Author: TEERAPONG TOVIRAT
Date: 03:04:50 07/02/00
Go up one level in this thread
>Yes, sure, you are right, and I think I do understand it quite well. >I just wanted to say, that the resulting approximate formula for the >error probability depends on N and H, and not on P (IMO). Thanks for your interest to my post. No one besides you seems to do so. Perhaps it's too difficult for me to express the third factor (P/H ) in detail(English). Let me try. Do you notice N and H are relatively constant or controlled by limitation of RAM ? In my hypothesis I base on assumption that every hash value has approximately equal number of positions,but I think in the actual situation is different. I don't know about the real distribution. I know only a lot of people here say " not all random numbers are suitable for hash value some of them leading to clash so often than others" . So, the third factor should be the real number of positions that have the same hash value not just by approximation. And now if my hypothesis is correct. Will it be useful? As I said above we cannot control N,H. I think we can detect a good series of random numbers by using these facts. In last few days I spent hours in experiment about this. I test my checkers program with 32 bit hash table. The results are quite the same as I expect. At the early stage of game the incidence of the type 1 error is so small. As the number of positions occupied on the table increase the incidence also increases until the table is fully occupied the incidence tend to be stable at some constant. My current problem is my incidence of error is higher than I expect. I got 0.5% instead of 1/4000. Do you think I can blame it on my random integers? or Do you have any idea to generate a good series of random number? I'm sorry it's quite a long post. Thanks, Teerapong
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.