Author: Uri Blass
Date: 09:12:07 08/03/03
Go up one level in this thread
On August 03, 2003 at 11:40:52, Ed Schröder wrote: >On August 03, 2003 at 08:48:16, Uri Blass wrote: > >>On August 03, 2003 at 08:32:35, Ed Schröder wrote: >> >>>On August 03, 2003 at 07:12:30, Uri Blass wrote: >>> >>>>On August 03, 2003 at 06:54:47, Jonas Bylund wrote: >>>> >>>>>On August 03, 2003 at 06:50:25, Uri Blass wrote: >>>>> >>>>>>On August 03, 2003 at 05:45:42, emerson tan wrote: >>>>>> >>>>>> >>>>>>> Of course given enough depth, most chess >>>>>>>programs can give good evaluation, but there are still a lot of positions that >>>>>>>are far beyond the search depth of top engines on the most powerful hardware, >>>>>>>and it is here where chess knowledge is needed. >>>>>> >>>>>>Based on my understanding increasing the chess knowledge in Rebel is not going >>>>>>to help it to understand these positions. >>>>>> >>>>>>I understood that >>>>>>Rebel does the full evaluation at every node except the leaves so I guess that >>>>>>if it can see something important in the tree before the leaves then it is going >>>>>>to see it also in the leaves. >>>>>> >>>>>>Cases when something important is generated only in the leaves can be detected >>>>>>with less chess knowledge (maybe even faster) thanks to deeper search. >>>>>> >>>>>>In order to change my mind >>>>>>I need to see a case when knowledge does not help to detect the problem one or 2 >>>>>>plies earlier but help to detect the problem 10 plies earlier. >>>>>> >>>>>>If Rebel with full knowledge say +1 for white at plies 5-10 when default Rebel >>>>>>say evaluations near +1 for black at plies 5-15 then it is going to be a >>>>>>convincing evidence that knowledge in Rebel is what you think about when you use >>>>>>that word. >>>>>> >>>>>>Uri >>>>> >>>>>But how can we know without trying it? >>>>> >>>>>This has the potential to be a very interesting experiment! >>>>> >>>>>Jonas >>>> >>>>I know that people tried personalities with Rebel century and found that >>>>reducing the knowledge of Rebel lead to better results. >>>> >>>>There was no difference that was observed between blitz and longer time >>>>control(Ed did not say based on tests to use knowledge=25 in blitz and >>>>knowledge=50 at long time control). >>>> >>>>I see no reason to believe that things changed. >>>>I do not claim that knowledge is not important but that I believe that the name >>>>knowledge in Rebel is simply misleading because Rebel does the full evaluation >>>>in every node when the remaining depth is positive based on Ed's page. >>>> >>>>I believe that the lazy evaluation at the leaves miss nothing big that was >>>>detected in the previous plies so there is no case that Rebel with knowledge can >>>>do clearly better than Rebel without knowledge(in the best case for the full >>>>knowledge evaluation it can only find that it is in trouble faster but there are >>>>more cases when it can find that it is in trouble faster when the default >>>>personality is used) >>>> >>>>I guess that it may also miss a small positional difference that it does not >>>>consider in the lazy evaluation but the important thing in the evaluation is not >>>>to miss a big thing and deeper search by 1 ply often more than compensate for a >>>>small positional difference. >>>> >>>>I am not Ed and I may be wrong but if you want to prove that I am wrong then it >>>>is better that you start by provifing a position when default Rebel show clear >>>>advantage for white at depthes 5-15 when Rebel with maximal knowledge shoe >>>>something completely different at depth 5-10. >>>> >>>>Uri >>> >>> >>>Hi Uri, >>> >>>While it is true the [Chess Knowledge] parameter is about tuning Lazy Eval you >>>shouldn't underestimate the sometimes disastrous effects done to the evaluation >>>especially in Q-search. Search and Lazy Eval bite each other, it is a matter of >>>feeling and taste to find the best combination represented by the value of the >>>[Chess Knowledge] parameter. >>> >>>Here are some disastrous examples, I am cherry-picking of cousre as these are >>>the big exceptions but it shows you how sensitive and sometimes fragile (thus >>>important) the whole concept is. >>> >>>I picked the Rebel-XP engine as I have the examples straight available. >>> >>>[d]1q2N3/3p1Q2/3p3K/p7/b5k1/8/7P/8 w - - bm Qf2; >>> >>>[Chess Knowledge = 100] -> not found after 5 minutes and 12 plies. >>>[Chess Knowledge = 500] -> 8 seconds, depth=9 >>> >>>00:00:08.4 1,81 9 7018724 Qf2 >>>00:00:08.9 2,38 9 7427943 Qf2 Qb3 >>>00:00:10.7 4,17 10 9016929 Qf2 Qb3 >>> >>>Second example... >>> >>>[d]2q4r/2p4p/kpnp2p1/p2b4/P6Q/1PR2NP1/2P3BP/2K5 w - - bm Rxc6; >>> >>>[Chess Knowledge = 100] -> found at depth 10 >>>[Chess Knowledge = 500] -> found at depth 8 >>> >>>Last example.... >>> >>>[d]r3b1nr/ppqn1k1p/4p1p1/1P1pPpP1/1B1N1P1P/R7/3Q4/R3KB2 w Q - bm Rc3; >>> >>>[Chess Knowledge = 100] -> 1:53 and 13 plies. >>>[Chess Knowledge = 500] -> 0:52 and 11 plies. >>> >>>Ed > >>I see that even in 2 of 3 of the big exceptions the difference is only 2 plies >>so my point is that in the best case it can help to detect things faster and >>there is no case that Rebel has a completely different evaluation with >>knowledge. >>This is the reason that I do not believe that it is productive for long time >>control. > >The first example is about 4 plies at least, if you are going to be simplistic >about that as the tenor of your answer certainly is then count me out of this >discussion. Mind you, if you have such a case one time in each game overlooking >that kind of things your program is dead meat for every decent chess engine. > > >>I could imagine that knowledge may be about detecting a fortress position(or >>detecting some dangerous passed pawn that is going to promote only 20 plies >>later) so rebel with knowledge is going to see it even after 5 ply search when >>default Rebel could not see it even after 15 plies but this is not what >>knowledge is about in Rebel. > >I think that cases of 2 plies are already bad enough, I think you are >underestimating what the algotithm is all about. On one hand you put a lot of >efforts to write a superb evaluation while on the other hand you are killing the >same chess knowledge using Lazy Eval. killing seems to me too big word because you use the knowledge if it is relevant before the positions of the qsearch. Especially when you are in Q-search (QS) >and there is a long trade off of pieces and/or a long series of checks (see the >first example) QS might stop too early returning a crippled score because of >Lazy Eval. Increasing the [Chess Knowledge] parameter might avoid that. I do not claim that there are no positions when it helps but I see no reason to assume that they are more important than the positions when it is the opposite. > > >>I do not claim that knowledge is not important for long time control but that >>the knowledge that is needed is not the difference between lazy and not lazy >>evaluation in Rebel. > >Well, I have experimented with the algorithm for more than 20 years and I can >safely say that you are wrong :) > >Ed Do you have some evidence that bigger knowledge parameter is better for longer time control? It is impossible to test correspondence chess because we cannot get enough games but it is possible to compare between blitz and rapid chess. Statistical evidence that the optimal knowledge for 30 minutes per game is higher than the optimal knowledge for 5 minutes per game may make me believe that the tendency continue at longer time control. Uri
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.