Author: Sune Fischer
Date: 17:17:24 12/06/01
Go up one level in this thread
On December 06, 2001 at 19:31:13, David Rasmussen wrote: >>Okay, then try and calculate the odds of getting 300 collisions out of 10000 >>using "danish: tilbagelægning" from a pool of 4 billion, those odds are >>astronomically small for sure, so something strange _is_ happening. >> > >I don't necesarily think that that is an accurate model for the situation here. >I trust my emperical data. It's the model that needs to be fixed IMO. I don't see why that should be any more accurate. The model is very simple, the program is not, I have one line of math, you have a hundred of lines of code. You need to point out at least one flaw in the model if you want to dispute it. The fact that it is a simple model is not enough to say it is incorrect. I've have used upper bounds on all the numbers, which mean they should describe the worst case senario. If you and Bob have 10 times as many pawn positions, then there is nothing strange at all with you having collisions, but since you won't test for that we have no way of knowing. >>>I don't. This is not hard evidence. This is theory based on false assumptions >>>and a wrong model. >> >>Feel free to name 1! :) >> >>>Hard evidence is hundreds (or thousands or millions if you >>>want) of pairs of positions that has the same pawn hash signature in Crafty. >>>That is by definition a collision. >> >>Maybe so, but you think you are proving that 32 bit keys are no good, when all >>you are proving is that you have _some_ bug IMO. >> > >Then a lot of us have the same bug. We see exactly the same behavior. Same >rates, bursts of clusters of collisions etc. Who is the rest of you? Can't any one of you test how many positions you get? That number is the key to the mystery. >>Me too, espicially since it was him who talked Bob into it in the first place;) > >He's not really sure. He got the idea from Bruce, yes? >>But you've seen my results and they confirm "my theory", so that would be a very >>strange double bug in any case. >> > >Your program stands out, then. From yours? Very small set to do statistics by. >>>Talk to Hyatt about that. As for me, I have tried many different PRNG's. All >>>with similar results. >> >>It is very strange indeed, but 32 bit seems to be working for me (for what ever >>reason), so I will change from 64 to 32 soon :) >> >>-S. > >That's understandable. Just be sure you don't have a buggy test, as you are >implying that I have. > >/David The test may show collisions, but then the question is why you get them. I can think of only two reasons: 1. You have a bad zobrist table 2. You have many more positions than me. I vote for number 2, but obviously I can't prove that since your numbers seem to be a well guarded secret. -S.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.