Author: Uri Blass
Date: 11:24:32 01/28/02
Go up one level in this thread
On January 28, 2002 at 12:59:19, Vincent Diepeveen wrote: >On January 28, 2002 at 08:03:08, Ed Schröder wrote: > >>On January 28, 2002 at 06:33:26, Amir Ban wrote: >> >>>On January 28, 2002 at 06:12:53, Ed Schröder wrote: >>> >>>>>Will TACTIC's eventually REFUTE! Positional play? >>>> >>>>In the end yes. >>>> >>>>It is my (new) opinion that the nature of chess is just search. >>>> >>>>Elo progress of (professional) chess programs... >>>> >>>>1990 - elo 2000 (average depth 6-8) (TC 40/2h) >>>>1995 - elo 2300 (average depth 8-10) >>>>2000 - elo 2500 (average depth 11-13) >>>>2002 - elo 2600 (average depth 12-14) >>>> >>> >>>This begs the question, because the programs are newer and play positionally >>>different. Will a 1990/1995 program perform 2600+ on today's hardware ? >> >>No. >> >>Improved search and improved chess knowledge make nowadays programs way better >>than the programs of 1990-1995. > >The first is utter nonsense: my search nowadays is much simpler than >in 1997. I do simple PVS with nullmove and a few extensions now that's >it. No big tricks there. Just well debugged search. No doubt that not doing stupid extensions is an improvement in the search rules relative to doing stupid extensions. > >Also i have less hashtables now. I used to have from the start > - transpositiontable > - evaluation table > - pawn table > - bishop table > - king safety table > >The king safety table only existed short it was removed soon as in >evaluation too much knowledge was added to use it. My bishop table >has gone too and most pawn patterns nowadays also can't get hashed >anymore so they are in other functions now. > >In short what i have is > > transposition table > evaluation table > pawn table (for only some basic patterns the majority can't be > stored in hashtable). > >In 1997 i had all kind of search enhancements, like internal iterative >deepening, i remember all kind of weird independant move ordering >techniques. Also some dubious things near the leafs were done. > >All that has been thrown out by now. > >Qsearch i limited back in 1997 way more than i do now. Though the >code still is very sophisticated, in 1997 i had more 'smart' rules >to limit qsearch, now the principle is very simple: "a move that can >go to the position being more quiet , LET'S TRY IT". The result is that >i need more nodes for qsearch now than i needed back in 1997. > >So it didn't get more efficient at all. > >Of course diep runs parallel now, that's just to get more cpu power, >it doesn't improve the search in itself, whatever big the effort was >to get it working good. > >The ONLY thing that really has become serious, huge, and put loads of >effort in, that is the evaluation function. > >Still it is far from perfect of course, but the real progress is in the >EVALUATION, and nowhere else. I believe that the progress in Diep is mainly in the evaluation but it does not mean that other programs do not do progress also in search. Evaluation and search are also connected. If you have better evaluation you can be better in tactics. > >The 1998 version from diep would get like 15 ply easily on todays hardware. > >In fact in 1999 in paderborn (IPCCC) i remember how i at a single cpu 450PII >searched like 11 to 12 ply easily at 3 minutes a move. IN the world >champs 1999 i searched 20 ply in the endgame at a quad xeon 400. that's >1.6Ghz in total. A dual 1.2 K7 gives way less than that in the same >positions now. > >To get that now i need a dual 1.x Ghz K7. plies mean nothing If you had bad evaluation and use null move pruning you had a good chance to miss 10 ply tactical trick thanks to null move pruning even if you search clearly more than 10 plies. 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.