Re: [computer-go] Re: Sharing Secrets

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

> Just for the record, I did not use the idea with Hamming distances. It has 
> been on my to try list for a while, but so far I have not tested it. The 
> main reason why I am a bit interested to try something like it is that I 
> might get a more uniform distribution over the available memory. However, 
> I expect only very small gains.
My pattern system needs very good (totally random) memory distribution, in 
order to work with very fine granularity of the target blocks. I use 
1024x1024 blocks with each 8 hashes. This allows me to utilize 20 bits for 
something else.

This would not work if there were some statistical "bumps" in the first 20 
bits of the hashes because I would loose important pattern during 
harvesting.

To David Fotland I would suggest to change the hash keys to a maximum of 2 
below attainable maximum Hamming distance (use a very fast bit population 
counter in MMX or be prepared to wait forever) and let that version play 
against the old version.

I expect an improvement in case there is deep/wide search going on with a 
ten million, hundred million hashes generated.

To speak for my reason why I need lowest collision: My pattern database is 
huge. Many millions of patterns. The possible search pattern input is near 
infinite. The system matches everything it sees on the board to the database 
and it matches hundreds of patterns per position, worst-case.

Even a single collision would be disastrous because a very bad move could 
easily get a very high score. This is much less a problem with tree search 
where things tend to balance eachother out.


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