Author: Robert Hyatt
Date: 08:37:37 10/15/99
Go up one level in this thread
On October 15, 1999 at 03:44:30, Ed Schröder wrote: >>Posted by Marcus Kaestner on October 14, 1999 at 04:01:40: >> >>in chessbits 3 i´ve made a test with the nunn-endgames (10 endgame-positions >>played with both sides). every program one time with, one time without tb´s >> >>after abaout 140 games >>h7.32 earned +5 elo for the tb-version >>n7.32 earned +12 elo >> >>far the best was crafty 16.13 with +65 elo, topping even the score of hiarcs >>7.32tb >> >>marcus > >Take this into account: Some programs may have so much chess knowledge in their >programs that (the current) TB's have zero effect in the end. This will change >as soon as the 6-man TB's (and up) will come, no doubt.. > >Other programs are developed to 100% rely on TB's and therefore have very few >(or nothing) chess knowledge inside because there is no need for that. I don't >know in to what extend this is true for Crafty but I am confident Bob will >tell you. Crafty has a significant amount of such endgame knowledge. Including special code for krp kr, which is a non-trivial ending. > >As a result you can not compare the value (elo gain) of TB's in this way as >it will depend of what is already inside a program (or not). > >Rebel knows the most important ones such as (by head): > >- KRK >- KNNK >- KBNK >- KPK >- KRPKR that is a tough one. You probably do just like I do and have some rules about the king in front is enough to draw, not advancing it too far with the rook in front, etc. But I have seen Rebel let Crafty trade into a won ending with krp vs kr. EA6PZ on ICC used to run rebel, and it would fall into that on multiple occasions. Because there are so many exceptions it is difficult to get them all right. Yet the tablebase never fails. >- KNPKN >- KBPKB >- KQPKQ > >I am sure this list is not complete. Of course the bad-bishop situation is >covered too. Some others are not covered such as KQKR (Hi Jim!) for which >Rebel will rely on TB's in the future. > >Ed I find that even though I have specific code for krp vs kr, that is a very common ending vs computers, and that tablebases are the only way to play this really well. At least I have seen _every_ commercial (and amateur that doesn't probe in the search) fall into dead lost positions where they didn't have to. IE in one case, a game KQR vs KQRPP saw the weaker side find a clever way to win one of the two pawns, but in doing so, ended up in a dead lost krp vs kr ending yet it thought this was better. The position prior to this was a dead draw as there was no way to stop the continual checks... A program with tablebases will have a real advantage over one without. Because in looking at that game, we all noticed that the move Crafty made seemed to say "go ahead, take that pawn, it really is free..." because it knew it was protected by the forced mate that resulted if it was taken (sort of like the way crafty offers the h2/a2 pawns to opponents that don't know that Bxa2 fails to b3 and the bishop is trapped.) by the way, kq vs kr is trivially solvable without tablebases. Mark tried this last year and reported that fritz (I think) could solve this at maybe 10 secs per move. I tried the experiment of playing 100 random positions using crafty with tablebases vs crafty without, and I was amazed to find that crafty won every game without using tablebases, against the perfect play by crafty with tablebases. Even down to 1 second per move, it could still force mate against a tablebase-user. I don't think kq vs kr is worth anything as a result, other than it lets you see the mate and know that going to kq vs kr is better than sticking with kqn vs krp... just sac the knight for the P and you instantly see a forced mate.
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.