Author: martin fierz
Date: 05:57:49 12/14/00
Go up one level in this thread
On December 14, 2000 at 06:50:45, Steve Maughan wrote: >My program doesn't use fail soft atr the moment but I have been reading about it >on the web. I hadn't realised that with fail soft the score, if less than >alpha, is an upper bound and if greater than beta is a lower bound. I can see >that this would be useful - but how useful is it in practice? > >A couple of things I can think of doing if this is the case: > >1) Back up the hash score with the returned value as the bound and NOT alpha or >beta. This would give more hash cutoffs but is it accurate when one has null >move & futility pruning etc? in my checkers program i initially returned alpha on a fail-low instead of the best fail-low-value - i did return the best value on fail-high, so i had some kind of hybrid fail hard/soft - of course this was merely by stupidity and not by design... when i changed to complete fail-soft i think i needed about 5% less nodes on average to depth 17 in my test suite. so i didn't get a huge improvement, but at least it was for free :-) maybe i could plug fail hard back in the program and give you the result some time, but then again i don't know how well the two games compare - it would surely be better to get a number of a chess programmer. i don't know how a null-move would affect this since the null-move concept doesn't translate well to checkers - too many zugzwang situations. OTH i'm using much more agressive pruning schemes in checkers than chess programmers, and i never noticed a problem with them. > >2) Use the score as a better bound for a fail high - obvious but I can't see >anyone doing it. As an example Shredder 5 seems to fail high at +0.25 above >current score then +0.50 above current score. Surely if this fail soft property >is robust there must be a better way. even though i use fail-soft i mostly get exactly the window value as return value - i suppose it would be some marvelous coincidence if EVERY refutation you find scores higher than the current beta - and if you get only one single one with beta, you will end up with beta at the root. >Are all of these insights correct? i'd say: yes. > >Regards, > >Steve Maughan cheers martin
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.