Re: [computer-go] Computer Go hardware

["Frank de Groot" <frank@xxxxxxxxxxxxx>]

> In my case, filling up the huge block of memory with learned interesting
> shapes is what takes the time.  The actual game play of a fully-learned
> system is not slow (except for the major L2 cache inefficiencies).

Yes, when I use the pattern system to play it takes 1 mS/move.
But filling up the memory with interesting shapes takes 25 hours for about
460,000 games. This is totally dependent on CPU speed. I could have sped it
up in 3 ways:

1. Do not use 8 classes with their own delta-update code but put all in one
data structure and do not do any updating more than once,

2. Convert more to MMX

3. Instead of initializing complex structures with every move, cache this
data.

I have experimented with all three and the speedup was 2x in total but there
were bugs in all 3 attempts so I gave up. Accuracy is better than speed in
this case.
I had a cases that I lost months and later weeks due to teeny weeny bugs,
resulting in much worse performance than possible so I had to re-do it.

You gain more by optimizing the learning algorithm instead of the learning
code.

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