Author: Peter Fendrich
Date: 07:24:06 03/17/04
Go up one level in this thread
On March 17, 2004 at 09:32:16, Sune Fischer wrote: >On March 17, 2004 at 09:21:43, Peter Fendrich wrote: > >>Oh, I see. index is pointing at the highest value. I've tried that too! >>I have worked a lot with my history table... > >yes index=FunctionOf(to,stm,depth); > >There is no need to check other entries than the one being updated :) > >>>It's a form of normalizing the table, if you scale you must scale everything >>>otherwise only the best entries gets cut down and not the bad ones. >> >>I really thought that you tried some scrambling technique here! >>It's maybe not be as bad as it sounds, the good moves will prove themself very >>quickly again. The random part shouldn't be too dominating though. > >"scrambling", "random"? Yes, my relation to randomness is probably a bit weird after working 5 years with gentecic algorithms in the past and also before that dealing with the SSDF rating system! I see random behaviour everywhere - is that paranoia? :-) >One small problem with rescaling often is that a lot of entries might zero out >completely, that turns into pure random ordering. > >I think a small improvement here is to seed with pcsq or similar, but use values >so small that the real history scores easily takes over. Probably a good idea for you but not for me with a stm-fr-to table and no pieces. That could also explain our different experiences of deeper searches. The same pieces will probably not want to go the same places after advancing or regrouping. I'm not sure of the implications of that to a fr-to table but my first thought is that it is a little bit more stable for these kind of things. I'm not claiming that fr-to is better than piece-to in overall! There are many other factors to count in of course. /Peter
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.