Author: Robert Hyatt
Date: 19:51:02 03/06/98
Go up one level in this thread
On March 06, 1998 at 20:41:44, Stuart Cracraft wrote: >For a long time, my program had been sparring with >another program. The result over 50 games was about >40 losses and 10 draws. Never a win. > >I suspected a misbalance of positional factors which >resulted in >1 pawn positional contribution to the >evaluation score as well as poor development and >poor analysis of backward pawns. > >It seemed likely that by improving the development and >backward pawn factors and minimizing the positional >contributions swaying the material score, this might >help. > >So I converted to millipawns (a 5 second change for >me and a 10 minute compile). Also I added David Levy's >development terms from "The Joy of Computer Chess" >and added analysis of backward pawns. > >Then I played it against its sparring partner and it >immediately won its first game. There were no outright >technical blunders and it kept the game together. > >I think another good approach is to analyze your >worst/largest positional contribution to the overall score >and find out what kinds of positions are producing these >big swings in the score that are not materially based and >start lessening the heuristic scores to bring the positional >score back within a narrow window. > >For example, a king scoring function that results in massive >changes in the positional score could be a blunder-creator >not a king-saver! By going to millipawns first it's possible >to see a formerly blundering program actually play half-decently >while still working the behind-the-scenes bugs. Later, centipawns >could be restored, if desired, so that positional sacrifices might >be made on a more conservative basis. > >I am curious regarding the other program's authors on this bulletin >board, what are the largest positional heuristics you award or >penalize positions for? And do you use centipawn or millipawn >and why? And have you experimented with the other and why not? > >--Stuart I started with centipawns, switched to millipawns, then back to centipawns again. I don't believe that *I* can accurately divide a positional score into fractions of 1/1000th... I don't even believe I can divide it into pieces of 1/100th... I also have significant scores... many are worth a pawn, several are worth way more... ie king safety can be up to +/- 3 pawns or so de- pending on how exposed and how many pieces are left for the opponent. Ditto for some endgame things like connected passers on the 6th/7th, or passer advanced with king support, and so forth. Only way I know to tune is play and watch and, if you are a good enough player, try to figure out where your program went wrong and adjust the weights to stop that...
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.