Author: Robert Hyatt
Date: 13:25:08 03/01/06
Go up one level in this thread
On March 01, 2006 at 16:13:47, Frank Phillips wrote: >Thanks for sharing this. I am at the preliminary stages and just thinking about >refinements. Initial indications are: >1. Basic implementation, not if move extended or check or threat..etc, improved >scores on tactical tests. >2. Crude use of history scores helped again In this simple case calculating the >max. history score after each root move and only reducing if move history < 50% >(say). (My history scores are not 2^depth and therefore do not grow ultra >fast.). I haven't used 2^depth since the early history testing back in 1995 when I first started with Crafty. I have been using depth*depth to keep counters from overflowing too quickly and corrupting history values... >3. Next stage was thoughts about trying explicit counters of when moves >fail/not_fail high rather than using the history score, or adding piece_type to >the history counters, which you just mentioned below :-) The counters are better for me. So far, after just testing WAC with my usual scheme that says "terminate a position after completing the first iteration where the key move is found") has suggested 50% as the best result there. I am now doing the same test on ecmgcp, 20 seconds per position but terminating as soon as the key move is found, and I am trying history % values from 10 through 90 by 10's for starters...
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.