Author: Robert Hyatt
Date: 19:11:31 09/08/99
Go up one level in this thread
On September 08, 1999 at 18:20:44, Alessandro Damiani wrote: >On September 08, 1999 at 14:30:12, Robert Hyatt wrote: > >>On September 08, 1999 at 13:37:31, William Bryant wrote: >> >>>On September 08, 1999 at 11:20:54, James Robertson wrote: >>>> >>>>I have had a terrible time trying to implement something like this, and finally >>>>gave up. How much does it help Crafty/Screamer? Is it worth it to try to get it >>>>to work? >>>> >>>>James >>> >>>I also had a terrible time until I finally worked out the errors, and there were >>>several. The problem lies in how to structure the aspiration windows, >>>the researches for fail-high and fail-low, and then adding PVS searching on top >>>of it. >>> >>>Bob and others were a lot of help. >>> >>>My problem now is that my SEE seems to be pruning to much. On rather quiet >>>lines, it does a great job of getting to the next ply. On positional issues, >>>it makes the program play much poorer. >>> >>>For Example, on LCT-II the POS part of the suite, the score drops in half, >>>missing several of the answers with the Delta-SEE pruning in place. It makes up >>>some of this in the tactical middle game portion. >>> >>>When playing the two versions against each other, at short time controls >>>(2s/move), the SEE version clearly dominates. At longer controls (5-10s/move or >>>greater), the slower more positional version does better. >>> >>>I am operating under the assumption that there is a bug somewhere and that >>>eventually the SEE will give me the speed without sacrificing the positional >>>skill. >>> >> >> >>I have done a lot of testing here, and I don't notice any particular positional >>'drop-off' with SEE on or off, other than the fact that it gets me almost one >>more ply since it reduces the q-search by over 50%. >> >>You might have your 'delta' window too narrow. Of course, making it too wide >>disables the pruning... so, like Goldilocks and the three bears, you have to >>get it "just right"... :) >> >> > >What is the 'delta" used for? Thank you. > >Alessandro > > It is a 'guestimate' of how far up the positional eval can pull the score in addition to the material being gained(lost) by the capture being considered. IE suppose alpha is 0 (dead equal game) and when you get to this node the material score is -400. If you play PxN, you gain 300, so the score is at -100, which is _still_ well below alpha. What do you assume your positional score can add? 100? If so, then you ought to try this capture. 200? Then you definitely want to try this capture? 50? then that won't pull the score above alpha and there is no point in trying, if you are pretty _sure_ that 50 is the biggest score you can get... > > >> >>>It is possible that the positional 'insight' is accidental, finding the right >>>move without knowing why. >>> >>>Any insights in getting you SEE to work would be appreciated. >>> >>>William >>>wbryant@ix.netcom.com
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.