Author: Vasik Rajlich
Date: 05:08:24 02/06/04
Go up one level in this thread
On February 06, 2004 at 06:26:20, Uri Blass wrote: >On February 06, 2004 at 05:54:29, Vasik Rajlich wrote: > >>On February 06, 2004 at 03:42:42, Uri Blass wrote: >> >>>On February 06, 2004 at 02:15:35, Tord Romstad wrote: >>> >>>>On February 05, 2004 at 15:15:47, Uri Blass wrote: >>>> >>>>>I think that you underestimate your engine. >>>>>It seems to get similiar depth to crafty. >>>>> >>>>>For example in the following position it got depth 11 even in blitz 4+2 >>>> >>>>Yes, 11 plies in blitz games is not unusual. But 11 plies in Gothmog and 11 >>>>plies in Crafty is not the same. I do much more forward pruning and depth >>>>reductions than Bob, and fewer extensions. In non-tactical positions like >>>>the one you give, my qsearch is also considerably smaller than Bob's (I think). >>>> >>>>Tord >>> >>>I do not think that there is a big difference. >>>Crafty searches bigger tree because it searches more irrelevant lines. >>> >>>I guess that the main advantage of Crafty relative to Gothmog when you use one >>>processor is superior evaluation(Gothmog's evaluation is more complex but bigger >>>is not always better and not having bugs or some too optimistic scores of >>>gothmog that lead to wrong sacrifices can be more important and it is possible >>>that Gothmog can get crafty level if you only reduce the big positional scores >>>that encourage it to sacrifice). >>> >>>I do not think that gothmog see less than crafty in the relevant lines(crafty >>>has bigger tree but it proves nothing). >>>I know that test suites are no proof but results of the gcp test suite give me >>>the impression that cases when Gothmog can see more than crafty are not rare. >>> >>>Uri >> >>I have the theory that the greater your search resources (ie combination of time >>and hardware), the less important is the search, and the more important is the >>evaluation. > >I do not agree with that theory. > >For example suppose a program has no tablebases. > >With deep search it may not need knowledge how to win KQ vs K when with small >search it may need the knowledge. > >If the hardware is fast enough the program can solve the game with only piece >square table evaluation. > >Of course we are not going to see it but with good hardware evaluation what win >is better in some endgames become unimportant because the program will not fail >to win thanks to search. > >Uri I'm afraid I didn't explain myself very well. As I see it, there are two benefits to more search depth: better tactical play, and better positional play. You can emphasize the tactical play by extending various forcing lines, but your branching factor suffers and your positional play gets worse. My testing setup emphasizes fast decisions - I run blitz matches on my 2600+ Athlon, and I run testsuites. In this setup, tactics dominate. It's quite different, I think, to run on let's say a dual Opteron with 2 minutes/move. In this case, the tactics may take care of themselves, and it's important to search the positional lines instead. This means working on your branching factor (move ordering, and some sort of positionally safe pruning) and your evaluation. Maybe by the time of the 2005 WCCC, everybody who is serious will be getting 10 million NPS, and all the time playing around with "tactical extensions" will have been more-or-less wasted. 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.