at 6th ply.

Or I am just the next who misses the point ? :)))))))

Janos Wagner


I said:

The time is linear in the depth of search. The width of the search
takes
virtually no wall clock time because different search paths are each
done on a different cpu in the cluster. At each branch new cpus are
brought into the calculation.



Marco said:



It was also my understanding that there is currently no branching

except on the first ply, that after the 1st ply the look ahead only

examine one move. There's no minimax involved yet. Is that correct?  
And

adding minimax would certainly be fun, but require many cpus!





Marco's paragraph correctly describes what we did to get more than 5  
stones

stronger against GNU Go (using only its own code) and to beat MFG, but  
not what

we are limited to doing. We have 72 cpus and can branch at will. We can  
overload

the cpus with multiple processes if we wish. The space of possibilities  
is very large,

but with only one summer intern and bringing up the cluster in the  
first place, there

has only been time for trying the most obvious and simple things.



Almost all of the Go playing knowledge and programming credit goes to  
the GNU

Go developers. The only thing we have added is a cluster and additional  
reliance

on the Go proverb "Your opponent's best move is often your best move."  
We use

GNU Go internal evaluation functions to get whole board evaluations of  
influence

and territory, and do move based weighting of those results with  
influence weighted

more at the beginning of the game and territory weighted more late in  
the game.



The main point is this: something very simple worked to get a few  
stones stronger.

These results are still preliminary and there are many many other  
things worth trying.



Snipped from the discussion on the GNU Go development list:

***** snip *****

Check out the diagrams at

	

http://cowboy.directgames.net:81/research/SlugGo/documentation/sluggo- 
system_overview.pdf



We have not yet implemented the hash table functions indicated on the  
diagrams.

These functions are being implemented in an attempt to save the time of  
recalculating

the lookahead moves we have previously generated.



The curves on page 5 indicate weighting functions that are used as  
multipliers for the

values returned by the influence and territory functions.



I want to emphasize that these results were obtained with very little  
tuning of the very

many parameters we have in our additions to GNU Go, and thus these  
results are still

preliminary. There are a large number of things we can do which we  
believe will make

the play stronger.



In the "Big Picture" sense, we have applied some simple ideas with lots  
of computing

power to see what we could find. With respect to the recent thread on  
the subject of

fast v.s. slow computing, I think this is only a glimpse of what can be  
done with plenty

of silicon and longer compute times.

***** end snip****



I hope that this makes things clearer.



I also hope that future computer Go tournaments loosen their rules to  
allow distant

cluster computers the opportunity to compete. I cannot pack my cluster  
for travel and

doubt that an arbitrary tournament site will be able to supply the  
cluster with the

amount of power it needs.



Cheers,

David



_______________________________________________

computer-go mailing list

computer-go@xxxxxxxxxxxxxxxxx

http://www.computer-go.org/mailman/listinfo/computer-go/