Author: Bruce Moreland
Date: 01:32:47 10/09/98
Go up one level in this thread
On October 09, 1998 at 03:31:34, James Long wrote: >On October 09, 1998 at 02:11:19, Jouni Uski wrote: > >>I have feeling, that using 3-5 pieces tablebases don't give any measurable >>rating gain - may be 2-10 points maximum = no real gain. >> >>Jouni > >It would depend on how well the program evals endgame positions. >Programs with little or no code to analyze things like controlling >a queening square would benefit more. > >I imagine this isn't the case with commercial programs. I think this is a very complex subject. I don't know why more commercial programs don't ship with endgame tables. It seems that users want these, given that many of the people who run Crafty on ICC use them. Perhaps there is the matter of distributing enough of them that they would be useful. You could fit all of the 3- and 4- man tables on a CDROM, but once you start going beyond that, you're talking serious space. A two-point improvement is too low. That number means that if you scored 50% against somebody before, now you will score 50.29%. So in a 400 game match, rather than scoring 200, perhaps you score 201, which probably means that two games got shifted a half-point in your favor. I think this would be a very disappointing result for endgame tables, I think they shift the result a lot more often than this, against equal opposition. It's hard to tell for sure. If they shift the result in one game in every fifty, that's 7 points, by the way. If it's one in forty, that's 9 points. I don't think they'll shift the result in many games against you. The program is already scoring massively well against the typical player, so the odds of even getting to a reduced endgame aren't very high. It's interesting though, to see the the effect of even a rare table shift against a weak player. Let's say you win 99.0% against somebody, and 99.1% with the tables. This is equivalent to one half-point shift every five hundred games. But this is actually 18 Elo points in this case. I think that regarding programs being "smart" in low-material situations, I think there are a few obvious endings where you can add a smarts and get a lot of the benefit of endgame tables. But this doesn't actually mean that it is easy or efficient to do it, and even if it is easy and efficient to do it, it doesn't necessarily mean that the commercial programs have this knowledge. I had a drawn KQ vs KP vs Fritz the other day, where Fritz scored it as +8. I doubt that any program handles that ending in the eval function, it is a pretty hard case. bruce
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.