Author: Robert Hyatt
Date: 12:14:47 09/04/03
Go up one level in this thread
On September 04, 2003 at 11:20:45, Uri Blass wrote: >On September 04, 2003 at 11:00:41, Robert Hyatt wrote: > >>On September 04, 2003 at 09:56:21, Anthony Cozzie wrote: >> >>>On September 03, 2003 at 13:59:40, Robert Hyatt wrote: >>> >>>>On September 03, 2003 at 13:51:34, Gian-Carlo Pascutto wrote: >>>> >>>>>On September 03, 2003 at 13:48:21, Robert Hyatt wrote: >>>>> >>>>>>Mine is tunable. I run about 100 positions and pick the minply value that >>>>>>optimizes all 100 the best. I just checked and I am using 3 at the moment. I >>>>>>have not tested this recently. I think 3 was the optimal value for my quad >700. >>>>>> >>>>>>I will probably test this again on my dual to be sure that I have already done >>>>>>this correctly. >>>>> >>>>>You ignored my point, which is that saying that if you'll just split at 4 ply >>>>>to alleviate that problem causes another problem. >>>>> >>>>>It's not 'free' >>>>> >>>>>-- >>>>>GCP >>>> >>>>I never said _anything_ is "free". However "optimal" != "free" in this >>>>case. And since I have no choice, I go for optimal, whatever that is. There >>>>is a cost for splitting. If that cost is high, you have to limit the number of >>>>splits. Also, for obvious reasons, move ordering near the root is _far_ better >>>>than move ordering near the tips. Splitting where ordering is better is >>>>always a good idea. >>>> >>>>Note that I don't "just split at 4 ply" I split when there are >= N plies >>>>remaining in the search, where N can be set. I use 3 for the moment. I have >>>>run with 4 and 2 as well in the past, but for my program, my hardware, 3 was >>>>best last time I tested. >>>> >>>>I have noted that hardware influences this. On my P6/200 quad box, N=2 was >>>>better. >>> >>>also note that Diep & Sjeng probably have better move ordering than crafty >> >>Based on what? My fail high percentage on the first move is as high as >>anybody's I have seen. That is the critical measure. How many times, when >>you fail high, does it happen on the first branch as opposed to the second or >>later? The higher the percentage, the better. > >I think that this may be misleading comparison because it is also important if >the second move fail high or the third move fail high. > >I can understand why slow searchers can have better move ordering. >They can let themselves to calculate information for better move ordering when >for you calculating the same information may be too expensive and the gain may >be smaller than the time that you spend for the calculations. > >I express no opinion about comparison between Crafty and other programs but only >explain a reason to believe that Crafty can get worse move ordering relative to >slower searchers. > >Uri The reason I asked the question is that it is to simple to just assume that there is an explanation that _might_ be good enough to explain some sort of behavior. IE I don't claim my move ordering is better or worse than anyone. But if I believe it is causing problems, I will certainly _measure_ that to see if it is true, rather than saying that it might be the explanation. In the world of parallel searching, "might be" is _never_ good enough if you want to really produce good results... You have to _know_ what is going on in your program, or you have no chance for reasonable speedups.
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.