Author: Robert Hyatt
Date: 21:04:52 01/16/99
Go up one level in this thread
On January 16, 1999 at 21:22:28, Chris Moreton wrote: >I am in the process of conduction a number of tests regarding the performance of >alpha-beta and its various enhancements using Rival. Clearly a wealth of >information and publications exist already on this subject and it may be >considered no longer interesting but my reasons are also to see how much effect >other search parameters have on the engine, notably, the quiescense search and >search extensions. The tests searching the 10 nunn positions to depths 2 thru 7 >for various combinations of AB enhancements. > >The question I would like to ask regards the history heuristic. Jonathon >Schaeffer's work on the HH (The HH and AB search enhancments in practice) shows >that it provides significant gains when applied to plain AB with no knowledge >move ordering and slightly more gains when used in conjunction with knowledge >ordering. JS mentions that in his test program that there was a certain amount >of move ordering even when no knowledge was applied (e.g. capture moves were >generated first). > >My results indicate that when the move order is fully randomised, the HH has no >positive effect on node count when added to plain AB, often exceeding the pure >AB figure. Other results from my tests regarding the HH seem consistent with >JS's observations and show that at least up until depth 7, HH outperforms a >transposition table with approx. 300,000 entries when applied to AB+Knowledge >ordering + Minimal Window searching. > >Am I implementing the HH incorrectly, or are these results consistent with >others observations? > >Chris The history heuristic is 'global' in nature, in that moves from all over the tree affect the history counts and hence move ordering. This should _never_ be better than the hash table move or non-losing captures. _ever_. If it is, something is wrong somewhere else, because the hash move is specific and not global. IE this move was best in _this_ position before... Random move ordering isn't interesting to test, because you get random results...
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.