Author: Gerd Isenberg
Date: 05:43:44 03/12/04
Go up one level in this thread
On March 11, 2004 at 13:55:45, Bob Durrett wrote: > >A few days ago, someone was drawing a distinction between knowledge in >evaluation and knowledge in searching. Domain dependent chess knowledge is immanent inside the evaluation. It is used to assign scores to the leave nodes, while the (alfa-beta, pvs, mdt...) search backups or minimax these scores to the root. Since move sorting is very important in alfa-beta, some standard heuristics are also dominated by chess knowledge, eg. the value of pieces to distinguish between good and bad captures for instance. Inside the evalscore there is no differentiation between several aspects of the evaluation. The eval may report a zero score for equal positions, for draws by insufficient material, for positions with material up or down but positional compansation, and mutual compansation of (huge) volatile positional aspects such as advanced passer against king safety... Depending on the implementation one may feed some other eval information (volatile, danger or uncertain flags) to influence the search. At interior nodes, static chess knowledge (whether implemented in eval or other special routines) may guide the search in doing extensions, reductions, foreward pruning and better move sorting. But there are, probably more important, domain independent heuristics inside the search, like nullmove heuristics or singular extensions - but of course that's knowledge too. > >I am intrigued and very interested in knowing more about "knowledge in >searching." Some of the questions are: > >(1) Is this commonly used or done in chess engines? > Yes, more or less. Even if it is to extend checks or moves out of check - or to try "good" captures before "bad". >(2) What sort of knowledge would be used in or for searching, other than >knowledge used in evaluation? Whether a move is "interesting" or not and whether further search depth should be extended or reduced. Does the move attack or defend something important, how does a move affect own and opposite attack maps considering several areas of the board. For nullmove observation it is usefull to consider some tactical threat heuristics a bit more, pieces en prise or hanging, pins, forks etc... > >(3) How much can searching be improved by using knowledge in searching? > Difficult to quantify. For move sorting it is essential... >(4) Any other insights into this topic = ? > >Sincerely, > >Bob D. Cheers, Gerd
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.