Author: Peter McKenzie
Date: 01:10:32 05/09/02
Go up one level in this thread
On May 09, 2002 at 02:07:03, Uri Blass wrote: >On May 08, 2002 at 19:31:50, Peter McKenzie wrote: > >>On May 08, 2002 at 04:27:43, Amir Ban wrote: >> >>>On May 07, 2002 at 14:17:51, Uri Blass wrote: >>> >>>>On May 07, 2002 at 07:44:16, Amir Ban wrote: >>>> >>>>>On May 06, 2002 at 18:06:47, Christophe Theron wrote: >>>>> >>>>>>On May 06, 2002 at 15:34:01, Amir Ban wrote: >>>>>> >>>>>>>On May 05, 2002 at 19:58:09, Christophe Theron wrote: >>>>>>> >>>>>>>> >>>>>>>>"Knowledge" in the sense of positional evaluation (that's what most people think >>>>>>>>about when they talk about knowledge) makes for 10% of the strength of a chess >>>>>>>>program. >>>>>>>> >>>>>>>>Chess is 90% about tactics (which is a concept close to "search"). >>>>>>>> >>>>>>> >>>>>>>Before strongly disagreeing (as I guess I will), what does this mean ? >>>>>>> >>>>>>>If I freeze my search engine and work only to improve the evaluation, how much >>>>>>>do you expect the total strength to improve ? Is it limited ? >>>>>> >>>>>> >>>>>>I expect the strength of your engine to improve, but not much in regard to the >>>>>>energy invested. Because you are going to focus your efforts on an area that >>>>>>does not have the biggest potential in strength. >>>>>> >>>>>>On the other hand people will love it more and more because it will have a much >>>>>>better playing style. >>>>>> >>>>>>People can forgive gross tactical blunders, but not slight positional mistakes. >>>>>>Go figure... >>>>>> >>>>>>Here I'm talking about current top engines of today, naturally. >>>>>> >>>>>>Building a chess engine with a broken evaluation to demonstrate that a better >>>>>>evaluation could improve it tremendously is not in the spirit of my idea. >>>>>> >>>>>> >>>>>> >>>>>>>I understand that you are saying that it will change the style but overall >>>>>>>strength will not be much changed. >>>>>> >>>>>> >>>>>>I do not know exactly how far we will be able to go with the 10% I attribute to >>>>>>positional evaluation. >>>>>> >>>>>>I'm not saying it counts for nothing and that overall strength will not benefit >>>>>>from research in this area. >>>>>> >>>>>>I believe that the positional evaluation is the part of a chess program >>>>>>responsible for only 10% of the strength, and that the rest is done by the >>>>>>search. >>>>>> >>>>>>I believe that the positional evaluation is responsible for most of what people >>>>>>perceive as the "playing style". >>>>>> >>>>>>Now you can strongly disagree, I do not have the absolute truth. >>>>>> >>>>> >>>>>Ok. I think this is wrong. Anyway I'm working for a long time under the >>>>>assumption that it's the evaluation rather than the search that needs work. >>>>> >>>>>The search engine of Junior7 is basically the same as Junior6. >>>>> >>>>>Junior5 was the last engine where I did extensive work on the search. Since then >>>>>in terms of effort it was at least 80% evaluation, no more than 20% search. >>>>> >>>>>Amir >>>> >>>>I wonder how much of it is testing to find the right weights in your evaluation >>>>and how much of it is adding new evaluation functions. >>>> >>> >>>Hard to say because they belong to completely different mental processes. New >>>elements are part of a creative process, which is not something that can be >>>regulated. Adusting and testing is more routine and automatic. There's no reason >>>why they can't take place at the same time. >>> >>> >>>>I find that in the endgame there is knowledge in the evaluation that Junior does >>>>not have when part of the top programs and even part of the amaturs have it. >>>> >>>>Here is one example: >>>> >>>>Junior7 does not know that the following position is a draw >>>> >>> >>>Junior doesn't know about billions of positions that they are draws. I wish I >>>could reduce the number by just 20%. This particular case doesn't seem very >>>important. >> >>Strange as it may seem, but in my experience this ending comes up quite >>frequently. Of course frequently is a relative term, but I have seen it a >>number of times while watching programs play on ICC. >> >>One time, I was discussing this particular ending with Peter Kappler and bang, >>it occurred in a game one of us was observing!! >> >>Peter > >When you say "this ending do you mean to the case when the side with the >material advantage has 2 pawns? I meant the general wrong coloured bishop ending, I guess the case with 2 pawns is alot more rare but I reckon if you're going to go to the trouble of coding the thing at all you may as well make it generic. > >Note that Junior knows that the side with the material advantage cannot win if >it has only one pawn. > >I agree that the knowledge that Junior does not have in this case may be worth >less than 1 elo but the point is that if you add to the evaluation a lot of >cases when you can earn less than 1 elo then the total result may be getting 50 >elo improvement. > >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.