Author: leonid
Date: 17:21:45 07/26/00
Go up one level in this thread
On July 26, 2000 at 18:26:39, ujecrh wrote: >I never understand these nps issues. How can we tell if one program processes >less or more nodes per second than another ? For example: >_ What says that Fritz is not counting leafs as well ? >_ How do we know that Hiarcs is not skipping nodes from quiescence search in its >count ? >_ Is there any reason why a program should count hash hits or not ? >_ Are cutoffs counted or not ? >etc. etc. > >These examples are a bit extreme but I mean that maybe we are just fool to put >any faith in the nps given by different engines. What's more it may happen that >two versions of a same program change the way nodes are counted, look for >example at the last Yace release notes: > >"Changed the method of counting nodes. Yace will now count less nodes >and will show less nodes/s. Actually it is even slightly faster, nevertheless." > >Souldn't we just say that nps does not mean anything or at least very few ? Like everywhere we can insiste on exceptions but it is important to see that general rule is simple. In each ply must be counted only those nodes that were used for finding the best move inside of it. If, for instance, we produced 35 legal moves for certain ply, but after using 3 we found already the best value for that ply, only 3 must be added to the NPS counter. High NPS could be sometime indicative for very low efficiency of search. This is how minimax search will give NPS around 4 time as big as in usual advanced search. Leonid.
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.