Author: Uri Blass
Date: 08:31:30 06/27/03
Go up one level in this thread
On June 27, 2003 at 11:13:37, Sune Fischer wrote: >On June 27, 2003 at 11:00:56, Uri Blass wrote: > >>On June 27, 2003 at 10:46:08, Sune Fischer wrote: >> >>>On June 27, 2003 at 08:13:53, Albert Bertilsson wrote: >>> >>>>Just wanted to share my latest findings and warn you from making the same >>>>mistake (or maybe I just point out the obvious). >>>> >>>>I've always been proud to have a very fast movegen, doing perft calculations >>>>very fast. I optimized the movegen alot, and always saw playing strength >>>>increasing. I use a legal move generator and always saw the cost of verifying >>>>legality to be minor. Lately however I've noticed that the engine performance >>>>isn't that great (compared to other engines) inspite of the fast movegen. So I >>>>started digging today and the numbers struck me as lightning... >>>> >>>>As I've improved moveordering very much with hashtable/iterative deepening, >>>>killer moves and history tables the effective BF is now much lower. The lower BF >>>>now means that very few moves of all the generated ones are actually ever made. >>>>This means that the cost of generating legal moves have increased incredibly >>>>much. >>>> >>>>It's really anoying to notice that something you though was very good now is on >>>>the list of improvements again =(. >>>> >>>>/Regards Albert >>> >>>What that saying with premature optimization being the root of all evil? :) >>> >>>Well it think it is correct, and probably a typical mistake to spend oceans of >>>time optimizing one thing only. >> >>I do not think that the Albert spent ocean of time in optimizing. >>Based on reading his posts I got the impression that he did not do it. > >Ah hmm, but I get the opposite impression. :) I remember that in the winboard forum he said that he did not work much about it. He did not add a lot of tables to do his makemove slower and to do his gen faster. I believe that it can help me to do my perft faster but the main advantage is the knowledge that I get from the tables and not speed. > >>His program is relatively fast in perft because his program has a different >>structure(he use legal move generator when most of the programmers do not do >>it). > >Perft is no good speed indicator anyway, not even for a move gen. >The perft tree is so different from the chess tree. > >Optimizing for perft is basicly like building a formula 1 car to use for >off-road rally. I think that it is dependent on the structure of the program. 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.