Author: Robert Hyatt
Date: 07:13:42 12/09/99
Go up one level in this thread
On December 08, 1999 at 20:03:41, Dann Corbit wrote: >On December 08, 1999 at 19:56:04, Bruce Moreland wrote: > >>On December 08, 1999 at 18:28:26, Dann Corbit wrote: >> >>>On December 08, 1999 at 17:16:28, Luis E. Alvarado wrote: >>> >>>>I agree with you 100%. Humans tend to be far more inconsistent than computers. >>> >>>You might be surprised. Computers will play a real stinker move some times. >>>Could be an unknown hash collision. Could be a bad opening book entry. Could >>>be a program bug or just an inferior algorithmic decision. But for whatever >>>reason, computers screw up now and again, in a big way. >> >>I have never seen evidence of a hash collision causing a game to be lost. I >>have seen people suggest this as a reason for losing when a particularly bad >>clunker is made, but this suggestion is always made before investigation and >>never after. >> >>In college I used to do a lot of programming in a large open area full of a >>bunch of crazy people, on a minicomputer that we all shared. I noticed that >>whenever one particular guy had a bug, he would tell us all to save our work, >>because he believed that the mini was behaving strangely and was about to crash. >> >>This is the same sort of deal I think. >> >>Most bad moves are produced as a result of the program working as designed, and >>responsibility for this falls on the shoulders of the programmer. >Hash collisions are so improbable that it would be very difficult to pinpoint >that as the problem. But if you shrink the hash tables to 10K or so, I suspect >the problems will show up frequently if not handled correctly [which brings up >the point -- how do you handle hash collisions?]. Every increase in size will >make collisions less likely, but even with a terabyte a collision is possible. >That collision could cause a key choice to be made incorrectly. It's just not >very likely. The "size" of the hash table isn't an issue. Whether you store two positions or two billion positions doesn't matter. It is the size of the hash signature that is all-important. A couple of us ran some nearly impossible simulations a few years ago, and found that 64 bits effectively eliminates all collisions. Yes you must get a collision every few gadzillion nodes, but even when you do, it also has to happen in a place where it affects the score of the best move, otherwise it has no affect at all. The probability is close enough to zero to call it zero without losing sleep you are wrong...
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.