Author: Uri Blass
Date: 14:55:11 08/07/05
Go up one level in this thread
On August 07, 2005 at 16:59:17, Stefano Gemma wrote: >On August 07, 2005 at 16:14:47, Ricardo Gibert wrote: > >... >> TC 1 min 2 min >> w/o mod: 45% 46% >> w mod: 40% 44% >> >>Apparently, an improvement with the mod as far as long TCs go. >> >>More questions: Does this idea have any practical value or would the results of >>any such testing be too often inconclusive to have any practical value? Has >>anybody already tried this? Has anybody tried this and found it wanting? What is >>wrong with this idea? > >Just today i've done a little test of the new release of Freccia, against the >very stronger program Delfi. There were no chance for Freccia, obviously. What >i've noticed is that at one minute per game, Freccia looses in no more than 18 >moves. In 30 minutes games, instead, Freccia looses in about 32 moves. What it >means, for your question? IMHO it means that a longer TC tends to make different >programs on the same level. So, if you modify your program, as long is your TC >and as difficult is to see any improvement; this is true not for the time you >have to spend, but for the fact that the two version tends to be equal when you >increase the TC. > >Stefano > >Stefano I do not agree. number of moves is not important and there are cases when it is more easy to see improvement at long time control. Imagine that you have a small improvement in the order of moves(no change in the evaluation or in pruning or extensions) but the program is slower in nodes per second because for better order of moves you need to calculate more information and calculating the information cost time. It is logical that if you test at blitz time control the program may be worse and it may be better at longer time control in this case. I think that it may be better to test improvement in test suites to test if a version is better at long time control. The idea is the following: Take test suite with many positions. define X as the minimal time that the program need to solve 50% of the test suite. How much problems the program solve in time of 2X,4X,8X,16X I think that programs that show better results are probably relatively better at long time control. Note that you need to be careful not to include times of initializing structure inside X in case that the program always start it's search by initializing some structures because in this case the program may be weak at very fast time control because of 0.1 seconds that it need to initialize the structures in every move but it does not tell you much about the difference between 20 seconds per move and 40 seconds per move when the 0.1 second is almost meaningless in both cases so it is better if the program counts the time and it starts to do it only when it is ready to search. Uri
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.