Author: Robert Hyatt
Date: 13:38:50 07/07/02
Go up one level in this thread
On July 07, 2002 at 11:42:54, Uri Blass wrote: >On July 07, 2002 at 10:10:07, Robert Hyatt wrote: > >>On July 07, 2002 at 01:44:27, Uri Blass wrote: >> >>>On July 06, 2002 at 23:31:15, Robert Hyatt wrote: >>> >>>>On July 06, 2002 at 18:48:07, Uri Blass wrote: >>>> >>>>>On July 06, 2002 at 17:19:21, Robert Hyatt wrote: >>>>><snipped> >>>>>> >>>>>>OK... first, me, in Cray Blitz. 1994. GCP modified Crafty and used the >>>>>>Hsu-definition of SE in doing so. >>>>> >>>>>And what was the result? >>>> >>>>It was tactically significantly stronger with than without. Unfortunately >>>>we had a severe bug in 1994 and did poorly. Harry had limited the max >>>>search depth to 64 plies to match the older cray vector length (newer machines >>>>had 128 word vectors but he wanted it to run on older machines too since he >>>>had access to several of them). He also took out the MAXPLY check. And it >>>>never caused a problem thru 1993 as the search extensions were pretty "sane" >>>>and didn't go that deep. But in a few cases in 1994, singular extensions >>>>drove the search beyond 64 plies with devastating effect on the chess board >>>>and the alpha/beta scores backed up as a result. We never had a chance to play >>>>real games to see how deep the SE stuff could extend, which was a common problem >>>>with cray access back then... >>>> >>>>Crafty I am not sure about yet. Mike Byrne has been playing with this >>>>further and seems to like the results he is getting. I have not yet looked >>>>at the changes he is using, but I will when I have time. It does seem to be >>>>very good tactically, producing some good WAC test scores in very short time >>>>limits. >>>> >>>>> >>>>>Is the new crafty with singular extensions better? >>>>>I guess it is worse because you do not use it in games. >>>> >>>>I never found something I like. However, I am not sure that aggressive >>>>null-move mixes well with cute extensions. They almost work against each >>>>other without some controls to limit this interaction... >>>> >>>> >>>> >>>> >>>>> >>>>> >>>>><snipped> >>>>>>>There are programmers who use singular extensions but I know about no programmer >>>>>>>of one of the top programs of today that use it in the way that they use it. >>>>>> >>>>>> >>>>>>So? They choose to implement a less-than-optimal version to control the >>>>>>computational cost. >>>>> >>>>>Deep thought already used singular extensions in the past and the top programs >>>>>of today that search similiar number of nodes do not use them in the same way >>>>>because they prefer to play better and not to use "optimal" extensions. >>>> >>>>How can you possibly say "play better"? The current SE approach used in Ferret, >>>>and the way I did it originally for Crafty, is simply a 'cheapo version" that >>>>misses things that deep thought would not miss, extension wise. But the cost >>>>was more palatable for the much slower hardware we have to use. However, HiTech >>>>used it at 150K nodes per second, so it worked for them. And it worked fine in >>>>Cray Blitz as well... >>> >>>You always give examples that are not about the default version of the top >>>programs of today(Cray Blitz or Hitech are history and Crafty does not use the >>>GCP version because you did not like the result). >>> >>>It is possible that some people believed that it is better and did not have time >>>to compare results with and without it. >>> >>>I want example of one of the top programs of today that use it and not some GCP >>>version of Crafty that is not the best crafty. >> >> >>Wchess from Dave Kittinger. Implemented 1/2 of Hsu's algorithm, namely the >>PV-singular test. >> >> >> >>> >>>My point is that if the algorithms that they used in deep thought are not >>>inferior then it is logical to expect one of the top programs of today to use >>>them(they already search similiar number of nodes). >> >> >>The "singular extension algorithm" is precisely defined. At any node in the >>tree, you simply prove that one move is better than all other moves by a window >>"S". This is very precise. What Bruce does, and what I tried, was a very >>limited sub-set of that. It was not as accurate as a proper implementation. It >>was also not as expensive. But the main point is that it was not as accurate. >> >> >>> >>>If the algorithm helps to find some tactics and miss other things that are more >>>important in games and not in test positions because of smaller depth then the >>>algorithm is inferior. >>> >>>Uri >> >>They tested SE enough to _prove_ that it didn't hurt, if you read their >>paper. > >It is possible that it did not hurt their program because they did not do >pruning. > >My point is that their algorithm does hurt the programs of today and it means >that their search is inferior relative to the search of the top programs of >today. > >Uri And I _still_ say that is simply a badly flawed assumption. One more time: You can accomplish the same thing with forward pruning or with selective extensions. They chose the latter. Others are choosing the former. Some of us do _both_. I see _nothing_ that suggests that any of the three approaches is better than either of the other two. They should, in theory, all be equivalent...
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.