Author: Vasik Rajlich
Date: 04:20:20 12/17/05
Go up one level in this thread
On December 16, 2005 at 15:49:39, Uri Blass wrote: >On December 16, 2005 at 15:35:55, Vasik Rajlich wrote: > >>On December 16, 2005 at 13:34:43, Uri Blass wrote: >> >>>On December 16, 2005 at 03:32:16, Vasik Rajlich wrote: >>> >>>>On December 16, 2005 at 00:56:04, Zappa wrote: >>>> >>>>>On December 15, 2005 at 17:05:00, Sergei S. Markoff wrote: >>>>> >>>>>>1) Fruit. >>>>>> >>>>>>Fruit search seems to be primitive. "History pruning" is a variation of >>>>>>well-known idea. After implementing such method in SmarThink some years ago I >>>>>>named it "history-based pruning" and then changed to "ordering-based pruning". >>>>>>The outcome of such methods very depends of whole search model, but anyway >>>>>>history pruning is not the key to Fruit strength. >>>>>> >>>>>>To my mind, the key of Fruit strength is that the "Chess is the art of >>>>>>exchange". So, Fabien's idea about flexible game stages looks to be a beautiful >>>>>>way to improve positional play. Fruit can effectively consolidate the position. >>>>>>It simply knows when to excange to improve position. I think that it's the main >>>>>>key (cumulative with very good tuning of evaluation function). I think Fruit is >>>>>>very perspective. The main line of progress for this project, to my mind, is to >>>>>>add more complicated knowledge and intellectualize a search. >>>>>> >>>>>>2) Rybka >>>>>> >>>>>>Some time ago we discussed with Gian-Carlo Pascutto an idea of create special >>>>>>"SET-tables" with sets of piece-square values indexed by 1) material on the >>>>>>board; 2) king position; 3) pawn structure. Such tables can be calculated by >>>>>>analyzing a lot of games. That time I delayed my work in this area because I >>>>>>found other perspective things. >>>>>>You can see that Rybka executable contains a lot of precalculated tables. And >>>>>>also we all know that Rybka plays positional style. My version is that Rybka >>>>>>uses some variation of SET-approach. At all cases it uses some precalculated >>>>>>positional knowledge, but what sort of it? ;) >>>>> >>>>>My personal opinion: >>>>> >>>>>Fruit wins by 3 things: deep PV checking, mobility, and correctness. I talked >>>>>about this with Fabien at Reykjavik. When you have mobility, you are very >>>>>sensitive to being "driven back". And when you can check your mainline 18-20 >>>>>ply and not lose any mobility, its very probably you're playing a good move. >>>>> >>>>>Rybka: I'm starting to think that a lot of Rybka's strength is tactical. Try >>>>>that baby out on a few test positions some time. For example, the rapid TC CEGT >>>>>list has Rybka 55 rating points ahead of Fruit, while the slower BFF list has >>>>>Rybka only 15 rating points ahead of Fruit. >>>>> >>>>>anthony >>>> >>>>We need more data. I'll try to put it together when it's all ready. >>>> >>>>One thing people tell me is that Rybka tends to stick with her moves from lower >>>>depths more than other engines. This would also suggest better blitz play. >>>> >>>>Vas >>> >>>I think that it may suggest simply that rybka is better because it needs less >>>time to find the correct moves unlike other engines that need big depth to find >>>the correct moves. >>> >>>I think that the only correct test is test with time handicap. >>>I think that it may be interesting to see how much time programs need to get 50% >>>against Rybka at 1 minute per game and the same for longer time control. >>> >>>I think that there is diminishing returns so if a program score better at blitz >>>but wins at all time controls then it is not fair to claim that it is better >>>blitz player. >>> >>>If you can prove that Rybka score 50% with time handicap of 3:1 at blitz against >>>engine X(ponder off) and score less than 50% with the same time handicap at >>>longer time control then you have a point. >>> >>>Uri >> >>Another topic we need some data about. Was it ever proven that engine >>differences are magnified at shorter time controls? It's not obvious to me. >> >>Vas > >I remember that there was some experiment with Fritz6 that showed diminishing >returns and it also logical. > >Previous experiments also supported diminishing returns but the results were not >significant. > >I think that it is obvious that some opening lead to forced draw and it is going >to be draw only if the sides search deep enough so after some time control there >is 0 returns in these openings. > >It means that in order not to have diminishing returns you need to have opening >with increasing returns. > >Uri In this case, there should be more draws at longer time controls. This we can check very easily (although I am too lazy :)). Vas
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.