Author: Robert Hyatt
Date: 21:35:22 03/02/06
Go up one level in this thread
On March 02, 2006 at 17:16:52, Dezhi Zhao wrote: >On March 02, 2006 at 11:40:51, Robert Hyatt wrote: > >>On March 02, 2006 at 00:48:59, Dezhi Zhao wrote: >> >>>On March 01, 2006 at 15:51:29, Robert Hyatt wrote: >>>>I increment the fail high count by 100 each time a move fails high. I also >>>>increment the tried count by 1 at the same time. Any time another move fails >>>>high, and this move was tried but did not, its tried count gets incremented, >>>>which lowers the "failed high percentage" significantly... >>>> >>>>sort of: >>>> >>>>if failed_high(x) >>>> x.count++ >>>> x.fh+=100 >>>> >>>>if failed_high (y) (later) >>>> y.count++ >>>> y.fh+=100 >>>> if (tried(x)) x.count++ >>>> >>>>Hope that makes what I am doing clearer, whether it is good or bad is not >>>>certain yet... >>>> >>>> >>> >>> >>>why not keep it simple?! >>> >>>Suppose x is any move we just searched. >>> >>>x.count++ >>>if failed_high(x) >>> x.fh+=100 >> >>That grossly biases the statistic downward. You can either count: >> >>(a) times a move was searched anywhere vs the number of times it failed high >>(that ratio will likely be very high, as fail-highs are not that common) >> >>(b) times a move was searched and didn't fail high when another move did, vs the >>number of times it was searched and did fail high. That seems to be a more >>useful comparison, because now we can say "something failed high, but not this >>move..." >> >>In case (a) there are plenty of times when all moves fail low, particularly when >>move ordering is very good. And the count value will be very high, but the fh >>count will be very low. That would say this is an ok move to reduce. In case >>(b) if something fails high but this move doesn't, then that is more useful to >>know how many times that happened vs the cases where this move failed high >>itself. >> >>Intuitively it seems better. Whether it actually is or not has not been tested >>by me (yet). > >As for (b) approach, it seems to me that you're trying to get some locality out >of a global killer (or history) table. > >This is somewhat foreign to history table. If results of (b) are not much better >than (a), you may have a good reason to consider the history killer table I >mentioned in another post in this thread:) I really wasn't aiming for "locality". I'm aiming for "understandability". The raw number of fail highs is not that interesting. The raw number of non-fail-highs is the same, at least IMHO on first glance. But if you know that a move failed high at some ply, any other moves that were tried and did not fail high deserve special notice. I simply increase their "use" counts which degrades their "fail high percentage". This makes it easy to compare a move that failed high 3 of 5 times to a move that failed high 200 out of 600 times, even though the "confidence" in the first set of numbers is necessarily low due to small number of observations...
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.