Author: Frank Phillips
Date: 02:57:10 08/13/00
Go up one level in this thread
On August 12, 2000 at 09:26:47, Robert Hyatt wrote: >On August 12, 2000 at 03:44:51, Frank Phillips wrote: > >>I am trying to improve my king safety term in the evaluation function, but >>without sufficient personal chess expertise to set a sensible benchmark. Maybe >>this is asking for information that people might not want to give, but with a >>full set of opponents pieces, white castled kingside, h, g, f files open and >>opponent castled queenside, how big is your kingsafety bonus. A knight, a pawn >>??..? >> >>Any thoughts on calibrating such terms are welcomed. My positional evaluation >>terms are an ad hoc set of numbers, bearing no particular relationship to each >>other. Generalised pieces values seem well established. Has the chess >>programming community evolved reference values for specific positional terms. >> >>Frank > >If you want to see what others have done, the best way is to simply set up >the position and see what the score is. IE in Crafty, just cut/paste a FEN >string directly into crafty, hit <enter>, then type "score" and it will give >you the total score, and then king safety score (although the king safety score >is only a 'partial score'). > >However, that is probably not the right way to tune, as in chess everything is >relative to everything else. You don't want to shred your own king safety just >to get a rook on the 7th or whatever.. Thanks. I have already done what you suggested with Crafty. You have tuned these values in play on ICC so they should be robust. Positional terms seem more sensitive to changes in piece configuration than piece values are (although the distinction between adding a bonus for a passed pawn and the pawn’s value is perhaps moot?) and there are more of them and therefore more interactions between them, but I would have thought that good chess players and programs have simple rules of thumb about how much wood to throw on the fire to expose the king, for example. I have experimented with a variety of king safety values. Sensible ranges seem not to make that much difference, but there are always some positions where you feel the computer should have seen the attack coming before its result was revealed too late by the search. Too high king safety values causing speculative sacrifices can be amusing against fallible, anxious humans, but fatal against Crafty. Frank
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.