Author: Bas Hamstra
Date: 05:04:10 09/28/99
Go up one level in this thread
Hi Ed, Some measure the quality of movesorting as follows: a) count the total nr. of fail highs b) count the total nr. of FH's at the first move b/a represents the goodness of movesorting. Bob reports 94%. A friend of mine reports > 90%. I still have to measure, will soon. But I am not sure this is the most advanced test there is. It seems that movesorting near the root is VERY important for the total nodes searched, while near the leafs this is much less important. I would like to know *how* much less, but have no idea. Anyway, if that is true, this b/a number doesn't take that into account. Since most nodes are near the leafs you don't get a clue about the movesorting of nodes near the root, where it is most important... Regards, Bas Hamstra. On September 27, 1999 at 07:03:59, Ed Schröder wrote: >>Posted by Bas Hamstra on September 27, 1999 at 04:23:24: >> >>Hello Ed, > >Hi Bas, > >>>From one thing came another, I had a critical look at move ordering again, >>>made some changes, did intensive testing and there it was: 10-15% faster in >>>the end-game and 25-30% faster in the middle-game. >> >>Congratulations. I too notice that small changes can have large consequences. >>However in some positions system a is better and in some system b is better. >>What exactly did you change? Just avoiding playing HOOGSTENS moves? > >Just zero the best-move in the hash table in case of an UPPER bound. It gave >8% for Rebel. The search then is forced to rely on normal move-ordering. > >>>After 18 years wrestling, fiddling and twiddling with move-ordering this is >>>remarkable as I expected move-ordering was close to perfect in Rebel. I >>wonder how much improvement there is still left in move-ordering. >> >>I too. Maybe for positions near the root use ETC and sort accordingly? So play >>each move, look if the resulting position is in the hashtable. If no cutoff is >>possible, you could improve sorting failsoftvalues. This is something I am >>planning. > >Tried ETC too, it gave no speed improvement for Rebel. I do keep a special >table for moves on ply=2. In cases of a fail-low (drop in score) on the best >move so far it has proven to be very powerful when finding a new move in >such cases. > >Ed > >>>I don't know what a 25% speed improvement means in terms of playing strength. >>>Opinions are divided. I would say it gives 5-10 elo points. Others will say >>>15-20 and they can be perfectly right too. >>> >>>Ed >> >>Regards, >>Bas Hamstra.
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.