Author: Robert Hyatt
Date: 15:32:29 05/07/01
Go up one level in this thread
On May 07, 2001 at 16:48:41, Tom King wrote: >Hi all, > >there's been a few postings recently about strategies for programs to >effectively order moves at the root of the search. One strategy which is quite >interesting is the "order by nodes searched" heuristic, used by Crafty (and >maybe others). Currently, my program, Francesca, uses a very simple "order by >eval()" heuristic for ordering root moves. > >As a test, I played a series of games between the standard version of my >program, and a version which uses a "order by nodes searched" heuristic for >ordering root moves. I set these two versions playing blitz, last night. When I >just checked, they had played 130 games, and each had scored 50%. Each had won >37, lost 37, and drawn 56. Amazing that it should end so even! > >It's difficult to draw any conclusions from this. > >Probably more games are needed to see if "order by nodes searched" really helps. >My hunch is that it may be a win at slower time controls - move ordering becomes >more important the deeper you search.. [this reminds me of the time back in 1995 >when the move ordering in my program was very poor.. didn't seem to matter too >much at blitz, but at slower time controls, the search just ground to a halt. I >remember (fondly) how I made some improvements to Francesca's move ordering, and >suddenly she was searching throught the deeper plies 10 times quicker! Ah! great >days..] > >Rgds, >Tom The way to test is fixed-depth searches with both approaches. One will give slightly smaller trees. this is not a _big_ win, just a win. (It _can_ be a big win in some cases, and once you do a parallel search it will be an even bigger win...
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.