Author: Heiner Marxen
Date: 06:09:44 03/24/04
Go up one level in this thread
On March 24, 2004 at 06:59:58, rasjid chan wrote: >On March 24, 2004 at 05:39:20, Tord Romstad wrote: > >>On March 24, 2004 at 05:17:56, Peter Fendrich wrote: >> >>>Uri didn't invent ETC if that's what you imply! >>> >>>Given your story about costly move/unmove functions it's possible that ETC gives >>>you some savings. Without ETC you will hit the cutoff anyway in the child node >>>and with smaller unmove costs ETC is not that effective IMHO. >> >>It seems to me that you miss part of the idea of ETC. You are right that >>you will get the cutoff in the child node even without ETC, but in which >>child node? If your move ordering is not perfect, there is a risk that >>you will have to search many moves before you get the cutoff. When you >>use ETC, you check the hash values for *all* child nodes before you >>start searching, which can sometimes save a lot of nodes. >> >>To me, ETC has always been a clear win. The last time I made any >>experiments, it reduced my tree size by about 10% at high search depths. >>I am fairly sure it is a technique which works better with MTD(f) than >>with more conventional search algorithms, though. >> >>Tord > >Thanks. > >I think there is much more to ETC. I'll see what can be done >like checking only good moves, etc.. . nothing is simple otherwise >all programs == Gothmog. > >An aside - immediately after I implemented my pseudo ETC, after >hash_cutoff_before_makemove()... I tested it against Gothmog at >60 moves / 40 secs. No need tournament time as Gothmog, Ktulu, Ruffian >almost never fail whatever time control. Pseudo or otherwise, it plays >very differently and I don't yet know why, sometimes making moves almost >instantly which was not noticeable in the past. Maybe you have a sign bug here. You do shortcut a makemove and (!) a recursive call to the search function, which for negamax-style implementation will be called with negating its result... ;-) This also influences the interpretation of upper/lower bound of the ETC data found in the TT. Cheers, Heiner > Gothmog was nearly caught >but my program finally conceded a draw from a clearly winning positions. >Very rare but can happen,mainly because my program has yet to reach anything >call strong.I mention this because I have some impressions about the 3 >programs;seemingly Ruffian is the strongest, Ktulu and Gothmog don't yet >seem miles apart. I think Gothmog has some lopeholes somewhere that allow >clearly weaker programs to occasionaly sneek pass. My guess is Gothmog seem >to have very aggressive kingsafety that sometimes backfire, or maybe the >assymetry you metioned before. > >Rasjid
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.