Author: martin fierz
Date: 14:59:16 02/26/04
Go up one level in this thread
On February 26, 2004 at 14:41:07, Dieter Buerssner wrote: >On February 26, 2004 at 06:59:37, martin fierz wrote: > >>another reason for not believing this stuff: your above graph shows *exactly* >>what happens when you go from a non EGTB position to an EGTB position (or, for >>that matter, what happens when you go into any position your program can >>recognize as a draw whether it has tablebases or not): your eval thinks it's >>doing great, but the exchange of something leads to a drawn position in your >>tablebases. are you going to claim that crafty plays better without TBs? >>:-) > >Some interesting related reading about this: E.A. Heinz. >Efficient interior-node recognition. >In ICCA Journal, Vol. 21, No. 3, pages 156-167, September 1998. >(Download at http://supertech.lcs.mit.edu/~heinz/ps/node_rcg.ps.gz) > >At page 3, they discuss, how to score certain known results (win or loss, but >not the distance to win/loss). 2 methods (my paraphrazing): Have a scoring range >outside the normal eval scores but lower (in magnitude) than the "real" mate >scores. Or just give rather normal eval scores. They started with the first >method, and went to the second (although the first method sounds better at first >sight). > >In my engine, I can compile TB-returns as fail hard (return beta or alpha) or as >fail soft (return the mate score or 0.0). Default is fail soft. Some time ago, I >found some positions, where the fail hard method used significantly smaller >trees (the engine uses fail-soft search, normally). I really could not explain >it well, but also did not invest lots of time, to try to understand it. > >I also remember some private mail from Ernst Heinz (about fail hard/soft). Heinz >(probably out of convincing arguments :-): "Fail hard läuft einfach besser". > >Regards, >Dieter hmm, interesting... BTW ernst's argument doesn't seem unconvincing if he has numbers to back it up! i find it hard to believe that this kind of stuff doesn't work as expected (i.e. fail-soft should by intuition work better). in any case, i have a nice checkers program with fail-soft and LOTS of EGTB accesses - i guess i'll have to try fail-hard for EGTB hits there in this case. i will be very surprised though if that really is better - but being an experimental scientist, i am willing to try anything at least once! cheers martin
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.