Author: Robert Hyatt
Date: 19:04:48 02/12/99
Go up one level in this thread
On February 12, 1999 at 17:22:19, Amir Ban wrote: > >On February 12, 1999 at 01:57:01, Komputer Korner wrote: > >>In Henri Arsenault's Games Domain review of Fritz 5-16 bit he states that "After >>each variation is analyzed, the hash table is emptied, so this is a much more >>efficient way to analyze a game." I say Oh REALLY ???!!!!!!!!!! >>Any correct hash table info (even a full one) is better than no hash info. When >>the hash table is full, programs implement a replace hash algorithm which >>determines ( many programs have 2 hash tables or even 3 if you count the pawn >>hash table) which hash entries get knocked out. If a hash entry is not there >>(got knocked out) and that position at the same depth comes up again, then the >>program of course has to calulate the score over again. BUT if it doesn't get >>knocked out then time is saved. That is the whole purpose of a hash table. I >>realize that initializing the hash table takes cpu time, but whether the hash >>table is full or empty at the beginning of the move should not affect this. >>Therefore I wonder why the Fritz and Junior engines do this. Don't they play a >>little weaker because of this practice? >>-- >>Komputer Korner > >Common sense tells me that if there's a loss, it's small. Assuming the >transposition table was generated when it was last the program turn to play, >when it's again our tun to play, we are now two plies deep in the old tree. >Assuming a branching factor of 6, I would expect then that only about 1/36 of >the old transposition table contents is still relevant, and even this part is of >limited use because the depth draft in those entries is now smaller (by >typically 2 ply) from what is needed in the new search. > >My conclusion from this is that the previous table is only about 1%-3% as >effective as it was on the search when it was generated (previous move). > >I think the transposition table is the element of computer chess that is most >consistently over-valued. (Second place goes to the opening book). > >Amir Here is a good test: pick a good middlegame position. with a time limit of 3 minutes, search it. Stop your program. Restart and set up the new position and search it for 3 minutes. Do this for 5-10 moves. Then back to the beginning, restart and search for 3 minutes, but don't exit and re-start, just continue the game with the old hash table. I generally find that this is close to 2x faster. The scores might not work, but the _move ordering_ certainly does. And if you do 12-13 plies, I find a lot of stuff is useful. I ran this test on Cray Blitz a few years ago and found saving everything was about 2x faster overall... YMMV of course... But I personally get _significant_ performance penalties when I clear things out.
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.