Re: [computer-go] Re: Sharing Secrets

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

> If your 64-bit keys give a colision probability p for searching, let's 
> say, 10^9 positions.
>
> How many bits extra would you need to get the same colision probability 
> for random keys?
Ah..
I don't know.

There are 2 aspects to this:

- Random has been discredited as being bad, varying from totally unusable to 
to just passable when the very best random generators are used.

- I proceeded with generating hash keys that had the maximum Hamming 
distance and I found that it was WORSE than using good random. I am sorry 
that 2 years ago I have not measured it (I never thought someone would ask 
about that).

But what happened was such bad performance that it became noticeable by many 
hash collisions as opposed to no hash collisions at all or the occasional 
one.

I then asked a few Chess programmers and I talked to some other guys and 
finally, experimenting a little, the best results are when you use a Hamming 
range for your hask keys and not try to maximise the Hamming distance and 
using only the keys with the highest mutual Hamming distance.

My Hamming range is between 23 and 42 (from the top of my head) with a 
certain Gaussian distribution. The internal representation of my Goban is a 
1-dimensional array with a length of 4096 so I need a very large number of 
keys, otherwise the range would have been narrower. 
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/