Author: Jaime Benito de Valle Ruiz
Date: 13:34:55 06/16/04
Go up one level in this thread
On June 16, 2004 at 15:43:29, Tom Likens wrote: >On June 16, 2004 at 10:54:18, Jaime Benito de Valle Ruiz wrote: > >>I'm re-writting the eval code to give consider different weights at different >>stages of the game... which is something I should have tried earlier. I've even >>considered having a late-midgame stage. >> >>I've tried a few methods for differenciating these stages of the game, all of >>them based on counting pieces on the board, but I suspect it has more little to >>do with the number of pawns, and it might not be so simple. >> >>Can anyone give me some suggestions on this? >>Thanks, >> >> Jaime > >Hello Jaime, > >In Djinn I scale a number of the evaluation terms based on the amount of >material left, excluding pawns. If you give the pieces pawn equivalent >values of 3, 3, 5 and 9 (used only for indexing purposes of course) then the >index can range from 0 to 31. I keep track of this number incrementally >during the search so it is always available. To use it I have a number of 32 >entry floating point arrays with names like "Scale3to1", "Scale1to6" etc., >(I wrote a C program to generate all the arrays and the floating point values). >These arrays scale linearly from 3.00 to 1.00 across 32 values or 1.00 to >6.00 and so forth. > >I also scale the values for white and black independently depending on >the material for each side. One caveat, with using this method- often a score >will be scaled depending on the amount of material the *opponent* has (in fact, >this is often the case). A good example is king safety, if your opponent has >little >material left then it may be okay to bring your king into the fray, regardless >of >how much material you have left. Once you start down this road, it's amazing >how many places it can be applied. I use this idea to scale passed pawns, >king attacks, center control etc. > >Another option, which I don't currently use, is to blend (i.e. interpolate) the >final >score from the middlegame evaluation and the endgame evaluation. I believe >Phanlanx, Gothmog and perhaps Fruit all use this technique so you might >inquire from them how they apply it. I've always shied away from this idea >because it seemed computationally expense, but perhaps not. > >regards, >--tom Thanks for such a detailed post, Tom. You're the second today telling me about this scale idea. It's given me a lot to think. Regards, jaime
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.