Author: Pauli Misikangas
Date: 05:26:38 10/28/98
Go up one level in this thread
On October 28, 1998 at 07:36:45, Ulrich Tuerke wrote: >On October 28, 1998 at 07:18:47, Pauli Misikangas wrote: > >>In my Shocky (which is mainly a shogi program, but can play also chess, although >>quite poorly at the moment) I use fully incremental evaluation. During every >>move, I update only those parts of the evaluation that have changed. So, when >>the "evaluation function" is called I only sum all evaluation-values (material, >>attack, ..., see the list below) together. Currently, the evaluation consists of >>the following parts: >> >>- material >>- attack >>- defense >>- board control >>- piece position >>- piece movement >>- castling (not as in chess, pieces around the king) >>- king safety >>- hand pieces (in shogi you can drop captured piece back to the board) >>- piece freedoms (number of safe squares to which pieces can move) >>- blocked pieces (penalty for pieces that cannot move safely) >>- plan success (for example, Shocky might have a plan to make a castle) >> >>I quess that many of those values would make my evaluation extremely slow if >>they were calculated each time 'evaluate' is called. With incremental evaluation >>they can be calculated at a very low cost. > >IMO, the drawback is that you have to calculate them always, even when you do >not want to evaluate. Yes, that's true. >E.g., you can't use an evaluation window technique to save the static evaluation >when the material is out of balance. Or for non-leaf positions you usually do >not have to evaluate. >In my program for instance, I ususally evaluate far less than half of the nodes >(depends on the type of position). On the other hand, I can use parts of the evaluation in the non-leaf nodes for move ordering and determining which moves are interesting (should be searched deeper) and which are boring (depth can be reduced). For example, I can extend those moves that cause big changes on the attack value. I've found these "tricks" very useful, because in shogi you might easily have 200 possible moves per position and you simply cannot do full-width search deeply enough. - Pauli Misikangas -
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.