Author: Stephen A. Boak
Date: 02:41:27 04/28/01
Go up one level in this thread
On April 28, 2001 at 03:34:38, Uri Blass wrote: >On April 28, 2001 at 03:29:46, Uri Blass wrote: > >>On April 28, 2001 at 02:08:00, Stephen A. Boak wrote: >> >>>On April 28, 2001 at 01:52:01, Stephen A. Boak wrote: >>> >>><snip> >>> >>>>This may indicate, for example, that there are relatively as many win, loss & >>>>draw nodes, generally speaking, at each fixed ply depth, no matter how many >>>>plies are calculated--even if those results are not precisely calculatable by >>>>the program! Therefore the program that calculates x plies more than its >>>>opponent will have approx the same increased chances to steer toward the >winning lines. [I hope you can understand the concept I am trying to >communicate.] >>> >>>I mean that the general percentage of calculated win nodes, arrived at by >>>calculating to a fixed N+X depth, which are *not* seen by calculating only to N >>>depth, may be approximately the same no matter what N is. At least during the >>>opening & middle games. >> >>I do not believe that it is so simple. >> >>I believe that there are positions when program A understands better than >>program B and if program A is lucky to get them then 10 plies may win against 15 >>plies but 5 plies does not win against 10 plies because 5 plies does tactical >>mistakes that are decisive. >> >>There are mistakes when some programs may do even at depth 15 when other >>programs may avoid them at depth 10. >> >>I expect to see more wins for the weaker side at higher depthes because of this >>reason. >> >>Unfortunately the depths that I played are not big enough and doing a match at >>depth 10 against depth 15 may take a long time and I expect average of at least >>some hours per game even after upgrading my hardware. > >I can add that my tests were only 3 against 8 plies and I even did not test 5 >against 10 plies. > >I did today one of the 3 plies against 9 plies and I got the first 50-0 result > >Deep Fritz(depth 9)-Tiger14(depth 3) 50-0 > >Uri I don't know how to assign a relative rating to a 50-0 score. Intuitively it just won't work, heh heh! But after you run the opposite test, DF (depth 3) - T14 (depth 9), and report the results, I'll certainly try. =(:0[) I sense that we have now reached a point of diminishing returns! With lot sizes of merely 50 games per ply vs. ply test, and all ply deltas of 6 or more producing perfect scores, we will be unable to distinguish any further improvement in program scoring--for some reason. :) As you proceed to test 7 ply and 8 ply deltas, we will be unable to discern a difference between, say, a relative rating difference of 5000 Elo points and 6000 Elo points. All will produce 50-0 scores, with the odd game in a million, billion or trillion (maybe more) producing a lucky draw. The difference between their Win Expectancies will be far less than the 0.5 score of a single draw in 50 games. Thus the test will fail, in practice, to discern diminishing returns--for lack of sufficient trials. Doesn't this say something about the intractability of testing effect of depth calculations on scoring (and hence Elo rating)? Since the number of possible positions grows exponentially, how can we ever test the notion of diminishing returns. We will rapidly run out of patience when we learn that it will take billions of games to possibly detect a diminishing return effect! Even our relentless (pathetically slow) technical growth in processor speeds will not handle this exponential medusa. We will need .0000000000000001 micron technology, and greater than speed of light transmissions without unwanted black hole tunnelling effects! Perhaps quantum computers, some day far in the future, will be able to test this idea. Or better fractal mathematics (whatever that is). Well, at least we have a fair number of results for ply deltas from 1 to 5. I guess that sometimes the question 'what is the next number in this series' is not solvable. Reminds me of some prime number generating formulas--they work for the first 5 or so examples, then after that the numbers are so huge we don't even know what they are, and don't have the time or resources to calculate them, to test whether the formula continues to generate prime numbers! Or the concept that a large army of monkeys could, given enough time, play around with keyboards until they finally generate enough text to exactly duplicate all of Shakespeare's plays. I guess if I have infinite patience (not to mention the monkeys too) I could test that idea also. These are the thoughts of a finite person, contemplating the relative infinity of chess! --Steve
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.