Author: Tord Romstad
Date: 02:41:18 03/02/06
Go up one level in this thread
Hi Bob, Thanks for starting this interesting discussion! It's a pity it appeared just now, when I don't have much time to read or post to the CCC. :-( On March 01, 2006 at 14:52:06, Robert Hyatt wrote: >My next question: Is anyone else doing the "to reduce or not reduce" decision >based on something similar? Or are there even better ideas waiting to be >discovered. If you are using this (I would be surprised if someone has not at >least tried this scheme, whether they improved it further or not) have you had >any particularly good threshold values you've discovered? While I plan on >arriving at mine based on my usual testing, it would be interesting to see if my >final number is close to others (Tord? Vas? Fabien? anyone else trying this?) My own experiments followed a pattern rather similar to yours, and I have tried all the schemes you describe (or at least something very similar). Unfortunately I have never been a very patient or systematic tester, and I haven't spent a lot of time searching for optimal threshold values or percentages. In the numerous short tests I did, however, there never seemed to be a big performance difference between the various parameter values I tried. The engine strength remained approximately constant. The small strength difference between the many schemes and parameter settings I tried made me suspect that history counters are too erratic to be useful as the main criterium for reduction decisions. The biggest recent leap in playing strength for my program recently (except for parallel search, of course) came when I finally decided to drop the use of history counters altogether, and use evaluation data and null move threat detection instead. This also makes sense from an intuitive point of view, IMHO. When deciding whether a move should be reduced, the impact of the move on the current position should be much more relevant than how the same move has performed in the often totally different positions we have visited earlier in the search. Removing history data from my ok_to_reduce() function seems to have given me about 50 Elo points. Not bad for such a small change ... Tord
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.