Author: Robert Hyatt
Date: 06:16:02 05/24/00
Go up one level in this thread
On May 24, 2000 at 01:53:48, blass uri wrote: >On May 24, 2000 at 00:17:35, Robert Hyatt wrote: > >>On May 23, 2000 at 22:54:45, Peter Kappler wrote: >> >>>On May 23, 2000 at 21:30:26, Dann Corbit wrote: >>> >>>>On May 23, 2000 at 18:24:21, Mark Young wrote: >>>>[snip] >>>>>Lets be generous and say a ply is worth 100 rating points, and we can take Fritz >>>>>6a for an example. Now you said it does not matter if it?s the 1st going to 2 or >>>>>14 ply going to 15 ply. Its pretty much the same, if I understand you >>>>>correctly. Lets assume Fritz 6a plays at a 2500 rating +/- 100 rating points, >>>>>with a average middle games search depth of 15 plies. >>>>> >>>>>15ply X 100 rating points = 1500 rating. Where are the extra 1000 rating points >>>>>coming from? Is Fritz really rated 1500, or are some plies worth much more the >>>>>others. >>>>> >>>>>It is clear that the early plies are worth much more the later plies, and if you >>>>>plot it out it?s a curve. I don't know of one program that does not exhibit a >>>>>curve. >>>> >>>>That is well established, as both Dr. Hyatt's and Dr. Heinz's experiments >>>>showed. However, as the depths increased, two very surprising things surfaced. >>>> >>>>At extreme depths, a linear model fits just as well as an exponential one. >>>>Hence, there may (or may not be) additional loss in the value of additional >>>>plies. >>>> >>>>Far more surprisingly (to me at least) is that the number of fresh ideas do not >>>>drop off. IOW, if the program liked one move at ply 10, and another at ply 11, >>>>and yet another completely different one at ply 12, they can just keep coming up >>>>with new moves that have not been considered best at deeper plies. This one is >>>>(to me at least) both astonishing and counter-intuitive. Obviously, it can't >>>>possibly find more fresh ideas than the number of possible moves! >>>> >>> >>>I need to go back and re-read the "Crafty/Dark Thought Go Deep" articles. >>> >>>I hope that "new best moves" were only counted if the evaluation also changed >>>significantly. If the change was just a few centipawns, then I think it's >>>misleading to report the new move as "better". >> >>Why would you want to call such moves "not better"??? _many_ moves made by >>today's engines are just "slightly" better than other moves that would be >>made at shorter time limits. But either you believe that .01 is better, or >>you have to ignore evaluations completely. >> >> >> >>> >>>This stuff interests me, because I fundamentally believe that playing strength >>>must diminish with each additional ply. This topic was discussed a couple of >>>months ago, and I was surprised that many people here don't agree. >>> >> >> >>The problem is that the only "old" evidence was self-play with varied depth. >>Which is probably not a good test, since small changes are often magnified in >>self-test play, while at other times small changes have no effect at all. >> >>New evidence suggests that at least thru 15-16 plies, things are still picking >>up with each additional ply... >> >>Which you believe is up to you... >> >>Neither is exactly overwhelming evidence... >> >>DB did make a statement however, with its 17 ply searches. DT at 10-12 plies >>got destroyed by Kasparov. DB at 15-18 plies played far differently. > >We cannot learn much from comparison between deep thought and deep blue because >they had not the same evaluation function and deep thought had search problems >and could not detect repetitions in the search. > >Uri It could detect repetitions, just not in the last 4 plies (which was done in the hardware).
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.