Author: Robert Hyatt
Date: 07:04:22 09/18/99
Go up one level in this thread
On September 17, 1999 at 22:31:57, Will Singleton wrote: >On September 17, 1999 at 12:33:46, KarinsDad wrote: > >>Robert, >> >>A very difficult problem to resolve. >> >>Maybe the best solution is to not resolve it heavily at all (i.e. only resolve >>it to the level you already have or maybe slightly moreso). >> >>In my program, I am attempting to make positional moves over tactical moves. >> >>So, I have multiple tactically sound PVs which I decide between. One of the >>criteria that can be used to decide between them is to determine pawn structure >>at the leaves of each "PV". If you had 30 different pawn structure "ideas" in a >>database (or hardcoded), you could have a preset evaluation for each of these >>structures (i.e. 3 white queenside pawns vs. 2 black queenside pawns where the >>black pawns are doubled is one pawn structure "idea"). >> >>You could then use this "evaluation" to positively or negatively affect the >>decision on which PV to choose (or at least I can do this in my code, you would >>have to change Crafty drastically to implement this type of model). >> >>Hence, if I have 5 PV scores all within 1/8th pawn of each other, I could pick >>one which SHOULD result in a favorable pawn structure within the next 16 ply as >>opposed to one which might not. >> >>And, the other advantage of this model is that most of this pawn evaluation >>"stuff" is handled just a few times for each "PV" just before deciding on a move >>as opposed to within the evaluation code for every node that gets evaluated. >>Hence, it could be relatively sophisticated (i.e. it could even take into >>account things such as how advanced the pawns are) and still not take a lot of >>time. >> >>So, the basic idea is to avoid queenside majorities and other endgame weaknesses >>if possible much earlier in the game. If I have two PVs of similar score, one >>which should result in a queenside majority for my opponent and one which should >>not, I pick the one which should not in order to avoid the potential weakness on >>move 13 as opposed to attempting to evaluate it at move 45. Of course, this type >>of solution will not work in all positions, but it should emulate better human >>strategic play as opposed to purely tactical play (i.e. attempt to avoid the >>pitfalls and cliffs before the position starts heading towards them). Part of >>the "plan" is to give the program's side the best pawn structure it can way >>before getting to the endgame. >> >>Just an alternative idea. >> >>KarinsDad :) > >So, as I understand it, you are suggesting that the time it takes to generate 2 >complete pv's (or more) is less than incorporating a pawn structure eval in a >normal (single pv) eval. > >It would seem to me that in order to get 2 good pv's to the same depth would >require twice the time as one pv. Of course, I'm on my second vodka & tonic, so >maybe I'm missing something. > >Will Remember that KD's program apparently isn't based on a traditional iterated alpha/beta search as we know it. He seems to be trying something more selective, and maybe more of a best-first search rather than a traditional depth-first search...
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.