Author: Gerd Isenberg
Date: 08:33:49 07/16/04
Go up one level in this thread
On July 16, 2004 at 05:42:09, Uri Blass wrote: >On July 16, 2004 at 04:53:44, Gerd Isenberg wrote: > >><snip> >>>I think that it is important to test and the problem of some participants is >>>lack of testing. >>> >>>I talked with Gerd(programmer of Icichess) >>>Gerd is not a strong chess player. >>> >> >>My german DWZ was slighty above 1600. > >I think that this level is not what Amir Ban meant by using the words:"strong >chess player" inspite of the fact that it may be better than Amir. It was not intended to point out my "strongness" ;-) >I do not say that you are a weak chess player but a lot of programmers are at >least in your level in playing chess. Yes. > >I guess that the only programmers who are strong chess players in WCCC are Jonny >and Vincent(I do not include myself because I am significantly weaker than >Jonny). > Yes, i don't think in general it is contra-productive to be a strong chess player - it might be true as well as for "too" good programmer or a theoretical scientists. Of course programming skills are important - but as you, as an not very experienced programmer, shows us, it is more important to have some math skills and to think one some abstract level, and some clue how to implement ideas. > > > >>>His program is even slower searcher than movei in nodes per second and >>>the programmer told me that most of the time is used on the evaluation. >>> >> >>See latest ICGA Journal, "The Tenth Commandment", Review of Dap Hartmann of >>"Advances in Computer Games 10", Chapter about Lines of Action, YL versus MONA: >> >>"If you are slow anyway, take advantage of it" >> >>Or for IsiChess, if you do expensive king eval, why not looking for static >>mates? > >The main question is if being slow means that you can add knowledge with smaller >price. Yes, the price is relative smaller. But anyway, i try to implement things as far as possible ;-) > >I can give a logical reason why not but I know almost nothing about computers >so I may be wrong. > >If you have a small evaluation then your program can use fast memory to do >something. Ok, faster, shorter code, less first level code cache pullution. I heared some programs code (search+eval) fit totally in former or todays 64K first level cache. Huge evals may as well access more constant data and store/load more locals and other veriables, so more data cache pollution as well. OTOH small and faster programs suffer probably more if probing huge hash tables, where i can already make some kogge stone mmx or sse2 stuff ;-) > >If you have a big evaluation then you may need to use the slow memory to do the >same thing so the relative speed demage from looking for static mate may be the >same. > >If my logic is wrong then what is wrong in it? > Yes, but the way i implemented king safety produce already most stuff needed for mate detection - a kind of waste product. > >> >>>I asked him if he checked to test if what he added in the evaluation was >>>productive and the surprising reply was negative. >>> >> >>I don't do any autoplayer matches between two versions with different eval, >>because i am not able to play automaticly ;-( >> >>During the last two years or so, i added a few more eval terms after i >>recognized lack of knowledge in some games IsiChess played in tournaments. >>Of course this helps in exactly this kind of positions. Depending on the >>generalization and interactions, i don't know exactly whether it improves the >>overall performance of the program. >> >>At least i debug each new eval terms i added ;-) >>And i check all kinds of symmetry. >> >>Gerd > >I do not do it and the question is also if it is important. Simply to be sure that the code works well as intended. I and probably others too introduced symmetry bugs of all kinds. Sometimes such bugs even help to win a game ;-) Ask Chrilly Donninger or Ed Schroeder remembering 93. If no more castles are possible, mirroring between king- and queen-wing should produce the same score. > >I remember that I read that a person who found symmetry bugs did not find >changes in the level of the program after he fixed the bugs. > >I test changes in the evaluation only by games. > Fine, but a lot of endgame related eval terms doesn't even occur in a game or are hided by alfa-beta algorithm. IMHO playing thousands of blitz games is not an appropriate way to tune for tournament games. But of course that may depend on the program. Gerd >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.