Author: KarinsDad
Date: 20:36:40 01/29/99
Go up one level in this thread
On January 29, 1999 at 23:06:53, Dann Corbit wrote: >On January 29, 1999 at 22:40:25, KarinsDad wrote: > >>On January 29, 1999 at 21:45:54, Dann Corbit wrote: >> >>>Now, about computers -- how is it that an 1800 program might beat a 2500 >>>program? Sounds pretty absurd does it not? >>>There are a large number of possibilities. For instance, a flaw in an opening >>>book might cause Mr. 2500 to start off at -6 when it falls out of book. Or a >>>bug in the evaluation function might cause a stupid move. Or someone might get >>>on the time shared machine and cause the 2500 program to swap to disk for 30 >>>seconds at some crucial time point. Now, you might think that the bug can be >>>fixed. And certainly it can. But as long as we are developing new ideas new >>>bugs will spring up. And you might think that we can fix the opening book, and >>>indeed, we can tag bad moves as we find them, but with so many possible moves we >>>will never find them all. And you might think that "well, when I swapped to >>>disk, that game did not really count." But on the ICC database it counts the >>>same as a game against ferret on the hottest available hardware. >>> >>>IOW ECC.9:11 >> >>Dann, >> >>You are quibbling. >> >>Stating that a computer who suddenly lost ability due to people using it could >>lose and hence lose points is silly. This is like saying that a player could get >>poisoned in the middle of the game and lose accordingly. This would not >>generally happen today in a tournament for a computer (even if it might on ICC). >> >>And yes, a bad opening book could get you in trouble. But why would any writer >>of a program not use the program to examine every leaf node in the opening book. >>Even at 1 minute per leaf node on a 100,000 leaf book (probably at least a >>150,000 node book or more with transpositions) would result in 100,000 minutes >>or about 70 days. You could probably even get away with 2 seconds a move or 2+ >>days to get rid of any glaring mistakes. It would be remiss to not do so. It >>would be like not testing. >I have 100+ computers operating 6 days a week on the C.A.P. project (12 hours on >weekdays and 50+ hours over the weekend). In several months we have analyzed >150,000 positions (analyzed twice, of course) at 12 minutes of PII 300Mhz time >per position. That is just the tiniest start of a beginning of a suitable >opening book. Now, find out how much computer time that is. In a few months, I >hope to have a few million positions analyzed. Yet there will still be 10^47 - >10^6 (still about 10^47) holes in the book. You think I am exaggerating, but Dr >Robert Hyatt has a *very* carefully prepared book from GM games, and his >computer learns all the time. And yet he still pops out of book at -3 once in a >while according to a recent message. We will *NEVER* get a full-proof opening >book. But you raise a good point. We can form books that are orders of >magnitude better than what we have now. > Are you analyzing only leaf nodes, or at you going through positions from the start of games? Please explain your numbers. 10^47 appears to be about 16 to 18 moves down for each side down considering all legal moves. Since about 9.999999999999999999999999999999999999999999 x 10^46 of those moves are inferior, it seems likely that you are more or less covering the bases as it is. What am I missing? If the opponent makes an inferior move, then let a program's tactical engine figure it out. If the opponent makes a theoretical novelty, let your learning routines record it and analyze it later. >>Stating that the up and coming players would be the ones to win is quibbling as >>well. There are only a few really good up and coming players that can beat >>someone 1000 points higher. Especially not 1 player in 400. Why? Because even if >>that up and coming player was really 650 in reality lower, he should still only >>win against the higher player 1 time in 50 (according to the formula, but a lot >>less in my opinion). Very rare. >Well, it was speculation on my part. But I will bet that I am right. Do a >statistical study on SuperGM players beaten by players with a lifetime ELO under >1800. I found none, zero, zilch in a database with over one million games in it >(hundreds of thousands by GM's). But I did find some losses by persons who were >on their way up. I was just wondering how these things come to pass. Now I >have beaten players that are a lot better than me. But my recent painful >experience with GNUChess on a PII 300 taught me something. I don't think I >would ever beat that game. I felt the superiority and knew I was dominated. >It's sort of like "Can a 5 year old who has learned the basic rules beat someone >who has played for 20 years?" >Of course not. >So where do wins come from by much weaker players over much stronger ones? >They must come from somewhere. > Yes, I was once in a tournament where I was playing in the finals, 0.5 points back. My opponent was only 100 points above me, but practically every move in the opening until about move 20 was a solid, multi-threatening move (I felt like I was playing someone about 500 points higher at least). It was probably just that he knew his openings cold, but man did he dominate me. As it turned out, I changed it around and with 5 minutes left on each clock (this was a G60 game), I had a won game but didn't know it. It helps to know your endgames. With 30 people standing around watching our game at close range, the position looked drawn, so I offered him a draw and took second place. I learned a lot of lessons from that game. >>If an up and coming 1800 ELO player is really playing at 2100 strength, then he >>is a 2100 player, regardless of his rating. So you can get a high percentage of >>up and coming 1800 players beating 2200 players since their strength is really >>2100, but that same set of 1800 players would NEVER win against the 2800 (i.e. >>Kasparov level) player at tournament times. Certainly not 1 time in 400 (or 1 >>time in 55 as per the real 700 point difference). Do you see now why I think the >>formula is skewed at larger ranges? >There rating may be 2100, but it is only 1800 at the moment of the game. The >statistics don't know that the player is on the rise, nor do they recognize a >2500 player in steep descent from old age. > >>Granted, if you took a statistical look at the USCFs database, you would >>probably find that the majority of wins by extremely lower rated players are >>based on up and coming players (or losses by down and going out players) and >>there would be next to 0 wins at a 700 delta, not 1 in 55. >> >>There are a lot of players mis-rated when they first come in as well. If there >>was a better way to estimate a player right away (or at any point within their >>career), then statistically, they would skew the results less. >Agreed. This is another source of how 1800 players can beat someone with a much >higher ELO. They really have a much higher ability than their rating suggests. > All of that stuff is *factored into the formula* because it is based purely on >measurement. > >>But I guess the bottom line is, until we can look at the statistics for a LOT of >>games (say millions), you and I will only be speculating. >Figures don't lie, but liars will figure, I guess. Yup :) KarinsDad
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.