Author: Dann Corbit
Date: 13:45:41 02/04/04
Go up one level in this thread
On February 04, 2004 at 14:15:33, Tom Likens wrote: > >Hello Everyone, > >I'm getting ready to implement some new features in my engine and >one of the items on my TODO list is to give the engine some rudimentary >planning capablities. The idea I've been toying with is to use the >features of the pawn structure (and the king locations) to guide what >the engine should try over the next few moves. The advantage of using >the pawn formation as a trigger or even index into a variety of evaluation >functions, (perhaps dynamically selected by a function pointer at runtime) >is that the information could be stored in the pawn hash table for >easy retrieval. > >As an example, if the center is blocked and white has a preponderance of >force and space on the king side then it may be okay to start a king- >side attack using the white king pawns to break open the black king's >castled location. At the very least moves like g4, h4 should be considered. >This would be especially true if the black queen was trapped on the >queenside and the black's king safety was already weakened. Other >factors would be control of the squares around the enemy king, but the >initial trigger would be the features of the pawn structure. > >Anyway, this is just one example other patterns could be the Stonewall, >the pawn chains that arise out of the French Defence, the Caro-Kan etc. >or even recognizing when a queenside minority attack made sense. > >In fact, I've starting using a variation of this idea in the king safety >routines. >I recently started saving the pawn state of the six squares directly in front >of the king (if a pawn is present then bit is set in an 8-bit index). This >value, once calculated, is used as an index into an array of function >pointers that point to functions tailored for the specific pawn structure >(e.g. if the pawn structure is a fianchetto structure then elements related >to the strength or weakness of the fianchetto bishop etc. are checked). >So far, the results have been encouraging but it needs a lot of tuning and >experimentation. > >Anyway, I was wondering has anyone else here given any of these >ideas a try yet? And if so, how successful have they been? > >regards, >--tom Suggestion: Make everything parameter driven (including whether to include an evaluation term or not). Read the setup stuff from a database or initialization file. Then you can do all kinds of interesting experiments without recompilation. You can look at Beowulf's personality code as an example of paramter driven evaluation.
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.