Author: John Merlino
Date: 19:07:04 10/06/01
Go up one level in this thread
On October 04, 2001 at 21:08:08, Dann Corbit wrote: >On October 04, 2001 at 14:40:47, John Merlino wrote: > >>On October 04, 2001 at 13:55:40, Kurt Utzinger wrote: >> >>>Is there any differnce in playing strength between these three version of >>>TheKing. My experience show no difference betwenn 3.12 and 3.13c but so far I >>>have not tested 3.12d. >>>Regards >>>Kurt >> >>There is no difference in playing strength/style/quality between the 3.12 and >>3.12c versions (and the versions in between them, obviously). However, 3.12d >>does improve blitz/bullet play under certain circumstances, specifically when >>using a very large hash table. >> >>The King clears its hash table on every move. In my testing with 3.12c on a >>PIII-600, it took about 1/3 to 1/2 second to clear a 32MB hash table. This is >>not significant during standard play, but losing 1/2 second per move in bullet >>is costly. The King 3.12d drastically improved the method in which the hash >>table is cleared, so this no longer becomes a problem. However, once again, this >>was only a problem if a large hash table was being used in a fast game. > >I find that really, really puzzling. When analyzing EPD records (for instance) >it makes great sense to clear the hash table between moves. > >When playing a game, you are throwing away a huge amount of computational >information. Let's suppose that they pick the move I was pondering... In that >case, I should already have computed the answer to quite a large depth. Why >solve them over from scratch? Stale entries should get over-written anyway. > >To throw away the hash table between moves by clearing it is bizarre. Why would >anyone do that? Sorry that it took so long to respond to this, but I decided to get an explanation from Johan. So, here it is (TT = transposition table = hash table): ----------------- The main reason is still reproducability. If the TT is preserved during a game, it is really hard to find out if odd behaviour is due to a real bug, an alleged bug, or a known-but-tolerated side effect of the TT (that's what TTs are famous for). There are other annoyances involved as well. Eg slightly more code and data. Centralization tables that change during a game render the evals in the TT invalid. Screwed up statistics for the early plies may misguide the root of the search. And the benefits are really not as good as Dann suggests. The data in the TT cover only a depth-2 search on average. And only if the opponent plays the ponder move, else the data is only partial due to the alpha-beta algortithm. So up to now, I've always considered the benefits to be outweighed by the (possible) problems. ----------------- There you go, jm
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.