Author: leonid
Date: 06:08:00 09/09/00
Go up one level in this thread
On September 09, 2000 at 02:02:46, TEERAPONG TOVIRAT wrote: >Hi, > >I want to estimate the efficacy of each method >of moveordering() . Some problems arise. > >Should I use branching factor to determine the >efficacy ? Yes! If so,the next problem is how to obtain >the factor ? It varies from ply to ply and from >game to game. By general statistics. 100 positions verified could give pretty good picture. I installed contant counters into all my plys. One number (for each ply) say how many times ply was accessed. Other indicate how many nodes in each ply was seen. This way branching factor is visible for each ply if asked. Total time count is decisive. NPS is useful for general orientation. Higher NPS and shorter time in search are good indicative of true speeding. Comparison of NPS between different programs is still good indicative of basic speed but it could be misleading. High NPS could indicate that program do its search inefficently. Only when your program do better in time that many other programs and keep its NPS as the best one, there are good chance that your speed is the best as well. In order to approximate it, should >I use geometric mean or arithmetic mean ? I find very often branching factor by finding the time for the same position after each two plys. If time for ply 6 is 1 sec. and for ply 8 is 25 sec, then 25 sec. : 1 sec. = 25. Branching factor is ^25=5. >Or it would be better to use total searching time ? >How u did it ? Total time is for me ideal factor. When I compare for my program, it is simple. I have many positions that I keep with the indication of total time, that it took at such a day to search this position at given number of plys. Shorter time, better search. When I compare with other programs, it is more that obscure and problematic. Extensions are the reason for this. To make my comparison as much credible as possible, I compare brute force search done at fixed depth. Total time spent for position and NPS represent some visible indicative. Since I don't use for now extensions, I consider that opposit side could go more slowly that mine between 3 and 5 times for the same position. 3 or 5 times when my NPS is 40% higher. If my NPS is 300% higher that in opposit program, then next program is just crazy about extensions. Now its time advantage should be put between 5 and 7 fold. Leonid. >Thanks, >Teerapong
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.