Author: Johan de Koning
Date: 01:32:19 08/17/03
Go up one level in this thread
On August 16, 2003 at 17:52:57, Dieter Buerssner wrote: >On August 16, 2003 at 03:59:23, Johan de Koning wrote: > >>However, the example you give has nothing to do with TT (or cleverness :-). >>IIRC Uri mentioned it about 2 years ago: the evaluation of The King decreases >>with increasing fifty_counter. This encourages it to take action after 10 moves >>rather than after 49 moves. It is a rather simple and old rule that applies only >>to the evaluation. Because of this it will fail occasionally due to perpetual >>postponement. Also because of this it hardly depends on the path, hence it >>rarely threatens TT consistancy. > >I tried the idea without too much success. I got search inconsistencies in the >few selected positions I tried. I used a very primitive scheme, that was not >tuned at all - more or less just to see how the idea works in selected test >positions. I saw decreasing scores (perhaps even fail lows, for the moves that >draw), and one ply later, it was increasing again. Critical TT positions were >overwritten. Some positions that are in the TT with higher draft can be forced >somewhere close to the leafs. An actual search there would have reduced the >score, but the search combined with TTs prevents this. > >This may be closely related to the replacement scheme used in TTs. Perhaps I >gave up too early. Actually I had an interesting discussion about this in ICC >chat with some other engine programmers, and we came up with some ideas (but I >did not try it yet). > >BTW. This can also be considered as another random factor in the engine (or at >least one, that amplifies some random effects). Dependent on the current filling >of the TTs, the engine can be lucky to have the right info inside, or it can be >unlucky, to have some info, that the search would now do differently (better). >Depending on the size of the TTs, things can become seemingly randomly. There >are already other things, that cause this, but path dependent scoring adds to >it. There is certainly a lot of luck involved in TTs. Matching positions with diffrent paths is unsound because different paths imply different draw opportunites. This is however unavoidable since the TT by definition matches them. Matching positions with different drafts is unsound but gains a lot if there's little action in the search. This problem is "solved" by assuming the score with a higher draft is wrong-but-better. And there is more implementation dependent stuff. Path and alfa-beta bounds may influence evaluation and extensions and pruning (eg q-search and null-search). All in all the shape of the tree will rarely be the same with varying TT size. >I fear, I don't understand "Also because of this it hardly depends on the path, >hence it rarely threatens TT consistancy." Because the score reduction applies only to the evaluation (= leaves). All the leaves in 1 search have roughly the same fifty_counter. Or they have a small fifty_counter that doesn't reduce the score. ... Johan
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.