Author: leonid
Date: 18:37:55 06/17/00
Go up one level in this thread
On June 17, 2000 at 15:03:21, blass uri wrote: >On June 17, 2000 at 14:57:35, blass uri wrote: > >>On June 17, 2000 at 14:40:47, leonid wrote: >> >>>On June 17, 2000 at 09:53:11, blass uri wrote: >>> >>>>On June 17, 2000 at 09:43:45, leonid wrote: >>>> >>>>>On June 17, 2000 at 08:40:11, Bas Hamstra wrote: >>>>> >>>>>>On June 16, 2000 at 05:41:35, leonid wrote: >>>>>> >>>>>>>On June 16, 2000 at 03:05:12, Bas Hamstra wrote: >>>>>>> >>>>>>>>For alphabeta, on a Celeron 466, doing only material: 800.000 positions /second. >>>>>>> >>>>>>>Thanks for your response that is perfect and clear! It correspond exactly to >>>>>>>what I am asking to know. You indicated speed for "doing material". My name for >>>>>>>this is "positional logic". If you still will be able to give some concret >>>>>>>position (or two positions) with concret numbers, it will make your response >>>>>>>even more complet. >>>>>>> >>>>>>>Thanks! >>>>>>>Leonid. >>>>>> >>>>>>You can use the "rule of thumb" that with more evaluation you can divide this >>>>>>number by at least 2, for a normal leaf processor. So with a normal eval I >>>>>>expect something between 200.000 and 400.000 NPS. It depends on how smart you >>>>>>want to make your program. >>>>>> >>>>>>Download Crafty and measure its NPS on your own CPU. If you program is not too >>>>>>dumb and NPS is in the same ballpark as Crafty with full eval, that's >>>>>>reasonable. If it has very little eval but is still 4x slower than Crafty you >>>>>>might want to redo the "core" routines and/or datastructures. Some interesting >>>>>>things to measure: >>>>>> >>>>>>- speed of make/unmake() >>>>>>- speed of a sorted GenCaptures() >>>>>>- speed of SquareAttacked() >>>>>>- speed of Static Exchange Evaluation (SEE) >>>>>> >>>>>>Of course speed isn't everything, but on the other hand it is "comfortable" to >>>>>>know your "core" is ok. >>>>>> >>>>>> >>>>>>Regards, >>>>>>Bas Hamstra. >>>>> >>>>> >>>>>Thanks! >>>>> >>>>>I probably already did what you have suggested. I tried Crafty and few other >>>>>best program that we have in Hiarcs package. There I cold see usual NPS for >>>>>those programs. Since you indicated before number of NPS for minimax (our >>>>>computers are almost identical) I could calculate curious factor for them. >>>>>Apparently mentined factor is the same for them and for me, around 5. This have >>>>>me some expectation that my moves ordering is already now close to the best one. >>>>> >>>>>Your number of minimax is astoundingly close to mine. On AMD 400 it is between >>>>>800000 and 1100000. Average number of NPS (normal logic) is around 200000. For >>>>>best games this number is around 150000. Probably still I must push a little bit >>>>>efficency of my move ordering to reach them. >>>> >>>>The number of nps is different for different top programs. >>>>You cannot learn from the number of nps if your program is good or bad. >>>> >>>>Uri >>> >>>Agree with you 100%! Only when you want compare positional logic (material >>>echange) you are in some borderless and strange place. Even some general >>>indication make you feel better. >>> >>>Perfect idea about speed of the program could be found only by solving mate >>>containing position. >>> >>>Your saying about nps make me think about Hiarcs numbers. They are actually >>>twice below others program figures. Enigmatic and beyond my explanation. As >>>"maybe" I see only two things: >>> >>>1) Perfect move ordering. Better is the move ordering lower is NPS. >>>2) Extensions. >>> >>>If somebody could explain this anomaly, it will be nice. >> >>I think that extensions is the reason. >>Another possible explanation could be the evaluation function but I read that >>hiarcs does not use most of the time about evaluation. >> >>Uri > >I remember that Amir Ban claimed that Junior is using only less than 20% of its >time for evaluiation and that other top programs also do not use more than 50% >of their time for evaluation. >I asked if hiarcs does not use about 90% of its time for evaluation(I suspected >that it is the case because of the nodes per second) and he replied that it does >not do it. > >Uri Uri, what exactly is the "evaluation"? Material exchange? Leonid.
This page took 0.02 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.