Author: Tony Werten
Date: 12:38:18 09/01/02
Go up one level in this thread
On September 01, 2002 at 14:11:01, Gian-Carlo Pascutto wrote: >On September 01, 2002 at 04:40:52, Tony Werten wrote: > >>>>>- FP I don't believe in. You? >>>> >>>>FP by definition can only save nodes when it's incorrect, so I threw it out. >>> >>>I do not understand. >>> >>>Do you mean that pruning based on evaluation,remaining depth and other >>>conditions is not FP. >>> >>>If it is not FP then what is FP? >> >>I'm quite sure (from previous talks with Bas ) he was talking about futility >>pruning wich is wrong by nature. >> >>FP prunes the last move ( before quiescence ) when the score is a certain margin >>below alpha. >> >>Now suppose FP is right but we don't prune. What happens ? We go to quiescence; >>evaluate; get a score above beta; and cutoff. So no nodes saved. >> >>FP can only save nodes when we would otherwise spend time in quiescence; ie >>score is not above beta ie fp was wrong. >> >>People who report big wins with fp probably have an "inc(nodecounter)" in the >>top of quiescence, wich doesn't get called when pruning. >> >>Tony >> >>PS I'm talking about nodecounts here not time. > >The difference in time should be minimal as well. You're saving a call >to quiescence, and a lazy evaluation. Your lazy eval should be no more >than just calculating the material situation. The difference is that >futility pruning only does that calculation once. > >What's your opinion on extended futility pruning and limited razoring? They can save nodes even if they are correct. It saves nodes in XiniX but I just don't like it. I see it happen that it takes longer to get high scores in testsets, or scores don't drop as hard as without pruning. So you do seem to be losing information. My idea is that pruning probably isn't wrong very often but when it is it's in crucial positions. Tony > >My experience with them is a love-hate relationship. I don't like the concept >very much, but each time I test them, I get 10% less nodes, and no tactical >problems. > >10% with no tactical problems is something I don't like to throw away, of >course. > >I suspect the real problem with it is that they reduce the amount of information >you get in your hashtables. Might cause extra problems in fail low situations. >And you don't want to worsen your worst case, of course. > >-- >GCP
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.