Author: Larry Coon
Date: 08:35:34 01/11/99
Go up one level in this thread
On January 10, 1999 at 10:12:43, Frank Phillips wrote: >Having concentrated more on the search and simply getting my chess program >working, my static evaluation function has grown in an adhoc fashion. It >contains the usual terms – two bishops, no bishop and pawns on both sides, >passed pawns, connected passed pawns, outside passed pawns, doubled pawns, >isolated pawns, rook on the seventh, rook on open file, and so on – but I have >no idea what the relative difference these all make let alone the importance of >their relative values. Clearly there are too many variables for rigorous >testing. Not a new problem. Any advice? Somewhat amusingly the program has >taken to flinging its h and a- pawns forward in the style of the rubbish I >remember from the seventies. Not quite the Kh8 style “null move”, but no doubt >similar. I'm definitely not an expert, but I did go through the same thing. I set all my values to initial guesses, and then played games. Then when it made moves that I thought weren't optimal, I tried to figure out what setting might have been wrong, changed the setting, reset the position, and tried again. Things like, "gee, it allowed checkmate in order to protect the bishop pair. Maybe my bishop pair value was excessively high." A couple of hundred games, and I think all my values are now fairly reasonable. But like you, mine will still move the outside pawns when it can't find something better. I suppose it's also possible to automate the process using the "hill climbing" method -- have your program tweak a value, play against itself several times and record the effect, and eventually narrow itself in on the right value. Of course, there's a danger of settling on local maxima rather than global maxima, but there are ways around that as well. Despite the fact that the second approach is automated, I think the first approach will still be faster and yeild better results -- plus it'll help you understand the way your settings affect your performance a lot better. Larry Coon
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.