Author: Robert Hyatt
Date: 07:49:52 10/30/03
Go up one level in this thread
On October 29, 2003 at 09:28:12, Uri Blass wrote: >On October 29, 2003 at 09:01:28, Anthony Cozzie wrote: > >>On October 29, 2003 at 04:47:31, Daniel Clausen wrote: >> >>>On October 29, 2003 at 03:15:23, Jorge Pichard wrote: >>> >>>>"Experiments in Chinook show that there comes a point where increased search >>>>depth provides diminishing returns." >>> >>>Many chess programmers agree that the search- and the eval- part of an engine >>>have to be tuned so they work optimal together. (like you can throw out certain >>>parts in the eval since they're now covered with a better/faster search etc) >>> >>>Now you take an engine, which is optimized for todays hardware to reach a >>>certain depth in typical middlegame positions and make the experiment of >>>increasing search depth. Why can't the effect of "diminishing returns" not be >>>explained by the fact that search and eval are no longer working together >>>optimal? >>> >>>It seems to me that in all these experiments which try to prove the effect of >>>deminishing returns, the errors bars are bigger than the effect they want to >>>prove. >>> >>>Sargon >> >>One experiment that both Hyatt and Heinz did was to run their programs through a >>"deep" (14 ply) search and record the frequency of new moves, that is, what % of >>the time the program changed its mind. Their conclusion was that we still >>haven't reached the point of diminishing returns. Of course, this was some time >>ago, and evals/pruning/extensions/etc have all improved since then . . . >> >>anthony > >You cannot get that conclusion from that data. You are asking for too much. There are two questions: Does diminishing returns happen? Does diminishing returns not happen. We can't prove the first question yet, because we can't search deeply enough to prove that it does happen. We can't prove the second question either for the same reason. All we can say, with significant accuracy, is that through a nominal search depth of 15 plies, diminishing returns does not seem to occur. It might at 16 or 17 for all we know, but nobody has been willing to spend that much time to search that deeply. In another year or two we can re-visit this experiment again and probably hit 16-17 plies in 24 hours. And then we will either prove it happens, or leave the question open. It is not going to be possible to prove it doesn't happen, because eventually it must be true, because once we see a forced mate from the starting position, another ply will not help at all. Fortunately, that is a few trillion years from happening. >The frequency of new moves proves nothing about diminishing returns because >it is possible that at small depth the new moves changes the result and at high >depth they do not change the result. > >You can also never get the conclusion that there is no diminishing returns >unless you get the conclusion that there is an increasing returns from time and >the reason is that it is possible that you need more games to prove diminishing >returns. > >Uri
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.